GAN-PD: generative adversarial networks for fabric defect generation towards precise detection Authors: Yuyang Xia, Yong Yin, Fengkai Luan, Junqi Liu, Jiaxing Yang Published: 2026 ABSTRACT Automated inspection of textile surfaces is challenged by the scarcity and geometric intricacy of real defect samples, especially the elongated, sub-millimetre scratches critical for quality grading. Existing generative approaches often prioritize image realism, thereby neglecting the geometric continuity and contextual fidelity essential for robust detection by modern systems. This paper introduces Generative Adversarial Networks for Fabric Defect Generation towards Precise Detection(GAN-PD), a task-oriented generative adversarial framework engineered to synthesize fabric anomalies with both high structural realism and detector relevance. GAN-PD’s generator employs Local-Feature Residual modules to adaptively focus convolutional resources on defect-salient channels, while its discriminator utilizes columnar pooling to enforce directional continuity along defect axes. This synergistic design preserves high-frequency weave details, enforces millimetre-scale linearity, and mitigates fragmentation artifacts typical of isotropic convolutions. Experimental results on benchmark textile dataset show GAN-PD produces sharper linear anomalies, achieves more consistent texture fusion, and leads to significant improvements in detection accuracy for both one-stage and two-stage detectors compared to state-of-the-art GANs. These results validate the efficacy of shifting the paradigm from mere “defect generation” to “generation for defect detection.” Ultimately, these findings demonstrate that reframing defect synthesis around downstream utility, rather than visual fidelity alone, can markedly enhance the robustness and accuracy of automated textile inspection systems. Introduction Automated surface-defect inspection constitutes a pivotal component of contemporary fabric manufacturing1,2;micrometre-scale anomalies–including yarn scratches, filament misalignments, and localised stains–compromise both product quality and operational safety. Early techniques relied on handcrafted descriptors and threshold-based segmentation3,4, yet their limited capacity to model complex textural dynamics renders them inadequate for the diverse patterns and defect morphologies present in modern fabrics. Deep learning has recently delivered notable advances in anomaly detection,owing to its powerful representational capability5.Nonetheless, the efficacy of these models is tightly coupled to large-scale, heterogeneous defect corpora. In industrial settings, the intrinsic sparsity of defect samples–particularly for subtle, elongated imperfections such as hairline scratches or filament offsets–poses a persistent obstacle, even though these defects are decisive for rigorous fabric quality assessment. Conventional data-augmentation routines, predicated on elementary geometric or chromatic transformations6,fail to substantially enrich data diversity and often incur information loss or distortion. Generative adversarial networks (GANs)7have therefore attracted significant attention for synthetic augmentation8;variants such as ACGAN9,WGAN10,Pix2Pix11,and StarGAN12demonstrate formidable image-generation prowess. However, these models predominantly pursue holistic photorealism while neglecting the localised textures and fine structures that typify fabric surface defects. Although GAN derivatives have been widely adopted in industrial applications13–22, fabric defects usually manifest as slender, anisotropic artefacts–sub-millimetre scratches traversing multiple texture repeats–demanding precise spatial continuity and seamless fusion with the underlying pattern. Existing methods employ isotropic convolutional kernels, thereby inadequately capturing directional dependencies; the resulting defects frequently appear as unnatural blotches rather than the requisite filiform structures. Moreover, the intricate texture of fabric substrates exacerbates blending difficulty, so visual artefacts and textural distortions emerge, especially in richly detailed regions. For instance, SDGAN17 introduces dual discriminators and cycle-consistency loss to balance diversity and fidelity; EID-GAN18leverages a norm-penalty to facilitate fine-grained anomaly synthesis; DT-GAN23learns defect-type representations independent of the background and applies defect-specific styles. Despite these improvements, perfect synthesis of elongated defects and their integration into complex backgrounds remains elusive. Prevailing evaluation regimes–peak signal-to-noise ratio (PSNR)24, structural similarity index (SSIM)25, and related metrics26,27–prioritise pixel-level fidelity while disregarding the utility of generated defects for downstream detection. A principled, task-aware assessment protocol linking synthetic defect characteristics to detection performance is conspicuously absent, creating a critical gap between synthetic data quality and industrial applicability. Consequently, because current defect-generation models are optimized for detection only indirectly–operating under the heuristic assumption that visual photorealism naturally translates to diagnostic utility–they struggle to synthesize explicitly task-oriented samples that simultaneously exhibit fine geometric specificity and contextual texture coherence on complex industrial surfaces, thereby precluding tangible improvements in downstream inspection. To surmount these obstacles, we introduce Generative Adversarial Networks for Fabric Defect Generation towards Precise Detection(GAN-PD), an architecture tailored to high-fidelity defect synthesis for detection-centric applications. The generator incorporates a Local Feature Residual module that deepens the network while employing adaptive convolutions to selectively process informative channels, concentrating computational resources on defect-relevant cues and mitigating redundancy. This design augments the network’s capacity to model fine-grained textures and intricate spatial dependencies. In the discriminator, we embed columnar pooling with stripe-shaped kernels to capture long-range spatial correlations along defect directions, explicitly resolving directional anisotropy and enforcing geometric consistency between generated defects and fabric background. By executing a paradigm shift from vision-driven “defect generation” to task-aware “generation for defect detection,” this framework transcends mere architectural modifications, thereby furnishing a scalable solution for industries constrained by data scarcity and computational budgets. Our contributions are threefold: We present the first GAN framework explicitly optimised for fabric defect synthesis, producing samples that unite structural realism with detection relevance and alleviating data scarcity in industrial quality control.We devise a deep residual generator with local feature convolutions; the LFR module mitigates computational inefficiency and mode collapse, enabling flexible capture of high-order defect distributions and faithful texture fusion within complex fabric backgrounds.We design a columnar-pooling discriminator that injects directional sensitivity into the adversarial process, precisely modelling elongated defects and overcoming isotropic-convolution limitations.We establish a novel task-driven evaluation protocol that directly correlates synthetic-data quality with detection performance, bridging the divide between conventional image metrics and industrial utility. Extensive experiments demonstrate that GAN-PD consistently surpasses state-of-the-art methods in elevating downstream defect-detection accuracy, validating the efficacy of task-aware synthesis and mixed-sample training.The remainder of this paper is organised as follows. Section Related work surveys related work on industrial defect-image synthesis. Section Methodology delineates the proposed GAN-PD architecture and its underlying principles. Section Generated defect evaluation provides the relevant ablation experiments for the proposed architecture, as well as detailed experimental results of the architecture’s comparative testing in downstream detection tasks. Finally, Section Defect Detection Verification Experimental Results and Analysis concludes the paper. We present the first GAN framework explicitly optimised for fabric defect synthesis, producing samples that unite structural realism with detection relevance and alleviating data scarcity in industrial quality control. We present the first GAN framework explicitly optimised for fabric defect synthesis, producing samples that unite structural realism with detection relevance and alleviating data scarcity in industrial quality control. We devise a deep residual generator with local feature convolutions; the LFR module mitigates computational inefficiency and mode collapse, enabling flexible capture of high-order defect distributions and faithful texture fusion within complex fabric backgrounds. We devise a deep residual generator with local feature convolutions; the LFR module mitigates computational inefficiency and mode collapse, enabling flexible capture of high-order defect distributions and faithful texture fusion within complex fabric backgrounds. We design a columnar-pooling discriminator that injects directional sensitivity into the adversarial process, precisely modelling elongated defects and overcoming isotropic-convolution limitations. We design a columnar-pooling discriminator that injects directional sensitivity into the adversarial process, precisely modelling elongated defects and overcoming isotropic-convolution limitations. We establish a novel task-driven evaluation protocol that directly correlates synthetic-data quality with detection performance, bridging the divide between conventional image metrics and industrial utility. Extensive experiments demonstrate that GAN-PD consistently surpasses state-of-the-art methods in elevating downstream defect-detection accuracy, validating the efficacy of task-aware synthesis and mixed-sample training. We establish a novel task-driven evaluation protocol that directly correlates synthetic-data quality with detection performance, bridging the divide between conventional image metrics and industrial utility. Extensive experiments demonstrate that GAN-PD consistently surpasses state-of-the-art methods in elevating downstream defect-detection accuracy, validating the efficacy of task-aware synthesis and mixed-sample training. Related work Recent advances in GAN-based defect synthesis have demonstrated promising results in addressing data scarcity for industrial inspection tasks. Early attempts, such as He et al.13, proposed semi-supervised frameworks combining GANs with residual networks to augment limited steel defect datasets, though their focus on global classification accuracy overlooked the geometric specificity of fine-grained anomalies. Subsequent works like SDGAN17and EID-GAN18introduced architectural innovations–dual discriminators and norm-penalty functions–to balance defect diversity and handle extreme class imbalances. While these methods improved anomaly detection accuracy in commutator cylinders and fabric datasets, their reliance on isotropic convolution kernels inherently limited their ability to model anisotropic defects like hairline scratches or thread misalignments. Liu et al.14attempted to address this by proposing regional training strategies with wavelet fusion, prioritizing defect area translation in surface inspection tasks. However, their method still struggled with seamless integration of synthetic defects into repetitive textile patterns, often introducing edge artifacts in texture-rich regions. Further refinements in spatial controllability emerged through defect masks and latent-space conditioning. Niu et al.21designed region- and strength-controllable GANs to localize weak defects in metal surfaces, while Wang et al.20leveraged StyleGAN for panel defect simulation using location-specific models. Cross-domain approaches, such as DT-GAN23, decoupled defect features from background textures via disentangled representations, enabling defect transfer across products. Despite these advances, a critical limitation persists: defects are treated as global stylistic attributes rather than locally coherent structures with directional dependencies, leading to fragmented artifacts when applied to fine-grained textiles. For instance, Lai et al.16employed generative models to construct texture patterns for novelty detection but failed to address the precise alignment of synthetic defects with background motifs–a requirement paramount for textile quality control.Fig. 1Generative adversarial networks for fabric defect generation towards precise detection. Generative adversarial networks for fabric defect generation towards precise detection. The evaluation of synthetic defects remains another unresolved challenge. Existing frameworks13–23predominantly rely on pixel-level metrics (e.g., PSNR, SSIM) or domain-agnostic scores like FID, which prioritize visual plausibility over task relevance. Guo et al.15demonstrated improved FID scores for bridge defect synthesis using online knowledge distillation, yet their evaluation ignored the downstream impact on detector decision boundaries. Similarly, Zaman et al.22applied Pix2Pix with multi-snapshot strategies to augment bone defect data but did not correlate synthetic quality with segmentation performance. This disconnect between synthetic realism and industrial utility is particularly acute in textile inspection, where defects must not only appear authentic but also align with detector-sensitive features to improve generalization–a gap unaddressed by prior methodologies. To address these limitations, we propose a novel framework that integrates task-aware defect generation and evaluation. Unlike existing approaches that optimize for standalone realism, our method employs a GAN architecture specifically tailored for textile defects, ensuring synthetic anomalies preserve both geometric fidelity and contextual consistency with complex background textures. Crucially, we establish a direct link between synthetic data quality and downstream detection performance, bridging the gap between augmentation and task efficacy. This work represents the first systematic effort to embed task-specific considerations into the defect generation pipeline, offering a pragmatic and scalable solution for industries grappling with data scarcity. By redefining defect synthesis as a detector-centric process, our approach ensures generated samples actively enhance real-world inspection scenarios rather than merely approximating them–a paradigm shift with profound implications for industrial quality control. Methodology The proposed GAN-PD framework addresses two fundamental challenges in fabric defect synthesis: the geometric anisotropy of sub-millimeter defects such as hairline scratches that span multiple texture repetitions, and the contextual coherence required for seamless integration of synthetic defects into complex woven patterns. Fundamental physical constraints are integrated into the synthesis pipeline to surpass conventional structural modifications. Explicit channel-level disentanglement is executed by the generator. Complex background topology is thus preserved. Concurrently, anisotropic geometric priors are enforced by the discriminator. The spatial continuity of elongated defects is rigidly guaranteed by this geometric enforcement. As depicted in Fig. 1, our architecture employs a dual-branch strategy where a deep residual generator progressively constructs defect morphologies through localized feature modulation, while a direction-sensitive discriminator enforces spatial consistency along defect axes. This symbiotic design ensures that synthesized anomalies not only mimic the visual characteristics of real defects but also align with the geometric priors critical for detection tasks.Fig. 2Network structure of the deep residual generator with local feature convolution. The DeepR Module denotes the deep residual component, while the LocalFC Module represents the local feature convolution. The LFR Module’s parameter N=3 balances feature extraction efficiency and computational cost, as validated through comparative experiments. Network structure of the deep residual generator with local feature convolution. The DeepR Module denotes the deep residual component, while the LocalFC Module represents the local feature convolution. The LFR Module’s parameter N=3 balances feature extraction efficiency and computational cost, as validated through comparative experiments. The generation process initiates by transforming random noise vectors into defect-embedded fabric textures through three hierarchical stages. Initial convolutional layers extract primitive texture motifs, establishing the baseline fabric pattern structure. Subsequent layers deploy cascaded Local Feature Residual (LFR) modules–a novel construct detailed in Section Methodology-A–that specialize in inducing directional defects through adaptive channel reweighting and dilated convolutions. These modules strategically amplify defect-related features while suppressing redundant background activations, thereby resolving the persistent trade-off between defect saliency and texture preservation. The final reconstruction phase employs texture-regularized upsampling to synthesize high-resolution images, explicitly mitigating edge artifacts in regions with dense repetitive patterns. Complementing this, the discriminator architecture introduces columnar pooling operations with elongated kernels to capture the directional persistence of filament misalignments–a critical advancement over conventional isotropic convolutions as elaborated in Section Methodology-B. Through adversarial training, the generator learns to synthesize defects that simultaneously deceive the discriminator’s global texture consistency checks and satisfy its anisotropic geometric constraints. This dual optimization pressure ensures that synthetic hairline scratches exhibit millimeter-scale continuity while maintaining phase alignment with underlying fabric weaves, a prerequisite for effective detector training. The following sections dissect the architectural innovations of the LFR generator and columnar-pooling discriminator, revealing how their coordinated operation overcomes the limitations of prior isotropic convolution-based approaches. Specifically, Section Methodology-A formalizes the defect induction mechanisms through localized residual learning, while Section Methodology-B demonstrates how directional sensitivity in the discriminator enforces physically plausible defect geometries. A. Deep residual generator for local feature convolution The proposed deep residual generator architecture addresses the dual challenges of gradient dissipation in deep networks and computational inefficiency in defect feature extraction through a hierarchical integration of residual learning and localized channel processing. As illustrated in Fig. 2, the generator employs cascaded Local Feature Residual modules that synergistically combine residual connections with adaptive channel-wise convolution–a strategic departure from conventional isotropic convolution paradigms. The residual pathways mitigate vanishing gradients by preserving information flow through identity mappings, while the local convolution mechanism dynamically allocates computational resources to defect-critical regions through channel segmentation. This dual mechanism enables the network to progressively construct defect morphologies while maintaining high-frequency texture details in complex woven substrates.Algorithm 1Local feature convolution algorithm Local feature convolution algorithm The generation pipeline commences with reflective-padded convolution layers that establish baseline texture representations through batch-normalized feature abstraction. Spatial downsampling via strided convolutions follows, systematically compressing resolution while expanding channel depth to capture multi-scale defect characteristics. At the core of the architecture, three LFR modules (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$N{=}3$$\end{document}, optimized through ablation studies) iteratively refine defect features through alternating residual blocks and localized convolutions. Each residual block, depicted in Fig. 3(a), employs twin convolution layers with batch normalization and ReLU activations, forming nonlinear transformations while preserving input identity through skip connections. The subsequent local feature convolution, governed by the formal mathematical operations detailed in Algorithm 1 and Fig. 3(b), implements channel-wise partitioning where a subset undergoes kernel-based feature transformation while the remainder retains original activations. Formally, for an input tensor \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^{C \times H \times W}$$\end{document}, the module splits channels into \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X_1 \in \mathbb {R}^{C_p \times H \times W}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X_1 \in \mathbb {R}^{C_p \times H \times W}$$\end{document}, applying learned convolution kernels \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$W_p$$\end{document} only to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X_1$$\end{document} before recombining with \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$X_2$$\end{document} through channel concatenation. Through empirical evaluation, the channel partitioning strategy was ultimately set to an equal-ratio split, achieving the best trade-off between semantic feature preservation and computational overhead. By assigning half of the channel dimensions to identity mapping, the network is able to strictly preserve the low-frequency topological priors of the fabric substrate. Meanwhile, the remaining half is routed through dedicated convolutional operations to model high-frequency, defect-salient morphological variations. This symmetric partitioning scheme not only prevents the degradation of repetitive background textures, but also effectively reduces the number of convolutional parameters by half.Fig. 3Network structures of deep residual and local convolution modules. Network structures of deep residual and local convolution modules. This selective processing concentrates parameter updates on defect-salient features while bypassing redundant background computations–a critical efficiency gain for offline data synthesis on industrial servers. During reconstruction, transpose convolutions with texture-regularized upsampling progressively restore spatial resolution, mitigating edge artifacts through learned prior constraints. The final output layer utilizes Tanh-activated convolutions with reflective padding to ensure pixel-level continuity across repetitive texture boundaries. Crucially, the LFR architecture’s directional sensitivity emerges from the interplay between localized channel processing and multi-scale residual learning, enabling precise modeling of anisotropic defects spanning multiple texture periods–a capability unattainable with isotropic convolution-based generators. B. Columnar pooling discriminator network The proposed columnar pooling discriminator directly addresses the limitations of conventional spatial pooling in modeling anisotropic defect geometries–a critical requirement in textile quality inspection where elongated scratches or filament misalignments exhibit strong directional dependencies. As illustrated in Fig. 4, the discriminator integrates columnar pooling operations into a deep convolutional framework, enabling targeted feature extraction along defect axes while suppressing interference from non-defective regions. Traditional pooling methods, which rely on isotropic square kernels, inherently struggle to capture linear structural patterns in texture-rich fabric substrates, often introducing noise from non-target areas and compromising defect localization accuracy. This inefficacy is particularly pronounced in scenarios requiring precise identification of sub-millimeter defects spanning multiple texture repetitions, where global or fixed-window pooling operations fail to resolve directional continuity.Fig. 4Network architecture of the columnar pooling discriminator. The module integrates horizontal and vertical columnar pooling after the third convolutional layer, with kernel geometries tailored to capture anisotropic defect features. Network architecture of the columnar pooling discriminator. The module integrates horizontal and vertical columnar pooling after the third convolutional layer, with kernel geometries tailored to capture anisotropic defect features. The discriminator’s architecture processes input images through a sequence of strided 4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document}4 convolutions, progressively downsampling spatial resolution while expanding feature depth. Each convolutional layer is followed by batch normalization and LeakyReLU activation to stabilize training dynamics and preserve gradient flow through nonlinear transformations. At the network’s midpoint–strategically positioned after the third convolutional layer–the columnar pooling module injects directional sensitivity by decomposing feature aggregation into orthogonal axes, with its formal mathematical operations explicitly defined in Algorithm 2. Horizontal pooling compresses features through width-wise averaging, while vertical pooling operates analogously along the height dimension. Furthermore, the strip pooling module employs adaptive average pooling to dynamically generate \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1 \times W$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$H \times 1$$\end{document} spatial feature representations, effectively circumventing the limitations inherent in fixed-size convolutional kernels. By aggregating global features along orthogonal axes, this mechanism achieves morphological invariance, enabling adaptation to diverse defect shapes without the need for hyperparameter tuning specific to defect categories. For instance, in the case of anisotropic defects such as elongated scratches, this cross-axis pooling operation precisely captures long-range structural dependencies along the defect trajectory. Conversely, for microscopic anomalies like tiny spots, the method establishes a robust global contextual baseline within the fabric feature domain, thereby enhancing the relative saliency of local high-frequency abnormal responses. These axis-specific descriptors are transformed via learnable weights and biases, generating spatial attention maps through sigmoid activation. The modulated feature maps amplify defect-related activations while attenuating background noise, effectively aligning synthetic anomalies with the directional persistence of real defects. This anisotropic pooling mechanism diverges fundamentally from isotropic approaches by employing elongated kernels that align with defect geometries. In horizontal pooling, the kernel’s extended shape along the width dimension captures long-range interactions between dispersed defect regions, while its narrow height concentrates on local texture coherence. Conversely, vertical pooling prioritizes height-wise continuity, ensuring synthetic scratches maintain phase alignment with underlying weave patterns. Such directional decomposition enables the discriminator to enforce geometric consistency–critical for suppressing artifacts in synthetic defects–while avoiding the computational redundancy of full-resolution attention mechanisms.Algorithm 2Columnar pooling algorithm Columnar pooling algorithm The discriminator culminates in a final 4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document}4 convolution layer that maps features to a single-channel authenticity prediction. Sigmoid activation constrains outputs to [0,1], providing pixel-level adversarial feedback that penalizes discontinuities in synthetic defect morphologies. This fine-grained discrimination capability ensures the generator synthesizes defects that not only mimic visual realism but also adhere to the physical constraints of textile substrates, bridging the gap between synthetic data quality and downstream detection efficacy. Generated defect evaluation A. Implementation details The experimental validation of GAN-PD’s defect synthesis capability was performed on the DAGM2007 benchmark dataset28, a widely adopted standard for texture surface defect analysis. This dataset encompasses ten distinct defect categories characterized by high morphological diversity and complex background textures, closely emulating industrial inspection scenarios through weakly supervised annotations. Prior to training, all raw image crops were uniformly resized to a spatial resolution of 224\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\times$$\end{document}224 pixels to ensure tensor consistency, with pixel intensities strictly normalized to the [−1, 1] interval to align with the Tanh activation bounded output space of the generator. To ensure reproducibility, the computational environment and hyperparameter configurations are systematically outlined in Table 1.Table 1Experimental environment.Required environmentVersion configurationOperating SystemWINDOWS 11CPU12th Gen Intel(R) i9-12900HGPUNVIDIA RTX 3070 Ti 16GProgramming Language VersionPython3.9Deep Learning FrameworkTorch2.0.1, Cuda11.7, Cudnn8.5.0 Table 1Experimental environment.Required environmentVersion configurationOperating SystemWINDOWS 11CPU12th Gen Intel(R) i9-12900HGPUNVIDIA RTX 3070 Ti 16GProgramming Language VersionPython3.9Deep Learning FrameworkTorch2.0.1, Cuda11.7, Cudnn8.5.0 Experimental environment. The training regimen comprised 150 epochs with a batch size of 1, initialized using the Adam optimizer (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\beta _1{=}0.9$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\beta _2{=}0.999$$\end{document}) and a learning rate of 0.0003. A linear learning rate decay was activated at the 50th epoch to stabilize convergence. Hyperparameters \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda _1{=}5$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda _2{=}0.5$$\end{document} were empirically determined through grid search to optimize the balance between adversarial training and texture preservation. B. Visual quality analysis of generated defects Visual assessment of fabric defects synthesized by GAN-PD using Classes 1–5 of the DAGM2007 dataset indicates that our model proficiently generates anomalies aligned with practical industrial inspection standards. Fig. 5 illustrates the robust performance of GAN-PD in synthesizing defects characterized by structural coherence and contextual fidelity, particularly excelling in linear scratches (e.g., Classes 2 and 4) and microscale anomalies (e.g., Class 3), seamlessly integrated within intricate fabric backgrounds. Linear scratches. Class 2 and Class 4 defects exhibit precise millimeter-scale continuity, effectively traversing multiple repetitive fabric textures. Notably, these synthetic anomalies preserve line-level textural intricacies even when superimposed on complex fabric patterns, thereby enhancing realism without diminishing the integrity of the underlying texture. Tiny Spots. Class 3 defects demonstrate pronounced luminance contrasts compared to genuine defect instances, enhancing visual distinguishability against intricate refractive textile backgrounds. This enhanced contrast facilitates clearer defect identification and underscores the model’s capability to amplify subtle anomaly characteristics. Fuzzy Coloring. Class 1 defects effectively reproduce authentic mosaic-like morphological anomalies, accurately capturing subtle boundary details and achieving a consistent transitional texture with the vibrant fabric substrate. Although these stains manifest realistically, their subtle features exhibit less pronounced visual enhancement compared to more salient defect types. Texture Imprints. Class 5 defects reflect an authentic depiction of imprint morphology against complex fabric backdrops. However, due to the moderate scale and inherent complexity of the texture, the visual differentiation offered by GAN-PD does not yield conspicuous improvements over the original defect representations. These observations underscore a crucial relationship between defect morphology and synthesis performance. For linear defects demanding precise directional alignment, the Columnar Pooling Discriminator adeptly mitigates anisotropic blurring–typical of isotropic convolution–thereby generating defects that maintain exact phase alignment with background weaves. Concurrently, the Deep Residual Generator leveraging Local Feature Convolutions distinctly amplifies the characteristics of low-saliency anomalies, ensuring their clear discernibility against heterogeneous textures. In contrast, the network exhibits measured stability when synthesizing defects with diffuse or stochastic characteristics, indicating balanced generalization rather than excessive optimization toward particular defect morphologies. While the subjective visual analysis confirms significant advances in synthesizing critical defect classes, comprehensive assessments of generation quality will be addressed quantitatively through objective metrics in subsequent analyses, and the direct utility of these synthetic defects for downstream detection tasks will be systematically evaluated in Section Defect Detection Verification Experimental Results and Analysis.Fig. 5Defective samples generated by GAN-PD in the DAGM2007 dataset. Defective samples generated by GAN-PD in the DAGM2007 dataset. C. Quantitative metrics and component ablation To rigorously assess the efficacy of architectural innovations in GAN-PD, systematic ablation studies were conducted across Classes 1–5 of the DAGM2007 dataset, evaluating both parametric efficiency and defect synthesis quality. Table 2 quantifies performance evolution through progressive integration of core components–Initial Generator (ING), Initial Discriminator (IND), Local Feature Convolution Generator (LFCG), and Columnar Pooling Discriminator (CPD)–using established metrics: PSNR, SSIM, Inception Score (IS), and Fréchet Inception Distance (FID). As shown in Table 3, the introduction of LFCG and CPD substantially reduces the generator, discriminator, and total model parameters, highlighting the improved parametric efficiency of GAN-PD.The complete GAN-PD configuration (LFCG+CPD) achieves parametric efficiency superior to baseline models, reducing trainable parameters by 42% (62.682M\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\rightarrow$$\end{document}36.267M) while substantially enhancing synthesis fidelity. Quantitative metrics exhibit systematic improvements: PSNR ascends from 23.91dB to 31.28dB, SSIM from 0.54 to 0.61, and IS from 1.23 to 1.95, coupled with a 34.8% FID reduction (172.67\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\rightarrow$$\end{document}112.69). These advancements confirm that localized feature processing and directional pooling synergistically eliminate redundant computations while preserving high-frequency defect characteristics. Notably, LFCG alone elevates PSNR by 4.82dB and reduces parameters by 28.1%, validating its capacity to prioritize defect-relevant features through channel-adaptive convolutions. Subsequent CPD integration further improves SSIM by 0.04 and FID by 27.3%, demonstrating anisotropic discrimination’s critical role in maintaining geometric consistency across multi-scale textures. The compounded enhancements confirm that LFCG’s residual learning captures defect morphology distributions, while CPD’s directional constraints enforce physically plausible integration with background patterns–a dual mechanism essential for industrial-grade synthesis. Concurrently, training stability analysis reveals accelerated convergence, attributable to the discriminator’s geometric priors regularizing generator updates. These results collectively substantiate GAN-PD’s capability to balance parametric efficiency with perceptual quality, overcoming the computational-intensity versus fidelity trade-off endemic to conventional GAN architectures, thereby facilitating its practical application in offline data-augmentation pipelines within standard industrial server environments.Table 2Comparison of image quality evaluation metrics in GAN-PD ablation experiments.INGLFCGINDCPDPSNR (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\uparrow$$\end{document})SSIM (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\uparrow$$\end{document})FID (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\downarrow$$\end{document})IS (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\uparrow$$\end{document})\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}—\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}—\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$23.91\pm 0.93$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$0.66\pm 0.15$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$172.67\pm 5.26$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1.23\pm 0.15$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}——\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$27.59\pm 0.74$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$0.77\pm 0.11$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$158.43\pm 3.62$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1.56\pm 0.14$$\end{document}—\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}—\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$28.64\pm 0.82$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$0.79\pm 0.13$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$154.78\pm 3.35$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1.61\pm 0.13$$\end{document}—\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}—\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}31.28±0.870.89±0.07112.69±2.711.95±0.09Table 3Comparison of parametric quantities in GAN-PD ablation experiments.INGLFCGINDCPDNumber of generatorparameters/MNumber ofdiscriminatorparameters/MNumber of modelparameters/M\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}—\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}—56.9475.73562.682\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}——\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}56.9472.83159.778—\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}—33.4365.73539.171—\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}—\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}33.4362.83136.267 Table 2Comparison of image quality evaluation metrics in GAN-PD ablation experiments.INGLFCGINDCPDPSNR (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\uparrow$$\end{document})SSIM (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\uparrow$$\end{document})FID (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\downarrow$$\end{document})IS (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\uparrow$$\end{document})\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}—\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}—\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$23.91\pm 0.93$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$0.66\pm 0.15$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$172.67\pm 5.26$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1.23\pm 0.15$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}——\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$27.59\pm 0.74$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$0.77\pm 0.11$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$158.43\pm 3.62$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1.56\pm 0.14$$\end{document}—\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}—\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$28.64\pm 0.82$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$0.79\pm 0.13$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$154.78\pm 3.35$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$1.61\pm 0.13$$\end{document}—\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}—\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}31.28±0.870.89±0.07112.69±2.711.95±0.09 Comparison of image quality evaluation metrics in GAN-PD ablation experiments. Table 3Comparison of parametric quantities in GAN-PD ablation experiments.INGLFCGINDCPDNumber of generatorparameters/MNumber ofdiscriminatorparameters/MNumber of modelparameters/M\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}—\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}—56.9475.73562.682\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}——\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}56.9472.83159.778—\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}—33.4365.73539.171—\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}—\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\checkmark$$\end{document}33.4362.83136.267 Comparison of parametric quantities in GAN-PD ablation experiments. Defect detection verification experimental results and analysis While conventional image quality metrics provide foundational insights into synthetic defect characteristics, their limitations in correlating with detection task performance necessitate complementary evaluation protocols. PSNR’s insensitivity to perceptual complexity and SSIM’s localized focus often misalign with human visual assessments, while FID/IS inherit biases from pretrained models, inadequately capturing industrial defect diversity. To holistically evaluate synthetic data utility, we benchmark GAN-PD against SDGAN17 and EID-GAN18 using three detection architectures (SSD, Faster R-CNN, YOLOv5) trained on distinct datasets: Real (100% authentic defects), Synthetic (100% generated defects), and Hybrid (1:1 real-synthetic mixture). This framework quantifies how synthetic defects influence model generalization, precision, and robustness under industrial inspection conditions. A. Experimental design and evaluation indicators Three training datasets—De-Train A (real), De-Train B (synthetic), and De-Train C (hybrid)—were constructed per defect class, each containing 700 defect-free samples and 100 defective instances (Table 4). Corresponding test sets (De-Test A/B/C) maintained identical class distributions with 120 defect-free and 60 defective samples. Training-validation splits followed a 7:3 ratio, with models trained for 100 epochs using Adam optimization (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\beta _1 = 0.95$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\beta _2 = 0.999$$\end{document}). Learning rates were fixed at 0.005 for Faster R-CNN/SSD and 0.01 for YOLOv5, decaying linearly after 30 epochs. A batch size of 8 and detection confidence threshold of 0.5 were uniformly applied.Table 4Dataset settings.DatasetNumber of defect-freesamplesSample size of truedefectsNumber of defectivesamples generatedDe-Train A700100—De-Train B700—100De-Train C7005050De-Test A12060—De-Test B120—60De-Test C1203030 Table 4Dataset settings.DatasetNumber of defect-freesamplesSample size of truedefectsNumber of defectivesamples generatedDe-Train A700100—De-Train B700—100De-Train C7005050De-Test A12060—De-Test B120—60De-Test C1203030 Dataset settings. In the experimental framework, performance evaluations of the detection models are conducted using three pivotal metrics pertinent to an actual industrial production setting: Data Detection Accuracy (DDA), Defect Detection Rate (DDR), and False Detection Rate. Data Detection Accuracy is defined as the proportion of correctly identified defective and non-defective samples relative to the total number of samples. Defect Detection Rate specifically refers to the percentage of actual defect samples that are correctly identified as defective. False Detection Rate involves the proportion of non-defective samples erroneously identified as defective relative to the total number of samples detected as defective. These metrics collectively assess how synthetic defects alter detector decision boundaries, balancing identification reliability against operational false alarms–a critical consideration for automated inspection systems. This paper establishes a direct mathematical relationship between synthesized defect characteristics and actual detection utility through task-specific quantitative detection metrics, rather than relying on generic image quality scores, thereby further verifying the explicit task-aware nature of the proposed framework. The relevant computational formulas are presented in Equations (1) to (3).1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} DDA = \frac{TP + TN}{TP + TN + FP + FN} \end{aligned}$$\end{document}2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} DDR = \frac{TP}{TP + FN} \end{aligned}$$\end{document}3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} FDR = \frac{FP}{TP + FP} \end{aligned}$$\end{document}During the testing phase, TP represents the number of defect samples correctly identified, TN is the count of non-defective samples correctly identified, FP denotes the number of non-defective samples erroneously identified as defective, and FN refers to the number of defect samples incorrectly identified as non-defective. B. Comparative experiments on defect generation Systematic visual comparison of synthetic defects generated by GAN-PD, SDGAN, and EID-GAN reveals critical insights into their respective capabilities and limitations, as exemplified in Figs. 6, 7. The SSIM localization maps, computed against consistent background references, expose fundamental disparities in defect fidelity and contextual integration. SDGAN-generated defects exhibit amorphous distributions across Classes 1, 2, and 5, with features appearing spatially incoherent–particularly detrimental for Class 4’s linear defects, which manifest as fragmented fold-like artifacts rather than continuous scratches. While SDGAN produces heterogeneous luminous spots in Class 3, these deviate from authentic dark micro-anomalies, indicating misalignment with real defect photometric profiles. EID-GAN marginally improves feature cohesion but suffers from mode collapse in Class 2 (identical linear defect repetitions) and inadequate defect-background fusion. Class 1 anomalies exhibit translucent artifacts revealing underlying textures, evidencing incomplete defect feature learning, while Class 5’s imprints display severe texture erosion under complex backgrounds. Both baselines struggle to preserve directional consistency in linear defects or photometric authenticity in micro-scale anomalies. GAN-PD addresses these limitations through coordinated generator-discriminator innovations. Linear defects in Classes 2 and 4 maintain millimeter-scale continuity, aligning precisely with underlying weave orientations. Class 1’s waxy smudges exhibit controlled blurring gradients that respect substrate texture boundaries, while Class 5 imprints retain high-frequency details despite complex backgrounds. Class 3 micro-spots achieve optimal luminance contrast, replicating the dark, crisp profiles of real defects. The SSIM maps further validate these improvements, showing tighter spatial correspondence between synthetic and real defect regions compared to baselines. These visual advancements stem from the framework’s architectural synergies: the LFR generator hierarchically amplifies defect saliency through channel-adaptive convolutions, while the columnar pooling discriminator enforces directional consistency via anisotropic feature aggregation. Such capabilities enable GAN-PD to overcome the isotropic convolution limitations inherent in SDGAN and EID-GAN, which prioritize global texture realism over localized defect fidelity. While qualitative analysis underscores GAN-PD’s perceptual superiority, Sections Defect Detection Verification Experimental Results and Analysis-C and Defect Detection Verification Experimental Results and Analysis-D quantitatively evaluate its industrial utility through task-driven detection metrics, correlating synthetic defect quality with downstream inspection performance. This dual validation paradigm bridges the gap between visual plausibility and operational efficacy–a critical requirement for industrial adoption.Fig. 6Comparative visualization of defect samples generated by GAN-PD, SDGAN, and EID-GAN across multiple defect classes. Comparative visualization of defect samples generated by GAN-PD, SDGAN, and EID-GAN across multiple defect classes. C.Baseline detection with real data In this experimental section, a thorough comparative analysis was conducted on defect samples generated by the GAN-PD model, as well as those produced by the SDGAN and EID-GAN models. The outcomes of these experiments are detailed in Fig. 6.Table 5Benchmark metrics for the three detection models in De-Train A.DatasetIndicator(mAP@0.5)YOLOv5SSDFaster-RCNNClass10.6670.5450.634Class20.7840.7350.791Class30.7230.6610.754Class40.5420.5960.552Class50.6540.6140.629 Table 5Benchmark metrics for the three detection models in De-Train A.DatasetIndicator(mAP@0.5)YOLOv5SSDFaster-RCNNClass10.6670.5450.634Class20.7840.7350.791Class30.7230.6610.754Class40.5420.5960.552Class50.6540.6140.629 Benchmark metrics for the three detection models in De-Train A. Through a visual analysis of the control experiment results, it is observed that in Classes 1 through 5, our proposed GAN-PD model demonstrates significantly enhanced defect feature recognition when generating defect samples, compared to the other two benchmark models. It exhibits a profound ability to deeply learn and accurately simulate the characteristics of industrial product surface defects, showing a certain degree of technical advancement over contemporary advanced models. While preliminary judgments from visual analysis suggest that our proposed models exhibit noticeable improvements in the diversity and finesse of defect generation compared to leading-edge models for new industrial product surface defects, it is imperative, considering that these generated defect samples are directly applied in the quality inspection stages of actual industrial production lines, to further obtain quantitative data through precise defect detection experiments. This is essential to provide robust scientific support for the superiority of our proposed models. Therefore, this section employs the real sample dataset De-Train A and its corresponding test set De-Test A as the basis for training and benchmarking all subsequent model evaluations to ensure the objectivity and scientific validity of the assessment process. The benchmark data for the three defect detection models, YOLOv5, SSD, and Faster-RCNN, are presented in Tables 5 and 6 respectively.Table 6Benchmark metrics for the three detection models in De-Test A.DatasetModelYOLOv5SSDFaster-RCNNIndicatorDDADDRFDRDDADDRFDRDDADDRFDRClass 189.4%68.3%0.0%81.7%46.7%3.4%87.8%63.3%0.0%Class 293.9%81.7%0.0%92.2%76.7%0.0%91.7%75.0%0.0%Class 389.4%73.3%6.4%88.3%71.7%8.5%90.6%75.0%4.3%Class 479.4%46.7%15.2%81.1%50.0%11.8%80.6%48.3%12.1%Class 586.7%68.3%10.9%85.0%66.7%14.9%86.1%68.3%12.8% Table 6Benchmark metrics for the three detection models in De-Test A.DatasetModelYOLOv5SSDFaster-RCNNIndicatorDDADDRFDRDDADDRFDRDDADDRFDRClass 189.4%68.3%0.0%81.7%46.7%3.4%87.8%63.3%0.0%Class 293.9%81.7%0.0%92.2%76.7%0.0%91.7%75.0%0.0%Class 389.4%73.3%6.4%88.3%71.7%8.5%90.6%75.0%4.3%Class 479.4%46.7%15.2%81.1%50.0%11.8%80.6%48.3%12.1%Class 586.7%68.3%10.9%85.0%66.7%14.9%86.1%68.3%12.8% Benchmark metrics for the three detection models in De-Test A. Fig. 7SSIM localization maps for synthetic defects, computed against identical real fabric backgrounds to emphasize spatial alignment and feature coherence. SSIM localization maps for synthetic defects, computed against identical real fabric backgrounds to emphasize spatial alignment and feature coherence. D.Baseline detection with real data To systematically evaluate the industrial utility of synthetic defects, we retrained three detection models (YOLOv5, SSD, Faster R-CNN) using two configurations: De-Train B (100% synthetic defects) and De-Train C (1:1 hybrid real-synthetic). Tables 7-8 detail the mAP@0.5 metrics across five defect classes, with corresponding performance visualizations provided in Figs. 8-9. Table 7 and 8 demonstrates the persistent superiority of GAN-PD in the synthetic-only training setting. Across almost every comparative scenario it delivers the highest accuracy, and in the few non-optimal cases the shortfall to the best score is confined to roughly 0.1%. Under the hybrid training regime, GAN-PD elevates class-2 defect detection by average margins of 5.0% and 3.5% over SDGAN and EID-GAN, respectively. These gains stem from its bespoke deep-residual generator with local-feature convolution and columnar-pooling discriminator, both of which preserve scratch linearity in repetitive fabric textures. Even for the geometrically challenging class-4 defects, GAN-PD surpasses SDGAN and EID-GAN by 4.2% and 2.6%, substantiating its capability to model anisotropic defects. To further elucidate the merits of GAN-PD, the optimal detection results under each training regime are consolidated in Table 9. When trained on the hybrid dataset, the detector’s mean accuracy increases by 5.0%, 3.6%, and 5.6% for YOLOv5, SSD, and Faster R-CNN, respectively, relative to the synthetic-only baseline. Compared with models trained solely on real defects, the corresponding gains reach 10.1%, 9.3%, and 10.4%. These findings suggest that synthetic data may approximate defect variations that are difficult to sample in practice and thus potentially enhance detector generalization, while also indicating that such samples complement rather than supplant real data and could regularize the detector to attenuate empirical distribution bias. To establish a genuinely task-oriented assessment framework and to quantify how much GAN-PD augments detector generalisation in realistic manufacturing lines, we designed two complementary evaluation scenarios, De-Test B and De-Test C. These test sets probe models optimised under distinct training regimes yet confronted with operational conditions that deviate from their learning distributions. De-Test A, already used for baseline benchmarking, is deliberately omitted here to eliminate evaluative redundancy and channel analytic bandwidth toward synthesis-dominated domains. Fig. 10 collates the outcomes for every fabric-defect class across four testing configurations, three synthesis engines, and three mainstream detectors, yielding twelve radar-plot traces per subpanel. For clarity, each trace is indexed by the triplet “training set–test set–generator”; for example, “B-B-GAN-PD” denotes a detector trained exclusively on De-Train B, evaluated on De-Test B, with synthetic samples provided by GAN-PD. This multi-axis experimental matrix is intentionally complex: only by intersecting heterogeneous training distributions with varied test distributions can we expose the latent strengths and weaknesses of synthetic augmentation under production-scale variability. Furthermore, aggregating complementary detectors within the same test cell reveals whether augmentation bias persists across architectural paradigms–an aspect seldom explored in prior textile-inspection literature. By jointly analysing the radial profiles and the surface area enclosed by each trace, practitioners can directly gauge which synthesis strategy delivers the optimal trade-off between sensitivity and precision at fixed operating points, thereby informing acceptance thresholds for automatic optical-inspection hardware and guiding downstream corrective-action policies on the shop floor.Table 7Detection metrics for three generative models in De-Train B.ModelIndicator(mAP@0.5)YOLOv5SSDFaster-RCNNSDGAN\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class1}}$$\end{document}170.6770.5610.644EID-GAN\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class1}}$$\end{document}180.7050.5680.658GAN-PD\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class1}}$$\end{document}0.7360.5920.673SDGAN\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class2}}$$\end{document}170.7940.7620.843EID-GAN\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class2}}$$\end{document}180.8220.7650.850GAN-PD\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class2}}$$\end{document}0.8440.7890.860SDGAN\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class3}}$$\end{document}170.7320.7120.791EID-GAN\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class3}}$$\end{document}180.7540.7370.800GAN-PD\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class3}}$$\end{document}0.7710.7360.811SDGAN\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class4}}$$\end{document}170.5480.6120.563EID-GAN\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class4}}$$\end{document}180.5770.6420.585GAN-PD\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class4}}$$\end{document}0.5920.6570.609SDGAN\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class5}}$$\end{document}170.6530.6120.618EID-GAN\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class5}}$$\end{document}180.6590.6350.624GAN-PD\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class5}}$$\end{document}0.6780.6610.642Table 8Detection metrics for three generative models in De-Train C.ModelsIndicator(mAP@0.5)YOLOv5SSDFaster-RCNNSDGAN\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class1}}$$\end{document}170.7150.6000.681EID-GAN\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class1}}$$\end{document}180.7730.6090.682GAN-PD\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class1}}$$\end{document}0.7950.6200.744SDGAN\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class2}}$$\end{document}170.8270.8030.846EID-GAN\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class2}}$$\end{document}180.8430.8140.863GAN-PD\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class2}}$$\end{document}0.8890.8440.893SDGAN\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class3}}$$\end{document}170.7910.7700.843EID-GAN\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class3}}$$\end{document}180.8020.7750.846GAN-PD\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class3}}$$\end{document}0.8080.7860.869SDGAN\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class4}}$$\end{document}170.5960.6690.631EID-GAN\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class4}}$$\end{document}180.6180.6880.636GAN-PD\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class4}}$$\end{document}0.6490.7120.660SDGAN\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class5}}$$\end{document}170.7050.6470.681EID-GAN\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class5}}$$\end{document}180.7120.6520.690GAN-PD\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class5}}$$\end{document}0.7310.6740.712 Table 7Detection metrics for three generative models in De-Train B.ModelIndicator(mAP@0.5)YOLOv5SSDFaster-RCNNSDGAN\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class1}}$$\end{document}170.6770.5610.644EID-GAN\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class1}}$$\end{document}180.7050.5680.658GAN-PD\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class1}}$$\end{document}0.7360.5920.673SDGAN\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class2}}$$\end{document}170.7940.7620.843EID-GAN\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class2}}$$\end{document}180.8220.7650.850GAN-PD\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class2}}$$\end{document}0.8440.7890.860SDGAN\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class3}}$$\end{document}170.7320.7120.791EID-GAN\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class3}}$$\end{document}180.7540.7370.800GAN-PD\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class3}}$$\end{document}0.7710.7360.811SDGAN\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class4}}$$\end{document}170.5480.6120.563EID-GAN\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class4}}$$\end{document}180.5770.6420.585GAN-PD\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class4}}$$\end{document}0.5920.6570.609SDGAN\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class5}}$$\end{document}170.6530.6120.618EID-GAN\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class5}}$$\end{document}180.6590.6350.624GAN-PD\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class5}}$$\end{document}0.6780.6610.642 Detection metrics for three generative models in De-Train B. Table 8Detection metrics for three generative models in De-Train C.ModelsIndicator(mAP@0.5)YOLOv5SSDFaster-RCNNSDGAN\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class1}}$$\end{document}170.7150.6000.681EID-GAN\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class1}}$$\end{document}180.7730.6090.682GAN-PD\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class1}}$$\end{document}0.7950.6200.744SDGAN\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class2}}$$\end{document}170.8270.8030.846EID-GAN\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class2}}$$\end{document}180.8430.8140.863GAN-PD\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class2}}$$\end{document}0.8890.8440.893SDGAN\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class3}}$$\end{document}170.7910.7700.843EID-GAN\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class3}}$$\end{document}180.8020.7750.846GAN-PD\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class3}}$$\end{document}0.8080.7860.869SDGAN\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class4}}$$\end{document}170.5960.6690.631EID-GAN\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class4}}$$\end{document}180.6180.6880.636GAN-PD\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class4}}$$\end{document}0.6490.7120.660SDGAN\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class5}}$$\end{document}170.7050.6470.681EID-GAN\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class5}}$$\end{document}180.7120.6520.690GAN-PD\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class5}}$$\end{document}0.7310.6740.712 Detection metrics for three generative models in De-Train C. Fig. 8Radar charts visualizing detection metrics for three generative models in De-Train B. Table 9The best metrics for detecting models in three training datasets.Training datasetIndicator(mAP@0.5)YOLOv5SSDFaster-RCNNDe-Train A\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class1}}$$\end{document}0.6670.5450.634De-Train B\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class1}}$$\end{document}0.7360.5920.673De-Train C\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class1}}$$\end{document} 0.795 0.620 0.744 De-Train A\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class2}}$$\end{document}0.7840.7350.791De-Train B\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class2}}$$\end{document}0.8440.7890.860De-Train C\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class2}}$$\end{document} 0.889 0.844 0.893 De-Train A\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class3}}$$\end{document}0.7230.6610.754De-Train B\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class3}}$$\end{document}0.7710.7370.811De-Train C\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class3}}$$\end{document} 0.808 0.766 0.869 De-Train A\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class4}}$$\end{document}0.5420.5960.552De-Train B\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class4}}$$\end{document}0.5920.6570.609De-Train C\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class4}}$$\end{document} 0.649 0.712 0.660 De-Train A\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class5}}$$\end{document}0.6540.6140.629De-Train B\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class5}}$$\end{document}0.6780.6610.642De-Train C\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class5}}$$\end{document} 0.731 0.674 0.712 Radar charts visualizing detection metrics for three generative models in De-Train B. Table 9The best metrics for detecting models in three training datasets.Training datasetIndicator(mAP@0.5)YOLOv5SSDFaster-RCNNDe-Train A\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class1}}$$\end{document}0.6670.5450.634De-Train B\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class1}}$$\end{document}0.7360.5920.673De-Train C\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class1}}$$\end{document} 0.795 0.620 0.744 De-Train A\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class2}}$$\end{document}0.7840.7350.791De-Train B\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class2}}$$\end{document}0.8440.7890.860De-Train C\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class2}}$$\end{document} 0.889 0.844 0.893 De-Train A\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class3}}$$\end{document}0.7230.6610.754De-Train B\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class3}}$$\end{document}0.7710.7370.811De-Train C\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class3}}$$\end{document} 0.808 0.766 0.869 De-Train A\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class4}}$$\end{document}0.5420.5960.552De-Train B\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class4}}$$\end{document}0.5920.6570.609De-Train C\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class4}}$$\end{document} 0.649 0.712 0.660 De-Train A\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class5}}$$\end{document}0.6540.6140.629De-Train B\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class5}}$$\end{document}0.6780.6610.642De-Train C\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class5}}$$\end{document} 0.731 0.674 0.712 The best metrics for detecting models in three training datasets. Fig. 9Radar charts visualizing detection metrics for three generative models in De-Train C. Radar charts visualizing detection metrics for three generative models in De-Train C. To establish a genuinely task-oriented assessment framework and to quantify how much GAN-PD augments detector generalisation in realistic manufacturing lines, we designed two complementary evaluation scenarios, De-Test B and De-Test C. These test sets probe models optimised under distinct training regimes yet confronted with operational conditions that deviate from their learning distributions. De-Test A, already used for baseline benchmarking, is deliberately omitted here to eliminate evaluative redundancy and channel analytic bandwidth toward synthesis-dominated domains. Figure 10 collates the outcomes for every fabric-defect class across four testing configurations, three synthesis engines, and three mainstream detectors, yielding twelve radar-plot traces per subpanel. For clarity, each trace is indexed by the triplet “training set–test set–generator”; for example, “B-B-GAN-PD” denotes a detector trained exclusively on De-Train B, evaluated on De-Test B, with synthetic samples provided by GAN-PD. This multi-axis experimental matrix is intentionally complex: only by intersecting heterogeneous training distributions with varied test distributions can we expose the latent strengths and weaknesses of synthetic augmentation under production-scale variability. Furthermore, aggregating complementary detectors within the same test cell reveals whether augmentation bias persists across architectural paradigms–an aspect seldom explored in prior textile-inspection literature. By jointly analysing the radial profiles and the surface area enclosed by each trace, practitioners can directly gauge which synthesis strategy delivers the optimal trade-off between sensitivity and precision at fixed operating points, thereby informing acceptance thresholds for automatic optical-inspection hardware and guiding downstream corrective-action policies on the shop floor. Across 540 experimental configurations encompassing five defect classes, three generative models, three detectors, two training regimes, two test sets, and three evaluation metrics, the hybrid implementation of GAN-PD consistently achieved the best performance. Under YOLOv5, it raised mean precision by 1.1% and 1.3% over EID-GAN and SDGAN, respectively, boosted mean recall by 2.3% and 2.7%, and cut the mean false-positive rate by 1.2% and 1.8%. In SSD experiments, mean precision increased by 1.3% and 1.1%, mean recall by 2.7% and 2.0%, while the mean false-positive rate declined by 1.9%. With Faster R-CNN, mean precision improved by 1.3% and 1.6%, mean recall uniformly rose by 3.4%, and the mean false-positive rate fell by 1.0% and 2.3%. These findings demonstrate that the generated defect samples confer measurable benefits to both one-stage and two-stage detectors.Table 10Best metrics for three detection models in the benchmark model and De-Train C training set.DatasetModelYOLOv5SSDFaster-RCNNIndicatorDDADDRFDRDDADDRFDRDDADDRFDRA-A\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class1}}$$\end{document}89.4%68.3%0.0%81.7%46.7%3.4%87.8%63.3%0.0%C-B\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class1}}$$\end{document}91.7%75.0%0.0%90.6%73.3%2.2%91.7%76.7%2.1%C-C\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class1}}$$\end{document}94.4%83.3%0.0%91.1%75.0%2.2%92.8%80.0%2.0%A-A\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class2}}$$\end{document}93.9%81.7%0.0%92.2%76.7%0.0%91.7%75.0%0.0%C-B\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class2}}$$\end{document}97.2%91.7%0.0%94.4%83.3%0.0%94.4%83.3%0.0%C-C\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class2}}$$\end{document}98.9%96.7%0.0%95.0%85.0%0.0%95.6%86.7%0.0%A-A\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class3}}$$\end{document}89.4%73.3%6.4%88.3%71.7%8.5%90.6%75.0%4.3%C-B\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class3}}$$\end{document}93.3%81.7%2.0%91.7%78.3%4.1%96.1%88.3%0.0%C-C\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class3}}$$\end{document}94.4%85.0%1.9%92.8%81.7%3.9%96.7%90.0%0.0%A-A\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class4}}$$\end{document}79.4%46.7%15.2%81.1%50.0%11.8%80.6%48.3%12.1%C-B\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class4}}$$\end{document}84.4%56.7%5.6%85.6%58.3%2.8%85.6%58.3%2.8%C-C\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class4}}$$\end{document}85.0%58.3%5.4%86.7%61.7%2.6%86.1%60.0%2.7%A-A\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class5}}$$\end{document}86.7%68.3%10.9%85.0%66.7%14.9%86.1%68.3%12.8%C-B\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class5}}$$\end{document}91.7%75.0%0.0%90.6%73.3%2.2%91.7%76.7%2.1%C-C\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class5}}$$\end{document}93.3%80.0%0.0%91.1%75.0%2.2%92.8%80.0%2.0% Table 10Best metrics for three detection models in the benchmark model and De-Train C training set.DatasetModelYOLOv5SSDFaster-RCNNIndicatorDDADDRFDRDDADDRFDRDDADDRFDRA-A\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class1}}$$\end{document}89.4%68.3%0.0%81.7%46.7%3.4%87.8%63.3%0.0%C-B\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class1}}$$\end{document}91.7%75.0%0.0%90.6%73.3%2.2%91.7%76.7%2.1%C-C\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class1}}$$\end{document}94.4%83.3%0.0%91.1%75.0%2.2%92.8%80.0%2.0%A-A\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class2}}$$\end{document}93.9%81.7%0.0%92.2%76.7%0.0%91.7%75.0%0.0%C-B\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class2}}$$\end{document}97.2%91.7%0.0%94.4%83.3%0.0%94.4%83.3%0.0%C-C\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class2}}$$\end{document}98.9%96.7%0.0%95.0%85.0%0.0%95.6%86.7%0.0%A-A\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class3}}$$\end{document}89.4%73.3%6.4%88.3%71.7%8.5%90.6%75.0%4.3%C-B\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class3}}$$\end{document}93.3%81.7%2.0%91.7%78.3%4.1%96.1%88.3%0.0%C-C\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class3}}$$\end{document}94.4%85.0%1.9%92.8%81.7%3.9%96.7%90.0%0.0%A-A\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class4}}$$\end{document}79.4%46.7%15.2%81.1%50.0%11.8%80.6%48.3%12.1%C-B\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class4}}$$\end{document}84.4%56.7%5.6%85.6%58.3%2.8%85.6%58.3%2.8%C-C\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class4}}$$\end{document}85.0%58.3%5.4%86.7%61.7%2.6%86.1%60.0%2.7%A-A\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class5}}$$\end{document}86.7%68.3%10.9%85.0%66.7%14.9%86.1%68.3%12.8%C-B\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class5}}$$\end{document}91.7%75.0%0.0%90.6%73.3%2.2%91.7%76.7%2.1%C-C\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$_{\text {Class5}}$$\end{document}93.3%80.0%0.0%91.1%75.0%2.2%92.8%80.0%2.0% Best metrics for three detection models in the benchmark model and De-Train C training set. Table 10 consolidates optimal performance metrics, contrasting the baseline established with exclusively real data (A-A configuration) against deployments utilizing synthetic-only (C-B) and hybrid (C-C) training datasets. These comprehensive results underscore GAN-PD’s efficacy in addressing the core challenges of synthesizing geometrically anisotropic defects and achieving their contextual integration within complex fabric patterns. For instance, in the detection of linear scratch defects, exemplified by Class 2 and Class 4, the YOLOv5 detector, when trained using the C-C configuration, achieved Defect Detection Rates of 96.7% and 58.3%, respectively. These figures represent absolute improvements of 15.0 and 11.6 percentage points over the A-A baseline DDRs of 81.7% and 46.7%. This marked performance enhancement directly validates the capability of our direction-sensitive columnar pooling discriminator to preserve millimeter-scale linear continuity, effectively mitigating the fragmentation artifacts typically produced by traditional isotropic convolutions, a limitation highlighted in Section Introduction. Critically, the False Detection Rate (FDR) for Class 4 defects, averaged across all detector architectures, was substantially reduced from approximately 12.1% in the A-A setup to 2.7% in the C-C setup, indicating that the local feature convolution mechanism within our LFR generator facilitates accurate texture fusion and ensures consistent defect integration with the surrounding fabric context. The utility of Table 10 extends beyond performance benchmarking; it systematically quantifies how our proposed task-aware evaluation protocol translates improvements in synthetic data quality into tangible industrial utility. Specifically, when employing the YOLOv5 model for the five defect classes, training augmentation with GAN-PD (C-C configuration) resulted in an average increase in DDA of 5.4 percentage points, an average DDR improvement of 13.0 percentage points, and an average FDR reduction of 5.0 percentage points, relative to the A-A baseline. Analogous enhancements were observed with the SSD detector (average DDA +4.9%, DDR +10.7%, FDR −6.0%) and the Faster R-CNN detector (average DDA +5.3%, DDR +12.7%, FDR −4.9%). This consistent uplift in performance is attributable to the generator’s proficiency in synthesizing features that are salient for detection, rather than merely mimicking visual realism, a principle central to our advocated paradigm shift from “defect generation” to “generation for defect detection.” The quantifiable decrease in FDR, in particular, offers a direct benefit by mitigating false alarms in production line inspection systems. Further analysis comparing the C-C and C-B configurations reveals that models trained exclusively on synthetic data (C-B) generally exhibit lower performance metrics than those trained on hybrid datasets (C-C). This observation suggests that GAN-PD-generated samples function as a potent complement to, rather than a complete substitute for, authentic data. This finding echoes the critique of over-reliance on conventional fidelity metrics, as discussed in Section Introduction: operational efficacy is not solely guaranteed by high visual fidelity unless the synthetic data distributions are appropriately anchored by, and aligned with, real-world examples. Consequently, orienting synthetic defect generation towards enhancing detection effectiveness, rather than pursuing standalone visual fidelity, is paramount. The substantial improvements in DDR and concurrent reductions in FDR achieved through GAN-PD underscore that such task-aware synthesis, not merely enhanced fidelity, is the critical driver for advancing defect detection accuracy in industrial textiles. This approach directly addresses the persistent gap between synthetic data quality and its practical industrial applicability.Fig. 10Radar-chart comparison of detector performance–DDA, DDR, and FDR–for the five DAGM2007 fabric-defect classes under four train/test configurations, three synthetic generators, and three detectors. Each trace is indexed by its training-set–test-set–generator triad (e.g., B-B-GAN-PD), enabling direct visualisation of how different augmentation strategies shift the precision–recall trade-off across heterogeneous operating conditions. Radar-chart comparison of detector performance–DDA, DDR, and FDR–for the five DAGM2007 fabric-defect classes under four train/test configurations, three synthetic generators, and three detectors. Each trace is indexed by its training-set–test-set–generator triad (e.g., B-B-GAN-PD), enabling direct visualisation of how different augmentation strategies shift the precision–recall trade-off across heterogeneous operating conditions. Conclusion This paper has introduced GAN-PD, a generative architecture expressly designed to produce fabric-surface anomalies that improve automated inspection rather than merely achieve visual plausibility. By integrating Local-Feature Residual modules within a deep residual generator and enforcing anisotropic consistency through a columnar-pooling discriminator, the framework synthesizes elongated scratches and fine-grained blemishes that blend seamlessly with complex woven textures. A task-aware evaluation protocol confirms that these detector-centric samples translate into higher recognition accuracy and fewer false alarms across both one-stage and two-stage detectors, while an efficient parameterization ensures the framework is computationally viable for large-scale offline augmentation on standard industrial servers. Despite these promising results, practical limitations include the current need for manual annotation of generated samples and the computational overhead associated with training, alongside the ongoing challenge of capturing the complete spectrum of defect morphologies. Future research will therefore concentrate on integrating automated annotation techniques, validating the framework on high-variability proprietary industrial datasets, and developing more computationally efficient, lightweight adversarial modules. These efforts aim to expand GAN-PD’s applicability, further propelling the development of comprehensive, real-time defect inspection systems for the textile industry through high-fidelity offline data augmentation.