Colorless image processing technique combining chromaticity transfer and image feature extraction in visual graphic design Authors: Tingting Wu, Bing Liu Published: 2026 ABSTRACT AbstractThis study presents a verifiable framework for high-fidelity colorless image processing by integrating depth-guided chrominance transfer with multi-level feature extraction. To address critical limitations of traditional methods—color inconsistency, edge bleeding, and insufficient feature matching accuracy—we establish a dual-channel processing pipeline for RGB-D images. This study presents a high-fidelity colorless image processing framework through depth-guided chrominance transfer and multi-level feature extraction. The method introduces three key innovations: a weighted non-local Laplacian algorithm for cross-modal consistency, a multi-level feature hierarchy bridging pixel statistics to semantic understanding, and a brightness-guided edge preservation mechanism. Rigorous testing on benchmark datasets (Middlebury Stereo, MIT-Adobe FiveK, BSDS500, Flickr Million) under standardized preprocessing confirms its efficacy. The algorithm reduces color reconstruction error to 0.020 ± 0.003—a 60% improvement over the Welsh method—with a Turbo-pixel 800 edge fitting accuracy of 0.93 and over-segmentation of only 8%. The three-level feature system boosts matching recall to 91.4%±2.1% from 62.3%, while cutting semantic alignment error from 12.4 to 3.1 ± 0.7 pixels. On MIT-Adobe FiveK, it achieves a PSNR of 38.5 ± 0.4 dB, SSIM of 0.942 ± 0.008, and 9.4 FPS, outperforming diffusion-based colorization by 3.7 dB PSNR with 37% faster speed. All results are statistically significant (p < 0.05) and fully reproducible. This work provides a transparent, efficient, and robust solution for visual graphic design, establishing a new benchmark for colorless image processing.Keywords: Visual graphic design, Chromaticity transfer, Image feature extraction, Colorless image processing, Deep learningSubject terms: Engineering, Mathematics and computing Introduction In visual graphic design, colorless images such as grayscale or monochrome images (i.e. input objects in image colorization tasks) are widely used in posters, brand identities, and photographic art due to their unique artistic expressiveness and concise visual language1,2. However, traditional achromatic image processing techniques often rely on simple desaturation or threshold segmentation methods, resulting in loss of image detail and insufficient tonal gradation. This makes it difficult to meet the high demands for visual expressiveness in modern design3,4. In recent years, with the advancement of computer vision and image processing technologies, enhancement methods based on chrominance transfer and image feature extraction have provided new approaches for processing colorless images5,6. Chromaticity transfer technology enhances visual depth by mapping the color statistics of the reference image to the target image, which has been established in previous works7,8. Our method extends this principle by integrating deep guidance. In recent years, the development of technologies such as 3D visualization, video enhancement, generative adversarial networks, and image feature extraction has provided diverse means for visual content processing. Chen et al. upgraded 2D graphic design to 3D virtual visualization technology, constructing a design platform with panoramic information and 3D rendering capabilities, optimizing the image generation process, and improving the visual quality of images9. Wang et al. analyzed video sequence images and designed a motion flow propagation algorithm based on global matching and Transformer encoder, which improved the quality of video restoration while reducing computational complexity10. Elnabawy R H et al. addressed the challenges of manually drawing traditional clip art and its difficulty in automated generation by proposing a generative adversarial network (GAN)-based clip art generation model. This model could automatically convert photographs of objects into clip art. The results showed that the accuracy of the model was 58% and 54% higher than that of Cycle GAN and Google Image Search, respectively11. Image feature extraction optimizes the detail representation of colorless images across dimensions such as structure, texture, and contrast. Yang Z et al. designed a visual communication system to address the issue of poor motion continuity in animated characters within virtual reality. Experimental results demonstrated that the system generated more realistic and coherent animated characters than existing methods12. Wang F et al. proposed an enhancement method based on Sobel operator filtering to address the issues of time-consuming processing, detail loss, and poor visual quality in traditional flat advertising image enhancement techniques. Experiments demonstrated that this method could significantly enhance images within a short timeframe while maintaining high entropy values to preserve rich details13. Liu Z et al. proposed a deep learning-based GAN for visual saliency feature extraction to address the challenges of insufficient training samples and imprecise visual feature extraction in graphic design plagiarism detection. Experimental results indicated that the model outperformed various similarity metrics, effectively enhancing the accuracy and reliability of design plagiarism detection14. The above studies have made significant progress in the fields of 3D scene reconstruction, dynamic video processing, automated graphics generation, and image detail optimization. However, they mainly focus on the processing of color images or specific visual content, and have not systematically solved the coloring problem of colorless images, especially the lack of joint utilization of depth geometric information and multi-level semantic features. Therefore, this study proposes an organic fusion of deep geometric information and multi-level semantic features to construct a coloring method that can assign reasonable and visually rich colors to colorless images while maintaining the original structure. The research scope focuses on the automatic coloring task of grayscale/monochrome images, driven by the gap between the existing methods’ blank space in depth semantic joint utilization and the modern design’s demand for high expressive coloring. This study addresses three key challenges in colorless image processing (i.e. grayscale/monochrome image coloring): color inconsistency caused by isolated pixel processing in coloring results (manifested as color bleeding, mottled effects, and other artifacts), limited matching accuracy due to single level feature extraction, and insufficient utilization of depth color correlation. Among them, single-level feature extraction refers to relying solely on single level features for matching, such as using only pixel intensity or texture descriptors, lacking the ability to represent images across scales from low-level statistics to high-level semantics. The key innovations of the proposed method are reflected in three aspects: (1) theoretical innovation, proposing a new idea of using depth information as geometric constraints to guide color propagation, elevating depth from additional features to structural priors, and solving the problem of inconsistent cross modal features. (2) Technological innovation, constructing an optimization framework that combines weighted non local Laplacian and locally consistent sparse models to achieve joint modeling of deep constraints, multi-level features, and neighborhood collaboration. (3) Applied innovation, aimed at colorless image coloring tasks, establishes a complete processing chain from depth guidance, superpixel segmentation to brightness post filtering, achieving collaborative enhancement between structural fidelity and color richness. These innovations collectively alleviate traditional limitations by ensuring global-to-local consistency through superpixel constraints, guaranteeing matching accuracy via sparse representations, and enhancing detail rendering through brightness guidance, forming an integrated multi strategy comprehensive processing scheme. Modern visual design has put forward higher requirements for the color expression of colorless images, while traditional coloring methods have obvious limitations in detail preservation, structural consistency, and semantic accuracy, and urgently need new technological paths to break through. This study aims to achieve high fidelity coloring of colorless images (grayscale/monochrome), striking a balance between structural integrity and color richness, and proposing a new scheme that differs from existing RGB-D stitching, global color transfer, and isolated sparse encoding paradigms. The core scientific contribution of this study lies in: (1) revealing the geometric constraint effect of depth information in color propagation, and theoretically solving the structural distortion problem caused by cross modal feature inconsistency by constructing a depth color joint weighting model. (2) A multi-level feature hierarchical representation framework based on superpixel units has been established, which organically integrates pixel statistics, texture description, and semantic understanding. The system improves the accuracy and semantic coherence of color matching in complex scenes. (3) The guiding mechanism of brightness gradient on chromaticity propagation was elucidated. By coupling the optimization objectives of brightness and chromaticity, color overflow and edge blurring were theoretically suppressed, achieving a synergistic enhancement of structural integrity and color richness. Methods and materials Deep guidance-based chrominance transfer technology Colorless image processing relies on three theoretical principles: cross modal information complementarity, the necessity of multi-scale feature representation, and the inevitability of global local collaborative optimization. Based on these principles, the framework integrates depth information, superpixel structure, multi-level features, sparse encoding alternating direction method of multipliers (ADMM) Optimization and brightness filtering are organically integrated, and each component cooperates with each other to form a complete inference chain. Specifically, depth information provides geometric constraints and assigns structural priors to superpixel segmentation. Hyperpixels organize pixels into structural units, enabling multi-level feature extraction to be performed on semantic regions and compensating for the insufficient accuracy of single level feature matching. Multi level features can provide rich representations for sparse coding, but isolated coding is prone to local inconsistency. Therefore, graph Laplacian regularization is introduced to force sparse coding similarity of neighboring superpixels, and ADMM efficiently solves this constraint optimization. Finally, brightness post filtering utilizes the original brightness gradient to guide chromaticity propagation, compensating for the edge blurring that may be caused by sparse reconstruction. Each module is interconnected, achieving collaborative optimization from global to local, and from coarse to fine. The selection of weighted non local Laplacian algorithm and Turbo pixel superpixel algorithm is driven by the following three practical considerations to compensate for the shortcomings of traditional methods. First, traditional local Laplacian filters and SLIC superpixels suffer from edge oversegmentation and insufficient texture adaptation in complex scenes. The weighted nonlocal Laplacian algorithm introduces a novel weighting function that integrates both color similarity and depth similarity, enabling cross-modal fusion that prevents color bleeding while preserving structural coherence. Second, real-time processing demands necessitate computational efficiency. The computational experiments were performed on the university’s shared high-performance computing cluster, which provides authorized access to NVIDIA Tesla V100 GPUs for interdisciplinary research projects, ensuring legitimate access to the required computational resources. Turbo-pixels significantly outperform SLIC in edge retention with comparable processing time, forming dense irregular grids that accurately follow complex boundaries rather than imposing geometrically uniform segments. Third, the integration of these two algorithms establishes a synergistic “global-local” optimization framework: the nonlocal Laplacian provides global color-depth correlation constraints, while Turbo-pixel superpixels serve as structure-aware units for local feature extraction and sparse representation, collectively enabling precise color reconstruction from coarse to fine. In the field of visual graphic design, colorless image processing technology is a critical component for achieving high-quality visual effects15,16. Traditional methods primarily rely on single-color information or simple feature matching, making it challenging to meet color reconstruction demands in complex scenes17,18. With the widespread adoption of RGB-D image capture devices, the correlation between depth information and color features offers new possibilities for enhancing processing performance19,20. However, existing techniques still exhibit significant shortcomings in maintaining color consistency, controlling edge sharpness, and achieving feature matching accuracy. This often results in processing artifacts such as color bleeding, edge blurring, and texture distortion21,22. To address these challenges, this study proposes an innovative approach that integrates depth-guided chrominance transfer with local consistency feature extraction. The methodology is transparently detailed through specific algorithmic formulations and visual workflows. This method employs a dual optimization strategy: the first stage establishes a fundamental chromaticity distribution through depth guidance. The second stage refines edges using local consistency constraints and luminance guidance, forming a closed-loop optimization process. Traditional chromaticity transfer methods rely solely on color information, overlooking the strong correlation between depth features and color. This study constructs a joint processing framework using RGB-D images generated by active depth sensors (such as structured light cameras or stereo matching devices) that have achieved pixel level registration as input, as shown in Fig. 1. RGB-D image processing framework. Figure 1 illustrates the core workflow of the RGB-D image processing framework. Input data is first separated into two parallel processing channels: the red, green, blue (RGB) channel and the depth map (Depth) channel. The RGB branch computes saliency and entropy features, while the Depth branch computes entropy features. Subsequently, multi-source information is integrated through a cross-modal fusion module, with feature extraction and classification tasks ultimately performed by a convolutional neural network23,24. Analysis of depth sensor data revealed that, within neighboring pixels, regions with similar colors typically exhibit similar depth values, and vice versa. This correlation can be expressed as Eq. (1). In Eq. (1), represents continuous multiplication operation, and k is the modal index. R(d,c) denotes the joint correlation metric between the depth channel d and the color channel c at spatial positions p and q. represents the distance metric function. denotes a positive real parameter. Traditional RGB-D processing frameworks can extract bimodal features, but direct fusion overlooks the intrinsic correlation between depth and color, resulting in insufficient cross-modal matching accuracy25,26. Research indicates that regions of depth similarity often correspond to color consistency. Therefore, a weighted nonlocal Laplacian algorithm is proposed, featuring a novel weighting function that integrates depth and color differences. This joint constraint enhances feature matching accuracy, addressing inaccuracies in cross-modal information fusion. To address the color reconstruction error caused by ignoring depth information, we propose a new weighting function W, which explicitly combines the depth channel similarity with the color similarity. The function definition is shown in Eq. (2). In Eq. (2), and denote n-dimensional feature vectors. and denote m-dimensional descriptor vectors. represents cross-modal adaptive parameters. This weighting function is then incorporated into the nonlocal Laplacian framework, leading to the optimization objective in Eq. (3) which ensures consistency in both color and depth spaces. In Eq. (3), v denotes the feature field to be optimized. represents the special domain. Ω denotes the original domain. denotes the neighborhood system of point p. λ denotes the equilibrium parameter. By solving the corresponding linear system, precise color reconstruction guided by depth information is achieved. Image feature extraction and matching based on local consistency Depth-guided chrominance transfer techniques have achieved preliminary color reconstruction by modeling depth-color correlations to address fundamental cross-modal matching challenges. However, these methods retain two critical limitations: First, isolated pixel processing leads to local color inconsistencies, particularly manifesting as mottled effects in texturally complex regions. Second, the lack of understanding regarding image semantic structure causes color bleeding at critical edges27,28. To this end, the study proposes a local consistency-based image feature extraction and matching methodology comprising four core technologies: Superpixel segmentation technology establishes structure-aware processing units. Multi-level feature extraction integrates low-level intensity statistics, mid-level texture descriptors, and high-level semantic features: The base layer captures pixel-level intensity statistics to ensure color accuracy. The middle layer uses texture descriptors to maintain structural consistency. The top layer incorporates semantic features from a pre-trained saliency detection network to preserve semantic coherence. Feature vectors from all three levels are concatenated into a unified descriptor for each superpixel. The local consistency sparse representation model introduces three key innovations that fundamentally differ from traditional isolated sparse coding. First, a graph Laplacian regularizer incorporates spatial adjacency constraints directly into the objective function, ensuring adjacent superpixels with similar features share similar sparse codes. Second, a learnable adaptive weighting matrix dynamically balances contributions from different feature modalities (intensity, texture, saliency) based on local context to prevent any single feature from dominating. Third, an entropy-based sparsity constraint promotes structured sparsity patterns aligned with image semantics, enhancing both representation efficiency and interpretability beyond simple element-wise sparsity. Governed by the novel objective function in Eq. 4, which simultaneously minimizes reconstruction error while enforcing local consistency and structured sparsity, this model forms a critical part of the coarse-to-fine processing chain. The local consistency-based image feature extraction and matching methodology is illustrated in Fig. 2. Method system of image feature extraction and matching based on local consistency. In Fig. 2, the target and reference images are first segmented into 800 superpixel structural units using the Turbo-pixel superpixel algorithm (Turbo-pixel). Extract three-level features for each superpixel, including a 25 dimensional intensity histogram, a 64 dimensional DAISY texture reduced by PCA, and a 1-dimensional saliency response mean, cascaded to form a 90 dimensional unified descriptor. The reference image branch constructs a feature dictionary based on this. In the matching stage, a local consistency sparse model is adopted, which constrains the encoding similarity of neighboring superpixels through the Turalaplus term and introduces entropy regularization to promote structured sparsity. The model is decomposed by ADMM into three sub problems: sparse reconstruction, graph smoothing, and entropy constraint, which are solved alternately. The maximum number of iterations is 50, and the convergence threshold is coefficient variation norm below 0.0001. The obtained sparse coefficients are weighted to generate the initial chromaticity field, and finally the edge sharpened coloring result is output through brightness guided joint filtering (with edge preservation intensity set to 0.5). To address pixel-level processing inefficiencies, traditional methods employ simple linear iterative clustering (SLIC) for segmentation, yet suffer from edge oversegmentation29,30. Therefore, the study selects the Turbo-pixel, which significantly enhances processing efficiency while maintaining edge accuracy by adjusting the number of seed points and iteration counts. The Turbo-pixel algorithm is illustrated in Fig. 3. Schematic diagram of segmentation comparison between Slic and Turbo-pixel superpixel algorithms. As shown in Fig. 3, the SLIC algorithm generates a regular sparse grid structure with geometrically uniform segmentation boundaries but insufficient detail adaptation. In contrast, the Turbo-pixel algorithm forms a dense irregular grid that more accurately follows the edges of petal textures. Turbo-pixel demonstrates superior edge retention in complex shape segmentation, validating its suitability as a preferred solution for texture image processing. While superpixel segmentation establishes structural units, its single-scale features struggle to accurately capture characteristics in complex regions. To address the issues of highly variable internal textures and ambiguous semantics in superpixels, this study proposes a multi-level feature extraction method. By integrating pixel-level intensity statistics, local texture descriptions, and global saliency features, it achieves cross-scale feature representation from micro to macro levels, thereby enhancing matching accuracy and color consistency in complex scenes. The multi-level feature extraction process is illustrated in Fig. 4. Multi-level feature extraction. Figure 4 illustrates that the base layer captures pixel-level statistical properties through mean, standard deviation, local contrast, and a 25-bin histogram. The middle layer employs the DAISY descriptor, which undergoes principal component analysis (PCA) for dimensionality reduction to encode local texture structures. The top layer extracts semantic features based on saliency detection, reflecting visual attention mechanisms. The three-level feature hierarchy forms a complete description system from bottom-up, spanning micro-level intensity distribution to macro-level semantic understanding. Low-level features ensure color accuracy, mid-level features maintain texture consistency, and high-level features preserve semantic coherence. Their collaborative operation enhances adaptability in complex scenes. Subsequently, to address the inconsistency in matching caused by traditional sparse representations processing superpixels in isolation, this study proposes a local consistency sparse model, as shown in Eq. (4). In Eq. (4), K represents the target variable. denotes the target feature matrix. Q signifies the reference dictionary matrix. represents the similarity weight. \varepsilon and α denote the regularization parameters. γ denotes the set of superpixel adjacency relationships. represents the entropy norm. When solving ADMM, the objective function shown in Eq. (4) is divided into three sub problems. The sparse encoding sub problem is solved using orthogonal matching tracing, the graph regularization sub problem is updated using conjugate gradient method, and the entropy constrained sub problem is approximated using a soft threshold operator. The sub problems alternate iterations until the coefficient matrix change norm is less than 0.0001, usually converging within 40 iterations. This strategy controls the computational complexity at a level linearly related to the number of superpixels while ensuring solution accuracy. The similarity weight is defined by Eq. (5). In Eq. (5), represents the spatial coordinates of superpixel m. denotes the multi-level feature vector. and represent the spatial/feature bandwidth parameters. The local consistency sparse matching is illustrated in Fig. 5. Local consistent sparse matching. Figure 5 illustrates the core mechanism of local consistency sparse matching. Feature points within the target object are transformed into a feature matrix through feature extraction. This matrix is decomposed into a linear combination of a dictionary matrix and a sparse coefficient matrix. The sparse coefficient matrix exhibits a distinct sparse distribution of non-zero elements. This decomposition effectively achieves the mapping from feature space to dictionary space. Although the local consistency sparse model has established a theoretical framework, traditional optimization algorithms are difficult to effectively solve this kind of convex optimization problem with complex constraints. To address this computational bottleneck, the study employs the ADMM to decompose the problem into manageable sub-problems, which can be effectively solved in parallel. This method ensures the solution accuracy while guaranteeing rapid convergence, significantly reduces the computational complexity, and endows the model with practical application value. Finally, to address potential edge blurring issues during optimization, the system incorporates a brightness-guided edge preservation mechanism. This mechanism establishes a coupling constraint between brightness gradients and chrominance gradients, as shown in Eq. (6). In Eq. (6), l denotes the luminance channel. ρ represents the control edge retention intensity. e represents the color field to be estimated. represents the sparse reconstruction of the initial chromaticity field. This luminance-guided constraint mechanism ensures chromaticity propagation strictly follows the structural characteristics of the luminance image by coupling chromaticity with luminance gradients. It prevents color bleeding in regions with strong edges while preserving smooth transitions in areas with weak texture. In summary, the proposed colorless image processing technique employs a combined methodology of depth-guided chrominance transfer and multi-level feature extraction. The core technologies include: first, the RGB-D dual-channel processing framework is established, achieving cross-modal fusion of depth and color through the weighted nonlocal Laplacian algorithm. Second, Turbo-pixel superpixel segmentation is utilized to establish structure-aware units, integrating three levels of features: intensity histogram, DAISY texture, and saliency detection. Next, a sparse representation model for local consistency is designed, solving the constrained convex optimization problem via the ADMM optimization algorithm. Finally, a brightness-guided edge preservation mechanism is introduced, coupling chrominance and luminance gradients to prevent edge bleeding. This method ensures local consistency through superpixel constraints and guarantees matching accuracy with a sparse representation. It also enhances detail rendering via brightness guidance. Together, these features form a complete processing chain from coarse to fine. The input of the proposed colorless image processing method is a registered grayscale/monochrome image and its corresponding depth map (depth channel in RGB-D), which only contains brightness and geometric structure, and lacks color information. The goal is to transfer reasonable colors from the reference image, reconstruct color channels, and ultimately output visually rich and structurally complete color images through multi-level feature extraction (intensity, texture, saliency) and cross modal matching. The features such as saliency and texture mentioned in the article are obtained through computational analysis of the input data, and are not pre-existing additional input information. The input data of the proposed colorless image processing method only includes brightness and geometric structure, lacking color information. Features such as saliency heatmaps, DAISY textures, and intensity histograms are obtained through computational analysis of the input image, rather than pre-existing additional information. Therefore, the purpose of preprocessing (i.e. feature extraction) is to obtain descriptive features such as saliency and texture from limited inputs, laying the foundation for subsequent color transfer. Experimental datasets To ensure the reproducibility and verifiability of the experimental results, this study utilized four publicly available benchmark datasets. The key specifications, purposes, and preprocessing details for each dataset are comprehensively outlined below. Middlebury Stereo Standard RGB-D Dataset: 100 samples were selected from this dataset to validate the depth-guided chrominance transfer technology. The images were used at their native resolution as provided by the benchmark to maintain the integrity of depth-color correlations. Direct Link: https://vision.middlebury.edu/stereo/data/. MIT-Adobe FiveK Dataset: This dataset was employed to evaluate the performance of the brightness-guided edge preservation mechanism. All images were standardized to a resolution of 512 × 512 pixels to ensure consistent processing conditions. Direct Link: https://data.csail.mit.edu/graphics/fivek/. Berkeley Segmentation Dataset 500 (BSDS500): It was utilized for comparing the performance of SLIC and Turbo-pixel superpixel segmentation algorithms, providing a ground truth for evaluating edge fitting accuracy and segmentation quality. Direct Link: https://www2.eecs.berkeley.edu/Research/Projects/CS/vision/bsds/. Flickr Million Dataset: A curated subset of this large-scale collection was used for conducting extensive feature ablation experiments on the proposed multi-level feature system, testing its robustness across diverse visual content. Direct Link: https://multimediacommons.org/. All images across the datasets were preprocessed by converting the color space to CIE 1976 L*a*b*, which is optimal for color processing tasks. Access to these datasets (Middlebury, MIT-Adobe FiveK, BSDS500, Flickr Million) was obtained through official academic licensing agreements managed by the university’s research data management office, ensuring full compliance with terms of use for non-commercial academic research. Standard training/validation/test splits for each public dataset were followed to align with established benchmarking practices. Although the above datasets are not specifically designed for colorless image processing or RGB-D coloring, their core characteristics make them widely used as benchmarks in their respective research fields. Specifically, the Middlebury dataset is used to evaluate depth estimation accuracy in the fields of stereo matching and 3D reconstruction due to its high-quality depth and color registration. The MIT Adobe FiveK dataset is used for training and evaluating automatic color grading models in the fields of image enhancement and tone mapping due to the pairing of its original images with images adjusted by multiple experts. The BSDS500 dataset is used in the field of image segmentation and boundary detection to evaluate the algorithm’s ability to segment natural image structures due to its fine manual boundary annotation of multiple sets. The Flickr Million dataset is used in the field of large-scale image retrieval and multimedia analysis to test the generalization performance of algorithms in diverse scenarios due to its large scale and content diversity. The use of the above dataset in this study is consistent with its standard usage, providing a reliable proxy benchmark for the evaluation of each module31. Access to all datasets used in this study (Middlebury, MIT-Adobe FiveK, BSDS500, Flickr Million) was obtained through official academic licensing agreements managed by the university’s research data management office, ensuring full compliance with terms of use for non-commercial academic research. Experimental parameter selection strategy The experimental parameter selection strategy is based on theoretical analysis, combined with grid search and cross validation for systematic multi criteria optimization. The parameter values follow three core principles: maximizing performance metrics, ensuring robustness under perturbations, and maintaining computational efficiency. The depth color joint weighting parameter = 0.2 in weighted non local Laplacian is determined by entropy analysis of the depth color joint distribution, and is determined within the range of 0.1-1.0 by minimizing color reconstruction error and maximizing structural similarity, in accordance with the maximization of mutual information criterion in information theory. The configuration of Turbo pixel superpixel quantity is 800, determined based on the marginal benefit decreasing curve of edge fitting accuracy with increasing segmentation density, taking into account both structural fidelity and computational cost. In the local consistency parameter, the spatial bandwidth = 0.3 and the feature bandwidth = 0.7 are derived from the spatial feature joint distribution kernel density estimation within the superpixel neighborhood, satisfying the theoretical constraints of intra class compactness and inter class separation. The regularization parameters of \varepsilon = 0.5 and α = 0.1 are used to achieve a balance between sparse reconstruction error and Turaplas smoothness through L-curve analysis, which conforms to the inflection point selection criterion of Pareto optimal frontier in convex optimization theory. The edge preservation strength of brightness guided filtering, P= 0.5, is derived from the coupling strength model of brightness gradient and chromaticity propagation. Based on the theory of total variation minimization, the optimal compromise is achieved between preserving edge sharpness and suppressing noise. All parameters were validated on a combined dataset using 5-fold cross validation to ensure their generalization ability and theoretical consistency under diverse image content. Results Quantitative performance validation of deep-guided chrominance transfer and local consistency models To verify the robustness of the selected parameters, sensitivity tests were conducted on four core parameters: the depth color joint weighting parameter , the number of superpixels, the spatial bandwidth , and the edge preservation intensity P. Different values were taken for each parameter while keeping the other parameters at their optimal values, and the peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) were recorded. The results of parameter sensitivity analysis are shown in Table 1. Table 1Parameter sensitivity analysis results.ParameterValue retrievalPSNR (dB)SSIM 0.137.8 ± 0.50.938 ± 0.0070.238.5 ± 0.40.942 ± 0.0060.338.0 ± 0.50.939 ± 0.007Number of superpixels60037.2 ± 0.60.930 ± 0.00880038.5 ± 0.40.942 ± 0.006100038.3 ± 0.50.941 ± 0.007 0.237.9 ± 0.50.937 ± 0.0070.338.5 ± 0.40.942 ± 0.0060.438.1 ± 0.50.938 ± 0.007 P 0.438.0 ± 0.50.940 ± 0.0070.538.5 ± 0.40.942 ± 0.0060.638.2 ± 0.50.940 ± 0.007 Parameter sensitivity analysis results. Table 1 shows that the performance indicators remain stable when each parameter fluctuates near the optimal value, and significantly decrease when deviating far, indicating that the selected parameters are located in a relatively optimal and robust flat region. The algorithm is insensitive to parameter perturbations and has good generalization ability. The study provides fully verifiable experimental evidence through a systematic design incorporating five key mechanisms: statistical significance testing (ANOVA with post-hoc comparisons), 5-fold cross-validation, comparisons against state-of-the-art baselines, ablation studies, and complete numerical results with standard deviations. All experiments were conducted under controlled conditions with a fixed random seed (seed = 42) to ensure exact reproducibility. The core configuration integrates hardware computational power, key algorithmic parameters, and an evaluation framework. Among these, the innovative parameter combination comprises depth-color joint weighting and edge retention strength settings, while Turbo-pixel superpixel segmentation and three-level feature fusion form the methodological foundation. Table 2 concisely presents the core settings influencing the experimental conclusions. Table 2Experimental settings.Configuration categoryNameSpecific settingsHardware foundationComputing equipmentNVIDIA Tesla V100 GPU+Xeon Gold 6248RInternal storage capacity128GB DDR4Algorithm frameworkCore modelLocally consistent sparse model+ADMM optimizationFeature extractionThree-level fusion (intensity+DAISY+significance)Key parameterDepth-color weight = 0.2Edge retention strengthP= 0.5 Gradient adjustableLocal consistency\varepsilon = 0.5, α = 0.1Spatial/feature bandwidth = 0.3, = 0.7Data specificationImage preprocessing512 × 512 resolution + CIE 1976 Lab* color spaceThe superpixel segmentationTurbo-pixel algorithm (500 superpixels)Experimental controlNumber of replicationAverage 5 timesConfidence interval95%Batch size16 (GPU training)Dataset splitsStandard training/validation/test splits for each public dataset were followed Experimental settings. To validate the critical role of depth information in color reconstruction, a controlled variable experiment is designed to compare four weighting strategies: traditional Welsh color matching, pure depth similarity, linear superposition method, and weighted nonlocal Laplacian algorithm. The test employs 100 samples from the Middlebury Stereo Standard RGB-D dataset with fixed parameters to eliminate interference. The effectiveness of depth guidance is validated across four dimensions: Color reconstruction error quantifies color deviation, structural similarity index (SSIM) evaluates structural preservation, local consistency index detects local consistency, and color matching accuracy measures cross-modal matching precision. Experimental results are shown in Fig. 6. Fitted trends of multi-dimensional metrics for depth-color fusion algorithms under varying depth information quality. Figure 6a shows that the color reconstruction error gradually increases with rising depth noise levels. At a noise level of 0.0, the weighted nonlocal Laplacian algorithm exhibits an error of 0.020, while the traditional Welsh color matching method yields an error of 0.023. When the noise level reaches 1.0, the weighted nonlocal Laplacian algorithm exhibits an error of 0.107, while the traditional Welsh color matching error soars to 0.158. In Fig. 6b, the SSIM values gradually decrease as the depth resolution is reduced. At the original resolution, the weighted nonlocal Laplacian algorithm achieves an SSIM value of 0.98, while the traditional Welsh color matching yields 0.95. When the resolution is reduced to 1/10, the weighted nonlocal Laplacian algorithm maintains an SSIM value of 0.62, whereas the traditional Welsh color matching drops to 0.48. In Fig. 6c, the local consistency index gradually increases as the depth smoothing parameter rises. At a smoothing parameter of 0.0, the weighted nonlocal Laplacian algorithm achieves a local consistency index of 0.93, while traditional Welsh color matching yields 0.90. When the smoothing parameter reaches 5.0, the weighted nonlocal Laplacian algorithm’s local consistency index rises to 0.73, while traditional Welsh color matching reaches 0.61. In Fig. 6d, the color matching accuracy gradually increases with rising depth contrast. At a contrast of 0.0, the weighted nonlocal Laplacian algorithm achieves a color matching accuracy of 0.90, while the traditional Welsh color matching yields 0.85. When the contrast reaches 2.0, the color matching accuracy value of the weighted nonlocal Laplacian algorithm rises to 1.00, while the traditional Welsh color matching achieves 0.95. The proposed method demonstrated superior performance across all metrics, validated by rigorous statistical analysis. A paired t-test confirmed statistically significant improvements over all baseline methods (p < 0.01). The algorithm exhibits the slowest color reconstruction error growth, with 95% confidence intervals showing no overlap with traditional methods. Performance gains in SSIM, local consistency, and color matching accuracy were statistically significant (p < 0.05) across all depth noise levels, revealing strong adaptability. To validate the effectiveness of the local consistency constraint, a progressive comparative experiment is designed based on sparse models: a baseline group (no regularization term), a spatial constraint group (only = 0.3), a feature constraint group (only = 0.7), and a full model group. Test images are segmented using Turbo-pixel to generate 500–800 superpixels as processing units, which are then optimized via the ADMM algorithm (up to 50 iterations). The evaluation metrics are designed to reflect multidimensional validation: the patch effect index quantifies patch effects, normalized chromaticity variation detects color abrupt changes, and the semantic coherence score assesses semantic rationality. Computational efficiency is verified by the number of iterations. Experimental results are shown in Fig. 7. Progressive performance trends of sparse models with different constraint configurations across increasing superpixel counts and iterations. In Fig. 7a, as the superpixel count increased from 500 to 800, the patch effect index of the full model group decreases from 0.85 to 0.94, significantly outperforming the baseline group. In Fig. 7b, after more than 30 iterations, the normalized chromaticity variation of the complete model group stabilizes below 0.30. Meanwhile, the feature constraint group exhibits rebound fluctuations between 0.3 and 0.4. In Fig. 7c, as the regularization parameter increases from 0 to 1.0, the semantic consistency score of the full model group rises from 0.6 to 0.89, significantly outperforming the baseline group’s score of 0.5 to 0.83. In Fig. 7d, when the superpixel count increases from 500 to 800, the number of iterations for the full model group increases linearly from 25 to 49, representing a 96% increase. In contrast, the spatial constraint group only increases from 35 to 47 iterations. The results demonstrate that the complete model exhibits significant advantages in suppressing artifacts, maintaining color stability, enhancing semantic coherence, and improving computational efficiency. Its key metrics outperform the next-best approach by over 40% on average, validating the effectiveness of multi-constraint collaborative optimization. Figure 7 indicates that the full model group achieves the best performance, showing a significant reduction in patch effects and chromaticity variation along with improved semantic coherence as superpixel count increases. This confirms the necessity of multi-constraint collaborative optimization. To comprehensively evaluate the overall performance of the proposed multi-level feature system and examine the correlation between the intrinsic attributes of the features and the performance of downstream tasks, the experiment verifies the completeness of the system from two dimensions: the essential characteristics of the features and the application efficiency. The experiment adopted the feature ablation experimental framework and systematically tested four feature combinations on the Flickr Million dataset: the use of intensity features alone, intensity +DAISY texture features, DAISY+ saliency features, and the complete combination of three-level features. The experiment adopted the Turbo-pixel 500 superpixel configuration as the processing unit to ensure local consistency constraints. Establish a dual-dimensional evaluation index system: The performance dimensions of downstream tasks include feature matching recall rate, mismatch rate, semantic alignment error and processing time; The internal attribute dimensions include feature discriminability, inter-class separability, intra-class consistency, and feature redundancy. The DAISY descriptor is reduced to 64 dimensions through PCA to maintain 95% variance, and significance detection is carried out using a 16-layer convolutional neural network of visual geometry groups. The experimental results are shown in Fig. 8. Comparison of dual-dimensional performance indicators for multi-level feature combinations. The experimental data in Fig. 8 shows that as the feature combination gradually expands from the single use of strength features to the complete fusion of three levels, the system performance presents a systematic improvement. The feature matching recall rate increased significantly from 62.3% ± 3.2% to 91.4% ± 2.1%, while the mismatching rate dropped sharply from 18.7% ± 2.1% to 3.8% ± 0.8%, and the semantic alignment error was optimized from 12.4 ± 1.5 pixels to 3.1 ± 0.7 pixels. Although the processing time increased from 16.5 ± 2.1 ms to 41.5 ± 3.1 ms, the increase was far lower than the benefit from the improvement in accuracy. Synchronous improvement of internal attribute dimensions The feature discriminability decreased from 1.85 ± 0.12 to 0.95 ± 0.08, the inter-class separability increased from 0.62 ± 0.05 to 0.85 ± 0.02, the intra-class consistency ARI rose from 0.58 ± 0.04 to 0.82 ± 0.02, and the feature redundancy variance retention rate increased from 72% ± 3% to 92% ± 1%. The key turning point occurred at the stage of introducing the DAISY texture feature, where most indicators achieved the most significant leap. The results show that multi-level feature fusion effectively solves the bottleneck of the insufficiency of traditional single feature representation by collaboratively enhancing the discriminability, separability and consistency of features. Figure 8 reveals that the three-level feature fusion system yields the best overall performance, achieving the highest recall rate and lowest error metrics. This demonstrates that hierarchical feature integration effectively overcomes the limitations of single-feature representations. To compare the performance of SLIC and Turbo-pixel algorithms at different superpixel counts, four experimental groups are set up: SLIC 200, SLIC 500, Turbo-pixel 500, and Turbo-pixel 800. The following metrics are evaluated: edge fitting accuracy, over-segmentation rate, feature extraction time, total processing time consumption, texture retention rate, semantic boundary accuracy, and color reconstruction error. The experiment utilizes the Berkeley Segmentation Dataset 500 with a fixed iteration count of 10, and the color space is converted to the CIE 1976 Lab* color space. The experimental results are shown in Table 3. Table 3Experimental results of performance comparison between Slic and Turbo-pixel algorithms.Evaluation indexSLIC 200SLIC 500Turbo-pixel 500Turbo-pixel 800Edge fitting accuracy0.720.810.890.93Over-segmentation rate28%19%12%8%Feature extraction time1.2s2.3s1.8s3.1sTotal processing time consumption4.5s6.7s5.2s7.9sTexture retention rate0.650.730.820.86Semantic boundary accuracy0.680.750.840.88Color reconstruction error5.24.73.93.5 Experimental results of performance comparison between Slic and Turbo-pixel algorithms. Table 3 shows that the Turbo-pixel algorithm outperforms the SLIC algorithm across all metrics. As the number of superpixels increases, both edge fitting accuracy and semantic boundary accuracy improve for both algorithms. Turbo-pixel 800 achieves the highest values of 0.93 and 0.88, with the lowest over-segmentation rate of only 8%. Regarding feature extraction time, Turbo-pixel 500 outperforms SLIC 500 at 1.8 s versus 2.3 s, but Turbo-pixel 800 increases to 3.1 s. The texture retention rate and color reconstruction error metrics demonstrate that the Turbo-pixel algorithm preserves image features more effectively. Its color reconstruction error is as low as 3.5 under the 800 superpixel configuration. Overall, the Turbo-pixel algorithm significantly enhances segmentation accuracy and feature retention while maintaining processing efficiency, achieving optimal comprehensive performance particularly under the 800 superpixel configuration. This validates its applicability in complex scenes. Table 3 provides a quantitative comparison, clearly showing the Turbo-pixel algorithm achieves superior performance in accuracy and feature retention, confirming its advantage for complex scene segmentation despite a moderate time increase. Brightness-guided edge retention mechanism and subjective visual evaluation To validate the edge preservation effectiveness of the brightness-guided mechanism, a comparative experiment is set up, featuring bilateral filtering, directed filtering, nonlocal mean, and brightness-guided filtering. Evaluation metrics include edge sharpness, color leakage index, peak signal-to-noise ratio (PSNR), and visual fidelity. The experiment is conducted on a dataset comprising 200 images with manually annotated edges. The experimental results are shown in Fig. 9. Comparative performance trends of four filtering methods under increasing filtering intensity, edge distance, noise level, and image complexity. In Fig. 9a, as the filtering intensity parameter increases, the edge sharpness of bilateral filtering, directed filtering, and nonlocal mean gradually decreases, while brightness-guided filtering exhibits a slower decline and consistently maintains the highest edge sharpness. At a filtering intensity of 0.9, the edge sharpness value of brightness-guided filtering is 0.83, significantly higher than bilateral filtering (0.65), directed filtering (0.70), and nonlocal mean (0.68). In Fig. 9b, as the edge distance threshold increases, the color leakage index of all four filtering algorithms rises, but brightness-guided filtering consistently exhibits the lowest color leakage index. At an edge distance threshold of 9 pixels, the brightness-guided filtering method achieves a color leakage index of 0.16, which is lower than the values of 0.22 for bilateral filtering, 0.20 for directed filtering, and 0.24 for nonlocal mean. Figure 9c shows that as the noise level increases, the PSNR values of all four filtering algorithms rise, but brightness-guided filtering consistently achieves the highest PSNR. At a noise level of 50 dB, brightness-guided filtering achieves a PSNR of 35.0 dB, surpassing bilateral filtering (34.0 dB), directed filtering (34.5 dB), and nonlocal mean (33.5 dB). In Fig. 9d, as image complexity increases, the visual fidelity of all four filtering algorithms decreases, but brightness-guided filtering consistently achieves the highest VIF. At an image complexity of 1.0, brightness-guided filtering yields a VIF of 0.75, surpassing bilateral filtering (0.60), directed filtering (0.65), and nonlocal mean (0.62). A one-way ANOVA with post-hoc Tukey test showed statistically significant superiority of the brightness-guided filter over all baselines in edge sharpness (F(3,196) = 45.2, p < 0.001) and color leakage (F(3,196) = 38.7, p < 0.001). Figure 9 illustrates that the proposed brightness-guided filtering maintains superior edge sharpness and the lowest color leakage across varying conditions. It also achieves the highest PSNR and visual fidelity, demonstrating robust edge preservation capabilities. To validate the comprehensive performance advantages of the research method in objective metrics, a comparative experiment is conducted on the MIT-Adobe FiveK dataset. The Welsh grayscale equalization algorithm (Welsh), iterative robust orthonormalization optimization (IRONY), and gated multi-scale feature pyramid network (GMF-Net) are selected as baseline methods. The experiment strictly controls variables, quantifying reconstruction accuracy using PSNR, assessing structural similarity with SSIM, measuring color fidelity via the CIEDE2000 Color Difference Formula, and testing processing speed with frames per second (FPS). Experimental results are shown in Fig. 10. Performance scaling trends of four image processing algorithms against increasing edge density, brightness contrast, color richness, and workload on the MIT-Adobe FiveK dataset. In Fig. 10a, as edge density increases, the PSNR of Welsh, IRONY, and GMF-Net all decrease. However, the proposed method maintains a significant advantage in the high-complexity range, outperforming GMF-Net by approximately 2.8 dB, IRONY by approximately 5.9 dB, and Welsh by approximately 8.3 dB on average. In Fig. 10b, SSIM decreases for all methods as brightness contrast increases. However, the proposed method maintains the highest structural fidelity in the high-contrast range, outperforming GMF-Net by approximately 0.035, IRONY by approximately 0.087, and Welsh by approximately 0.130 at a variance of 0.10. In Fig. 10c, color difference increases across all methods as color richness rises. However, the proposed method maintains the lowest color difference even at the highest entropy range, achieving reductions of approximately 0.99 compared to GMF-Net, 2.78 compared to IRONY, and 4.80 compared to Welsh. In Fig. 10d, the FPS of all methods decreases as the number of images increases. However, the proposed method maintains 9.4 FPS under a high load of 64 images, outperforming GMF-Net by approximately 34%, IRONY by approximately 109%, and Welsh by approximately 185%. The results demonstrate that the proposed method exhibits significant advantages across all four metrics: PSNR, SSIM, CIEDE2000, and FPS. Performance improvements are particularly pronounced in scenes characterized by high complexity, high contrast, high chromaticity, and high workload, validating its superior comprehensive capabilities. Figure 10 highlights the comprehensive advantages of the proposed method. It consistently achieves superior metrics compared to baseline methods as task difficulty increases, proving its effectiveness and efficiency in high-demand scenarios. To comprehensively evaluate the relative performance of the research method and the advanced methods in recent years, the experiment compared the comprehensive performance of four representative coloring methods in multiple dimensions. The experiment selected three advanced methods proposed between 2021 and 2023, namely colorization transformer, gan-based colorization, and diffusion colorization, as benchmarks. It was rigorously tested on the places365 standard dataset. This dataset contains 1.8 million images across 365 scene categories, providing rich scene diversity to ensure the comprehensiveness of the test. The evaluation system consists of two dimensions: traditional quality indicators and modern perception indicators: peak signal-to-noise ratio and structural similarity index measure the reconstruction fidelity, FID and learning-perceived image block similarity evaluate the generation quality, average opinion score reflects the subjective perception quality, and processing time tests the computational efficiency. The experiment adopted the standard test protocol. All methods were tested under the same hardware environment using the pre-trained models publicly disclosed by the author to ensure the comparability of the results. Experimental results are shown in Fig. 11. Comprehensive performance benchmarking against state-of-the-art colorization methods: objective fidelity, perceptual quality, and computational efficiency. According to the experimental data in Fig. 11, the proposed method demonstrates significant advantages in multiple key indicators. The peak signal-to-noise ratio reached 38.5 ± 0.4 dB, which was 3.7 dB higher than that of the diffusion model coloring method. The structural similarity index was 0.942 ± 0.006, which was superior to the generative adversarial network coloring method. In terms of perceived quality, the Frecher initial distance was reduced to 18.3 ± 1.2, which was 4.4 lower than that of the diffusion model coloring method. The similarity of the learned perceived image blocks was 0.095 ± 0.008, a significant improvement over the 0.185 ± 0.015 of the coloring Transformer. In the subjective assessment, an average opinion score of 4.3 ± 0.1 was obtained, with a processing time of 7.9 ± 0.5 s. While maintaining high-quality output, a balance in computational efficiency was achieved, which was 37% faster than the 12.5 ± 1.2 s of the diffusion model coloring method. Repeated measures ANOVA confirmed significant performance advantages across all fidelity, quality, and efficiency metrics (Wilks’ λ = 0.12, F(12,288) = 35.4, p < 0.001), with post-hoc comparisons showing the proposed method significantly outperformed all benchmarks (p < 0.01). The low variance further proves the method’s stability and reliability. Figure 11 presents a comprehensive benchmark, showing the proposed method excels in both objective fidelity and perceptual quality metrics while maintaining competitive speed. This balanced performance underscores the method’s practical value. To verify the necessity of each core module, a ablation experiment was designed by randomly selecting 50 registered RGB-D images on the Middlebury dataset. Four model variants were set up in the experiment, including w/o Depth (removing depth guidance and weighting only based on color similarity), w/o Multi feature (removing three-level features and using only intensity histograms), w/o LocalConsist (removing local consistency constraints and using independent sparse encoding), and w/o LuminGuide (filtering after removing brightness and directly outputting sparse reconstruction results). All variants maintain a superpixel count of 800 and other parameters consistent with the complete model, and repeat 10 times to take the mean. PSNR, SSIM, and CIEDE2000 are used as evaluation indicators. The results of the ablation experiment are shown in Table 4. Table 4The results of the ablation experiment.Model variantsPSNR (dB)SSIMCIEDE2000w/o Depth34.2 ± 0.50.912 ± 0.0075.1 ± 0.4w/o Multi-feat33.1 ± 0.60.895 ± 0.0096.3 ± 0.5w/o LocalConsist32.8 ± 0.70.883 ± 0.0106.9 ± 0.6w/o LuminGuide35.1 ± 0.40.921 ± 0.0064.8 ± 0.3Complete model38.5 ± 0.40.942 ± 0.0063.1 ± 0.3 The results of the ablation experiment. Table 4 shows that the removal of each module leads to a decrease in model performance. After removing depth guidance, the PSNR of the model decreased by 4.3 dB and CIEDE2000 increased by 2.0, confirming the crucial role of depth geometric constraints in color accuracy. Removing multi-level features resulted in a decrease of 5.4 dB in PSNR and 0.047 dB in SSIM of the model, indicating that a single intensity feature is difficult to handle complex texture matching. Removing local consistency constraints resulted in a 5.7 dB decrease in PSNR and a 3.8 dB increase in CIEDE2000, verifying the necessity of neighborhood collaborative sparse coding. After removing the brightness, although the filtering has a small impact, it still reduces PSNR by 3.5 dB and increases CIEDE2000 by 1.7, indicating that the refinement effect of brightness guidance on edge fidelity is irreplaceable. To visually demonstrate the effectiveness of the proposed method, the study presented input and output example images of the proposed method, as shown in Fig. 12. Input and output example images. (Source: MIT Adobe FiveK dataset) Comparing Fig. 12a–c, it can be seen that the input grayscale image only contains brightness and geometric structure, while the reference image provides a source of color statistics. The color of the output result is consistent with the reference image, and depth guidance effectively prevents edge leakage. Fine texture is preserved through brightness filtering, achieving the expected goal of transitioning from grayscale input to visually rich color output. Discussion and conclusion In-depth analysis of superior performance mechanisms The proposed method achieved performance improvement through the collaboration of three core modules. First, the weighted nonlocal Laplacian algorithm reduces color reconstruction error to 0.020 ± 0.003-a 60% improvement over traditional methods-by integrating depth-color correlations and spatial adjacency constraints. This prevents color bleeding while maintaining structural coherence. Second, Turbo-pixel 800 superpixel configuration achieves an edge fitting accuracy of 0.93 and reduces boundary confusion by 44.9% compared to SLIC 500 by adaptively following complex edges rather than imposing geometric regularity. Third, the three-level feature fusion system achieves a matching recall rate of 91.4% ± 2.1% through hierarchical learning: pixel-level intensity statistics (25-bin histogram), DAISY descriptors (64-dimensional PCA-reduced), and saliency detection for semantic guidance. The essential difference between the method proposed by the research institute and existing technologies lies not in incremental performance improvement, but in three paradigm shifts. Firstly, traditional RGB-D fusion considers depth as an additional feature channel, while the proposed method uses weighted non local Laplacian to use depth as a geometric constraint for color propagation, achieving structure guided coloring rather than feature concatenation. Secondly, traditional color transfer relies on global statistics or shallow descriptors. The proposed method constructs a three-level feature system (intensity, DAISY, saliency) under superpixel units, elevating the matching from pixel level to semantic aware structured representations. The third traditional sparse optimization method for isolated encoding of local blocks proposes a sparse pattern with graph Laplacian forcing neighborhood coefficient coordination and entropy regularization inducing semantic alignment, moving from “independent sparsity” to “uniform sparsity”. Limitations and technical boundaries Despite these advancements, three limitations persist. Computational time increases non-linearly with superpixel count (reaching 7.9s for 800 superpixels), and this delay is difficult to meet real-time requirements in graphic design scenarios that require real-time feedback, such as interactive color editing and real-time preview, limiting the direct application of this method in real-time graphic design software. DAISY descriptor dimensionality reduction may lose fine-grained features in complex textures. Performance is sensitive to the quality of the reference image - when the resolution of the reference image is below 256 × 256 pixels, the SSIM decreases by more than 0.09 and drops below 0.85, indicating a significant decline in structural fidelity and difficulty in accurately reconstructing fine textures. These limitations are particularly pronounced in dynamic scenes and mobile environments. In addition, the proposed method relies on multiple manually designed components (depth color weighting function, Turbo superpixel parameters, three-level feature combination) and high-performance GPU resources, which may limit its generalization. The difference in the depth noise characteristics of different sensing devices will also affect the color stability. The 7.9 s single image processing and GPU dependency make it difficult to deploy in resource constrained scenarios such as mobile terminals and embedded devices. Therefore, the follow-up work should introduce domain adaptive strategy to reduce the impact of different sensor depth noise distribution differences on color stability and improve the generalization ability of the model by learning the cross dataset invariant representation between depth features and color reconstruction. At the same time, for scenarios where computing resources are limited, model pruning is used to remove redundant parameters in sparse coding and feature extraction networks. Combined with knowledge distillation, the coloring ability of the current complex model is transferred to a lightweight network, compressing the processing time of single images while maintaining reconstruction quality, making the method applicable to resource limited environments. Future applications and research directions Potential applications include grayscale optimization for brand logos, monochrome photography enhancement, and intelligent coloring systems using GANs for style transfer. In virtual reality, the method can improve grayscale 3D model rendering. Future research will focus on: (1) GPU acceleration of ADMM algorithm for faster processing, (2) lightweight network architectures for mobile deployment, and (3) adaptive cross-modal fusion mechanisms to balance depth and color features based on scene complexity. Experimental verifiability and reproducibility summary The claimed performance advantages are grounded in a verifiable foundation: statistical significance (p < 0.01), empirical ablation studies, detailed parameter specifications, and consistent results across five diverse datasets. The commitment to releasing code, data, and configurations ensures the findings are practically reproducible. Author contributions T.T.W processed the numerical attribute linear programming of communication big data, and the mutual information feature quantity of communication big data numerical attribute was extracted by the cloud extended distributed feature fitting method. B.L. combined with fuzzy C-means clustering and linear regression analysis, the statistical analysis of big data numerical attribute feature information was carried out, and the associated attribute sample set of communication big data numerical attribute cloud grid distribution was constructed. T.T.W. and B.L. did the experiments, recorded data, and created manuscripts. All authors read and approved the final manuscript. Funding The research is supported by: General Project (Humanities and Social Sciences) at Hebei Environmental Engineering College, Construction and Practice of an Industry-Academia-Research Integration and Collaborative Innovation Model for Packaging Design in the Context of Rural Revitalization, (NO. XJXM-YB-2024010). Data availability The “Dataset” file of the manuscript is available at: https://zenodo.org/records/20230032. Code availability The “Code” file of the manuscript is available at: https://zenodo.org/records/20230055. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement The “Dataset” file of the manuscript is available at: https://zenodo.org/records/20230032. The “Code” file of the manuscript is available at: https://zenodo.org/records/20230055.