Complex-Domain Semantic Segmentation of Spacecraft Directly from ISAR Echoes Authors: Aoxiang Pan, Yonghua He, Yonggang Li, Jiahao Wang, Ruitao Shen, Weigang Zhu Published: 2026 ABSTRACT Highlights What are the main findings? We propose a complex-domain semantic segmentation framework OSS, which directly processes ISAR echoes to avoid scattering and phase information loss in traditional imaging pipelines.We design AIL, which enables automatic annotation of spacecraft masks based on ISAR scattering characteristics, effectively reducing manual labeling costs. What are the implications of the main findings? We introduce a novel paradigm for spacecraft semantic segmentation, which effectively addresses the challenges of high annotation costs and information loss inherent in conventional imaging processes.We significantly reduce application costs by enabling automatic label generation and shortening the data processing chain. AbstractSemantic segmentation technology based on Inverse Synthetic Aperture Radar (ISAR) images can provide crucial perception and analytical capabilities for intelligent safety maintenance of on-orbit spacecraft. However, conventional semantic segmentation methods suffer from three main limitations: firstly, the lack of modeling for radar physical characteristics in the “image first, segment later” pipeline leads to loss of scattering information and phase details; secondly, reliance on extensive pixel-level manual annotation increases application costs; thirdly, ineffective utilization of spacecraft structural priors fails to guide networks to focus on the main body and edges of spacecraft segmentation. To address these issues, this paper proposes a complex-domain semantic segmentation framework named One-Stop Segmentation (OSS) based on ISAR echoes. The framework incorporates two innovative modules: an Automatic ISAR Labeling (AIL) method designed based on ISAR scattering characteristics to generate labels corresponding to ISAR echoes, and a complex-domain semantic segmentation network named One-Stop Segmentation Network (OSSNet) that performs semantic segmentation directly on echoes, avoiding information loss from imaging while shortening the data processing chain. Core contributions of OSSNet include: (1) a Domain Alignment Module (DAM) to effectively mitigate domain mismatch caused by data distribution differences between raw echo signals and labels; (2) a Multi-Perspective Attention (MPA) framework incorporating a Sliding Correlation Attention (SCA) module and a Subdomain Balanced Attention (SBA) module, lever-aging spacecraft structural priors to guide the network’s focus on main structures and edge details from complementary perspectives, significantly improving segmentation ac-curacy. Experimental results on a simulated ground-based radar dataset demonstrate that the proposed OSS framework achieves a mean Intersection over Union (mIoU) of 92.13% and a mean F1-score of 95.75% in ISAR spacecraft semantic segmentation tasks, outperforming existing methods. 1. Introduction In recent years, the rapid development of global civil space missions has led to a significant increase in the number of on-orbit satellites. Concurrently, the risk of space collisions continues to rise, with occasional on-orbit impact events that can cause structural damage to satellite platforms or key payloads, severely threatening the safe on-orbit operation and long-term service capability of spacecraft. Against this backdrop, achieving fine-grained perception and state identification of space targets has become a core issue urgently requiring solution in the field of on-orbit spacecraft safety maintenance. Inverse Synthetic Aperture Radar (ISAR) [1], with its advantages of all-weather and all-day capabilities, long operating range, and high resolution, has emerged as a crucial technological means for maintaining the safe on-orbit operation of space targets. ISAR-based spacecraft semantic segmentation can not only accurately estimate component dimensions and identify main structures [2] but also further assess their operational status and enable precise spacecraft positioning [3]. This provides substantial information support for collision damage assessment and fault diagnosis. Traditional semantic segmentation methods often struggle to achieve precise segmentation of complex targets. With the advancement of deep learning, its powerful nonlinear modeling and feature extraction capabilities have provided new solutions for ISAR semantic segmentation. As one of the core tasks in computer vision, deep learning-based semantic segmentation has been widely applied in fields such as object recognition [4], remote sensing interpretation [5], and pose estimation [6]. These methods employ end-to-end pixel-level classification and leverage large-scale data training to automatically learn and extract high-level semantic features from images, thereby significantly distinguishing between different targets and backgrounds. Compared to traditional algorithms, deep neural networks can fully utilize high-order semantic information, maintaining high segmentation accuracy even in complex environments and with target variations [7]. Currently, numerous scholars have conducted extensive research on ISAR semantic segmentation. Zhu et al. integrated semantic segmentation with mask matching [8] to enhance the model’s generalization capability. Kou et al. introduced contrastive learning and a non-local U-Net [9], first using binary semantic labels for coarse segmentation of the target contour, then employing a non-local self-attention mechanism with global perception to guide the model to focus on the structural symmetry of ISAR images, thereby improving segmentation accuracy by enhancing feature discriminability. Coe et al. improved the Statistical Region Merging (SRM) algorithm [10], adapting its original three-channel optical-image mechanism to single-channel ISAR intensity data for effective image segmentation. Zhong et al. proposed the Scattering Characteristics Guided Network (SCGN) [11], which incorporates scattering features of various components into the mask segmentation process and models long-range semantic dependencies through a spatial attention mechanism to achieve fine-grained component segmentation. Ju et al. leveraged the inherent multi-scale stochastic structures of ISAR images to propose a Spatially Variant Mixture Multiscale Autoregressive (SVMMAR) model [12] for efficient ISAR image segmentation. Zheng et al. proposed an unsupervised segmentation method for ISAR images [13], using a set of multi-scale autoregressive models to characterize the inherent multi-resolution stochastic structures in ISAR images for reliable image segmentation. Wang et al. proposed a segmentation concept based on the CLEAN algorithm [14], designing an adaptive region-growing method to segment the main body, with a focus on analyzing and solving the threshold setting problem in region-growing methods. It is noteworthy that most existing research on ISAR semantic segmentation predominantly leverages amplitude information [15], often neglecting the rich phase information inherent in radar signals. In parallel with the ISAR-specific methods discussed above, attention mechanisms have become a key component of modern semantic segmentation architectures. Representative designs include the Squeeze-and-Excitation Network (SE-Net) [16] for channel-wise recalibration, the Convolutional Block Attention Module (CBAM) [17] for joint channel-spatial attention, non-local networks [18] for long-range contextual aggregation, and wavelet-based attention mechanisms [19] for multi-frequency feature decomposition. Despite their success in natural image segmentation, these mechanisms are fundamentally designed for real-valued feature representations and have not been adapted to exploit the complex-valued nature of radar echoes. Although the aforementioned ISAR-specific methods and general attention mechanisms have advanced segmentation performance in their respective settings, their direct application to complex-domain spacecraft segmentation is hindered by three limitations. First, the predominant “image-first, segment-later” paradigm relies on complex imaging that projects complex-valued echoes onto real-valued amplitude images, irreversibly discarding phase information critical for characterizing target characteristics. Existing attention mechanisms, being architected for real-valued features, inherit rather than resolve this information bottleneck. Although Li et al. attempted to incorporate echo-domain features for few-shot segmentation [20], this was limited to a preliminary real-domain mapping without constructing a full complex-domain network. Second, standard self-attention computes pairwise feature correlations globally and uniformly. In ISAR data, this biases attention toward isolated strong scattering points while suppressing the weaker yet structurally continuous responses that delineate component boundaries. Meanwhile, these data-driven mechanisms lack embedded priors about spacecraft geometry and radar scattering physics which are essential for distinguishing genuine structural edges from coherent sidelobe artifacts. Third, label acquisition remains dependent on costly manual annotation [21], and the imaging preprocessing step introduces both information loss and computational latency unsuitable for time-sensitive space situational awareness tasks. These factors collectively hinder the model’s capability to achieve high-precision segmentation of space targets. To address the aforementioned issues, this paper proposes an end-to-end complex-domain ISAR echo semantic segmentation framework named One-stop Semantic Segmentation (OSS). The OSS framework comprises an automatic labeling method Auto ISAR Labelling (AIL) and a complex-domain segmentation network One-stop Semantic Segmentation Network (OSSNet), achieving direct processing from echo signals to semantic segmentation results. Experimental results demonstrate that the proposed method significantly outperforms existing ISAR semantic segmentation approaches. The main contributions of this paper are summarized as follows:An Automatic ISAR Labeling (AIL) method that generates semantic masks from complex ISAR images using multi-scale gradient fusion, eliminating manual annotation costs.An end-to-end One-Stop Segmentation Network (OSSNet) that directly maps raw ISAR echoes to segmentation masks, bypassing traditional imaging pipelines.A Domain Alignment Module (DAM) that bridges the data distribution gap between raw echoes and image-domain labels via learnable domain transformation.A Multi-Perspective Attention (MPA) framework (SCA + SBA) that integrates spacecraft structural priors with full-dimensional scattering characteristics, enabling precise segmentation of main structures and edges. An Automatic ISAR Labeling (AIL) method that generates semantic masks from complex ISAR images using multi-scale gradient fusion, eliminating manual annotation costs. An Automatic ISAR Labeling (AIL) method that generates semantic masks from complex ISAR images using multi-scale gradient fusion, eliminating manual annotation costs. An end-to-end One-Stop Segmentation Network (OSSNet) that directly maps raw ISAR echoes to segmentation masks, bypassing traditional imaging pipelines. An end-to-end One-Stop Segmentation Network (OSSNet) that directly maps raw ISAR echoes to segmentation masks, bypassing traditional imaging pipelines. A Domain Alignment Module (DAM) that bridges the data distribution gap between raw echoes and image-domain labels via learnable domain transformation. A Domain Alignment Module (DAM) that bridges the data distribution gap between raw echoes and image-domain labels via learnable domain transformation. A Multi-Perspective Attention (MPA) framework (SCA + SBA) that integrates spacecraft structural priors with full-dimensional scattering characteristics, enabling precise segmentation of main structures and edges. A Multi-Perspective Attention (MPA) framework (SCA + SBA) that integrates spacecraft structural priors with full-dimensional scattering characteristics, enabling precise segmentation of main structures and edges. 2. Proposed Method 2.1. OSS-Based ISAR Semantic Segmentation High-precision ISAR semantic segmentation is crucial for on-orbit spacecraft maintenance. However, its advancement faces constraints due to the limited availability of high-quality annotated data and insufficient data utilization. Furthermore, traditional methods often struggle to address challenges such as discontinuous target scattering points and sidelobe interference in ISAR imagery, while generally neglecting phase information and prior structural knowledge of spacecraft. These limitations hinder effective modeling of complete physical scattering characteristics, consequently restricting segmentation accuracy. To overcome these issues, this paper proposes an end-to-end complex-domain semantic segmentation framework termed OSS, which achieves direct signal-to-segmentation mapping in the raw echo domain, thereby circumventing information loss inherent in conventional imaging-segmentation cascade processing. The OSS framework comprises two core innovative modules: the AIL method for automatic high-quality mask generation based on ISAR scattering characteristics, and the segmentation network OSSNet. Figure 1 illustrates the overall framework of OSSNet, while Figure 2 depicts the architectures of the encoder and decoder, respectively. The network takes raw ISAR echo signals as input. First, the DAM applies a learnable orthogonal transformation to the echo data, mapping the original signal domain to a more discriminative feature domain, effectively resolving the domain mismatch between echoes and mask labels. The transformed features are then fed into a multi-branch encoder integrated with the MPA framework. This encoder extracts scattering features at different hierarchical levels through parallel paths, and incorporates the SBA and SCA modules to enhance the representation capability for spacecraft main structures and detailed features. The decoder adopts a multi-branch structure without attention mechanisms, progressively performing upsampling and leveraging skip connections to fuse multi-scale features from the encoder. This design fully utilizes shallow high-resolution information to improve spatial localization accuracy, while combining deep semantic features for precise classification, ultimately outputting high-precision semantic masks. The AIL method is detailed in Section 2.2, and implementation specifics of OSSNet are provided in Section 2.3, Section 2.4 and Section 2.5. 2.2. Automatic ISAR Labeling (AIL) In the field of ISAR semantic segmentation, target segmentation masks are typically obtained through manual annotation. However, manual annotation methods are often accompanied by substantial labor and time costs. Furthermore, the high structural complexity of spacecraft targets, combined with characteristics such as discontinuous scattering point distribution, significant intensity variations in scattering points at different positions, and edge blurring, makes it difficult for manual annotation to meet precision requirements. Therefore, achieving automatic and accurate data annotation based on ISAR scattering characteristics has become a critical issue requiring urgent resolution in this field. To address these challenges, this paper proposes an automatic labeling method called AIL based on ISAR scattering characteristics. The core of AIL lies in constructing multi-scale fused complex gradient images: it enhances edge completeness by capturing multi-resolution edge features and employs a multi-scale feature weighting fusion mechanism to prevent small-scale details from being overwhelmed by large-scale structures. Furthermore, morphological post-processing is introduced to improve the geometric accuracy and topological completeness of the segmentation masks. This method essentially represents the mapping of target electromagnetic scattering physical characteristics into image space, providing theoretical support for generating high-precision spacecraft segmentation masks. Unlike traditional annotation workflows—where manual labeling is performed on ISAR images after imaging transformation and imaging compensation, as shown in Figure 3 (left)—AIL directly achieves automated labeling in the complex ISAR image domain according to its electromagnetic scattering characteristics, as illustrated in Figure 3 (right). In ISAR imaging, the physical edges of spacecraft produce significant gradient responses due to variations in electromagnetic scattering characteristics. Specifically, metal structural components such as the spacecraft main body and solar panels form strong scattering centers in complex ISAR images, where the complex signals undergo sharp variations at boundary regions, resulting in high gradient magnitudes. Structural junction areas exhibit medium-to-high gradient magnitudes due to scattering discontinuity, while background regions generally demonstrate low gradient magnitudes. This characteristic provides a physical basis for distinguishing target structures from background regions using gradient magnitude thresholds, thereby enabling the generation of effective target segmentation masks. Region of Interest=(x,y)|∇G(x,y)≥τ, where τ is the threshold. Inspired by the Laplacian pyramid method [22], a K-layer Gaussian pyramid is first constructed through Gaussian smoothing filtering and progressive downsampling, thereby forming a multi-scale complex image space. Let the complex-valued image matrix at the k-th layer be denoted as Gk∈ℂMk×Nk. Then, Gk(x,y)=Rk(x,y)+Ik(x,y)⋅j, where j is the imaginary unit. The complex convolution operation is separable, allowing convolution operations to be performed separately on the real and imaginary parts:(1)Rk−1smooth(x,y)=∑i=−rr∑j=−rrRk−1(x+i,y+j)⋅gσ(i,j)Ik−1smooth(x,y)=∑i=−rr∑j=−rrIk−1(x+i,y+j)⋅gσ(i,j) where gσ(x,y) denotes the discrete Gaussian kernel function, and σ denotes the standard deviation, which varies across pyramid levels: σk=σ1⋅2k−1, r=3σ. Define the complex downsampling operator D:ℂX×Y→ℂX/2×Y/2, D(G)(x,y)=G(2x−1,2y−1). Namely,(2)Rk(m,n)=Rk−1smooth(2m−1,2n−1)Ik(m,n)=Ik−1smooth(2m−1,2n−1) where m=1,2,…,M/2,n=1,2,…,N/2. The recurrence formula for complex images at each layer is as follows:(3)Gk=DRk−1∗gσkIk−1∗gσkσk=σ1⋅2k−1 where ∗ denotes the two-dimensional Gaussian convolution operation, with k=2,3…,K. Subsequently, multi-scale gradient magnitude calculation is performed. Starting from the original-size complex image G1, the finite difference method is used to calculate its real part gradient and imaginary part gradient along the horizontal and vertical directions of the image, thus obtaining the original-scale gradient magnitude image M1; for the k-th layer (k≥2) complex image matrix Gk, its gradient calculation uses the Sobel operator. Using its horizontal and vertical convolution kernels, the gradient responses of the image in two orthogonal directions are calculated respectively, thus obtaining the k-th layer gradient magnitude image Mk.(4)M1=∇xR1+j∇xI12+∇yR1+j∇yI12Mk=∇xsobelRk+j∇xsobelIk2+∇ysobelRk+j∇ysobelIk2 where ∇ denotes the gradient computation. The BiCubic interpolation algorithm [23] is employed to upsample the k-th layer gradient magnitude image to the original scale image resolution, i.e., Mk∈ℝM12k−1×N12k−1⇒ Mk∈ℝM1×N1. Weighted fusion is performed on the multi-scale gradient magnitude images:(5)Mfinal=∑k=1KwkMk,wk=1,k=112k,k=2,3,…,K At the original scale level, the second-order accuracy characteristic of the central difference algorithm: ∇xA=A(x,y+1)−A(x,y−1)2Δy+O(Δy2) enables sub-pixel localization of spacecraft edges, ensuring clear segmentation boundaries. Meanwhile, the Sobel operator provides high computational efficiency while maintaining linear characteristics by combining smoothing and differentiation operations. By integrating these two gradient calculation methods in multi-scale space, the localization accuracy of spacecraft structural edges can be ensured while maintaining low computational complexity. Finally, morphological post-processing is applied to the fused multi-scale complex gradient image to generate the final segmentation mask. The closing operation is first applied to effectively fill internal cavities caused by weak scattering regions and bridge edge discontinuities; subsequently, the opening operation is employed to smooth boundaries and suppress isolated artifact points. The complete workflow of the AIL method is shown in Algorithm 1. Algorithm 1: Auto ISAR LabellingInput: Complex image matrix G1∈ℂM×NOutput: Segmentation mask S∈ℝM×NPart 1: Gaussian pyramid construction for i={2,3,…,K} do Gi=reduce(Gi−1), where reduce(⋅) represents Gaussian smoothing and downsampling endPart 2: Multi-scale complex gradient fusion M1=∇G1 M=M1 for i={2,3,…,K} do Mi=resize(∇SobelGi), where resize(⋅) represents Bicubic interpolation M=M+12iMi endPart 3: Morphological processing Z=0,M≤τ1,M>τ, where τ represents Threshold Z=closing(Z,rect), where closing represents closing operation S=opening(Z,disk), where opening represents opening operationreturn S 2.3. Domain Alignment Module (DAM) Conventional ISAR segmentation masks are generated based on target-background edge localization in the image domain, sharing spatial domain characteristics with ISAR imagery. Consequently, neural networks can directly learn pixel spatial relationships to produce masks. However, ISAR echoes require complex imaging transformations to form images, creating significant data distribution differences between the echo domain and mask domain—a phenomenon termed “domain mismatch.” Directly transferring image segmentation algorithms to the echo domain leads to model incompatibility due to this domain mismatch, consequently limiting segmentation accuracy. Furthermore, while ISAR images essentially represent mappings of echo data into the pixel domain, existing image-based segmentation methods utilize only amplitude information while discarding the phase component, resulting in insufficient modeling of original scattering characteristics. Since both amplitude and phase in the echoes carry target scattering properties, the loss of phase information causes critical physical features to be overlooked, thereby constraining further improvements in segmentation accuracy. To address the aforementioned challenges of echo-mask domain mismatch and incomplete utilization of physical scattering characteristics, this paper proposes a DAM, whose structure is illustrated in Figure 4. The module constructs an adaptive feature transformation function through learnable parameter matrices {P,A,X} and learnable parameters {M,N}, enabling dynamic mapping of input echo data into a feature space compatible with the mask domain. This effectively bridges the data distribution gap between the echo and mask domains while significantly accelerating model convergence by compressing the solution space, ultimately improving segmentation accuracy and enhancing model generalization capability. Let the input of DAM be the ISAR echo E∈ℂ1×H×W, where dimensions 1,H,W correspond to the number of channels, azimuth samples, and range samples, respectively. The implementation workflow of DAM is as follows: First, by introducing learnable transformation matrices {P,A,X∈ℝH×W}, a feature space mapping is constructed. The tanh(⋅) function is employed to constrain matrix values within the range (−1,1). This range normalization eliminates amplitude scale disparities to enhance optimization stability, while the sign preservation mechanism maintains the discriminability of phase relationships. Ultimately, a robust basis for feature representation is established in the transformed domain. The corresponding mathematical operations are expressed as follows:(6)P˜,A˜,X′=tanh(P,A,X) Based on this, unit modulus constraints are applied to the transformation matrices P and A, as shown in (7), forcing the basis vectors to lie on the unit circle in the complex plane to maintain the numerical stability of the complex transformation. Building upon this, a learnable rotation transformation is performed on the input echoes. This process extracts global features while preserving the signal’s phase characteristics. This design enables the model to adaptively learn the optimal complex basis space, achieving optimization of feature representation.(7)P′A′=1P˜⊙P˜+A˜⊙A˜+ε⊙P˜A˜ where ⊙ denotes the Hadamard product, and ε is an extremely small positive number used to maintain numerical stability. The ISAR echo is intrinsically a single-channel complex-valued matrix E∈ℂ1×H×W. After channel dimension compression, E∈ℂH×W can be decomposed into real and imaginary components, i.e., E=R+Ij, where R,I∈ℝH×W. Through adaptive transformation, the features of the real and imaginary parts are recombined and reconstructed, achieving an optimized representation of complex-domain information, as shown in (8).(8)RtIt=∑i=1H∑j=1WP′ij−A′ijA′ijP′ij︸Rotation matrixRijIij where Let Qij=P′ij−A′ijA′ijP′ij, for every (i,j):(9)QijTQij=QijQijT=P′ij2+A′ij200P′ij2+A′ij2 As derived from (7),(10)P′ij2+A′ij2=1QijTQij=QijQijT=I2 where I2 denotes the identity matrix, and matrix Qij is an orthogonal matrix. Hence each position-wise complex weight Wij=Pij+jAij lies on the unit circle in the complex plane, and the local transformation constitutes a rotation in ℝ2. Simultaneously, feature space modulation is implemented through the learnable spatial weight matrix X, enhancing the model’s feature selection capability for critical regions. Learnable scaling factors {M,N∈ℝ} are introduced to dynamically adjust the feature amplitudes in the transformed domain, thereby maintaining numerical stability.(11)R′I′=M⋅N⋅RtIt⋅X′ Finally, adjust the dimensions of the transformed features, i.e., R′,I′∈ℝH×W⇒R′,I′∈ℝ1×H×W, and combine the transformed complex features to obtain the transformed domain feature E′=R′+I′j. In the DAM, the learnable matrices P, A, and X serve distinct roles. P and A are normalized element-wise to form complex weights of unit modulus, which apply a pure phase rotation to each spatial position of the input echo. This step is locally strictly orthogonal—it preserves the instantaneous power of every resolution cell and avoids information loss or distortion in the phase component. X, by contrast, operates after the global summation: it modulates the compressed scalar back to the original spatial size and provides location-selective re-scaling. Because the overall transformation involves cross-spatial summation followed by an outer-product reconstruction, the equivalent mapping matrix is not an orthogonal matrix. Therefore, DAM as a whole does not constitute a strictly orthogonal transformation. Nevertheless, the locally orthogonal design remains crucial. Constraining complex weights to the unit circle bounds the dynamic range of the parameters and prevents unbounded amplification during training, which markedly improves the numerical stability of gradient propagation and accelerates convergence. More importantly, the element-wise rotation retains the complete phase information of each echo sample, providing a stable representation that preserves full scattering characteristics for the subsequent encoder. 2.4. Complex-Domain Encoder Radar-target relative aspect variations induce amplitude and phase perturbations in echoes, which manifest as edge blurring and strong scattering point sidelobe interference in ISAR imagery. These artifacts fundamentally originate from distortions in scattering characteristics within the echo domain. Consequently, integrating spacecraft rigid-body structural properties with ISAR scattering mechanisms to extract spatially discriminative, efficient, and robust features from echoes constitutes a critical challenge for enhancing semantic segmentation accuracy. To address the aforementioned challenges and construct discriminative feature representations for precise semantic segmentation, this paper proposes a complex-domain encoder. The network first employs a cascaded structure of 3 × 3 Complex-domain Convolutional (CConv) layer, Complex-domain Batch Normalization (CBN), and Complex-domain ReLU (CReLU) activation function [24] to extract key local features in the transformed domain while enhancing nonlinear modeling capabilities. The features are subsequently fed into a Multi-branch Encoding Module (MEB) to hierarchically capture multi-scale scattering characteristics, where shallow layers preserve detailed information while deeper layers encapsulate high-level semantic information. The processed features then pass through a Complex-domain Fully connected layer (CFc) for global feature integration and compression. Finally, a 1 × 1 CConv is applied during the encoding stage to enrich feature representation, outputting a scattering point feature representation fused with global semantics. The complex-domain encoder is primarily constructed through the synergistic integration of CConv/CBN/CReLU, MEB, and CFc components. The MEB adopts a multi-branch feature fusion architecture, significantly enhancing the diversity and discriminability of scattering features through complementary multi-perspective feature integration within hierarchical layers. The MPA framework proposes dual modules—SCA and SBA—capable of dynamically balancing the weights of amplitude and phase subdomains in foreground and background regions. These modules guide the network to focus on critical features of main structures and edge details across different dimensions, thereby improving semantic segmentation accuracy and model interpretability. The MEB adopts a triple-branch heterogeneous feature fusion architecture. Through parallel extraction of original transformed features, correlation attention-enhanced features, and subdomain attention-enhanced features, complementary fusion of scattering representations is achieved. This design breaks through the limitations of single-path representations, effectively preserving the spatial continuity of target structures while ensuring edge localization accuracy, all while enhancing feature diversity. Let the input feature of MEB be Ein∈ℂCin×Hin×Win, where Cin,Hin,Win represent the input channel number, feature map height, and feature map width of MEB, respectively. The original feature transformation branch Br consists of a single 3 × 3 CConv layer, expressed as:(12)Eorigin=BrEin The correlation attention feature enhancement branch Bc consists of two 3 × 3 CConv layers, one CReLU activation, and one SCA module. Its operational workflow is expressed as:(13)Ecorrelation=BcEin The subdomain attention feature enhancement branch Bs consists of two 3 × 3 CConv layers, one CBN layer, one CReLU activation, and one SBA module. Its computational flow is expressed as:(14)Esubdomain=BsEin The triple-branch heterogeneous features are concatenated along the channel dimension to generate the combined features Etri, which are then processed through a 1 × 1 CConv layer for cross-channel feature fusion and dimensional compression, outputting the optimized representation Eout of MEB.(15)Etri=concatenate(Eorigin,Ecorrelation,Esubdomain)Eout=Complex_Conv(Etri) 2.4.1. Sliding Correlation Attention (SCA) On-orbit spacecraft exhibit highly regular geometric structures, with key components arranged in strongly symmetric patterns [9]. The complex interconnections between components and uneven distribution of scattering point intensities cause models to perceive the overall spacecraft structure weakly and blur boundaries. To address the aforementioned issues of structural awareness weakening and edge localization inaccuracy, inspired by the autocorrelation function—which, by the Wiener-Khinchin theorem, relates the autocorrelation of a signal to its power spectral density and thus provides a principled means of detecting periodic or symmetric structural patterns—this paper proposes the SCA module. This module employs a local autocorrelation operator to compute the spatial structural correlations of feature maps, generating a structural spatial attention map to selectively enhance feature responses of key components. This enables the model to precisely focus on highly correlated regions to extract structural main bodies and edge features, significantly improving structural perception accuracy. The structure of the SCA module is shown in Figure 5. First, let the input feature map of the SCA module be J∈ℂC×H×W, and construct its complex conjugate feature map J¯∈ℂC×H×W. Next, generate the shifted feature map Gc,i,j=J¯c,fh(i),fw(j). The height dimension shift mapping relationship is:(16)fh(i)=i−win_hif i≥win_hH+(i−win_h)if i