Structural-semantic double-graph collaborative SAR image quality evaluation method

Abstract

The invention relates to a structure-semantic double-graph collaborative SAR image quality evaluation method, which comprises the following steps of separating an SAR image, and constructing a double-channel quality measurement set for independent modeling; indexes are obtained, correlation calculation is carried out, an index map is constructed, and each node corresponds to one index; on the basis of the index map, the dependency relationship between indexes is calculated from the structure dimension and the semantic dimension through self-supervised learning, the dependency relationship of the two dimensions is integrated, and an image level score based on the integration result is output; and constructing a measurement index shared with system analysis and a measurement index special for information retrieval, and carrying out quality measurement evaluation: after extracting a structure sub-graph, integrating an IR special index which is not contained in SAR prior into a global graph. The method is superior to a reference model in the aspects of recognition accuracy, recall rate, precision and efficiency, and a solution with high universality and good interpretability is provided for cross-modal image quality evaluation.

Description

Technical Field

[0001] This invention belongs to the field of radar technology, and in particular relates to a SAR image quality assessment method based on structure-semantic dual-graph collaboration. Background Technology

[0002] SAR imagery technology plays a crucial role in remote sensing fields such as target detection and land cover classification due to its all-weather operation capabilities and quantitative Earth observation potential. However, SAR images often suffer from quality issues such as speckle noise and geometric distortion. Since downstream tasks are highly dependent on image quality, low-quality SAR data can reduce accuracy or lead to mission failure; therefore, SAR image quality assessment (IQA) is essential. Despite the significant impact of SAR image quality, mainstream global scoring methods generally ignore the distinction between foreground and background. This often results in a clean background masking a degraded foreground, leading to evaluation bias. While pure foreground training can improve performance in controlled environments, it loses contextual / structural cues and impairs generalization ability. Furthermore, unfiltered background content can mislead recognition and degrade performance. These limitations expose the dual flaws of a unified evaluation paradigm (ignoring regional differences) and an isolated evaluation paradigm (ignoring foreground-background interactions).

[0003] Most existing SAR IQA methods neglect complex inter-indicator relationships. This application categorizes these relationships into two types: statistical correlation (e.g., in low-texture regions, the covariance of gradient and texture energy (two metrics) indicates degradation of shared perceptual features) and semantic dependence (indicators share perceptual features such as information strength / sharpness / anti-interference, reflecting human-centered semantic meaning). Fusing metrics in the absence of structural / semantic modeling reduces the accuracy of image quality assessment; therefore, relationship modeling is crucial for robust assessment. To the best of our knowledge, existing technologies have not established complex inter-indicator relationship models to improve SAR IQA performance.

[0004] Despite progress, most SAR IQA methods still neglect generalization ability, resulting in limited applicability across datasets / modalities. Although some studies have noted common sensing features (such as blurring, contrast loss, and noise) among radar / optical / infrared modalities, existing research has not yet utilized these features to improve the generalization ability of SAR IQA. Summary of the Invention

[0005] Synthetic Aperture Radar (SAR) is crucial in remote sensing tasks such as target detection and recognition, and its image quality directly affects task performance. However, existing SAR image quality assessment methods have three major limitations: they rely heavily on overall image assessment, which can lead to blurring and degradation of small target areas due to high-quality backgrounds; although various quality indices have been proposed, most methods treat these indices in isolation and ignore their inherent correlations; and limited generalization ability restricts adaptability across datasets and imaging modalities. To address these challenges, this application proposes a Dual Image Collaborative Quality (DGCQ) assessment framework, an innovative solution specifically designed for SAR image quality assessment. Its core features are: (1) foreground / background decoupling; (2) indicator relationship modeling; and (3) cross-modal adaptation. Specifically: DC-DQM decouples foreground and background by region segmentation and specific region indicator extraction, thereby mitigating the masking effect of global scoring on local degradation. The SS-DGCN model represents metric relationships using a dual-graph representation: a structural metric graph and a semantic metric graph. Nodes represent metric indicators, while edges encode statistical correlations or semantic dependencies. A graph convolutional network (GCN) propagates node features, followed by node-level gating fusion to integrate structural / semantic features. The fused features are aggregated into a graph-level representation for self-supervised IQA scoring. SPAM enhances generalization capabilities through cross-domain plug-and-play structural prior transfer without modifying the backbone network.

[0006] DGCQ comprises three core modules: (1) The Dual-Channel Decoupled Quality Modeling (DC-DQM) module segments the image into foreground and background channels and designs differentiated quality indicators accordingly; (2) The Structure-Semantic Dual-Graph Collaborative Network (SS-DGCN) models the relationship between indicators through dual-graph propagation: the structure graph establishes connections based on the statistical correlation of indicator values, while the semantic graph not only uses perceptual attributes (such as sharpness and robustness) to infer the semantic association between indicators, but also avoids statistical coincidences of semantically irrelevant indicators in the structure graph through these attributes; (3) The Structure Prior Adaptive Module (SPAM) integrates structure alignment, perturbation consistency and semantic regularization to improve cross-domain generalization ability.

[0007] To achieve the above objectives, this application discloses a structure-semantic dual-graph collaborative SAR image quality assessment method, which includes the following steps: S1: Separate the SAR image into foreground (FG) and background (BG) regions, and construct a dual-channel quality metric set for independent modeling; S2: Obtain indicators and perform correlation calculations to construct an indicator graph, where each node corresponds to an indicator; based on the indicator graph, calculate the dependencies between indicators from both structural and semantic dimensions through self-supervised learning, integrate the dependencies from these two dimensions, and output an image-level score based on the integration result; S3: Construct shared metrics for system analysis and information retrieval-specific metrics for quality evaluation: After extracting the SAR structure sub-graph, integrate IR-specific metrics not included in the SAR prior into the global graph, including: for IR-specific metrics, use pixel-level similarity to address pixel-to-pixel abrupt changes; for IR-specific metrics and general metrics, use image-level statistical methods for correlation analysis.

[0008] Further, step S1 includes: Multiple candidate regions are generated using the global segmentation model SAM; A structure-noise-aware selection strategy is adopted. Candidate regions are scored by integrating Sobel gradient energy, regional contrast and regional signal-to-noise ratio. The region with the highest score is retained as the foreground channel and the remaining regions constitute the background channel, thereby constructing a dual-channel system to achieve decoupled quality modeling. Construct separate sets of quality assessment indicators for the foreground and background; The dual-channel structure is processed using a set of quality assessment metrics.

[0009] Further, step S2 includes: Each node or quality metric i is embedded through a ternary embedding group. The details are as follows: Where d is the embedding vector The total dimension 1 represents the foreground, and 0 represents the background. It is a normalized metric vector. It is the total number of sampled training images. It is a semantic embedding from a dynamic lookup table. It is a dimension derived from the semantic information of the lookup table. It is obtained by calculating the quality metric i of the SAR image, i.e. Each dimension represents a quality metric calculated for a specific image; Constructing a graph model based on structural similarity: Construct a composite structural graph containing three subgraphs: an undirected graph inside the foreground. Undirected graph inside the background And cross-channel directed graphs connecting the background and foreground. ; Semantic attribute-guided graph modeling: provides corrective optimization guidance for the generated structure graph, and enables joint reasoning between the two types of graphs through cross-graph interaction; The semantic attributes include four perceptual categories or indicators: information intensity, clarity, anti-interference, and spatial scale; each indicator is assigned a category label. And obtain its semantic embedding vector through a dynamic semantic lookup table. : Two independent lookup table workflows These correspond to the foreground and background areas, respectively. Constructing the semantic graph: First, a fully connected semantic graph is initialized. During training, by optimizing the weight parameters, it was found that the similarity between categories was consistently low. Finally, the semantic graph evolved into four independent subgraphs, each corresponding to a semantic category. Within each subgraph: metric pairs within the same region have fixed edge weights; metric pairs across regions but belonging to the same category have their weights calculated using cosine similarity. It is the i-th semantic embedding vector; The semantic graph is then sparsified using the Top-k algorithm, where each node retains only its top k neighbor nodes, ultimately yielding the semantic adjacency matrix. ; Graph structure correction: Adjusting the adjacency matrix of the original structure. Using semantic masking: It is the structural adjacency matrix after semantic masking; This operation integrates semantic constraints into the structural topology, pruning connections between semantically irrelevant nodes caused by statistical coincidence; In constructing the structure diagram semantic graph Then, graph convolution operations are performed on each graph separately; subsequently, in the node representation fusion stage, structural and semantic information are jointly modeled through a gating mechanism. Specifically, a two-layer GCN network is used for each graph to obtain the structure-aware representation and semantic-aware representation of node i, respectively: in, The initial node features of the model are represented by N, which is the number of nodes in either the FG or BG region. Each node's feature vector is composed of three parts: its region type, metric, and semantic embedding. ; Let i represent the structural awareness features and semantic awareness features of node i, respectively. These are the hidden dimensions used in each branch; then, a soft-gate mechanism is used to fuse the features of the two branches: here It is the Sigmoid function, [ || ] indicates vector concatenation. It is used to calculate the fusion weights Learnable gated vectors; To construct the training signal, a batch of original SAR images were downgraded in multiple stages to generate image pairs with known relative quality levels. These image pairs were used to define the ranking loss function, thereby driving the model learning. Integrating foreground and background node features and Average pooling is performed on the two sets of features respectively to obtain the pooled foreground and background node features. and The merged pooled features are then input into a multilayer perceptron (MLP) to generate an image score; the higher the score, the higher the perceived image quality. in This represents the feature matrix of FG / BG nodes, which contains the features of all FG / BG nodes. For each image pair ( When the image The quality is better than the image At that time, a margin-based ranking loss function is adopted, which enables the network to autonomously learn quality-aware and transferable node-level representations: in As a confidence threshold, the prediction score for high-quality images must be at least higher than that for low-quality images. , These are images and The prediction quality score.

[0010] Furthermore, and All are constructed using a multi-head metric interaction strategy; for each head Node similarity Computation is performed using projection metric embedding: Here, and Let i and j represent the metric vectors of nodes i and j, respectively. It is the learnable projection matrix of the h-th attention head, where Indicates the embedding dimension in the projection space; The final intra-channel adjacency relationship is obtained by calculating the average value of each end, followed by Top-K filtering: The final adjacency matrix Where N represents the number of nodes in the foreground (FG) or background (BG) region, and its form is: or Specifically depends on type It is the total number of heads; Cross-channel graph construction: A residual gating mechanism is introduced to filter BG features before calculating directional features. Specifically: Here, Let these represent the metric matrices for nodes BG and FG, respectively. It is the number of BG nodes. It is the number of FG nodes. It is the final cross-channel graph matrix, representing the information interaction and transmission after being filtered by the gating mechanism. Map BG features to a learnable gating space. It is a dimension of the gating space, used to control the complexity and learning ability of the gating mechanism; then through Map the features back to the metric space to achieve alignment. It is a non-linear activation function; it yields sparse and meaningful results. ; Final structure diagram This is constructed by concatenating the three subgraphs into a block-like adjacency matrix: The above formula models both intra-channel and inter-channel dependencies.

[0011] Further, step S3 includes: Construct metrics shared with system analysis and metrics specific to information retrieval; Structure Preservation Transfer of Shared Metrics: Extracting SAR Structure Subgraphs ,in Indicates a shared metric node. Represent their connectivity; during the structural migration process, construct the corresponding IR subgraphs. ; It is the set of edges in the IR subgraph, representing the nodes in the IR graph. The connectivity relationships with other nodes are re-estimated using the Pearson correlation coefficient. Both IR and SAR maps are processed using GCN models with the same architecture but independent parameters, with input features... And generate node embedding vectors using their respective adjacency matrices. and To bridge the gap in residual distributions, the maximum mean difference loss function is used to align the SAR and IR feature embeddings. The correlation analysis among IR-specific indicators employs a gradient-based similarity model to capture local spatial variations and establish the correlation relationships among IR-specific indicators; for indicator i, its image response is represented as a matrix. H is the image height, W is the image width, and each element... The index value representing the pixel position; the matrix This is the image response of the metric; at each pixel location, the spatial gradient is calculated. To quantify directional changes; for any pair of metrics and We compare their gradient maps in the image set and calculate the normalized gradient difference to measure the similarity of their local changes, denoted as . This index reflects the overall similarity of their spatial distribution patterns; specifically, for each image, the following is calculated: Where Z is the normalization factor, ensuring that the similarity score is within the range [0,1], and its similarity... At that time, an edge is established between nodes i and j. Threshold, edge weight ; The correlation analysis between IR-specific metrics and shared metrics is as follows: Shared metrics are image vectors. Each dimension represents a global statistical descriptor; a two-dimensional response plot is used to bridge the gap in IR-specific metrics. To assess the differences in representation between features, the mean, standard deviation, entropy, and skewness of each image are calculated, and each feature vector is assigned a specific value. Convert to a four-dimensional vector Apply Z-score normalization across the entire image set; calculate ( The Pearson coefficients between the edges are used as edge weights. Attached Figure Description

[0012] Figure 1 This is a flowchart of the image quality assessment method provided in the embodiments of this application.

[0013] Figure 2 These are the results of the comparative experiments in this application. Detailed Implementation

[0014] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0015] In one embodiment, dual-channel decoupled quality modeling (DC-DQM) includes: This step separates the SAR image into foreground (FG) and background (BG) regions and constructs a dual-channel quality metric set for independent modeling.

[0016] Foreground regions typically correspond to core targets (such as ships and aircraft). Although they occupy only a small portion of the image, they carry crucial structural identification information. However, the overall score is often affected by clean but irrelevant background regions, making damaged targets appear to be of good quality. To address this issue, this application employs a global segmentation model (SAM) (Kirillov et al. 2023) to generate multiple candidate regions (step 1). However, SAM relies on saliency prior features, leading to inaccurate SAR masks. Therefore, this application proposes a structure-noise-aware selection strategy (step 2), which scores candidate regions by fusing a comprehensive metric of Sobel gradient energy, regional contrast, and regional signal-to-noise ratio. The highest-scoring region is retained as the foreground channel, and the remaining regions constitute the background channel, thus constructing a dual-channel system to achieve decoupled quality modeling. Step 3 constructs independent quality evaluation index sets for the foreground and background respectively (see Table 1 for details). Step 4: These indices will be processed through a dual-channel structure. It is worth noting that background evaluation is not discarded. Even with advanced segmentation tools, boundary loss and segmentation errors are still prevalent, leading to the loss of key background structures during the identification process. The ablation studies in this application confirm that removing the background metric significantly reduces performance.

[0017] Table 1 Quality Assessment Indicator Set In one embodiment, the Structure-Semantic Bi-Graph Collaboration Network (SS-DGCN) includes: SS-DGCN first calculates the correlation of indicators obtained from DC-DQM to construct an indicator map, where each node corresponds to one indicator. Based on these indicator maps, the system calculates SAR IQA through self-supervised learning. DGCN consists of four tightly coupled modules: the structure module (SSGM), the semantic module (SAGM), the fusion module (CGRM), and the output module (SSQSB). The structure and semantic modules calculate the dependencies between indicators from structural and semantic dimensions, respectively. The fusion module integrates these dependencies and outputs an image-level score based on the integrated result.

[0018] Each node (metric) i is embedded through a ternary embedding group. The specific steps are as follows: (Step 5) in (1: Foreground, 0: Background) It is a normalized metric vector. It is a semantic embedding from a dynamic lookup table (see section B for details). It is obtained by calculating the quality metric i of a large number of SAR images, i.e. Each dimension represents a quality metric calculated for a specific image. This construction This method enables the subsequent calculation of the statistical correlation between the two metrics, as shown in Equation (2). It is worth noting that the vector dimension... This corresponds to the total number of sampled training images. Through experiments described in the supplementary materials, this application found that when... The model achieves optimal performance when the number of available training images equals 75% of the total number of images.

[0019] Structural Similarity-Based Graph Model (SSGM) To comprehensively model the statistical correlations within an image, SSGM constructs a composite structure graph containing three subgraphs: an undirected graph within the foreground. Undirected graph inside the background And cross-channel directed graphs connecting the background and foreground. .

[0020] Diagram inside the passageway. and All are constructed using a multi-head metric interaction strategy. For each head... Node similarity is calculated using projection metric embedding: Here, and Let i and j represent the metric vectors of nodes i and j, respectively. It is the learnable projection matrix of the h-th attention head, where This represents the embedding dimension in the projection space.

[0021] The final intra-channel adjacency relationship is obtained by calculating the average value of each end, followed by Top-K filtering: The final adjacency matrix (Where N represents the number of nodes in the FG or BG region), its form may be: or Specifically depends on The type.

[0022] Cross-channel graph construction. This application first assumes that there is a fully connected bipartite graph between BG nodes and FG nodes, because each BG metric may affect the FG metric. To optimize this overly dense structure, this application introduces a residual gating mechanism to filter BG features before calculating directional features. Specifically: Here, These represent the metric matrices for nodes BG and FG, respectively. BG features are mapped to a learnable gating space, and then... Features are mapped back to the metric space to achieve alignment. It is a non-linear activation function (e.g., sigmoid). This gated realignment method maintains interpretability while suppressing irrelevant background signals through a data-driven approach, ultimately yielding sparse and meaningful signals. .

[0023] Diagram integration. Final structure diagram. This is constructed by concatenating the three character images into a block-like adjacency matrix: This design enables SSGM to model both intra-channel and inter-channel dependencies simultaneously. However, due to graph heterogeneity—that is, structural connections exist between nodes with significant semantic differences—some edges may contain incorrect connections. To address this issue, subsequent step 8 uses a semantic graph constructed by SAGM and employs semantic masking techniques to... Optimization can both suppress semantically inconsistent connections and preserve task-related structural features.

[0024] Semantic Attribute Guided Graph Modeling (SAGM) To address the issue mentioned in step 6, this application introduces the SAGM module in step 7. SAGM performs a dual function in SS-DGCN: it provides corrective optimization guidance for the structure graph generated by SSGM, and it enables joint reasoning between the two types of graphs through cross-graph interaction in CGRM.

[0025] Semantic attribute definition. This application defines four perception categories based on domain knowledge: (1) information intensity; (2) clarity; (3) anti-interference ability; and (4) spatial scale. Each indicator is assigned a category label. And obtain its semantic embedding vector through dynamic semantic lookup table (DS-LT). : Two independent lookup table workflows Corresponding to the foreground and background regions respectively, parameter sharing is achieved through the DS-LT mechanism, so that 20 metric nodes can complete the task using only 8 semantic embeddings.

[0026] Semantic Graph Construction. This application first initializes a fully connected semantic graph. During training, by optimizing the weight parameters of DS-LT, it was found that the inter-category similarity was consistently low. Ultimately, the semantic graph evolved into four unrelated subgraphs, each corresponding to a semantic category. Within each subgraph: (1) metric pairs within the same region (foreground / background) have a fixed edge weight of 1. (2) metric pairs across regions but belonging to the same category have their weights calculated using cosine similarity: Subsequently, this application employs the ST-Top-k algorithm (Lu and Zhang 2021) to sparsify the semantic graph, retaining only the first k neighbor nodes for each node, ultimately obtaining the semantic adjacency matrix. .

[0027] Graph structure correction. Semantic graphs not only support collaborative reasoning but also act as filters to suppress false positives in structural connections. This is applied to the original structural adjacency matrix. This application employs semantic masking operations: This operation integrates semantic constraints into the structural topology, pruning connections between semantically irrelevant nodes caused by statistical coincidence. Therefore, the model is more robust to noise, redundancy, and local statistical artifacts.

[0028] SAGM ensures that quality metrics are based on both structural affinity and semantic relevance through a dual mechanism of structural semantic optimization and dual-graph reasoning.

[0029] Cross-Graph GCN and Image Quality Scoring Cross-Graph Inference Module (CGRM) In constructing the structure diagram semantic graph Subsequently, CGRM first performs graph convolution operations on each graph separately. Then, in the node representation fusion stage, structural and semantic information are jointly modeled through a gating mechanism (step 9). Specifically, this application employs a two-layer GCN network for each graph to obtain the structure-aware representation and semantic-aware representation of node i, respectively: in, This represents the initial node features of the DC-DQM model. Each node's feature vector is composed of three parts: its region type, metric, and semantic embedding. (As shown in formula (1)). Let i represent the structural awareness features and semantic awareness features of node i, respectively. These are the hidden dimensions used in each branch. By processing the graph structure separately, its unique informational features are preserved, and the impact of various relationships on quality prediction can be clearly assessed. Subsequently, this application employs a soft-gate mechanism to fuse the features of the two branches: here It is the Sigmoid function, [ || ] indicates vector concatenation. It is used to calculate the fusion weights The learnable gated vector, in its specific form It will be selected according to the corresponding region type. or ( or ).

[0030] Self-Supervised Quality Scoring Branch (SSQSB). SSQSB is a learnable module that predicts image quality through a self-supervised learning strategy (step 10). To construct the training signal, this application performs multi-level downgrading processing on a batch of original SAR images (such as adding noise, reducing contrast, etc., specific downgrading levels are described in the supplementary materials) to generate image pairs with known relative quality levels. These image pairs are used to define the ranking loss function, thereby driving model learning.

[0031] Based on CGRM fusion node features and This application performs average pooling on the two sets of features separately, and defines them as follows: and The merged pooled features are then input into a multilayer perceptron (MLP) to generate an image score. A higher score indicates a higher perceived image quality. in This represents the feature matrix of FG / BG nodes, which contains the features of all FG / BG nodes.

[0032] For each image ( ),when quality due to In this application, a margin-based ranking loss function is adopted, which enables the network to autonomously learn quality-aware and transferable node-level representations without the need for manual annotation.

[0033] in As a confidence threshold, the prediction score for high-quality images must be at least higher than that for low-quality images. .

[0034] In one embodiment, the structural pre-adaptation module (SPAM) includes: Existing SAR IQA methods exhibit poor generalization ability on infrared images, stemming from the fundamental difference between thermal correlation metrics and their spatial distribution. However, information retrieval (IR) and structured natural language understanding (SAR) still share a large number of quality metrics, providing a feasible basis for the transfer of structural prior knowledge. But due to significant differences between SAR and IR in semantic categories and perceptual features, effective transfer of semantic prior knowledge is difficult, making reliable alignment between the two challenging.

[0035] Based on this discovery, this application proposes a plug-and-play module—SPAM. This application constructs a complete quality measurement system for information retrieval, including: (1) measurement indicators shared with system analysis; and (2) measurement indicators specific to information retrieval.

[0036] Structure-preserving migration of shared metrics (step 11). This application extracts SAR structure subgraphs. ,in Includes shared metric nodes. This indicates their connection relationships. It should be noted that, due to the weak cross-modal alignment effect, semantic vectors... Not included. During the structural migration process, this application constructed the corresponding IR subgraph. Although SAR and infrared images share the same topology, differences in imaging physics, texture features, and noise levels alter the metric dependency, rendering direct edge weight transfer methods unreliable. Therefore, this application uses the Pearson correlation coefficient based on the IR metric to re-estimate... Both graphs were processed using GCN models with identical architectures but independent parameters (input features). And generate node embedding vectors using their respective adjacency matrices. and To bridge the gap in residual distribution, this application employs the maximum mean difference (MMD) loss function (Gretton et al. 2012) to align SAR and IR feature embeddings.

[0037] For the construction of hybrid similarity for IR-specific indicators. This application integrates IR-specific indicators not included in the SAR prior into the global graph through two connection methods: (1) In terms of IR-specific indicators: pixel-level similarity is used to deal with abrupt changes between pixels; (2) Between infrared-specific indicators and general indicators: image-level statistical methods are used to model stable correlation.

[0038] Correlation analysis among IR-specific indices. (Step 12) This application uses a gradient-based similarity model to capture local spatial changes and establish the correlation between IR-specific indices. For index i, its image response is represented as a matrix. Each element This represents the index value at that pixel location. This matrix can be viewed as the image response to that index. At each pixel location, this application calculates the spatial gradient. To quantify directional changes. For any pair of metrics and This application measures the similarity of their local variations by comparing their gradient maps in the image set and calculating the normalized gradient difference, denoted as... This metric reflects the overall similarity of their spatial distribution patterns. Specifically, for each image, this application calculates: Where Z is the normalization factor, ensuring that the similarity score is within the range [0,1]. Its similarity... At that time, an edge is established between nodes i and j, and the edge weight is... .

[0039] The correlation between IR-specific metrics and shared metrics. (Step 13) Shared metrics are image vectors. (Each dimension represents a global statistical descriptor). A two-dimensional response plot to bridge the gap in IR-specific metrics. To address the differences in representation between images, this application: (1) calculates the indices (mean, standard deviation, entropy, and skewness) for each image and assigns each feature vector to the corresponding feature vector. Convert to a four-dimensional vector (2) Apply Z-score normalization to the entire image set; (3) Calculate ( The Pearson coefficients between the edges are used as edge weights.

[0040] This application uses the DGCQ algorithm to conduct comparative tests with 12 SAR IQA benchmark methods, including some similar methods based on near-infrared (ISAR) technology, to ensure the fairness and transferability of the experiments. Experimental data covers eight publicly available SAR datasets D1-D8, detailed in [link to relevant documentation]. Figure 2 .

[0041] Figure 2This is a performance comparison of eight SAR datasets. Each cell displays accuracy / inference time (% / ms). The best result is shown in bold, and the second best result is indicated by an underline. The datasets include: D1 (SDFSD-v1.0 (Cai et al. 2024)), D2 (SARDet-100K (Li et al. 2024b)), D3 (SSDD (Zhang et al. 2021)), D4 (SRSDD-V1.0 (Lei et al. 2021)), D5 (HRSID (Wei et al. 2020)), D6 (SARAIRcraft-1.0 (Zhirui et al. 2023)), D7 (SADD (Zhang et al. 2023)), and D8 (ADCOS (Hu et al. 2024b)). DGCQ generates quality scores through unlabeled self-supervised training. Since there is a lack of real-world benchmarks for SAR image quality, this application uses downstream recognition performance as the practical evaluation metric—which aligns with the goal of IQA to filter harmful, low-quality images. To achieve this filtering function, this application sorts all data in the dataset by quality score (highest score first) and filters high-quality samples using an adaptive threshold. This threshold is dynamically adjusted based on the score distribution within each dataset.

[0042] Accuracy and Inference Time Comparison Analysis. The DGCQ model proposed in this application consistently outperforms all benchmark methods on the D1-D8 datasets, achieving the highest recognition accuracy in most scenarios while maintaining the shortest inference time. Benchmark methods not only have lower accuracy and longer inference times but also exhibit significant performance fluctuations across different datasets. This instability indicates limitations in their generalization ability when applied to diverse SAR scenarios or imaging conditions. DGCQ, on the other hand, maintains stable performance across all datasets, demonstrating strong cross-dataset adaptability and robustness.

[0043] Recall and Precision Analysis. To evaluate whether DGCQ can effectively retain high-quality images beneficial to downstream recognition tasks, this application conducts a detailed analysis of recall and precision. An image is considered high-quality when it is correctly recognized by at least 8 recognition models on the D1-D8 dataset. This application calculates recall and precision for DGCQ and 12 IQA benchmark models at different Top-Q% thresholds. DGCQ consistently outperforms all benchmark models at different Q values. Even at low Q values, it maintains high recall, indicating its effectiveness in prioritizing key samples; its precision remains stable, showing only a slight decrease with increasing Q value. In contrast, the benchmark methods exhibit greater volatility and lower global reliability. When the recall Q is in the range of 60% to 65%, the DGCQ model achieves the optimal balance between recall and precision.

[0044] Efficiency-Accuracy Cost Analysis. This application further compares DGCQ with all 12 SARIQA benchmark models used in previous experiments, analyzing aspects such as floating-point operations, accuracy, and parameter size. The bubble size corresponds to the number of parameters; the DGCQ model achieves the best accuracy while maintaining low computational cost and a moderate model size.

[0045] This application proposes the DGCQ framework, aiming to overcome the core limitations of existing SAR IQA methods in terms of regional sensitivity, metric relevance, and generalization ability. Through foreground-background decoupling (DC-DQM), structure-semantic graph reasoning (SS-DGCN), and plug-and-play adaptable modules (SPAM), DGCQ achieves state-of-the-art performance on eight SAR datasets and generalizes well to the IR image domain. Experiments show that the model significantly outperforms 12 benchmark models in terms of recognition accuracy, recall, precision, and efficiency, while ablation experiments validate the importance of each component. Overall, DGCQ provides a versatile, interpretable, and efficient solution for cross-modal image quality assessment.

[0046] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A SAR image quality assessment method based on structure-semantic dual-graph collaboration, characterized in that, Includes the following steps: S1: Separate the SAR image into foreground (FG) and background (BG) regions, and construct a dual-channel quality metric set for independent modeling; S2: Obtain indicators and perform correlation calculations to construct an indicator graph, where each node corresponds to an indicator; based on the indicator graph, calculate the dependencies between indicators from both structural and semantic dimensions through self-supervised learning, integrate the dependencies from these two dimensions, and output an image-level score based on the integration result; S3: Construct metrics shared with system analysis and metrics specific to information retrieval for quality evaluation: After extracting the SAR structure sub-graph, integrate IR-specific metrics not included in the SAR prior into the global graph, including: In terms of IR-specific metrics, use pixel-level similarity to deal with abrupt changes between pixels; Correlation analysis was performed using image-level statistical methods between IR-specific and general indicators.

2. The SAR image quality assessment method based on structure-semantic dual-graph collaboration according to claim 1, characterized in that, Step S1 includes: Multiple candidate regions are generated using the global segmentation model SAM; A structure-noise-aware selection strategy is adopted. Candidate regions are scored by integrating Sobel gradient energy, regional contrast and regional signal-to-noise ratio. The region with the highest score is retained as the foreground channel and the remaining regions constitute the background channel, thereby constructing a dual-channel system to achieve decoupled quality modeling. Construct separate sets of quality assessment indicators for the foreground and background; The dual-channel structure is processed using a set of quality assessment metrics.

3. The SAR image quality assessment method based on structure-semantic dual-graph collaboration according to claim 1, characterized in that, Step S2 includes: Each node or quality metric i is embedded through a ternary embedding group. The details are as follows: Where d is the embedding vector The total dimension 1 represents the foreground, and 0 represents the background. It is a normalized metric vector. It is the total number of sampled training images. It is a semantic embedding from a dynamic lookup table. It is a dimension derived from the semantic information of the lookup table. It is obtained by calculating the quality metric i of the SAR image, i.e. Each dimension represents a quality metric calculated for a specific image; Constructing a graph model based on structural similarity: Construct a composite structural graph containing three subgraphs: an undirected graph inside the foreground. Undirected graph inside the background And cross-channel directed graphs connecting the background and foreground. ; Semantic attribute-guided graph modeling: provides corrective optimization guidance for the generated structure graph, and enables joint reasoning between the two types of graphs through cross-graph interaction; The semantic attributes include four perceptual categories or indicators: information intensity, clarity, anti-interference, and spatial scale; each indicator is assigned a category label. And obtain its semantic embedding vector through a dynamic semantic lookup table. : Two independent lookup table workflows These correspond to the foreground and background areas, respectively. Constructing the semantic graph: First, a fully connected semantic graph is initialized. During training, by optimizing the weight parameters, it was found that the similarity between categories was consistently low. Finally, the semantic graph evolved into four independent subgraphs, each corresponding to a semantic category. Within each subgraph: metric pairs within the same region have fixed edge weights; metric pairs across regions but belonging to the same category have their weights calculated using cosine similarity. It is the i-th semantic embedding vector; The semantic graph is then sparsified using the Top-k algorithm, where each node retains only its top k neighbor nodes, ultimately yielding the semantic adjacency matrix. ; Graph structure correction: Adjusting the adjacency matrix of the original structure. Using semantic masking: It is the structural adjacency matrix after semantic masking; This operation integrates semantic constraints into the structural topology, pruning connections between semantically irrelevant nodes caused by statistical coincidence; In constructing the structure diagram semantic graph Then, graph convolution operations are performed on each graph separately; subsequently, in the node representation fusion stage, structural and semantic information are jointly modeled through a gating mechanism. Specifically, a two-layer GCN network is used for each graph to obtain the structure-aware representation and semantic-aware representation of node i, respectively: in, The initial node features of the model are represented by N, which is the number of nodes in either the FG or BG region. Each node's feature vector is composed of three parts: its region type, metric, and semantic embedding. ; Let i represent the structural awareness features and semantic awareness features of node i, respectively. These are the hidden dimensions used in each branch; then, a soft-gate mechanism is used to fuse the features of the two branches: here It is the Sigmoid function, [ || ] indicates vector concatenation. It is used to calculate the fusion weights Learnable gated vectors; To construct the training signal, a batch of original SAR images were downgraded in multiple stages to generate image pairs with known relative quality levels. These image pairs were used to define the ranking loss function, thereby driving the model learning. Integrating foreground and background node features and Average pooling is performed on the two sets of features respectively to obtain the pooled foreground and background node features. and The merged pooled features are then input into a multilayer perceptron (MLP) to generate an image score; the higher the score, the higher the perceived image quality. in This represents the feature matrix of FG / BG nodes, which contains the features of all FG / BG nodes. For each image pair ( When the image The quality is better than the image At that time, a margin-based ranking loss function is adopted, which enables the network to autonomously learn quality-aware and transferable node-level representations: in As a confidence threshold, the prediction score for high-quality images must be at least higher than that for low-quality images. , These are images and The prediction quality score.

4. The SAR image quality assessment method based on structure-semantic dual-graph collaboration according to claim 3, characterized in that, and All are constructed using a multi-head metric interaction strategy; for each head Node similarity Computation is performed using projection metric embedding: Here, and Let i and j represent the metric vectors of nodes i and j, respectively. It is the learnable projection matrix of the h-th attention head, where Indicates the embedding dimension in the projection space; The final intra-channel adjacency relationship is obtained by calculating the average value of each end, followed by Top-K filtering: The final adjacency matrix Where N represents the number of nodes in the foreground (FG) or background (BG) region, and its form is: or Specifically depends on type It is the total number of heads; Cross-channel graph construction: A residual gating mechanism is introduced to filter BG features before calculating directional features. Specifically: Here, Let these represent the metric matrices for nodes BG and FG, respectively. It is the number of BG nodes. It is the number of FG nodes. It is the final cross-channel graph matrix, representing the information interaction and transmission after being filtered by the gating mechanism. Map BG features to a learnable gating space. It is a dimension of the gating space, used to control the complexity and learning ability of the gating mechanism; then through Map the features back to the metric space to achieve alignment. It is a non-linear activation function; Obtain sparse and meaningful ; Final structure diagram This is constructed by concatenating the three subgraphs into a block adjacency matrix: The above formula models both intra-channel and inter-channel dependencies.

5. The SAR image quality assessment method based on structure-semantic dual-graph collaboration according to claim 4, characterized in that, Step S3 includes: Construct metrics shared with system analysis and metrics specific to information retrieval; Structure Preservation Transfer of Shared Metrics: Extracting SAR Structure Subgraphs ,in Indicates a shared metric node. Represent their connectivity; during the structural migration process, construct the corresponding IR subgraphs. ; It is the set of edges in the IR subgraph, representing the nodes in the IR graph. The connectivity relationships with other nodes are re-estimated using the Pearson correlation coefficient. Both IR and SAR maps are processed using GCN models with the same architecture but independent parameters, with input features... And generate node embedding vectors using their respective adjacency matrices. and To bridge the gap in residual distributions, the maximum mean difference loss function is used to align the SAR and IR feature embeddings. The correlation analysis among IR-specific indicators employs a gradient-based similarity model to capture local spatial variations and establish the correlation relationships among IR-specific indicators; for indicator i, its image response is represented as a matrix. H is the image height, W is the image width, and each element... The index value representing the pixel position; the matrix This is the image response of the metric; at each pixel location, the spatial gradient is calculated. To quantify directional changes; for any pair of metrics and We compare their gradient maps in the image set and calculate the normalized gradient difference to measure the similarity of their local changes, denoted as . This index reflects the overall similarity of their spatial distribution patterns; specifically, for each image, the following is calculated: Where Z is the normalization factor, ensuring that the similarity score is within the range [0,1], and its similarity... At that time, an edge is established between nodes i and j. Threshold, edge weight ; The correlation analysis between IR-specific metrics and shared metrics is as follows: Shared metrics are image vectors. Each dimension represents a global statistical descriptor; a two-dimensional response plot is used to bridge the gap in IR-specific metrics. To assess the differences in representation between features, the mean, standard deviation, entropy, and skewness of each image are calculated, and each feature vector is assigned a specific value. Convert to a four-dimensional vector Apply Z-score normalization across the entire image set; calculate ( The Pearson coefficients between the edges are used as edge weights.

Images