CN-121998902-A - SAR image quality evaluation method based on structure-semantic dual-graph cooperation

Abstract

The application relates to a SAR image quality evaluation method based on structure-semantic dual-graph cooperation, which comprises the following steps of separating SAR images, constructing a dual-channel quality measurement set for independent modeling, acquiring indexes, performing correlation calculation, constructing an index map, wherein each node corresponds to one index, integrating the dependency relationship between two dimensions of a structure and a semantic respectively through self-supervision learning based on the index map, outputting an image level score based on an integration result, constructing a measurement index shared with system analysis and a measurement index special for information retrieval, performing quality measurement evaluation, and integrating IR special indexes which are not contained in SAR priori into a global graph after a structural subgraph is extracted. The application is superior to the reference model in the aspects of recognition accuracy, recall rate, accuracy and efficiency, and provides a solution with strong universality and good interpretability for cross-mode image quality evaluation.

Inventors

• WANG RUIQI
• HAN RUIXUAN
• LU MINGMING
• Tao chao
• LI HAIFENG
• SONG WEI
• YANG SHUO
• ZHANG ZHEXUAN

Assignees

• 中南大学
• 航天科工智能运筹与信息安全研究院(武汉)有限公司
• 北京机电工程研究所

Dates

Publication Date

20260508

Application Date

20251215

Claims (5)

1. 1. The SAR image quality evaluation method based on the structure-semantic dual-graph cooperation is characterized by comprising the following steps of: S1, separating an SAR image into a foreground FG region and a background BG region, and constructing a dual-channel quality measurement set for independent modeling; S2, acquiring indexes and performing correlation calculation to construct an index map, wherein each node corresponds to one index; based on the index map, calculating the dependency relationship between indexes from two dimensions of the structure and the semantics respectively through self-supervision learning, integrating the dependency relationship of the two dimensions, and outputting an image grade score based on an integrated result; And S3, constructing a measurement index shared with system analysis and a measurement index special for information retrieval, and performing quality measurement evaluation, namely, after extracting a SAR structural subgraph, integrating IR special indexes which are not included in SAR priori into a global graph, wherein the IR special indexes adopt pixel-level similarity to deal with inter-pixel mutation, and an image-level statistical method is used for performing correlation analysis between the IR special indexes and the universal indexes.
2. 2. The SAR image quality evaluation method of structure-semantic dual graph synergy of claim 1, wherein step S1 comprises: Generating a plurality of candidate areas by adopting a global segmentation model SAM; scoring candidate areas by adopting a structure-noise perception selection strategy and integrating comprehensive measures of Sobel gradient energy, area contrast and area signal to noise ratio, wherein the highest partitioned area is reserved as a foreground channel, and the rest areas form a background channel, so that a dual-channel decoupling quality modeling is realized; respectively constructing independent quality evaluation index sets for the foreground and the background; the dual channel structure is processed using the quality assessment index set.
3. 3. The SAR image quality evaluation method of structure-semantic dual graph synergy of claim 1, wherein step S2 comprises: Each node or quality metric i passes through a ternary embedding set The representation is as follows: where d is the embedded vector Is used to determine the overall dimensions of the (c) device, 1 Represents a foreground, 0 represents a background, Is a normalized metric vector of the values of the metric, Is the total number of sampled training images, Is semantic embedding from a dynamic look-up table, Is the dimension of the semantic information from the look-up table, Obtained by calculating a quality measure i of the SAR image, i.e Representing a quality metric calculated for a particular image; constructing a graph model based on structural similarity, namely constructing a composite structure diagram, wherein the composite structure diagram comprises three subgraphs, namely an undirected graph in the foreground Undirected graph inside background A cross-channel directed graph connecting background and foreground ; Providing revised optimization guidance for the generated structure diagram, and realizing joint reasoning of the two types of diagrams through cross-diagram interaction; The semantic attributes comprise four perception categories or indexes, namely information intensity, definition, anti-interference performance and spatial scale, wherein each index is endowed with a category label And obtain its semantic embedded vector through dynamic semantic look-up table : Two independent look-up table workflows Respectively corresponding to a foreground region and a background region; The method comprises the steps of firstly initializing a full-connection semantic graph, finding out that similarity among categories is always low through optimizing weight parameters in a training process, finally evolving the semantic graph into four mutually uncorrelated subgraphs, wherein each word graph corresponds to one semantic category, and in each subgraph, measuring pairs in the same area have fixed edge weights, measuring pairs in cross-area but same category are obtained through cosine similarity calculation, wherein the weights of the measuring pairs are obtained through cosine similarity calculation: Is the i-th semantic embedded vector; then, the Top-k algorithm is adopted to carry out sparse processing on the semantic graph, each node only keeps the first k neighbor nodes, and finally, a semantic adjacency matrix is obtained ; Correction of graph structure, i.e. adjacency matrix to original structure Semantic masking operations are employed: Is a structure adjacency matrix after semantic mask operation; the operation integrates semantic constraints into the structure topology, pruning connections between semantic independent nodes generated by statistical coincidences; in the construction of the structural diagram And semantic graph Then, in the node representation fusion stage, the structure and semantic information are jointly modeled through a gating mechanism; Specifically, a two-layer GCN network is adopted for each graph, and the structural perception representation and the semantic perception representation of the node i are respectively obtained: Wherein, the Representing initial node characteristics of the model, wherein N is the node number of FG or BG regions, and the characteristic vector of each node is formed by splicing three parts of region type, metric value and semantic embedding ; Respectively representing structural and semantic perceptual features of the node i, Is the hidden dimension used in each branch, and then the features of the two branches are fused by a soft door mechanism: Here, the Is a Sigmoid function [ || And is representative of the vector concatenation, Is used for calculating the fusion weight Is a learning gating vector; To construct a training signal, a batch of raw SAR images is subjected to a multi-stage degradation process, generating image pairs with known relative quality levels, which are used to define a ranking loss function, thereby driving model learning; Fusing foreground and background node features And Respectively carrying out average pooling treatment on the two groups of features to obtain pooled foreground and background node features And And inputting the merged pooled features into a multi-layer perceptron MLP to generate a picture score ŷ, wherein the higher the score is, the higher the image quality perception level is: Wherein the method comprises the steps of Representing FG/BG node feature matrix, including features of all FG/BG nodes; For each image pair ) When an image is Is better than the image When the method is used, a ranking loss function based on edge margins is adopted, and the mechanism promotes the network to autonomously learn quality-aware and migratable node level representation: Wherein the method comprises the steps of For the confidence threshold, the predictive score for a high quality image is required to be at least higher than that for a low quality image , Respectively, are images And Is included.
4. 4. The SAR image quality assessment method based on the structure-semantic dual graph cooperation as set forth in claim 3, wherein, And All adopt multi-head measurement interaction strategy construction, and for each head Node similarity Calculation by projection metric embedding: here the number of the elements is the number, And Representing the metric vectors of nodes i and j respectively, Is a learning projection matrix of an h-th attention head, wherein Representing an embedding dimension in the projection space; the final intra-channel adjacency is obtained by the average value calculation of each head, followed by Top-K filtering: The resulting adjacency matrix Where N represents the number of nodes in the foreground FG or background BG region in the form of Or (b) Depending on Is of the type(s) of (a), Is the total number of heads; The cross-channel diagram construction, namely introducing a residual gating mechanism, and filtering BG features before calculating directivity features, specifically: here the number of the elements is the number, Metric matrices representing BG nodes and FG nodes respectively, Is the number of BG nodes and, Is the number of FG nodes and, Is the final cross-channel graph matrix, represents the information interaction and transmission after the filtering by the gating mechanism, BG features are mapped to a learnable gating space, Is the dimension of the gating space for controlling the complexity and learning ability of the gating mechanism, followed by Features are mapped back into the metric space to achieve alignment, Nonlinear activation function, sparse and meaningful ; Final structure diagram The construction is performed by stitching the three sub-graphs into a block adjacency matrix: the above equation models both intra-and inter-channel dependencies.
5. 5. The SAR image quality evaluation method according to claim 4, wherein step S3 comprises: constructing a measurement index shared with system analysis and a measurement index specific to information retrieval; Structure preserving migration of shared metrics, extraction of SAR structure subgraphs Wherein Representing the nodes of the shared metric, Representing the connection relation, constructing corresponding IR subgraph during the structure migration process ; Is the set of edges in the IR graph, represented in the IR graph by nodes Connection relationships with other nodes using Pearson correlation coefficients to re-estimate The IR spectrum and SAR spectrum are processed by GCN model with same architecture but independent parameters, and the characteristics are input And generating node embedded vectors by using respective adjacent matrixes And For bridging the residual distribution gap, adopting a maximum mean difference loss function to align SAR and IR characteristic embedding; Correlation analysis between IR specific indexes, capturing local spatial variation by adopting a similarity model based on gradient, and establishing a correlation between IR specific indexes, wherein for index i, the image response is expressed as a matrix H is the image height, W is the image width, where each element An index value representing the pixel position, the matrix Is the image response of the index, calculates the spatial gradient at each pixel location To quantify the directional change for any pair of metrics And Comparing their gradient maps in the image set and calculating normalized gradient differences to measure their local variation similarity, noted as The index reflects the overall similarity of their spatial distribution pattern, in particular, for each image, the calculation: Wherein Z is a normalization factor that ensures that the similarity score is in the [0,1] range, the similarity thereof When an edge is established between nodes i and j, As threshold value, edge weight ; The correlation analysis of the IR specific index with the shared index is as follows, the shared index is an image vector Each dimension represents a global statistical descriptor, a two-dimensional corresponding graph for bridging IR-specific indicators The characterization difference between the two images is calculated, the mean value, standard deviation, entropy and skewness of each image are calculated, and each characteristic vector is calculated Conversion to four-dimensional vectors Z-score normalization is applied to the whole image set, and the calculation is performed ) The Pearson coefficient in between as the edge weight.

Description

SAR image quality evaluation method based on structure-semantic dual-graph cooperation Technical Field The invention belongs to the technical field of radars, and particularly relates to a SAR image quality evaluation method based on structure-semantic dual-graph cooperation. Background SAR image technology has key effects in the aspects of remote sensing fields such as target detection, land coverage classification and the like by virtue of all-weather operation capability and quantitative earth observation potential. However, SAR images often have quality problems such as speckle noise and geometric distortion. Since downstream tasks are highly dependent on image quality, low quality SAR data can reduce accuracy or cause task failure, SAR Image Quality Assessment (IQA) is critical. Although SAR image quality is of significant impact, the global mainstream scoring approach generally ignores foreground versus background differentiation. This often results in a clean background that masks the degraded foreground, thereby producing an assessment bias. While pure foreground training can promote effects in a controlled environment, context/structural cues are lost and generalization ability is compromised. Furthermore, unfiltered background content may mislead recognition and degrade performance. These limitations expose the dual disadvantage of a unified assessment paradigm (ignoring regional differences) with an isolated assessment paradigm (ignoring foreground-background interactions). The existing SAR IQA method mostly ignores complex inter-index relationships. The application classifies metric relationships into two categories, namely statistical correlation (for example, in a low texture region, common variation fluctuation of gradient and texture energy (two metrics) shows degradation of perception characteristic sharing and semantic dependence (indexes share perception characteristics such as information intensity/definition/anti-interference performance and represent semantic meaning based on people)), fusion indexes can reduce accuracy of image quality assessment under the condition of lacking of structure/semantic modeling, so that the relationship modeling is crucial to robust assessment. Despite advances, most SAR IQA methods ignore generalization capability, resulting in limited applicability across datasets/modalities. Although some studies have noted the existence of common perceptual features (e.g., blurring, contrast loss, noise) between multiple modes, such as radar/optical/infrared, the current studies have not utilized these features to enhance the generalization ability of SAR IQA. Disclosure of Invention Synthetic Aperture Radar (SAR) is critical in remote sensing tasks such as target detection and recognition, and its image quality directly affects task performance. However, the existing SAR image quality assessment method has three major limitations that the overall image assessment is relied on, so that a high-quality background can mask the fuzzy degradation phenomenon of a small target area, although various quality indexes are proposed, most methods process the indexes in an isolated mode and ignore the inherent association between the indexes, and the limited generalization capability limits the adaptability of the cross-data set and the imaging mode. To solve these challenges, the application proposes a dual-graph collaborative quality (DGCQ) assessment framework, which is an innovative solution designed specifically for SAR image quality assessment, and is characterized in that (1) foreground/background decoupling, (2) index relationship modeling, and (3) cross-modal adaptation are performed. Specifically: The DC-DQM realizes foreground and background decoupling through region segmentation and specific region index extraction, so that the masking effect of global scores on local degradation is reduced. The SS-DGCN model represents the measurement relation through double graphs, namely a structure measurement graph and a semantic measurement graph, wherein nodes represent measurement indexes, and edges encode statistical correlation or semantic dependency relation. The graph rolling network (GCN) propagates node features followed by node-level gating fusion to integrate the structure/semantic features. The fusion features are aggregated into a graph-level representation for self-supervising IQA scoring. SPAM enhances generalization capability through cross-domain plug and play architecture prior migration without modifying the backbone network. DGCQ comprises three core modules, namely (1) a dual-channel decoupling quality modeling (DC-DQM) module divides an image into a foreground channel and a background channel, and designs differentiated quality indexes according to the foreground channel and the background channel, (2) a structure-semantic dual-graph collaborative network (SS-DGCN) propagates the relationship among modeling indexes through dual graphs, namely, a structure gra