Method and device for interpretability analysis of SAR image classification network, and computer readable storage medium

Abstract

The invention belongs to the technical field of SAR image classification network interpretation, and particularly relates to an interpretability analysis method and device of an SAR image classification network and a computer readable storage medium. The method comprises the following steps: firstly, calculating a Shapley value of a feature of an SAR image input into an SAR image classification network by adopting an SHAP algorithm so as to obtain an original SHAP saliency map; then, adopting a gradient-based attribution algorithm for the features of the SAR image to obtain a gradient saliency map for evaluating feature importance; and finally, fusing the original SHAP saliency map and the gradient saliency map to obtain a fused SHAP saliency map, and carrying out interpretability analysis on the SAR image classification network by using the fused SHAP saliency map. According to the method, the SHAP method is improved from the angle of increasing the gradient weight, the problem of misinterpretation encountered by a traditional SHAP and the interpretability challenge faced by the traditional SHAP in an SAR recognition task are solved, the intuition and comprehensiveness of model interpretation are further improved, and a new view angle is provided for deep analysis and optimization of an image classification model.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of SAR image classification network interpretation, and particularly relates to an interpretable analysis method and device for a SAR image classification network and a computer readable storage medium. BACKGROUND

[0002] Synthetic Aperture Radar (SAR) is a kind of microwave active observation sensor that can realize high resolution, and SAR target classification can obtain the category information of the target of interest from the SAR image, while the deep learning network can automatically learn the features from the data set, so the SAR target classification method based on the deep learning network becomes an important research direction in the field. However, the feature extraction of the deep learning network has the black box characteristic, which leads to the lack of interpretability of the decision-making process and the difficulty in determining the causal relationship between the input and the output, so it is necessary to perform interpretability analysis.

[0003] A common method in network interpretation is to attribute the prediction to the input feature set. For an image classifier, the importance weight is obtained by activating a specific pixel alliance. The first to use this idea is some perturbation-based algorithms, which measure the importance of features by adding feature masks and inputting the network to obtain confidence reduction. These algorithms often optimize the network interpretation, but for complex high-dimensional data, the convergence of the optimizer is difficult to guarantee, and these algorithms cannot directly link the feature attribution and confidence in numerical terms. From another perspective, some current works have proposed distortion, sensitivity, and Taylor interaction effect to objectively evaluate network interpretation methods, but these algorithms have been proven to be unable to give reasonable explanations from the perspective of mechanism soundness.

[0004] A production business process timeout prediction method based on interpretable artificial intelligence is disclosed in Chinese patent application CN118822028A published on October 22, 2024. In order to improve the interpretability of the TCN model, the SHAP value analysis is used to determine the specific contribution of each feature to the prediction result. The complex prediction result and feature contribution are presented in a visual way through the construction of the SHAP summary graph, so that business experts can intuitively understand the prediction logic of the model, thereby improving the credibility of the model and the effectiveness of decision support. Although the SHAP value analysis is used to improve the interpretability of the TCN model in the patent, the SHAP value itself may encounter computational efficiency and interpretability limitations when dealing with high-dimensional data, especially when the number of features is large, which leads to biased final interpretation.

[0005] Yang Zhiyuan of the University of Electronic Science and Technology of China's 2025 master's degree thesis "Research on deep network explainability method for SAR target classification task" proposes an explainable algorithm framework for SAR image classification task based on the occlusion method. In order to accurately reflect the different effects of different regions in the SAR image on the overall classification result, a multi-variable interaction and SAR clutter distribution based occlusion method is further proposed, which eliminates the causal information of the original region as much as possible without introducing additional information. This method ensures the reliability of the SHAP method used in the attribution process by estimating the optimal baseline value of the Shapley value calculation. Through qualitative and quantitative methods, the reliability of the baseline value is established.

[0006] SHAP has a solid mathematical foundation, allowing researchers to explain the network from a deep mechanism, with high fidelity and robustness, and has a significant advantage in providing model prediction explanation. However, the SHAP method is usually based on the assumption that the features of the model have no correlation, so there will be problems of insufficient representation accuracy and target mismatch; in addition, the SHAP method mainly provides local explanation (i.e. explanation for a single prediction), although SHAP values can be aggregated to obtain a global perspective, but it may mask the subtle differences in model behavior, which in turn limits the comprehensive understanding of the overall behavior and decision-making process of the model. SUMMARY

[0007] The purpose of the present application is to provide an explainability analysis method and device for SAR image classification network, and a computer readable storage medium, to solve the problem of insufficient accuracy and target mismatch caused by the application of the existing SHAP method to the SAR image target classification task.

[0008] To solve the above technical problems, the present application provides a technical scheme of an explainability analysis method for SAR image classification network, which specifically comprises:

[0009] An explainability analysis method for SAR image classification network, comprising the following steps:

[0010] The SHAP algorithm is used to calculate the Shapley value of the features of the SAR image input into the SAR image classification network, thereby obtaining the original SHAP saliency map;

[0011] A gradient-based attribution algorithm is used to obtain a gradient saliency map for evaluating the importance of the features of the SAR image;

[0012] The original SHAP saliency map and the gradient saliency map are fused to obtain a fused SHAP saliency map, and the fused SHAP saliency map is used for explainability analysis of the SAR image classification network.

[0013] The beneficial effects of the above technical solutions are: the explainability analysis method of the SAR image classification network is a gradient-based SHAP explainability method, the gradient-based attribution algorithm can reduce the average fluctuation of a single gradient point, so that important features are more prominent, therefore, the gradient-based attribution algorithm is combined with the SHAP method to overcome the limitations of the traditional SHAP method, the SHAP method is improved from the perspective of increasing gradient weight, the problem of incorrect interpretation encountered by the traditional SHAP is solved, the explainability challenge faced by the traditional SHAP in the SAR identification task is solved, the intuitiveness and comprehensiveness of model explanation are further improved, a new perspective is provided for in-depth analysis and optimization of the image classification model, and the important value in improving model transparency and decision explainability is shown.

[0014] Further, the way to obtain the fused SHAP saliency map is: the original SHAP saliency map and the gradient saliency map belonging to the same predicted category are pixel-by-pixel multiplied to obtain an operation result of the same predicted category, and then the operation results of different categories are added to obtain the fused SHAP saliency map; the predicted category is a predicted category of the SAR image classification network.

[0015] Further, the gradient-based attribution algorithm is an integral gradient algorithm or a smooth gradient algorithm.

[0016] Further, when the gradient-based attribution algorithm adopts the smooth gradient algorithm, the way to use the fused SHAP saliency map to perform explainability analysis on the SAR image classification network is: the SAR image is directly fused with the fused SHAP saliency map to obtain a comprehensive explanation map of the SAR image fused with feature importance and input, and the comprehensive explanation map is used to perform explainability analysis on the SAR image classification network.

[0017] Further, when the gradient-based attribution algorithm adopts the smooth gradient algorithm, the way to use the fused SHAP saliency map to perform explainability analysis on the SAR image classification network is: the SAR image is subjected to regularization processing, the SAR image subjected to the regularization processing is fused with the fused SHAP saliency map to obtain a comprehensive explanation map of the SAR image fused with feature importance and input, and the comprehensive explanation map is used to perform explainability analysis on the SAR image classification network.

[0018] Further, the way to fuse the SAR image with the fused SHAP saliency map is: the SAR image and the fused SHAP saliency map belonging to the same channel are pixel-by-pixel multiplied to obtain an operation result of the same channel, and then operation results of different channels are added.

[0019] Further, the manner of fusing the normalized SAR image and the fused SHAP saliency map is: performing pixel-by-pixel multiplication operation on the normalized SAR image and the fused SHAP saliency map belonging to the same channel to obtain an operation result of the same channel, and then adding the operation results of different channels.

[0020] Further, the manner of obtaining the fused SHAP saliency map is: performing pixel-by-pixel multiplication operation on the original SHAP saliency map and the gradient saliency map belonging to the same prediction category to obtain an operation result of the same prediction category, pre-processing the operation result of the same prediction category to only focus on positive information, and then adding the pre-processed operation results of different categories to obtain the fused SHAP saliency map; the prediction category is a prediction category of the SAR image classification network.

[0021] Further, the preprocessing is ReLU activation function processing.

[0022] To solve the above technical problems, the application further provides a technical scheme of a SAR image classification network explainability analysis device, specifically comprising:

[0023] A SAR image classification network explainability analysis device comprises a processor, and the processor is used to execute a computer program to realize the steps of the following method:

[0024] The SHAP algorithm is used to calculate the Shapley value of the feature of the SAR image input into the SAR image classification network, so as to obtain an original SHAP saliency map;

[0025] A gradient-based attribution algorithm is used to obtain a gradient saliency map for evaluating the importance of the feature of the SAR image;

[0026] The original SHAP saliency map and the gradient saliency map are fused to obtain a fused SHAP saliency map, and the fused SHAP saliency map is used for SAR image classification network explainability analysis.

[0027] The beneficial effects of the above technical solutions are: the SAR image classification network explainability analysis device provides hardware support for the SAR image classification network explainability analysis method, the SAR image classification network explainability analysis method is a SHAP explainability method based on gradients, the attribution algorithm based on gradients can reduce the average fluctuation of a single gradient point, so that important features are more prominent, therefore, the attribution algorithm based on gradients is combined with the SHAP method to overcome the limitations of the traditional SHAP method, the SHAP method is improved from the perspective of increasing gradient weights, the problems of incorrect interpretation encountered by the traditional SHAP and the explainability challenges faced by the traditional SHAP in the SAR identification task are solved, the intuitiveness and comprehensiveness of model explanation are further improved, a new perspective is provided for in-depth analysis and optimization of the image classification model, and important values in improving model transparency and decision explainability are shown.

[0028] Further, the way to obtain the fused SHAP saliency map is: the original SHAP saliency map and the gradient saliency map belonging to the same predicted category are pixel-by-pixel multiplied to obtain an operation result of the same predicted category, and then the operation results of different categories are added to obtain the fused SHAP saliency map; the predicted category is a predicted category of the SAR image classification network.

[0029] Further, the attribution algorithm based on gradients is an integral gradient algorithm or a smooth gradient algorithm.

[0030] Further, when the attribution algorithm based on gradients is the smooth gradient algorithm, the way to use the fused SHAP saliency map to perform the explainability analysis of the SAR image classification network is: the SAR image is directly fused with the fused SHAP saliency map to obtain a comprehensive explanation map of the SAR image fused with feature importance and input, and the comprehensive explanation map is used to perform the explainability analysis of the SAR image classification network.

[0031] Further, when the attribution algorithm based on gradients is the smooth gradient algorithm, the way to use the fused SHAP saliency map to perform the explainability analysis of the SAR image classification network is: the SAR image is subjected to regularization processing, the SAR image subjected to the regularization processing is fused with the fused SHAP saliency map to obtain a comprehensive explanation map of the SAR image fused with feature importance and input, and the comprehensive explanation map is used to perform the explainability analysis of the SAR image classification network.

[0032] Further, the way to fuse the SAR image with the fused SHAP saliency map is: the SAR images belonging to the same channel are pixel-by-pixel multiplied with the fused SHAP saliency map to obtain operation results of the same channel, and then the operation results of different channels are added.

[0033] Further, the manner of fusing the normalized SAR image and the fused SHAP saliency map is: performing pixel-by-pixel multiplication operation on the normalized SAR image and the fused SHAP saliency map belonging to the same channel to obtain an operation result of the same channel, and then adding the operation results of different channels.

[0034] Further, the manner of obtaining the fused SHAP saliency map is: performing pixel-by-pixel multiplication operation on the original SHAP saliency map and the gradient saliency map belonging to the same prediction category to obtain an operation result of the same prediction category, pre-processing the operation result of the same prediction category to only focus on positive information, and then adding the pre-processed operation results of different categories to obtain the fused SHAP saliency map; the prediction category is a prediction category of the SAR image classification network.

[0035] Further, the preprocessing is ReLU activation function processing.

[0036] To solve the above technical problems, the application further provides a technical scheme of a computer readable storage medium, specifically comprising:

[0037] The computer readable storage medium of the application has a computer program stored thereon, and the computer program realizes the steps of the following method when executed by a processor:

[0038] The SHAP algorithm is used to calculate the Shapley value of the feature of the SAR image input into the SAR image classification network, so as to obtain an original SHAP saliency map;

[0039] A gradient-based attribution algorithm is used to obtain a gradient saliency map for evaluating the importance of the feature of the SAR image;

[0040] The original SHAP saliency map and the gradient saliency map are fused to obtain a fused SHAP saliency map, and the fused SHAP saliency map is used for explainability analysis of the SAR image classification network.

[0041] The beneficial effects of the above technical solutions are that the computer readable storage medium of the application stores a software program code corresponding to the explainability analysis method of the SAR image classification network of the application, and provides hardware support for implementing the method. The explainability analysis method of the SAR image classification network is a gradient-based SHAP explainability method. Considering that the gradient-based attribution algorithm can reduce the average fluctuation of a single gradient point and make important features more prominent, the gradient-based attribution algorithm is combined with the SHAP method to overcome the limitations of the traditional SHAP method, the SHAP method is improved from the perspective of increasing gradient weight, the problems of incorrect interpretation encountered by the traditional SHAP and the explainability challenges faced by the traditional SHAP in the SAR identification task are solved, the intuitiveness and comprehensiveness of model explanation are further improved, a new perspective is provided for in-depth analysis and optimization of the image classification model, and important value in improving model transparency and decision explainability is shown.

[0042] Further, the way to obtain the fused SHAP saliency map is that the original SHAP saliency map and the gradient saliency map belonging to the same predicted category are pixel-by-pixel multiplied to obtain an operation result of the same predicted category, and then the operation results of different categories are added to obtain the fused SHAP saliency map; the predicted category is a predicted category of the SAR image classification network.

[0043] Further, the gradient-based attribution algorithm is an integral gradient algorithm or a smoothing gradient algorithm.

[0044] Further, when the gradient-based attribution algorithm adopts the smoothing gradient algorithm, the way to use the fused SHAP saliency map to perform explainability analysis on the SAR image classification network is that the SAR image is directly fused with the fused SHAP saliency map to obtain a comprehensive explanation map of the SAR image fused with feature importance and input, and the comprehensive explanation map is used to perform explainability analysis on the SAR image classification network.

[0045] Further, when the gradient-based attribution algorithm adopts the smoothing gradient algorithm, the way to use the fused SHAP saliency map to perform explainability analysis on the SAR image classification network is that the SAR image is subjected to regularization processing, the SAR image subjected to the regularization processing is fused with the fused SHAP saliency map to obtain a comprehensive explanation map of the SAR image fused with feature importance and input, and the comprehensive explanation map is used to perform explainability analysis on the SAR image classification network.

[0046] Further, the way to fuse the SAR image with the fused SHAP saliency map is that the SAR images belonging to the same channel are pixel-by-pixel multiplied with the fused SHAP saliency map to obtain operation results of the same channel, and then the operation results of different channels are added.

[0047] Further, the way of fusing the regularized SAR image and the fused SHAP saliency map is: performing pixel-by-pixel multiplication operation on the regularized SAR image and the fused SHAP saliency map belonging to the same channel to obtain an operation result of the same channel, and then adding the operation results of different channels.

[0048] Further, the way of obtaining the fused SHAP saliency map is: performing pixel-by-pixel multiplication operation on the original SHAP saliency map and the gradient saliency map belonging to the same prediction category to obtain an operation result of the same prediction category, pre-processing the operation result of the same prediction category to only focus on positive information, and then adding the pre-processed operation results of different categories to obtain the fused SHAP saliency map; the prediction category is the prediction category of the SAR image classification network.

[0049] Further, the pre-processing is ReLU activation function processing. BRIEF DESCRIPTION OF DRAWINGS

[0050] Figure 1 is a method flowchart of embodiment 1 of the explainability analysis method of the SAR image classification network of the present application;

[0051] Figure 2 is a method flowchart of embodiment 2 of the explainability analysis method of the SAR image classification network of the present application;

[0052] Figure 3 is a comparison chart of various index results under different network models;

[0053] Figure 4 is a comparison of the precision of each class of four network models on MSTAR;

[0054] Figure 5 is the experimental result of VGG-19, GoogleNet and ResNet-18;

[0055] Figure 6 is the network structure of ResNet-18;

[0056] Figure 7 is the saliency map of ten SAR target categories corresponding to different explainability methods;

[0057] Figure 8 is an IOU comparison chart of five explanation methods under ten categories;

[0058] Figure 9 is a recovery image of SG-SHAP under different insertion percentages θ;

[0059] Fig. 10(a) is a network confidence confusion chart of the original SAR image;

[0060] Fig. 10(b) is a confusion matrix diagram of the SHAP explainable method;

[0061] Fig. 10(c) is a confusion matrix diagram of the Grad-CAM explainable method;

[0062] Fig. 10(d) is a confusion matrix diagram of the Ablation-CAM explainable method;

[0063] Fig. 10(e) is a confusion matrix diagram of the Integrated-Grad explainable method;

[0064] Fig. 10(f) is a confusion matrix diagram of the Smooth-Grad explainable method;

[0065] Fig. 10(g) is a confusion matrix diagram of the IG-SHAP explainable method of the present application;

[0066] Fig. 10(h) is a confusion matrix diagram of the SG-SHAP explainable method of the present application. DETAILED DESCRIPTION

[0067] The core idea of the present application is to overcome the limitations of the traditional SHAP method by combining the gradient-based attribution algorithm with the SHAP method. The specific process is as follows: first, the SHAP algorithm is used to calculate the Shapley value of the features of the SAR image input into the SAR image classification network, thereby obtaining the original SHAP saliency map; then, the gradient-based attribution algorithm is used to obtain the gradient saliency map for evaluating the importance of the features; finally, the original SHAP saliency map and the gradient saliency map are fused to obtain the fused SHAP saliency map, which is used for explainable analysis of the SAR image classification network. This is a gradient-based SHAP explainable method, which combines SHAP and gradient information as the core, not only overcoming the limitations of the traditional SHAP method, but also enhancing the understanding of the model decision-making process. In order to make the purpose, technical scheme and advantages of the present application more clear and explicit, the present application will be further described in detail below with reference to the accompanying drawings.

[0068] Embodiment 1 of a SAR image classification network explainable analysis method

[0069] The present SAR image classification network explainable analysis method embodiment is an integral gradient-based SAR image classification network explainable method embodiment, which will be referred to as IG-SHAP hereinafter.

[0070] The SAR image classification network is used for target classification in SAR images. Various networks can be used, including existing deep learning networks or improved versions of existing deep learning networks. Generally, the SAR image classification network is trained using a SAR image dataset with target classification result labels. After training, the SAR image classification network can be used to obtain the target classification result in the SAR image.

[0071] Integrated-Grad is a method that quantifies the contribution of different features to the prediction by calculating the integral of the gradient of the input image with respect to the prediction score of the target class. It interpolates the path between the input image and the baseline image, calculates the gradient at each interpolation point, and then integrates these gradients to obtain a measure of the importance of the features.

[0072] Compared with other gradient methods, Integrated-Grad has the following advantages: ① Completeness. Integrated-Grad is a complete gradient method that considers the contribution of each pixel in the input image to the model prediction. By calculating the integral of the gradient of the input image with respect to the prediction score of the target class, it can quantify the overall contribution of each feature to the prediction. This makes it provide more comprehensive and detailed feature importance measures, and better capture the subtle changes in model predictions and interactions between features. ② Consistency. Integrated-Grad uses path integration in the gradient calculation process, which means it considers the continuous change process of the input image from the baseline image (black and white image) to the target image. This consideration of continuous change makes the method's results more stable and consistent. Compared with the naive gradient method, Integrated-Grad can better handle noise and uncertainty in the input image, thus providing more reliable interpretive results. ③ Interpretability measure. The interpretive saliency map generated by Integrated-Grad can be used as a measure to measure the importance of each feature in the model prediction. By accurately calculating and ranking the importance of each feature, researchers can gain insight into how the model makes decisions based on different feature information. Such a measure can provide quantitative support for the interpretability of the model, helping to evaluate the reliability and stability of the model. ④ Model independence. Integrated-Grad is a model-independent explanation method that can be applied to various types of models, including deep learning models and traditional machine learning models. This makes it have wide applicability and can be used to explain various models in different fields.

[0073] Therefore, the present embodiment integrates the idea of integral gradient into the SHAP algorithm. First, the SHAP method is used to quantify the importance of each feature by iteratively calculating the marginal contribution of each feature, thereby obtaining the contribution value of each feature to the model output. Second, the Integrated-Grad is used to integrate in the feature space, starting from a reference point (such as zero feature value or average feature value), gradually changing the input of the model to the current input to calculate the contribution of each feature to the gradient, and obtaining the comprehensive gradient change of each feature to the model prediction result by integrating these contributions. Then, the gradient information is combined with the SHAP contribution value to obtain the final SHAP graph. Finally, the ReLU function is applied to the final SHAP graph, setting all negative values to zero. Through this process, a more interpretable SHAP graph can be obtained, which will better consider the importance relationship between features and only focus on positive class information. This method has a positive contribution to the decision of target area, interference spot, shadow area, etc. in the SAR recognition network, and helps to understand the decision basis of the model in the positive class prediction. As shown in Figure 1 , the specific process is as follows:

[0074] Step 1: Use the SHAP method to calculate the Shapley value of each feature (N) in the input SAR image, thereby obtaining the original SHAP significant graph.

[0075] The Shapley value of the feature can be represented as:

[0076]

[0077] In the formula, S represents a subset of the member set N , S is a subset of N and does not contain feature i; v represents the value function, which is used to measure the value obtained by S; φ v (i) is the Shapley value, which represents the expected marginal benefit of feature i. By calculating the Shapley value of all features, the value distribution of different features can be obtained.

[0078] Step 2: Calculate the integral gradient of the features of the input SAR image to obtain the gradient significant graph (specifically, the integral gradient significant graph) for evaluating the importance of the features.

[0079] Formally, let F: R n →[0,1] be a deep network; x∈R n is the input image, is the baseline input (set to a black image in the present embodiment). Consider the straight line path between the baseline and the current input, and calculate the gradient of all points on the path, input x and baseline along i thThe integral gradient is defined as follows:

[0080]

[0081] For convenience of calculation, the integral gradient can be effectively approximated by the cumulative method, which simply sums the gradients of the points at small intervals on the route path between the baseline and the input, expressed as:

[0082]

[0083] In the formula, represents the number of steps of the Riemann approximation of the integral.

[0084] Step 3: Fuse the gradient saliency map obtained in Step 2 with the original SHAP saliency map obtained in Step 1 to obtain a fused SHAP saliency map, and use the fused SHAP saliency map to perform explainability analysis of the SAR image classification network.

[0085] In this step, the fusion operation is a pixel-by-pixel multiplication operation, that is, the gradient saliency map and the original SHAP saliency map belonging to the same prediction category are multiplied pixel by pixel to obtain the operation result of the same prediction category, and then the operation results of different categories are added to obtain the fused SHAP saliency map. Moreover, before the addition operation, the operation results of the same prediction category need to be properly normalized to make the visualization result of the image more intuitive and easy to explain. The whole process is expressed as:

[0086]

[0087] In the formula, S ig represents the fused SHAP saliency map, S C is the original SHAP saliency map of the image prediction category c; is the integral gradient saliency map of category c; represents Hadamard product; S ig represents the fused SHAP saliency map after fusion with the integral gradient saliency map; ReLU represents the ReLU activation function, which sets all negative values to 0, so that only the information of the positive class is focused on. It should be noted that the category here refers to the prediction category of the SAR image classification network.

[0088] Embodiment 2 of the SAR image classification network explainability analysis method:

[0089] The SAR image classification network explainability analysis method is based on the smooth gradient SAR image classification network explainability analysis method, which will be referred to as SG-SHAP in the following.

[0090] Of course, the SAR image classification network targeted in this embodiment is consistent with the SAR image classification network introduced in Embodiment 1 of the “explanation method for a SAR image classification network”, and this embodiment will not repeat the SAR image classification network.

[0091] Although the IG-SHAP explanation method is an effective method for explaining model predictions, it also has some drawbacks. In the selection of reference points, the Integrated-Grad method may have an impact on the selection of reference points, and different reference points may lead to different feature importance evaluation results; when the integral path approaches the reference point, the problem of gradient disappearance or explosion may occur. This situation may lead to inaccurate estimation of feature importance.

[0092] Research has found that the Smooth-Grad method has some advantages: ①Smooth explanation map: the Smooth-Grad method smooths the explanation map by adding random noise to the input image and repeatedly running the model. This helps to reduce noise and uncertainty, producing a smoother and more stable explanation map. ②Easy to implement and apply: the implementation of SmoothGrad is relatively simple and can be added as a supplementary step to the existing gradient calculation process without fundamentally changing the model or training process. This makes it easy to integrate into various deep learning frameworks and tools.

[0093] Therefore, this embodiment adopts the Smooth-Grad method combined with the original SHAP method to reduce noise and uncertainty, provide a smoother and more stable explanation map, and combine the final significant map with the original image to obtain a comprehensive explanation result that combines feature importance and the original image. This method can provide more intuitive and comprehensive model explanation, helping to understand the model's prediction results and decision-making process. As shown in Figure 2 the specific process is as follows:

[0094] Step 1: Use the SHAP method to obtain the feature importance SHAP map, i.e. the original SHAP significant map.

[0095] The calculation steps for obtaining the SHAP significant map are the same as the Shapley value calculation method mentioned in Step 1 of Embodiment 1 of the “SHAP explanation method for a SAR image classification network based on gradient”, which will not be repeated here.

[0096] Step 2: Use the Smooth-Grad method to calculate the smoothed gradient information to enhance the accuracy and stability of feature importance evaluation, to obtain a gradient significant map (specifically a smooth gradient significant map) for evaluating feature importance.

[0097] For the input image x, add a random noise N noisy images x+ε are obtained by adding Gaussian random noise to the original images x, and the model is run multiple times, each time using the noisy input features to obtain multiple prediction results. The prediction results of multiple runs are averaged to reduce the impact of noise and obtain a more stable feature map importance saliency map:

[0098]

[0099]

[0100] where S c (x) represents the score of the predicted class c for x, where the predicted class is the predicted class of the SAR image classification network.

[0101] Step 3: Fuse the gradient saliency map obtained in Step 2 with the original SHAP saliency map obtained in Step 1 to obtain a fused SHAP saliency map.

[0102] where the fusion operation in this step is a pixel-by-pixel multiplication operation, i.e., the gradient saliency map and the original SHAP saliency map belonging to the same predicted class are multiplied pixel by pixel to obtain the operation result of the same predicted class, and then the operation results of different classes are added to obtain the fused SHAP saliency map. Moreover, before the addition operation, the operation result of the same predicted class needs to be properly normalized to make the visualization result of the image more intuitive and easy to interpret. The entire process is expressed as:

[0103]

[0104] where S sig represents the fused SHAP saliency map, S C is the original SHAP saliency map of the image prediction class c, is the smoothed gradient saliency map of class c, represents the Hadamard product, S sig represents the fused SHAP saliency map after fusion with the smoothed gradient saliency map; ReLU represents the ReLU activation function, which sets all negative values to 0, so that only the information of the positive class is focused on.

[0105] Step 4: Fuse the SAR image (SAR image input to the SAR image classification network) with the fused SHAP saliency map to obtain a comprehensive explanation map that combines feature importance and input SAR image, and use the comprehensive explanation map to perform explainability analysis of the SAR image classification network.

[0106] In this step, there are two ways to fuse the SAR image with the fused SHAP saliency map: the first way is to directly fuse the SAR image with the fused SHAP saliency map; the second way is to first perform regularization processing on the SAR image, and then fuse the regularized image with the fused SHAP saliency map generated in Step 3. Both of these two ways can make the saliency map only focus on the places with target objects and ignore other irrelevant places, and obtain a comprehensive explanation result that combines feature importance and the original image.

[0107] Moreover, the fusion method in the above two ways can be pixel-by-pixel multiplication. The specific process of the above second way is to perform pixel-by-pixel multiplication operation on the regularized SAR image belonging to the same channel and the fused SHAP saliency map to obtain the operation result of the same channel, and then add the operation results of different channels. The process is expressed as:

[0108]

[0109] In the formula, S self represents the comprehensive explanation map, Normalize( ) represents the regularization processing, is the a-th channel of the input image.

[0110] The following experiments are designed to explore the use of IG-SHAP and SG-SHAP methods to visualize and analyze the image classification results of pre-trained classification models when processing image inputs, highlight the image regions that the model focuses on, and enhance the understanding of the decision-making process. Through comparative analysis, combined with qualitative and quantitative evaluation, the performance of these two methods in terms of explanation ability is comprehensively evaluated.

[0111] 1) Experimental parameter setting.

[0112] The experiment uses Anaconda 4.13.0 integrated development environment, and is trained under the Pytorch deep learning framework, with GPU as the training platform and MSTAR validation set as the data set. The VGG-19, AlexNet, GoogleNet and ResNet-18 network models used on the MSTAR data set are trained and fitted, and the precision (Precision), recall (Recall), F1-Score value and Acc (accuracy) under different models are calculated. The calculation formulas of the four indicators are as follows:

[0113]

[0114] In the formula, TP (True Positive) represents the number of true samples predicted as true samples, FP (False Positive) represents the number of false samples predicted as true samples, FN (False Negative) represents the number of true samples predicted as false samples, and TN (True Negative) represents the number of false samples predicted as false samples. The results of each indicator are compared as follows: Figure 3 As shown, the accuracy for each class is as follows: Figure 4 As shown.

[0115] Depend on Figure 3 Experimental results show that the ResNet-18 model achieves over 98% accuracy across all metrics. Figure 4 The accuracy of each class is generally better than other models, with the best performance on the MSTAR dataset. Since the results for VGG-19, GoogleNet, and ResNet-18 are quite similar, to fully compare the differences in model performance, Figure 5 Saliency maps of the SG-SHAP method under three network models: VGG-19, GoogleNet, and ResNet-18. The first column shows randomly selected SAR images of three types: T62, ZIL131, and SNC71. The second column shows the saliency map for VGG-19, and the third and fourth columns show the saliency maps for GoogleNet and ResNet-18, respectively. Comparison shows that the VGG-19 and GoogleNet models are more sensitive to noise, eliminating noise interference while also excluding some features related to the target. ResNet-18 shows the best visualization results; therefore, the CNN network used in subsequent experiments is ResNet-18. ResNet-18 is a deep convolutional neural network architecture whose core components are multiple residual blocks. Each residual block is carefully designed to contain two consecutive convolutional layers and a crucial skip connection mechanism, allowing information to be transmitted faster within the network. ResNet-18 has 18 layers, including 16 convolutional layers and 2 fully connected layers. A schematic diagram of the network structure is shown below. Figure 6 As shown.

[0116] 2) Qualitative analysis.

[0117] The following section randomly selected ten SAR images of different categories from the test set of the MSTAR dataset and calculated their saliency maps for the corresponding methods. The results are as follows: Figure 7The SHAP method can highlight the target region of the input image, but the result is usually rough and the details of the target features are not fine enough. The Integrated-Grad method is extremely sensitive to input perturbations, resulting in a large number of irrelevant regions in the positioning results, thereby affecting the stability of the result map. The Smooth-Grad method can find the relevant regions of the network by propagation and remove the noise effect, but the explanation result often only contains the shadow area, which is difficult to accurately identify the SAR target information. In contrast, the IG-SHAP method combines the saliency map generated by the SHAP value and the gradient information, not only focusing on the features that play a decisive role in the target object, but also considering the secondary features extracted by the image classification network for multiple target features, providing more comprehensive feature coverage and more accurate positioning. The SG-SHAP method can better depict the complex edges of the target, and its positioning ability and visual coherence are the best, significantly improving the matching degree between the visual interpretation result and the target. The result obtained by SG-SHAP is very similar to the original image shape, considering the correlation between pixels, and can accurately position according to the outline of the target.

[0118] These observations show that the IG-SHAP and SG-SHAP methods have obvious advantages in improving the visualization quality, accuracy and robustness of model explanation, especially the SG-SHAP in optimizing the details of the target object, which shows a high degree of fineness, and its positioning effect is more accurate than IG-SHAP, without too many redundant areas.

[0119] 3) Quantitative analysis.

[0120] Since the data images adopted are SAR images, compared with optical images and human intuitive perception, the complex nature of SAR images makes it not easy to understand the network decision. Qualitative analysis provides intuitive perception, which is an effective way of visual evaluation method, but the evaluation result lacks certain logical basis. The quantitative evaluation method follows a pre-defined logical framework, first calculates the effect score of each sample in the data set after visual processing, and then summarizes and calculates the average value of all sample scores, thereby quantitatively comparing and judging the advantages and disadvantages of different visualization techniques. The following uses three classical quantitative evaluation techniques.

[0121] ① Weakly supervised localization.

[0122] Weakly Supervised Localization (WSL) is a method that uses image-level labels (such as image categories) to locate target objects in images. The specific process is as follows: for the generated saliency map, binarize it according to the pre-set threshold (threshold value 0.15 in the experimental stage), then draw the bounding box for the area greater than the threshold, and compare the drawn bounding box with the manually annotated true box to calculate the Intersection of Union (IoU) to measure the similarity between the original SAR image and its corresponding saliency map. If the IoU is greater than 0.5, it is considered that the weakly supervised localization of the explanation result is successful. The formula for calculating IoU is:

[0123] IoU = Area_overlap / Area_union

[0124] In the formula, Area_overlap represents the overlapping area of the highlighted area in the saliency map and the corresponding target area in the SAR image, and Area_union represents the union of the two parts. As can be seen from the formula, the higher the IoU value, the more accurate the weakly supervised localization. In Table 1, the IoU results of each category image in Figure 5 are calculated. Each row of data in the table represents the IoU of a category under five methods. The optimal case of each algorithm under each category is highlighted.

[0125] Figure 8 The figure shows the IoU (Intersection over Union) comparison of the five explanation methods under ten categories. This figure presents the data in Table 1 in the form of a bar chart, making the result comparison more intuitive. It is obvious that only SG-SHAP has an IoU value higher than 0.5 under most categories, successfully locating the target. The positioning effect of IG-SHAP is similar to that of SHAP, but it is still competitive compared to Integrated-Grad and Smooth-Grad methods. This experiment verifies the effectiveness of the SG-SHAP method, which is more accurate in positioning and has a higher matching degree between the saliency map and the target compared to other algorithms.

[0126] Table 1

[0127]

[0128] ②Insert experiment.

[0129] To better evaluate the performance of various explainable methods, the following will be inserted into the significant map generated in the qualitative analysis. Insertion inspection refers to starting from a full zero image and gradually restoring the content according to the corresponding significant map. Specifically, replace 1% of the pixels of the full zero image, respectively, with 1%, 5%, 10%, 20%, 40%, 60%, 80% of θ insertion until the image is restored. For example Figure 9 The restored image of SG-SHAP for different insertion percentages θ. The first column is the original SAR image, and the second to sixth columns are the inserted images when θ is 1%, 5%, 10%, 20%, 40%, 60%, 80%.

[0130] From Figure 9 It can be seen that as long as a very small θ (θ≤20%), the shape of the target can be restored. This further proves that the pixels with the highest value in the significant map can be accurately concentrated on the target area.

[0131] Without loss of generality, all 2636 verification images are sent to the network model, and the area under the curve (AUC) of the classification score after Softmax is calculated for different θ from θ=5% to θ=80%. The average AUC results under different insertion rates are shown in Table 2 (the best AUC value is shown in bold). AUC is the area under the receiver operating characteristic (ROC) curve. For a binary classification problem, ROC refers to the curve of each point, plotted with the false positive (FR) rate as the abscissa and the true positive (TR) rate as the ordinate. AUC can reflect the performance of the model, that is, AUC=1, the performance of the model is the best; AUC=0.5, the model is a random classifier; AUC<0.5, the model is usually worse than the random classifier. This concept can be extended to multi-classification problems, considering that the true label is true and other labels are false classification.

[0132] Table 2

[0133]

[0134] As can be seen from Table 2, AUC generally increases with the increase of θ, but when θ increases to a certain extent, AUC will suddenly decrease. This may be because when θ is very small, the re-introduced pixels are concentrated in the target area, which represents the most distinctive features of the target; when θ becomes larger, some sharp or "strange" edges will be introduced, resulting in a lower AUC. Therefore, the early arrival of the maximum AUC means that the brightest pixels in the significant map are concentrated on the target.

[0135] When the insertion ratio is 5%, the AUC value of SG-SHAP is the highest, reaching 0.158, while the AUC values of other methods are all below 0.04. This data shows that SG-SHAP can significantly distinguish target features at very low insertion ratios, indicating its strong sensitivity and positioning ability for key features. When the insertion ratio increases to 10%, the AUC value of SG-SHAP increases significantly to 0.429, which is much higher than other methods, further proving its superior performance. In the insertion ratio of 15% to 20%, the AUC values of IG-SHAP and SG-SHAP increase rapidly, and reach 0.541 and 0.703 respectively at 20%, which shows that these two methods can effectively identify and utilize key features to improve model performance when the key features are inserted. When the insertion ratio continues to increase to more than 30%, the AUC values of most methods begin to decrease, but the performance of SG-SHAP is still good. It is speculated that higher insertion ratios introduce more background information or noise, which negatively affects the model performance.

[0136] Overall, SG-SHAP performs better than other methods at all insertion ratios, especially at lower insertion ratios, with AUC values much higher than other methods, showing its excellent model explanation ability and efficient feature utilization efficiency.

[0137] ③ Confidence evaluation.

[0138] The confidence evaluation index is widely used to evaluate the influence of the occlusion of a certain decision part of the sample image pixel on the model classification result, and is an important index to verify whether the saliency map annotation is accurate. The confidence evaluation index includes the average drop (AD) and the average increase (AI) of the confidence. The calculation methods of the two indexes are as follows:

[0139] For N input images participating in the calculation, the output score of the image in the category is represented as , and the saliency map is generated by the saliency map generation method (such as CAM and RISE). In the case of , the original image and the saliency map are integrated to obtain a mask as the network input to produce the confidence level . The results after processing the original image are shown in Figure 9 . The calculation formula of the average confidence drop is:

[0140]

[0141] The calculation formula of the average confidence increase is:

[0142]

[0143] In the formula, Sign( ) represents the sign function.

[0144] To fully measure the effectiveness of the explanation method, the accuracy of the saliency map of each method under the confidence is also compared. Mathematically, it can be expressed as:

[0145]

[0146] In the formula, N represents the number of categories, M j represents the number of samples of a single category. According to the above formula, the average AD, AI and accuracy of each method under all categories are calculated as shown in Table 3.

[0147] Table 3

[0148]

[0149] It is worth noting that the smaller the AD, the closer the predicted score of the original image and the saliency map, that is, it can be considered that the class information captured by the highlighted area of the saliency map obtained by the explainable algorithm is more distinguishable. Similarly, the larger the AI, the more significant the difference between the semantic information contained in the original image and the saliency map, that is, the discarded area contains more class information.

[0150] As can be seen from Table 3, the average decline value of IG-SHAP and SG-SHAP methods is the lowest and the average rise value is the highest compared with other explainable methods, which fully proves its good performance. In addition, the accuracy value of IG-SHAP and SG-SHAP among the five methods is also significantly greater than the first three methods, among which the accuracy mean of SG-SHAP is the largest, which fully illustrates that the saliency map obtained by SG-SHAP contains the most information of the corresponding category and the positioning target is more accurate.

[0151] In order to further study the recognition accuracy for each category, the confusion matrix of seven explainable methods is given in Figure 10(b)~Figure 10(h) Fig. 10(a) is the network confidence confusion chart of the original SAR image, Figure 10(b) to Figure 10(h)The network confidence confusion chart is Grad-CAM, blation-CAM, SHAP, Integrated-Grad, Smooth-Grad, IG-SHAP and SG-SHAP in turn, and each row in the confusion matrix chart represents the average network confidence output of the corresponding class data, and the diagonal line represents the correctly identified samples. It can be observed from the original data set that the network confidence output is concentrated on the diagonal line. Most of the activation-based methods (Grad-CAM, Ablation-CAM) and gradient-based methods (Integrated-Grad and Smooth-Grad) all show poor performance. In this test, SHAP, IG-SHAP and SG-SHAP show greater diagonal element values, and it is worth noting that IG-SHAP and SG-SHAP show the best performance, and the significance map generated by them is closely consistent with the actual class. It is fully verified that the gradient-based SHAP explainable method is effective.

[0152] In summary, the SHAP method is improved based on the gradient strategy of the Grad-CAM method, and a gradient-based SHAP explainable method is proposed. In method implementation 1, the method quantifies the overall contribution of each feature to the prediction by calculating the integral gradient of the input image on the target class prediction score. Not only does it maintain the completeness and consistency of the SHAP method, but it also provides a more comprehensive feature importance view by measuring the contribution of each feature in detail. In method implementation 2, in order to reduce the noise in the gradient saliency map and enhance the reliability of the gradient information, a SHAP explainable method based on smoothed gradient is proposed. By introducing smoothed gradient information into the SHAP map, a high-resolution explainable saliency map that integrates with the original image can be generated, further improving the intuitiveness and comprehensiveness of model explanation. By combining SHAP with gradient information as the core, not only does it overcome the limitations of traditional SHAP methods, but it also enhances the understanding of the model decision-making process. The core of the two method implementations is to use the gradient-based attribution algorithm to obtain the gradient saliency map for evaluating feature importance, and to fuse the original SHAP saliency map and the gradient saliency map to obtain the fused SHAP saliency map. The fused SHAP saliency map is used for explainable analysis of the SAR image classification network. These improved methods provide a new perspective for in-depth analysis and optimization of image classification models, and demonstrate the important value in improving model transparency and decision explainability.

[0153] An embodiment 1 of a SAR image classification network explainable analysis device:

[0154] The embodiment of the SAR image classification network explainability analysis device of the application comprises a memory, a processor, an internal bus and a computer program stored in the memory, the processor, the memory and the internal bus complete mutual communication and data interaction through the internal bus. The processor executes the computer program to realize the steps of the method introduced in embodiment 1 of the SAR image classification network explainability analysis method of the application. Wherein, the processor can be a microprocessor MCU, a programmable logic device FPGA and the like processing device; the memory can be various memories for storing information by using electric energy, such as RAM, ROM and the like, and can also be a memory using other ways.

[0155] The specific method flow is referred to Figure 1 , and the details are as follows:

[0156] Step 1: using SHAP method to calculate the Shapley value of each feature (N in total) in the input SAR image, so as to obtain the original SHAP saliency map.

[0157] Step 2: calculating the integral gradient of the features of the input SAR image to obtain the gradient saliency map (specifically, the integral gradient saliency map) for evaluating the importance of the features.

[0158] Step 3: performing fusion operation on the gradient saliency map obtained in Step 2 and the original SHAP saliency map obtained in Step 1 to obtain a fused SHAP saliency map, and using the fused SHAP saliency map to perform the explainability analysis of the SAR image classification network. Wherein, the fusion operation in this step is a pixel-by-pixel multiplication operation, that is, performing pixel-by-pixel multiplication operation on the gradient saliency map and the original SHAP saliency map belonging to the same prediction category to obtain the operation result of the same prediction category, and then adding the operation results of different categories to obtain the fused SHAP saliency map. Moreover, before the adding operation, the operation result of the same prediction category needs to be appropriately normalized to make the visualization result of the image more intuitive and easy to explain.

[0159] Embodiment 2 of the SAR image classification network explainability analysis device

[0160] The embodiment of the SAR image classification network explainability analysis device of the application comprises a memory, a processor, an internal bus and a computer program stored in the memory, the processor, the memory and the internal bus complete mutual communication and data interaction through the internal bus. The processor executes the computer program to realize the steps of the method introduced in embodiment 1 of the SAR image classification network explainability analysis method of the application. Wherein, the specific processor and memory can refer to the content introduced in embodiment 1 of the SAR image classification network explainability analysis device.

[0161] A specific method flow is referred to Figure 2 , as follows:

[0162] Step 1: Use the SHAP method to obtain the feature importance SHAP map, that is, the original SHAP saliency map.

[0163] Step 2: Use the Smooth-Grad method to calculate the smoothed gradient information to enhance the accuracy and stability of the feature importance evaluation, so as to obtain the gradient saliency map (specifically, the smoothed gradient saliency map) for evaluating the feature importance.

[0164] Step 3: Fuse the gradient saliency map obtained in Step 2 and the original SHAP saliency map obtained in Step 1 to obtain a fused SHAP saliency map. The fusion operation in this step is a pixel-by-pixel multiplication operation, that is, the gradient saliency map and the original SHAP saliency map belonging to the same prediction category are multiplied pixel by pixel to obtain the operation result of the same prediction category, and then the operation results of different categories are added to obtain the fused SHAP saliency map. Moreover, before the addition operation, the operation results of the same prediction category need to be appropriately normalized to make the visualization result of the image more intuitive and easy to explain.

[0165] Step 4: Fuse the SAR image (SAR image input into the SAR image classification network) and the fused SHAP saliency map to obtain a comprehensive explanation map that fuses the feature importance and the input SAR image, and use the comprehensive explanation map to perform explainability analysis on the SAR image classification network.

[0166] In this step, there are two ways to fuse the SAR image and the fused SHAP saliency map: the first way is to directly fuse the SAR image and the fused SHAP saliency map; the second way is to first regularize the SAR image, and then fuse the regularized image and the fused SHAP saliency map generated in Step 3. Both of these two ways can make the saliency map focus only on the places with target objects and ignore other irrelevant places, and obtain a comprehensive explanation result that fuses the feature importance and the original image. Moreover, the fusion method in the above two ways can adopt the pixel-by-pixel multiplication method.

[0167] Embodiment 1 of a computer-readable storage medium:

[0168] An embodiment of the computer-readable storage medium of the present application, on which a computer program is stored, the program being executed by a processor to implement the SAR image classification network explainability analysis method proposed in embodiment 1 of the SAR image classification network explainability analysis method described above. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, or flash memory.

[0169] The specific method flow is referred to Figure 1 , as follows:

[0170] Step 1: Calculate the Shapley value of each feature (N in total) in the input SAR image using the SHAP method, thereby obtaining the original SHAP saliency map.

[0171] Step 2: Calculate the integral gradient of the features of the input SAR image to obtain the gradient saliency map (specifically, the integral gradient saliency map) for evaluating the importance of the features.

[0172] Step 3: Fuse the gradient saliency map obtained in Step 2 with the original SHAP saliency map obtained in Step 1 to obtain a fused SHAP saliency map, and use the fused SHAP saliency map to perform the explainability analysis of the SAR image classification network. The fusion operation in this step is a pixel-by-pixel multiplication operation, i.e., the gradient saliency map and the original SHAP saliency map belonging to the same predicted category are multiplied pixel by pixel to obtain the operation result of the same predicted category, and then the operation results of different categories are added to obtain the fused SHAP saliency map. Moreover, before the addition operation, the operation results of the same predicted category need to be appropriately normalized to make the visualization result of the image more intuitive and easy to explain.

[0173] An embodiment of the computer-readable storage medium of the present application, on which a computer program is stored, the program being executed by a processor to implement the SAR image classification network explainability analysis method proposed in embodiment 1 of the SAR image classification network explainability analysis method described above. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, or flash memory.

[0174] An embodiment of the computer-readable storage medium of the application, which stores a computer program, the program is executed by a processor to realize the SAR image classification network explainability analysis method proposed in the embodiment 2 of the SAR image classification network explainability analysis method as described above. The storage medium can be realized by any type of volatile or non-volatile storage device or their combination, and specific examples can refer to the content introduced in the embodiment 1 of the computer-readable storage medium.

[0175] The specific method flow refers to Figure 2 , details as follows:

[0176] Step1: Use SHAP method to obtain feature importance SHAP map, that is, original SHAP saliency map.

[0177] Step2: Use Smooth-Grad method to calculate the smoothed gradient information to enhance the accuracy and stability of feature importance evaluation, to obtain the gradient saliency map (specifically, the smoothed gradient saliency map) for evaluating feature importance.

[0178] Step3: Fuse the gradient saliency map obtained in Step2 with the original SHAP saliency map obtained in Step1 to obtain a fused SHAP saliency map. The fusion operation in this step is a pixel-by-pixel multiplication operation, that is, the gradient saliency map and the original SHAP saliency map belonging to the same prediction category are multiplied pixel by pixel to obtain the operation result of the same prediction category, and then the operation results of different categories are added to obtain the fused SHAP saliency map. Moreover, before the addition operation, the operation results of the same prediction category need to be appropriately normalized to make the visualization result of the image more intuitive and easy to explain.

[0179] Step4: Fuse the SAR image (SAR image input into the SAR image classification network) with the fused SHAP saliency map to obtain a comprehensive explanation map that integrates feature importance and input SAR image, and use the comprehensive explanation map to analyze the explainability of the SAR image classification network.

[0180] In this step, there are two ways to fuse the SAR image with the fused SHAP saliency map: the first way is to directly fuse the SAR image with the fused SHAP saliency map; the second way is to first regularize the SAR image, and then fuse the regularized image with the fused SHAP saliency map generated in Step3. Both of these two ways can make the saliency map only focus on the place with target objects and ignore other irrelevant places, and obtain a comprehensive explanation result that integrates feature importance and original image. Moreover, the fusion method in the above two ways can adopt the pixel-by-pixel multiplication method.

[0181] In summary, the present application proposes a gradient-based SHAP explainability method to solve the problem of misinterpretation of traditional SHAP and the explainability challenge in complex scenarios. From the perspective of increasing gradient weight, the SHAP method is improved. Firstly, an integral gradient-based SHAP explanation method is proposed. This method introduces the integrated gradient change information of each feature on the model prediction result into the original SHAP graph to obtain a saliency map that can more accurately capture the details inside the image. Since the effectiveness of SAR feature extraction depends not only on the reliability of the extraction algorithm, but also on the quality of the image itself. When the SAR image is seriously disturbed by noise, the reliability of feature extraction will be affected. Therefore, the present application further proposes a smooth gradient-based SHAP explainability method. This method considers using denoising, enhancement and other preprocessing methods to improve the recognition performance of the image before SAR image target recognition. Then, the smoothed gradient and the original SHAP graph are multiplied by Hadamard product, and the original image is retained. The fused SHAP graph and the original image are self-matched to obtain the final explainable saliency map. The experimental results confirm that the present method has significant superiority in visualization, axiomatic verification and quantitative evaluation compared with the traditional SHAP.

Claims

1. A method for explainability analysis of a SAR image classification network, characterized in that, The method comprises the following steps: a SHAP value of a feature of a SAR image input into a SAR image classification network is calculated by using a SHAP algorithm, so as to obtain an original SHAP saliency map; a gradient saliency map for evaluating the importance of the feature is obtained by using a gradient-based attribution algorithm on the feature of the SAR image; the original SHAP saliency map and the gradient saliency map are fused to obtain a fused SHAP saliency map, and the fused SHAP saliency map is used for explainability analysis of the SAR image classification network.

2. The method of claim 1, wherein the method is performed by a computer system. The fused SHAP saliency map is obtained by multiplying the original SHAP saliency map and the gradient saliency map belonging to the same predicted category pixel by pixel to obtain an operation result of the same predicted category, and then adding the operation results of different categories to obtain the fused SHAP saliency map; the predicted category is a predicted category of the SAR image classification network.

3. The method of claim 1, wherein, The gradient-based attribution algorithm is an integral gradient algorithm or a smooth gradient algorithm.

4. The method of claim 3, wherein, When the gradient-based attribution algorithm uses the smooth gradient algorithm, the explainability analysis of the SAR image classification network by using the fused SHAP saliency map is performed in the following manner: the SAR image is directly fused with the fused SHAP saliency map to obtain a comprehensive explanation map in which the feature importance and the input SAR image are fused, and the comprehensive explanation map is used for the explainability analysis of the SAR image classification network.

5. The method of claim 3, wherein, When the gradient-based attribution algorithm uses the smooth gradient algorithm, the explainability analysis of the SAR image classification network by using the fused SHAP saliency map is performed in the following manner: the SAR image is subjected to regularization processing, the SAR image subjected to the regularization processing is fused with the fused SHAP saliency map to obtain a comprehensive explanation map in which the feature importance and the input SAR image are fused, and the comprehensive explanation map is used for the explainability analysis of the SAR image classification network.

6. The method of claim 4, wherein, The SAR image is fused with the fused SHAP saliency map in the following manner: the SAR image and the fused SHAP saliency map belonging to the same channel are multiplied pixel by pixel to obtain an operation result of the same channel, and then operation results of different channels are added.

7. The method of claim 5, wherein the method further comprises: The SAR image subjected to the regularization processing is fused with the fused SHAP saliency map in the following manner: the SAR image subjected to the regularization processing and the fused SHAP saliency map belonging to the same channel are multiplied pixel by pixel to obtain an operation result of the same channel, and then operation results of different channels are added.

8. The method of claim 1, 4 or 5, wherein, The fused SHAP saliency map is obtained by multiplying the original SHAP saliency map and the gradient saliency map belonging to the same predicted category pixel by pixel to obtain an operation result of the same predicted category, and then adding the operation results of different categories to obtain the fused SHAP saliency map; the predicted category is a predicted category of the SAR image classification network.

9. The method of claim 8, wherein, The preprocessing is a ReLU activation function processing.

10. An apparatus for explainability analysis of a SAR image classification network, comprising a processor, characterized in that, The processor is configured to execute a computer program to implement the steps of the method in any one of claims 1-9.

11. A computer readable storage medium having stored thereon a computer program, characterized in that, The computer program, which is executed by a processor, implements the steps of the method according to any one of claims 1 to 9.

Images