CNN Model Visualization Method for SAR Image Target Classification

Click the blue“Radar Journal”to follow

Video playback address on computer (Radar Journal Bilibili Learning Platform):

Short Video of the Paper | CNN Model Visualization Method for SAR Image Target Classification

Background Introduction

Deep neural network models represented by convolutional neural networks (CNN) are widely used in the task of target classification in synthetic aperture radar (SAR) images. However, deep network models have certain application risks and trust crises in practical SAR target detection and recognition tasks due to their lack of interpretability and decision opacity. On one hand, the imaging mechanism of SAR images is complex and easily influenced by target characteristics, leading to certain cognitive limitations of the human visual system. On the other hand, the “black box” nature of deep neural network models results in prediction uncertainty and certain vulnerabilities in specific scenarios. Currently, the performance evaluation and analysis of deep models mainly rely on recognition rates, while the actual application fields require that the model’s prediction results have high credibility. Therefore, relying solely on recognition rates may lead to issues such as insufficient problem description and model errors.

The research on the interpretability of deep network models can be roughly divided into self-explanatory and post-hoc explanations. The former constructs models that are structurally simple and highly interpretable or embeds the interpretability of knowledge from fields such as physics and semantics into specific model structures. The latter mainly focuses on analyzing the decision-making behavior of trained deep network models through visualization and other interpretable methods. The visualization techniques for CNN models can intuitively display the neuron features and decision features autonomously mined by the network from data, and the knowledge contained within can inspire people’s interpretation of SAR images, thereby assisting in further enhancing the model’s credibility and interpretability, guiding the evaluation, improvement, and updating of the model. Class activation mapping methods, as a typical post-hoc visualization method, visualize the decision areas of CNN models by generating heatmaps, offering advantages such as simplicity of use, strong generality, and class discriminability.

Team Work

In recent years, Professor Chen Bo’s team at Xi’an University of Electronic Science and Technology has conducted in-depth research on radar target detection and recognition, intelligent perception recognition, deep learning, and probabilistic modeling.

In terms of neuron feature extraction capability, a neuron visualization method based on maximum activation values has been proposed for trained CNN models, demonstrating the target recognition focus of specific key neurons in certain layers of the network. Regarding the basis for network decisions, a new class activation mapping method (CS-CAM) based on mixed channel and spatial class activation weights has been proposed, which confirms through qualitative and quantitative experiments and sufficiency evaluation results that CS-CAM can more accurately reflect the model’s decision-making behavior and basis, thereby locating important areas in input SAR images. Based on the above research foundation, this paper analyzes the CNN model from multiple perspectives, including neurons, similar and dissimilar targets, model recognition accuracy, and model initialization in an offline manner. Experimental results show that the proposed CNN model visualization method for SAR images can clearly visualize the key neuron feature recognition focus and core decision areas of the model, further enhancing the interpretability and robustness of the SAR image target recognition model.

Figure 1 Flowchart of CNN Model Visualization Method for SAR Images

This work is proposed to be published in the 2024 Issue 2 of the special issue “Microwave Vision and Intelligent Interpretation of SAR Images” with the paper titled“CNN Model Visualization Method for SAR Image Target Classification” (Li Miaoge, Chen Bo, Wang Dongsheng, Liu Hongwei)

Paper Introduction

This paper first proposes a CNN model neuron visualization method based on maximum activation values from the neuron level, and demonstrates the feature extraction effect of neurons on input SAR images by loading a trained network model in experiments.As shown in Figure 2, low-level neurons have a small receptive field and mainly focus on partial structural features of the target; with increasing network depth, high-level neurons gradually have larger receptive fields, and some neurons can observe the entire target completely. CNN Model Visualization Method for SAR Image Target Classification Figure 2 SAR Image Neuron Visualization ResultTo analyze the distribution shape of clutter in the distance-Doppler map in actual scenarios, this paper tests the proposed algorithm using measured data as prior information for clutter. The experimental results show that the generated cross-ambiguity function has depressions in the distance-Doppler regions where the corresponding clutter distribution is relatively concentrated, further proving the effectiveness of the proposed algorithm in more complex scenarios.Next, based on the existing class activation mapping methods, which have flaws, as shown in Figure 3, a class activation mapping method based on mixed channel-spatial class activation weights is proposed. The effectiveness of the proposed method is verified through qualitative analysis, quantitative analysis, and sufficiency evaluation.Finally, based on the proposed visualization method, a comparative analysis is conducted from the perspectives of model accuracy, initialization methods, and class discriminability.