Research on Interpretability of SAR Image Component Recognition Based on Residual Attention Mechanism
Binqian Wu, Gong Zhang, Yuhua Sun · Proceedings of the 2020 4th International Conference on Electronic Information Technology and Computer Engineering · 2020
The interpretability of SAR target images has always been an important and challenging subject in the field of SAR image research. In recent years, deep learning technology has been successfully applied to the recognition of SAR target images, and significantly surpasses the performance of traditional methods. However, its internal working mechanism is opaque and lack of interpretability, which restricts the reliable and credible application of SAR target images recognition technology. Therefore, we make a research on the interpretability of SAR target images recognition from the perspective of component recognition. First, based on the VGG16 model structure, we construct a new convolutional neural network as the backbone model. Then, to obtain highly discriminating component features in the SAR target images and improve the interpretability of the network, the residual attention module is embedded between the convolution blocks of the backbone network. Finally, we use class activation map (CAM) to visualize the features, thus enabling a visually intuitive understanding of the model's decision logic and rationale for input samples. We have verified the performance of the proposed model on the MSTAR datasets. Experiments show that the proposed model can effectively identify the components of the SAR target images under both the Standard Operating Conditions (SOC) and the Extended Operating Conditions (EOC).