Feature extraction analysis method of pre-trained CNN model for SAR target recognition

Benefited from the latest advances in deep learning, convolutional neural network (CNN)-based SAR target recognition has made an excellent breakthrough. However, the most previous works pay attention to improve accuracy but neglect the working process analysis of the CNN models. Here, feature extraction of CNN plays an important role in recognition task. Moreover, in image processing field, frequency analysis is a classical way, which can directly reflect the frequency components of images. In view of that, we provide a feature extraction analysis method of pre-trained CNN in frequency domain in this paper. There are two analytical perspectives, i.e. amplitude and phase spectrum. Firstly, we can observe direction and range of pass-frequency through amplitude spectrum of convolution kernels, which represents information of raw SAR image and feature map captured from the convolution layer. Secondly, phase spectrums of SAR images store contour information of targets. We can understand the role of contour information from raw SAR images and feature maps in recognition task by phase spectrum analysis. In experiment part, according to the proposed feature extraction analysis method, we discuss the working process of several pre-trained CNN models. To some extent, it explains why the joint CNN has strong robustness against speckle noise in SAR image target recognition task. In short, the proposed feature extraction analysis method can improve the transparency and credibility of pre-trained CNN used for SAR target recognition.