2D-Temporal Convolution for Target Recognition of SAR Sequence Image

Abstract

Although deep learning has greatly improved the target recognition accuracy of synthetic aperture radar (SAR), the characteristics of SAR continuous imaging are not fully utilized in available methods. This paper proposes a SAR sequence image target recognition network based on two-dimensional (2D) temporal convolution. The proposed network includes three stages: feature extraction, sequence modeling and classification. Firstly, convolutional networks are utilized to extract features of each image and obtain a sequence of feature vectors. Secondly, the sequence is fed into the 2D temporal convolution network and sequence modeling is performed. Finally, recognition result of the SAR sequence image is inferred by the softmax classifier. Compared with available methods, the proposed network shows better recognition accuracy on the moving and stationary target acquisition and recognition (MSTAR) dataset.