Target Recognition of HRRP Based on CNN with Multi-Axis Attention and Residual Connections

In view of the translation sensitivity of highresolution range profile (HRRP) caused by random sea conditions and the change of relative position between target and radar, a typical hierarchical network that is more suitable for HRRP target recognition is proposed. Firstly, the raw HRRP is represented by a designed feature representation block, which includes operations such as convolution, batch normalization (BN) and pooling. Secondly, a new basic block containing MultiAxis Attention is constructed to extract and weight the deep physical structure features of the data. In terms of memory access, the proposed network is 0.4 times of the advanced Maxvit network, and experimental results based on test data indicate that the method proposed in this article has higher recognition accuracy and stronger resistance to sample shift.