JSSAA-Net: A Joint Sparse Sampling and Attention Augmented ISAR Image Classification Network

ISAR images, when observed under varying conditions, can undergo deformations due to both affine and feature transformations. These distortions can significantly impair the recognition accuracy of automatic target recognition systems in ISAR contexts. While contemporary deep-learning methods have targeted the impact of affine transformations, the influence of feature transformations is neglected. Additionally, current network structures fail to take full advantage of the sparsity characteristic of ISAR imagery. This article designs an innovative joint sparse sampling and attention augmented network. The network includes three modules: the affine transformation network (ATN) module, the sparse sampling network (SSN) module, and the attention augmented network (AAN) module. Specifically, the ATN module employs random transformations on the input images to bolster the network generalization capabilities. The SSN module facilitates swift feature extraction while maintaining the sparsity of ISAR images through a sparse downsampling technique. Concurrently, the AAN module accurately extracts the depth features using a joint convolution and self-attention mechanism. The integration of these modules into recognition networks is demonstrated, and their efficacy is validated using a variety of test datasets. Simulation results indicate that our proposed structure has better classification performance compared with traditional processing efficiency and recognition precision approaches.