Attention Mechanism and Sparse Low-Rank Modeling Based Neural Network for Data Augmentation of Through-the-Wall Radar Human Motion Recognition

Abstract

In order to better address the non-adaptability of image recognition algorithms on through-the-wall radar human motion data and the low signal-to-noise ratio (SNR) caused by microwave penetration through walls, an attention mechanism and sparse low-rank modeling based neural network is proposed in this paper. The method combines information from physical background of moving target and vision characteristics of imaging to achieve effective suppression of wall clutter and noise as well as enhancement of motion feature. The super resolution of human motion feature is achieved by sparsely encoded learning iterative shrinkage thresholding (LISTA) module. The location of human behind the wall is obtained by an improved coordinate attention mechanism, which automatically calibrates the regions associated with human motion characteristics. Chunked output super-resolution information after attention mechanism and LISTA module is finally weighted and aggregated by the parallel adaptive weight module. Experiments demonstrate that a better SNR is achieved for image feature extraction, while the accuracy and the convergence speed of the existing classification algorithms is effectively raised.