Learning salient features in radar micro-Doppler signatures using Attention Enhanced Alexnet

This work introduces an attention mechanism that can be integrated into any standard convolution neural network (CNN) to improve model sensitivity and prediction accuracy with minimal computational overhead. We introduce the attention mechanism in a lightweight network-Alexnet and evaluate its classification performance for human micro-Doppler signatures. We show that the Alexnet model trained with an attention module can implicitly learn to highlight the salient regions in the radar signatures whilst suppressing the irrelevant background regions and consistently improve the network predictions by more than 4% in most cases. We further provide network visualizations through class activation mapping, providing better insights into how the predictions are made.