Target Recognition in Noisy SAR Images Based on Dual Network

Abstract

Synthetic aperture radar (SAR) images are inevitably contaminated by speckle noise, which may severely affect the accuracy of target recognition in SAR images. This paper proposes an automatic target recognition model, called despeckling and classification coupled dilated residual attention network (DCC-DRAN), for improving the target recognition accuracy of SAR images with strong speckle noise. With respect to despeckling sub-network, the noise information is learned by leveraging multi-scale dilated convolution and attention mechanism with the residual learning architecture. To adaptively extract more discriminative features, an effective attention mechanism named triplet attention module is embedded into convolutional layer in the classification sub-network, which can further promote target recognition performance. The experimental results on the Moving and Stationary Target Acquisition and Recognition (MSTAR) data set show that the recognition performance is improved by classification subnetwork with the help of despeckling sub-network, and ablation experiment also demonstrates that the attention module is effective for enhancing SAR target recognition performance.