Multi-Scale Capsule Network with Coordinate Attention for SAR Automatic Target Recognition

Deep learning-based approaches have made impressive achievement in synthetic aperture radar automatic target recognition (SAR-ATR), due to its powerful ability to extract hierarchical features. However, data-driven neural network requires a large amount of labeled data, which is hardly obtained especially in SAR target recognition. Recently capsule network shows its potential to perform well with few training samples. Thus, this paper proposals a novel neural network based capsule network to achieve comparable performance with limited data. It is composed of an encode network and a decode network. we deploy multi-scale feature extractor and coordinate attention to extract robust features by multiple dilated convolution layers. Then we adopt capsule module to encode and preserve spatial information. The decoder network consisting of three fully connected layers is used to reconstruct input images and encourage capsules to learn discriminative features. The MSTAR 10-class dataset is used to evaluate and demonstrating the effectiveness of the proposed framework. Our experimental result shows that it can achieve high overall accuracy up to 99.59% using all training samples and 89.92% even if only 20% training samples are applied without data augmentation.