SAR vehicle recognition via scale-coupled Incep_Dense Network (IDNet)

Abstract

ABSTRACT Vehicle target recognition of synthetic aperture radar (SAR) is one of the critical tasks of image interpretation. Convolutional neural networks (CNNs) are increasingly used in SAR images and make great progress in single-scale recognition under standard operating conditions (SOCs). However, the classical CNNs have the limitation of single model design and single scale, which is unfavourable for the recognition of multi-scale targets. This paper puts forward the design idea of scale coupling with both parallel convolution and dense connection, and designs an Incep_Dense module, which can learn and express scale information from different convolution paths and different feature levels simultaneously. Based on the Incep_Dense module, we built Incep_Dense network (IDNet) for SAR image vehicle target recognition. Furthermore, we add special layers to make IDNets more standardized and practical. Experiments show that the proposed IDNet achieves the optimal accuracy of 99% in the ten classifications of MSTAR and less than 1.5% accuracy degradation on the constructed multi-scale dataset. It has excellent multi-scale learning and representation capabilities with better performance than classical CNNs. It also performs well under the extended operating conditions (EOCs) of MSTAR and Gotcha with good generalization ability and robustness.