End-to-End Target Detection and Classification with Data Augmentation in SAR Images

While applying traditional algorithm to synthetic aperture radar automatic target recognition (SAR-ATR) is facing difficulties, deep learning-based end-to-end object detection algorithms are becoming better options due to the automatic feature extraction and availability of high-quality data. In this paper, both single-staged and two-staged end-to-end models are experimented. We proposed modified Faster R-CNN models and SSD models to address SAR-ATR. Data augmentation techniques including random flipping, multiplying, rotation, translation, and flipping are applied to MSTAR SAR dataset to solve problems related to limited training samples. Transfer learning of SSD models and Faster R-CNN models on COCO dataset are utilized. Both existing algorithms and proposed algorithms are tested in ten-class MSTAR dataset. Experimental results show that SSD-Inception with widened network and MobileNet-SSD with light weight structure perform with much faster speed and cheaper computational cost, hundreds of times faster than Faster R-CNNs. MobileNet-SSD is especially suitable for mobile devices with 0.028 second per batch*step. Faster R-CNN with ResNet-101 and Inception ResNet perform in slightly higher accuracy than SSDs, reaching 99.4% mAP. MobileNet-SSD and SSD-Inception reach 96.79% and 99.16% mAP respectively.