An Improved Algorithm for Ship Target Detection in SAR Images Based on Faster R-CNN

Ship detection is of great value for fishing activity control, military defense, maritime transport, etc. Satellite-based synthetic aperture radar (SAR) can provide high-resolution images, allowing surveillance over massive water bodies to be possible. However, traditional ship detection algorithms like CFAR (Constant False-Alarm Rate), cannot produce convincing results. In recent years, detectors based on convolutional neural networks have made great progress, and among them, Faster R-CNN is one of the best in performance. In this paper, we propose an improved algorithm based on Faster R-CNN for ship target detection and adopt four strategies to improve the performance. The four strategies are replacing the VGG16 backbone network with ResNet101, implementing online hard example mining, replacing the default non-max suppression algorithm with soft-nms, and changing the aspect ratio of the anchors. Experimental results show that our improved algorithm can boost the performance of the traditional Faster R-CNN detector (72.7% mAP) by about 5% in mAP (mean average precision) on the HRSC2016 (high Resolution Ship Collection) dataset, showing the effectiveness of the proposed approach.