A scale transfer convolution network for small ship detection in SAR images

Abstract

Synthetic aperture radar (SAR) ship detection is a heated research topic. Traditional CFAR-based methods are not ideal when the background condition is complex. At present, the methods based on deep learning, such as Faster RCNN, have occupied a dominant position in the field of target detection. However, the Faster RCNN method is not suitable for SAR graphic ship target detection. In this paper, we design a scale transfer module for SAR ship detection. Instead of using a single feature map, the scale-transfer module connect to several feature map to get multiscale features. In addition, we use the method of RoIAlign to calibrate the accuracy of the bounding boxes. In detection subnetwork, we add the context features to assist the detection of complex targets. Experiments on SAR ship detection dataset (SSDD) show that the proposed method achieve superior performance in detection accuracy.