Saliency-Based Centernet for Ship Detection in SAR Images

Abstract

Ship detection in SAR images plays a significant role in civilian and military fields. The CNN-based detection algorithms have been widely used for ship detection in SAR images and the CenterNet has received extensive attention as an anchor-free detector. However, for the detection method based on CNN, the diversity of data can determine the generalization ability of the detection algorithm. Saliency detection can highlight target regions and suppress background clutter which can produce high-quality SAR images. In this paper, we integrate saliency detection with the CNN network and achieve an end-to-end detection network to complete the ship detection in SAR images. Our proposed saliency-based CenterNet contains two parts, saliency detection and CNN-based detection. The saliency detection can produce high quality slices in which the ship region is retained well, while most sea and land regions around the ship are suppressed. The following CNN-based network CenterNet can automatically extract the features of the ships without human intervention and can effectively distinguish ships and the strong scattering targets on land. Combining saliency detection with CNN-based detection methods can obtain a more accurate target detection model. The experimental results on SSDD demonstrate the effectiveness of the proposed method. Compared with the original CenterNet network, our method improves the mAP by 1.11% on all images of SSDD, and especially improves the mAP by 6.71% for inshore ships detection.