Ships Detection in SAR Images Based on Anchor-Free Model With Mask Guidance Features

Ship targets in synthetic aperture radar (SAR) images have various scales. The detection model based on anchor boxes requires manual design of candidate boxes, which are fixed and cannot completely match all kinds of targets. Instead, large of anchor boxes with different sizes also result in large amounts of computing resources being consumed. Another potential issue comes from complex background information of near-coast scenes, which leads to ship targets being unrecognized because the background contains similar appearing objects. Therefore, this article proposes an anchor-free detection model based on mask guidance features, which achieves detection mainly through three modifications. First, feature maps of multiple scales are fused to obtain high-resolution feature maps containing rich semantic information. Second, a transformer encoder module is introduced to focus on the context relationship between the target object and the global image and to enhance the dependence between ship targets. Third, the mask guide feature is used to highlight the positions of the target in the feature map, and a loss function in the mask guide mechanism is designed to optimize the mask feature map to reduce false detections and missed detections. Testing the model on the public dataset SAR ship detection dataset, the model's detection accuracy reached 96.17%, with its accuracy on small-size ships reaching 96.11% and 97.84% on large ships.