| Summary: | Ship detection in synthetic aperture radar (SAR) images has gained great attention in civil and military fields. Anchor-based detection algorithms usually rely on preset candidate boxes, and a large amount of anchor boxes with different sizes will result in a large amount of computing resources being consumed. Recently, anchor-free algorithms have found wide applications in ship detection from SAR images. However, there are still some problems which limit the ship detection performance to a certain extent, such as how to effectively fuse salient features and unbalanced distribution of positive samples. In order to tackle the above problems, we propose a novel anchor-free model named salient feature fusion (SFF)–YOLOX with SFF mechanism. First, we redesign the network of YOLOX to obtain the best balance between detection accuracy and running speed. Second, a saliency region extraction module is introduced to generate the corresponding salient guide map of the input image. Besides, the SFF mechanism is proposed by fusing deep features and salient features to better enhance the discrimination of the multiscale targets. Finally, we improve the SimOTA mechanism by combining the predicted intersection over union (IoUs) and the anchor IoUs to the ground truth bounding boxes to instruct label assignment. We evaluate the detection accuracy and running speed of SFF–YOLOX on the public dataset single shot detector and test the generalization ability on HRSID and two complex large-scale SAR images, and the experimental results prove the model's effectiveness for ship detection task in SAR images.
|
|---|