Learning an anchor-free network with angle attention for remote sensing object detection

Rotation object detection has achieved significance progress in the community of remote sensing image processing. In contrast to horizontal bounding boxes, rotating bounding boxes require an accurate angle parameter to position the object. However, angle regression accuracy remains an urgent challenge due to the arbitrary orientation, large aspect ratio, and dense alignment of aerial images. To solve this problem, we propose an anchor-free network with angle attention (FAA-Net), which delicately develops two core designs, i.e, angle feature extraction (AFE) module and angle prediction head (APH) module. Specifically, the AFE module performs feature extraction and angle selection to reduce the negative effects of angle adaptation. In addition, APH module is further employed to better predict the object orientation by introducing the frequency attention and integral operation. Extensive experiments on two classical benchmark datasets have demonstrated that our method can effectively boost detection performance against several start-of-the-art approaches.