GEO SAR Rotated Target Detection based on Edge Enhancement and Feature Alignment Convolution

With the advantages of large imaging swath, long observation time and short revisit period, geosynchronous orbit synthetic aperture radar (GEO SAR) plays a significant role in large-scale observation and maritime surveillance. To figure out the issues of low range-resolution, image distortion and target defocus in GEO SAR ship target detection, this paper proposes a GEO SAR rotated target detection method based on edge enhancement and feature alignment convolution. Firstly, the edge features are extracted and fused with context features by channel priority attention mechanism. To register the object with arbitrary pose with the horizontal convolutional kernels, the rotated anchors are predicted by feature alignment convolutional network, where the feature maps are adjusted according to the anchor poses. Finally, the high-precision bounding box and classification results are obtained by a two-stream detection head. In this paper, we produced a simulated GEO SAR ship target detection data set. The experiment shows that the proposed method improves the average precision for 1.8% on focus target and 6.1% on defocus target compared with the traditional methods.