In view of the high noise of synthetic aperture radar images and inconspicuous imaging features, especially in complex scenes such as sea and land boundaries, ports, and coastal reefs, it is difficult for common detection algorithms to extract target features from SAR images, resulting in low detection accuracy and leak detection, etc. This paper designs a rotating target detection method based on YOLOv5, and proposes that the multi-branch attention module can be used for cross-dimensional information fusion, which can better extract the location information and semantic information in SAR image targets. In addition, the boundary discontinuity will be caused by rotating target detection, which will affect the regression of the bounding box. Therefore, the circular smooth label method is used to transform the angle parameter from regression problem to classification problem, thus improving the accuracy. Finally, experiments are carried out on HRSID and SSDD+ datasets, and the accuracy reaches 84. 98% and 90. 13%, respectively, which is 1. 29% and 2. 57% higher than the original YOLOv5 algorithm, respectively. Experimental results prove the effectiveness of the proposed algorithm.