In view of the complex maritime background, ship targets with large scale changes and noise interference, the ship detection accuracy of Synthetic Aperture Radar (SAR) is low and the missed detection is serious. An improved YOLOv7 model is proposed to solve these problems. Firstly, the Receptive Field Enhancement Feature Extraction Module (RFEFM) is designed to reconstruct the backbone network, enhance the receptive field and improve the multi-scale target feature extraction ability. Secondly, a High-Low Dimensional Feature Fusion Pyramid Network (HLF-FPN) is proposed to filter the noise and background information of interference and efficiently fuse the information of different scales. Then, a new F-MPDIoU loss function is proposed, which accelerates the convergence of the model and improves the problems of missed detection and false detection. Finally, the experiment on HRSID dataset shows that compared with the original YOLOv7 model, the proposed model improves the mAP@0.5, accuracy and recall by 4.9, 9.4 and 13.4 percentage points respectively, with the value of FPS reaches 68 frames per second, which meet the requirements of real-time detection.