Ship Detection in SAR Image Based on Switchable Cavity Convolution

This paper presents an enhanced SAR image detection algorithm for YOLOv7 to address the issue of low detection accuracy of small-target ships and ships in complex backgrounds in synthetic aperture radar images. The proposed algorithm replaces the traditional convolutional layer in the backbone network of YOLOv7 with a convolutional SAConv, thus enhancing the network's capability to detect small targets and large objects. The introduction of the new metric, Wise IoU, into the loss function aims to improve the convergence speed and robustness while mitigating the adverse gradient effects caused by low-quality examples, thereby enhancing the detection of small targets. In the experimental comparison using the SSDD dataset, the improved method is employed to enhance the precision value of the model by increasing the number of model parameters and computational resources compared to the original YOLOv7. The mean Average Precision increases to 96.59%, showing an improvement of 9.33% over the original YOLOv7. Additionally, the accuracy improves by 3.81%, and the recall rate improves by 16.36%. These results demonstrate the superiority of the proposed method, resulting in enhanced precision compared to the original approach. The experimental results demonstrate that the enhanced algorithm significantly improves the detection accuracy of SAR images. Moreover, it effectively mitigates the issues of false positives and missed detections of ships in complex backgrounds.