Improved Ship Target Detection in SAR Images Based on YOLOv7

In SAR ship detection, deep learning-based methods have emerged as the mainstream approach. However, challenges including complex marine-terrestrial backgrounds, multi-scale ship targets, and densely distributed small objects significantly increase the difficulty of SAR ship detection. To address these issues, this paper proposes an improved YOLOv7 algorithm to enhance detection accuracy for multi-ship targets in complex backgrounds. Firstly, to mitigate the impacts of complex backgrounds and noise interference, a Dual-Channel Pooling Attention module (DCPA) is developed. This module employs average pooling and strip pooling operations to capture long- and short-range semantic and positional information from feature maps respectively, thereby enhancing feature representation and alleviating overfitting caused by noise and complex backgrounds. Secondly, a Spatial Weight Balancing Module (SWBM) is designed to replace the original feature fusion components in the neck network. This module addresses feature inconsistency between upper and lower layers within the pyramid structure through balanced alignment of multi-scale feature maps. Finally, the proposed method achieves 96.54% mean Average Precision (mAP) on the SAR Ship Detection Dataset (SSDD), demonstrating an improvement of approximately 5 percentage points over the baseline model. Comparative experimental results validate that the enhanced algorithm outperforms other mainstream detection approaches in terms of detection effectiveness.