YOLOv7-SIMAM: An effective method for SAR ship detection

Synthetic Aperture Radar (SAR) is an all-weather sensor extensively employed in military investigations, maritime rescue operations, and maritime transport management. However, the accuracy of ship detection in SAR images significantly influences the success of these tasks. To improve ship detection performance, we modify its structure and deploy the attention module on the YOLOv7 network. In deep learning, an attention-based model has better results than a convolution neural network. To solve the problem that embedding attention mechanism into convolutional neural network (CNN) has limited ability to extract spatial features and increased computational amount, we adopted SIMAM attention mechanism not only avoids introducing extra parameters but also concentrates on the feature space distribution across different channels, facilitating the efficient extraction of ship positions in high-resolution SAR images. Experimental results on the publicly measured SAR ship detection dataset (SSDD) demonstrate a 3.3% improvement in accuracy for the proposed SIMAM-YOLOv7 over the original YOLOv7. In summary, the proposed method offers a better alternative or complementary approach to ship detection in SAR images than the benchmark YOLOv7 model.