EMO-YOLO: a lightweight ship detection model for SAR images based on YOLOv5s

Enhancing maritime alert capabilities relies on effective Synthetic Aperture Radar (SAR) ship detection, often achieved through deep learning techniques. However, existing SAR ship detection models face challenges due to their large sizes, rendering them impractical for deployment on resource-constrained devices. Moreover, the complexity of ship backgrounds and the small size of ship targets contribute to decreased detection accuracy. In response, this paper proposes a lightweight ship detection model based on YOLOv5s. Our approach involves restructuring the original model, with a focus on an Efficient Model as the backbone. We achieve model lightweighting by employing a simple stacked Inverted Residual Mobile Block. Additionally, we introduce an enhancement feature extraction module, SCConv_C3, which utilizes Spatial and Channel Reconstruction Convolution (SCConv) to eliminate channel and spatial redundancies in the image while enhancing feature representation capabilities. Furthermore, we integrate Triplet Attention after feature fusion to enhance detection capabilities for small ship targets. Experimental evaluations were conducted on four public datasets. The results demonstrate that our proposed lightweight model maintains high detection accuracy even in challenging scenarios, including complex backgrounds and small ship targets. Notably, on the SSDD dataset, the AP \(_{50}\) ${}_{50}$ value reaches 97.8%, surpassing other advanced detection models such as YOLOv5s.