Based on YOLOv5 lightweight submarine target detection algorithm

Abstract Submarine recognition is of paramount importance for maritime security and military defense. To address the limitations of traditional submarine recognition algorithms, which are restricted by limited feature representation capability and poor robustness, as well as the deployment bottleneck of deep learning methods on embedded and mobile platforms, an improved YOLOv5-based lightweight submarine automatic recognition detection algorithm is proposed. By employing the Feature Pyramid based on MobileNetV3 and the C3_DS module, the model's computation and parameter complexity are reduced while ensuring high precision in submarine recognition. The problem of missed detections is addressed through the combination with the adaptive neck of the SA-net strategy, thereby enhancing the accuracy of submarine target detection and recognition. Testing the improved model on a submarine dataset demonstrates significant improvements in Precision, Recall, and mAP0.5, with respective increases of 8.54%, 6.02%, and 3.36%. Moreover, a reduction of 34.1% in parameter quantity and 67.9% in computational complexity is achieved, showcasing notable lightweight effects.