Enhanced Ship Target Detection with YOLOv5s Utilizing Lightweight MobileNetv2 and Inverted Residual Block Structures

In response to the complex navigation environment, low model detection accuracy, and low efficiency of waterborne vessels, we propose an improved YOLOv5s object detection algorithm in this paper. Firstly, the original backbone network was replaced with a lightweight MobileNetv2, and the inverted residual block structure was improved to prevent feature loss during information transmission. Secondly, introducing deformable convolution into the original feature fusion module improves the spatial interaction ability between features at different levels. Finally, the introduction of GIOU solved the problem of mismatched directions between predicted and real boxes in object detection. Through validation on a ship dataset, the results showed that the mAP, mRecall, and AVGFPS of the improved YOLOv5s reached 0.963, 0.918, and 92, respectively, providing strong technical support for ship safety monitoring and management.