YOLOv5 ship: An improved ship target detection algorithm study

Abstract

Aiming at the problem of difficult detection of maritime ship targets, which is prone to false detection, missed detection, and low recognition accuracy. In this paper, an improved ship target detection algorithm, YOLOv5, is proposed. firstly, the attention mechanism CBAM module is added to feature enhancement of image targets, and irrelevant interference information is suppressed based on retaining key features; secondly, the fusion structure of image feature extraction is improved, and the BiFPN structure is used instead of FPN+PAN structure, which realizes bi-directional multi-scale feature fusion and reduces target information loss and missed detection rate; finally, the loss function is optimized, and WIoU v3 with position and size focus factors is used to optimize the loss function, and the relative position and size of the prediction frame and the rear frame are learned more accurately, which effectively enhances the detection capability of small target ships. The experimental results show that the improved algorithm improves the target detection efficiency, the detection accuracy mAP50 value reaches 93.5%, and the FPS (Frames Per Second) goes to 51 frames per second, which can better meet the requirements of real-time detection and improve the target detection ability of ships in different environments with better robustness.