GST-YOLO: a lightweight visual detection algorithm for underwater garbage detection

Underwater cleaning work primarily relies on human labor, but applying computer vision technology to Autonomous Underwater Vehicles can enhance cleaning efficiency. Considering that existing vision detection algorithms are difficult to deploy on resource-constrained embedded devices, this paper introduces a lightweight vision detection algorithm based on an improved YOLOv8-GST-YOLO. This algorithm integrates the lightweight Ghost network and prunes the model, overcoming the drawbacks of YOLOv8’s high computational parameters and large size. It also features a GTR module and a bi-directional path aggregated feature pyramid guided by SimAM attention to enhance detection accuracy and global feature extraction capabilities. Experiments on a specially collected underwater trash image dataset show that GST-YOLO, while reducing the model size by 51% and increasing computational efficiency by 54%, improves the accuracy rate to 95.4%, surpassing the YOLOv8 algorithm. This demonstrates its potential as a crucial detection tool for underwater unmanned cleaning tasks, offering broad application prospects.