The Object Detection of Underwater Garbage with an Improved YOLOv5 Algorithm

Abstract

Litter deposition in aquatic environments has devastating effects on marine ecological environment and poses a threat to a sustainable economy. Autonomous Underwater Vehicles (AUV) could solve the issue nicely by detecting and clearing litter. A good object detection algorithm is very important in the process of AUV detection and garbage collection. In this research, YOLOv5 was applied as the detection algorithm of the detector and the prediction side of the algorithm was improved. The anchor boxes of the model are re-clustered by using the improved KMeans++ algorithm, the loss function was optimized and the box loss function of the original model was replaced by CIoU. When detecting the trash_ICRA19 dataset, the results demonstrated that the improved model achieved a detection accuracy of 88.7%, a mean average precision (mAP) of 90.6%. The mean average precision of the research work was 9.6% higher than previous studies. The results showed that the improved model could realize the detection and identification of plastic waste in water.