An improved YOLOv8-based method for litter identification in drone shoreline images

This paper utilizes drone-based orthophoto images of beach litter in Tangshan city, Hebei province. A recognition model (YOLOv8-b) is proposed for beach litter detection, based on the YOLOv8 model, incorporating methods such as the lightweight upsampling operator DySample, the CBAM attention mechanism, and the Powerful-IoU loss function to improve model detection accuracy. Experimental results on a self-constructed dataset show that, compared with the base YOLOv8 model, the YOLOv8-b model achieves a 2.1% increase in accuracy, a 1.9% increase in recall, and a 1% improvement in mAP. The model demonstrates strong recognition capabilities for four types of beach litter, with recognition accuracy for all categories exceeding 90%. Furthermore, the model was applied to test a beach in Tangshan bay, yielding favorable results in litter recognition. The YOLOv8-b model meets the requirements for garbage localization and recognition in drone imagery, providing technical support for the operational monitoring of beach litter.