Small-target water-floating garbage detection and recognition based on UNet-YOLOv5s

Abstract

To address the problem of low recognition accuracy caused by the interference of illumination, water waves and complex backgrounds in the process of detecting and recognizing small-target of water-floating garbage, a water-floating garbage recognition algorithm that fuses image semantic segmentation network and target detection network is proposed. Firstly, using UNet network, the water floating garbage is segmented from the complex background and the image is cropped into a uniform size image. Then, a dark channel prior algorithm is used for noise removal, which cuts down the interference of illumination and water waves. Finally, eight categories of water floating garbage are defined from the perspective of the relationship between the degree of contamination and recyclability of water-floating garbage, and according to this classification, the improved YOLOv5s network is trained to achieve the detection and recognition of water-floating garbage. The results show that the proposed algorithm has 89.80% detection and recognition accuracy, 85.76% recall and 87.73% F1score, which can effectively improve the detection and recognition accuracy of water-floating garbage and provide a new and effective method for the management of water-floating garbage.