Detection of River Floating Waste Based on Decoupled Diffusion Model

In recent years, the conservation of water resources has attracted widespread attention. The development and application of water surface robots can achieve efficient cleaning of floating waste. However, limited to the small size of floating waste on the water surface, its detection remains a great challenge in the field of object detection. Existing object detection algorithms cannot perform well, such as YOLO (You Only Look Once), SSD (Single-Shot Detector), and Faster R-CNN. In the past two years, diffusion-based networks have shown powerful capabilities in object detection. In this paper, we decouple the position and size regressions of detection boxes, to propose a novel decoupled diffusion network for detecting the floating waste in images. To further promote the detection accuracy of floating waste, we design a new box renewal strategy to obtain desired boxes during the inference stage. To evaluate the performance of the proposed methods, we test the decoupled diffusion network on a public dataset and verify the superiority compared with other object detection methods.