An Attention Mechanism Based YOLOv5DS for USV Floating Object Detection and River Channel Segmentation

Nowadays, the environmental crisis facing humanity has become increasingly severe. One of the concerns is the pollution caused by floating debris in rivers. With the rapid development of AI technology, there has been a shift towards utilizing Unmanned Surface Vehicles (USV) to identify floating objects and water areas. However, many complex network models can only address a single visual task with limited efficiency and performance. In this paper, we propose an Improved- YOLOv5DS method based on attention mechanism to achieve detection and segmentation tasks simultaneously in USV. It is composed of one encoder for feature extraction and two decoders to handle the specific tasks. Our model with fewer parameters performs extremely well on the challenging Flow-Img dataset, achieving state-of-the-art on two tasks in terms of accuracy. Specifically, Improved-YOLOv5DS achieves an average precision (AP) of 94%, which represents a significant improvement of nearly 7% over the original model. Additionally, the mean Intersection over Union (mIOU) of the segmentation results is increased from the original 91.7% to 96.2%. Besides, we also verify the effectiveness of different modules via ablative studies.