YOLO-DS: research on lightweight algorithm for underwater small target detection by introducing dynamic upsampler

Complex underwater environments, redundant backgrounds, and overlapping aggregations of underwater organisms cause many problems of missed and false detections in underwater small target detection. To address these problems, and considering the high demand of underwater detection hardware, it is especially important to design a lightweight network for underwater target detection. Therefore, we design the YOLO-DS network detection algorithm based on YOLOv8n. Firstly, we replace the traditional upsampler in the neck network with DySample to enhance the ability of the detection network to extract image features; secondly, we modify the loss function in the original YOLOv8 and use Wise-IoUv3 as the new bounding-box loss function to enhance the training effect of the dataset. Finally, we add the CA attention mechanism in front of the detection head for better analyzing the spatial information of the extracted features. Compared with YOLOv8, the accuracy of the improved model YOLO-DS is increased by 1.5%, the mAP is increased by 1.7%, and the F1 score is increased by 0.5% in the URPC2020 dataset. At the same time, we have outstanding performance compared to the latest YOLOv9 and YOLOv10. It is finally proved that the underwater small target detection algorithm YOLO-DS proposed in this paper meets the current underwater detection task requirements.