Abstract
Unlike many other fields of object detection, underwater object detection presents some unique challenges. Underwater object detection is a key technology that enables underwater robots to explore aquatic environments. This task is affected by many unavoidable factors, including poor image quality, high environmental randomness, and concealment of fish. These factors make it difficult to perceive and detect fish underwater. To address these complex issues in underwater environments, this paper proposes a fish detection model named YOLO-Starfish and an Underwater Freshwater Fish Dataset (UFFD). The UFFD contains 16,904 unduplicated images of 19 species, encompassing photographs of a variety of complex underwater environments. YOLO-Starfish builds upon YOLOv8, specifically, the C2Star module and the Attention-driven Enhancement Module (ADEM) are proposed. The C2Star leverages the "star operation" (element-wise multiplication) to achieve high-dimensional feature distribution in a physically consistent manner, mimicking the modulation characteristics of underwater optical degradation. Meanwhile, the ADEM mitigates the impact of image channel imbalance by adaptively enhancing channel-driven features, thereby improving the model's robustness in underwater environments. Experimental results demonstrate that YOLO-Starfish not only performs well on underwater object detection datasets (RUOD and our UFFD) but also achieves excellent performance on the common object detection dataset benchmark COCO2017. The source code is available at https://github.com/Sdafah/YOLO-Starfish .