Search for a command to run...
Integral to marine ecological research and conservation, underwater biological object detection technology provides precise, efficient data essential for resource management and biodiversity studies. This capability establishes it as a fundamental methodology for promoting the sustainable stewardship of oceanic resources. Recent advancements in deep learning have enabled the application of sophisticated algorithms, such as YOLO, to underwater biological detection. Although these methods demonstrate considerable potential, their detection accuracy often falls short of practical requirements. This performance gap is primarily due to adverse underwater conditions, the high density and clustering of organisms, and the predominance of small-scale targets. To overcome these persistent issues, we developed YOLO-CHS, a customized algorithm built upon the YOLOv11 architecture for robust underwater detection. First, by incorporating the feature complementary mapping module into the C3k2 block, we design C3FNet, which integrates spatial positional information more deeply into the network, facilitates alignment with high-level semantic features, enhances small-object localization capability, and simultaneously reduces model complexity. Second, we develop the HDM-FPN feature pyramid structure, which leverages hypergraph computation to integrate multiscale features, enabling high-order information propagation across layers and spatial positions, thereby significantly strengthening the neck network’s high-order feature extraction ability. Finally, we introduce a spatially enhanced attention module into the detection head. In this module, reinforced responses from nonoccluded regions compensate for the loss of responses in occluded areas, thereby improving detection robustness under occlusion. The proposed method was benchmarked on the RUOD and Aquarium datasets, where it exhibited superior detection capabilities.