A detection method for big-belly seahorse bacterial enteritis based on dynamic sampling and hierarchical feature enhancement

Accurate detection of bacterial enteritis in seahorses (Hippocampus abdominalis) is important for health monitoring in aquaculture. However, dense distribution, overlap, and partial occlusion can weaken fine-grained postural cues, cause feature interference between adjacent individuals, and reduce discriminative information from key regions. To address these challenges, this study proposes RHDW-DETR, a seahorse bacterial enteritis detection method based on dynamic sampling and hierarchical feature enhancement. Based on RT-DETR-R18, Lightweight Re-parameterized Partial Convolution (LRPConv) is introduced into the backbone to enhance local feature representation while reducing redundant computation. Dynamic Sampling (DySample) replaces fixed upsampling to improve spatial alignment and reduce feature interference among adjacent individuals. A Hierarchical Attention Fusion Block (HAFB) integrates local structural and multi-scale semantic information to improve feature representation under overlap and partial occlusion. WIoU replaces GIoU for bounding-box regression to improve localization stability. Experiments were conducted on a custom underwater dataset containing 1,500 images and 12,315 annotated seahorse instances. RHDW-DETR achieved a precision of 79.7%, a recall of 76.9%, an mAP@0.5 of 85.1%, and an mAP@0.5:0.95 of 67.3%. Compared with the baseline, these metrics increased by 0.5%, 0.6%, 2.3%, and 1.4%, respectively, while the number of parameters decreased by 10.8% to 17.91 million. Ablation experiments confirmed the effectiveness of the four modules, and comparative experiments showed that RHDW-DETR achieved the highest precision, mAP@0.5, and mAP@0.5:0.95 among the compared models. These results demonstrate that RHDW-DETR improves feature representation, reduces interference between adjacent individuals, and enhances detection and localization under overlap and partial occlusion, providing technical support for intelligent detection of seahorse bacterial enteritis in aquaculture environments.