To achieve early and precise detection of various diseases in Carassius auratus , a lightweight multi-disease detection model, EFM-YOLO-DD, was proposed in this study. First, we designed EFM, an efficient lesion-aware backbone network for feature extraction, by embedding the Focal Modulation (FM) module into EfficientNetV1-B0. This design enabled cross-channel feature interaction and enhanced the backbone network's ability for fine-grained representation of small lesions. Second, a deformable dynamic fusion architecture, YOLO-DD, was developed. By reconstructing the FPN+PAN hierarchical topology, we customized a neck network that integrated a deformable attention mechanism (DAT) with a dynamic upsampling operator (DY) to dynamically adjust sampling positions and establish a spatially adaptive feature fusion strategy, thus achieving efficient multi-scale lesion detection and improving the model's feature fusion ability. Subsequently, an integrated YOLO detection head was employed to detect multiple diseases in Carassius auratus . Following this, a scale-aware Inner-IoU loss function was introduced to optimize multi-scale bounding box regression by dynamically adjusting the scaling factor, thereby effectively balancing the localization errors of lesions at different scales. Evaluated in controlled infection experiments, our model achieves a 93.0% mAP@0.5, 59.2% mAP@0.5:95, 90.7% recall, and 87.9% F1-score, while operating in real-time at 76.3 FPS with a computational cost of only 1.5 GFLOPs. Comparative studies with 9 existing mainstream detection methods demonstrate that EFM-YOLO-DD achieves the best overall performance in terms of precision, recall, mAP, and F1-score, requiring only 21.7% of the computational budget of the best-performing baseline (6.9 GFLOPs).
Wu et al. (Fri,) studied this question.