Fine-grained tuna detection in complex dynamic monitoring scenarios: A lightweight framework based on feature decoupling and partial convolution

Deploying fine-grained object detection models on resource-constrained shipborne edge devices remains a significant challenge for marine electronic monitoring (EM) systems. Although lightweight network architectures reduce computational overhead, they exacerbate the task conflict between object localization and fine-grained classification. This conflict degrades detection accuracy for visually similar species in long-tailed distributions, such as yellowfin tuna. To address this issue, we propose PSE-YOLO, an end-to-end framework designed to decouple features and mitigate task conflicts. Rather than utilizing generic decoupled heads, we introduce EfficientHead, a detection head based on deep parameter sharing tailored specifically for edge constraints. To prevent the feature confusion typically induced by parameter sharing, a Supervised Contrastive Learning (SupCon) mechanism is integrated into the training phase. This explicitly maximizes fine-grained inter-class margins without incurring any additional inference latency. Furthermore, Partial Convolution (PConv) is employed within the shallow layers of the backbone network to preserve high-frequency local details while pruning spatial redundancy. Evaluations on a real-world tuna dataset demonstrate that PSE-YOLO reduces the total parameter count by 16.3% (to 2.10 M) and computational complexity by 31.0% (to 4.0 GFLOPs). By effectively mitigating task conflicts, the proposed framework improves the overall mAP@0.5 from 88.5% to 89.0% and increases the specific accuracy for yellowfin tuna from 76.3% to 78.4%, all while maintaining a processing speed of 48.07 FPS.