DE-YOLOv8n: Detection of Dead Fish in Net-Pen Aquaculture in the Complex Environment of the Deep and Distant Sea

In deep-sea net-pen aquaculture, the real-time and rapid detection of dead fish is crucial for early risk warning and enhancing aquaculture efficiency. However, the complexity of the real-world environment and the variability in fish morphology present significant challenges to dead fish detection. To address these challenges, this paper proposes DE- YOLOv8n, a lightweight and high-precision dead fish detection model based on deformable convolution and an efficient multi-scale attention mechanism. First, we establish a dataset for underwater dead grouper snapper detection in a simulated real-world net-pen aquaculture environment. Next, we improve the C2f module in YOLOv8 to C2f _ DCNv2, which enhances the model's ability to capture and represent the complex shapes of dead fish, thereby improving overall detection accuracy. Finally, we introduce an efficient multi-scale attention mechanism (EMA) based on cross-space learning into the model's backbone network, which enhances its ability to detect dead fish across different scales. Experimental results show that the proposed method achieves 95.5%, 71.9%, 93.2%, 86.4%, and 400 for mAP50, mAP50-95, Precision, Recall, and FPS, respectively, on the UDOD dataset. These results outperform detection algorithms such as YOLOv7, Fast R-CNN, and YOLOv8s. Furthermore, compared to the baseline model YOLOv8n, the proposed method improves the AP50 by 4.9% and the frame rate (FPS) by 43. In conclusion, the DE-YOLOv8n model facilitates real-time, high-precision dead fish detection and provides effective technical support for underwater robots in deep-sea net-pen aquaculture.