In the Erimo region of Hokkaido, Japan, the Ministry of the Environment has promoted the installation of rope-grid deterrents in salmon set-net fisheries to mitigate fishery damage caused by harbor seals. Since 2016, continuous underwater video monitoring has been conducted to evaluate the effectiveness of these measures and to support local seal population management. This study presents a lightweight underwater object detection model developed by integrating an Efficient Channel Attention (ECA) mechanism into the YOLOv4-tiny framework. The model was trained and evaluated using a dataset derived from surveillance videos recorded between 2019 and 2021, encompassing seals, various fish species, and other marine organisms. The optimized model demonstrates strong performance in complex underwater environments, achieving a seal-class Average Precision (AP) of 98.26% and an inference speed of 28.62 FPS under challenging water conditions, making it well-suited for real-time applications. Validation on continuous video segments reveals high consistency between automated detections and manual annotations, indicating substantial potential to significantly reduce the labor intensity of manual monitoring and enable efficient, automated assessment of seal intrusions in set-net aquaculture systems.
Zhang et al. (Sun,) studied this question.