Application of Deep Learning Techniques for Accurate Fish Segmentation

Aquaculture is a vital component of the global food security apparatus. However, it is faced with significant challenges in relation to the management of cage aquaculture. These difficulties are largely attributed to the prevailing mon-itoring techniques, which have been observed to induce stress in fish and lack efficiency. This study proposes an innovative solution by integrating deep learning techniques, specifically the YOLOv8 architecture, to enhance fish segmentation in complex underwater environments. A bespoke dataset was constructed, comprising images of seabass and seabream captured under diverse conditions, in order to address issues pertaining to background interference, visual similarities between species, and the rapid movement of fish. Following rigorous training over 100 epochs, the model exhibited excellent performance, with a precision of 96.7%, a recall of 96.5%, and a mean average precision (mAP50) of 98.5 %. These results exceed those of existing methods and demonstrate the model's capacity to accurately identify and segment the targeted species across di-verse contexts. This research paves the way for the development of advanced automation in aquaculture monitoring, facilitating ecological tracking, stock assessment, and sustainable resource management.