YOLOViT‐DETECT: A Hybrid Approach for Real‐Time Fish Species Detection and Classification in Aquaculture

ABSTRACT The ability to differentiate between different species of fish plays an essential role in aquaculture, as it helps to protect their populations and monitor their welfare. However, conventional learning methods are performance‐limited, especially under complex submarine conditions such as similarity of fish body and background, low light levels and poor water quality. This paper proposes a new method titled YOLOViT‐DETECT, which has practical significance for aquaculture monitoring by enabling accurate, automated tracking of fish species, supporting fish welfare, efficient farm management, and sustainable production. The approach is distinguished by its hybrid and attentional architecture which jointly optimizes the detection and classification of fish in complex underwater conditions. YOLOViT‐DETECT uses an improved YOLOv8 model for the detection phase and a ViT‐B/16 model for fish classification. To reinforce the model's focus on relevant areas of the image, even in the presence of a complex background, we have added a spatial attention module (SAM) to YOLOv8. Detection results from YOLOv8 Improved and classification results from ViT were merged into a single output. Experiments have shown that YOLOViT‐DETECT is a robust method whose results exceed those achieved with YOLOv8 and ViT‐B/16 separately. The precision rate reached 0.9033 in the detection phase and 0.9524 in the classification phase.