An Efficient Method for Underwater Fish Detection Using a Transfer Learning Techniques

Detecting fish species is crucial in aquaculture, playing a vital role in safeguarding populations and monitoring their health and nutritional systems. However, traditional machine learning methods struggle to identify objects in images with complex backgrounds, especially in low-light conditions. This paper aims to enhance the performance of a YOLO NAS model for fish recognition. Employing transfer learning, our model leverages a pre-trained model on the COCO(Common Objects in COntext) dataset and is subsequently tested in various scenes. The experimental results demonstrate the model’s effectiveness. Using Recall and mAP50 evaluation metrics, our novel model achieves precision rates of 0.973 and 0.922, respectively.