A CNN-Based Semi-supervised Self-training Method for Robust Underwater Fish Recognition

Recent AI advances have revolutionized automation in diverse fields. However, despite object detection research progress, underwater fish species identification remains underexplored. Underwater fish recognition is challenged by the unique underwater environment, fish diversity, and limited labeled data. This study introduces a semisupervised self-training method, using YOLOv5 as the foundation. Our approach iteratively refines the model with labeled and unlabeled data, enhancing accuracy in data-scarce scenarios. Random data augmentation is also adopted to bolsters model robustness, addressing the complexities of underwater environments.