Aqua Vision: Few-Shot Learning Based Efficient Fish Identification in Challenging Aquatic Habitats

Abstract

Aquatic ecosystems play a vital role in marine biodiversity and coastal protection, yet monitoring these habitats remains a significant challenge due to the scarcity of labeled data for training robust detection models. Traditional approaches often rely on extensive labeled datasets, which are costly and time-consuming to obtain, leading to a critical research gap in effective fish detection methodologies. This study introduces an innovative approach to fish detection by leveraging few-shot learning and pseudo-labeling techniques. We employ SimCLR, a contrastive learning framework, to pre-train a ResNet50-based encoder on unlabeled Deep Fish images, thereby extracting robust feature representations. These features are then utilized to train a Faster R-CNN object detection model using a limited set of labeled sea grass images. To further enhance the model’s performance, we incorporate pseudo-labeling, a semi-supervised learning technique that generates additional training data from unlabeled images based on a confidence threshold. Our methodology demonstrates significant improvements in fish detection accuracy. The final model achieves an average precision of 0.8167 and recall of 0.7967, outperforming other state-of-the-art models such as YOLOv5 and RetinaNet. These results highlight the effectiveness of combining few-shot learning with pseudo-labeling in addressing the challenge of limited labeled data, paving the way for more efficient and accurate marine ecosystem monitoring.