Probabilistic Model-Based Active Learning with Attention Mechanism for Fish Species Recognition

Accurate fish species identification is essential for stock assessments, production management, document ecosystem changes, and protection of endangered species. Image processing and computer vision techniques have been widely employed for fish species detection, classification, and tracking, reducing human efforts in these tasks. However, these methods often rely on extensive training data with correct annotations. Annotating many images captured from marine environments poses a significant challenge. This work proposes a deep-learning model designed for fish detection and classification. The model incorporates an attention mechanism named Convolutional Block Attention Module (CBAM) to improve detection performance. A popular Deep Active Learning approach with cost-efficient annotation is employed, which selects the most informative samples from the unlabeled set. The proposed method utilizes probabilistic modeling based on mixture density networks to estimate probability distributions for localization and classification heads. This study uses the Southeast Area Monitoring and Assessment Program Dataset 2021 (SEAMAPD21). Our model is compared with the conventional supervised algorithm. Experimental results demonstrate superior detection accuracy, achieving a mean average precision (mAP) of 41.6% with minimal labeled data, compared to traditional supervised approaches (mAP-36.7%) that rely on larger labeled datasets. The active learning method with the attention module effectively reduces annotation costs while maintaining excellent detection accuracy. Overall, our proposed deep active learning model with an attention mechanism proves to be highly effective for fish species recognition, providing significant advancements in accuracy and cost efficiency for fish detection tasks.