Crab species recognition method based on incremental learning

Abstract

Crab species are highly diverse, and accurate identification is vital for production management, maturity prediction, yield estimation, automated sorting, and quality grading. Existing deep learning methods typically require all category samples to be input simultaneously, which is impractical because of the challenges of obtaining sufficient training samples and the long-term nature of sample collection. Incremental learning offers a solution by enabling models to learn new categories while retaining knowledge from previous ones. This study proposes a crab species recognition method based on an improved adaptive aggregation networks learning a unified classifier indirectly via rebalancing (AANets-LUCIR) framework. AANets combined with LUCIR serves as the baseline. The ResNet18 backbone is improved by replacing the activation function with ReLU6 to preserve image features and introducing an efficient pyramid split attention mechanism in each BasicBlock module to enhance feature extraction. A k-means clustering-based sample replay strategy is integrated to improve adaptation to new data while retaining old knowledge. Experiments on a self-constructed dataset yielded excellent results, with average scores of 92.30% for adaptability, 89.80% for base, and 91.66% for the final evaluation. Compared with iCaRL, Bic, LwF, EWC, and Replay, the proposed method improved the final scores by 10.15, 5.83, 21.76, 7.2, and 17.61 percentage points, respectively. This approach effectively addresses the incremental crab species classification challenges and supports the development of intelligent aquaculture.