This paper addresses the challenge of limited sample size in fish species identification by proposing a few-shot learning method. The method combines an attention module with prototypical networks to enhance model accuracy. The attention module, which includes channel and spatial attention mechanisms, improves feature extraction by focusing on key information and locating the primary regions of fish images. The prototypical networks are used to calculate similarity across various fish datasets, achieving an accuracy improvement of 2% to 10% compared to traditional prototypical networks. The proposed method is evaluated on three public datasets: Croatian Fish, Fish4Knowledge, and WildFish, demonstrating effective performance in small sample sizes and complex scenes. The method provides valuable technological tools for fisheries management and the preservation of fish biodiversity.This paper addresses the challenge of limited sample size in fish species identification by proposing a few-shot learning method. The method combines an attention module with prototypical networks to enhance model accuracy. The attention module, which includes channel and spatial attention mechanisms, improves feature extraction by focusing on key information and locating the primary regions of fish images. The prototypical networks are used to calculate similarity across various fish datasets, achieving an accuracy improvement of 2% to 10% compared to traditional prototypical networks. The proposed method is evaluated on three public datasets: Croatian Fish, Fish4Knowledge, and WildFish, demonstrating effective performance in small sample sizes and complex scenes. The method provides valuable technological tools for fisheries management and the preservation of fish biodiversity.