In order to solve the problems of insufficient accuracy of fish target recognition in complex underwater environment, high complexity and slow recognition speed of traditional target recognition models, an underwater fish target recognition model based on improved YOLOv8n-fish was proposed. In this paper, a lightweight double-convolutional module C2f-DualConv is proposed to improve the feature learning ability of the C2f module in YOLOv8n. Based on the idea of efficient structure reparameterization, a new neck network EffQAFPN was designed to balance the recognition accuracy and speed of the target model. A two-stage fine-tuning method was used to improve the recognition accuracy of the fish target recognition model in the underwater low light and interference environment. The experimental results show that the average accuracy of the YOLOv8n-fish model in the test set is 97.47%, which is 1.07% higher than that of the traditional YOLOv8n model. However, the number of parameters, floating-point arithmetic and model memory occupation of the improved model are only 56.1%, 82% and 66.7% of the original model. The recognition speed of the YOLOv8n-fish model is second only to that of YOLOv5n-P6, which can reach 121 f/s. Experimental results show that the YOLOv8n-fish model can significantly reduce the computational cost while maintaining high recognition accuracy, and provide effective technical support for intelligent monitoring of aquaculture.