A Novel Lightweight Underwater Fish Individual Recognition Method Based on Local Feature Guidance Training for Precision Aquaculture

Abstract

Highlights This article proposed a novel lightweight underwater fish individual recognition method. The proposed method can fully learn the characteristic information of fish. This method can guide training through local features. During prediction, only the global branch is utilized, resulting in a small number of parameters. Abstract. Individual recognition of fish within aquaculture fish groups is crucial for monitoring the growth status of each fish and is a key technology for precision and green aquaculture. However, accurately recognizing fish faces challenges due to the significant variations in fish postures and the influence of complex environments. This article aims to present a highly accurate method for underwater fish individual recognition, offering technical support for intelligent fish farming. To achieve this goal, a novel lightweight underwater fish individual recognition method based on local feature guidance training was proposed for precision aquaculture in this study. The method designs a multi-scale network based on deep learning and utilizes SE_ResNext for feature extraction. This network consists of a branch representing the global feature and a branch representing the local feature. During training, the full image of the fish individual is inputted as a coarse scale to the global branch (G-branch) to learn the global feature and calculate the ID loss. The image of each part formed after segmenting the fish individual is inputted as a fine-scale input to the partial branch (P-branch) to learn the local feature. Subsequently, the global feature and local feature are combined to form the fusion feature, and the Triplet loss and ID loss are calculated on the fusion feature. Finally, the loss calculated in the global feature and the loss calculated in the fusion feature are weighted and summed to obtain the total loss. Through the design of the loss function, the global feature is consistently influenced by the local feature, and the local feature serves to guide the global feature. During prediction, the global feature, having learned the information from the local feature, allows for the sole utilization of the G-branch to recognize fish individuals. This approach significantly reduces the number of parameters and computational requirements, resulting in a lightweight network. The method achieves high accuracy, with both Rank-1 and Rank-5 accuracy rates surpassing 95%. Remarkably, even with low image resolution and slight obscuration of the fish body, the Rank-1 accuracy remains above 90%. Keywords: Deep learning, Multi-scale network, Object detection, Recognition of fish individual.