Abstract
The classification of fish species has important practical significance for both the aquaculture industry and ordinary people. However, existing methods for classifying marine and freshwater fishes have poor feature extraction ability and do not meet actual needs. To address this issue, we propose a novel method for multi-water fish classification (Fish-TViT) based on transfer learning and visual transformers. Fish-TViT uses a label smoothing loss function to solve the problem of overfitting and overconfidence of the classifier. We also employ Gradient-weighted Category Activation Mapping (Grad-CAM) technology to visualize and understand the features of the model and the areas on which the decision depends, which guides the optimization of the model architecture. We first crop and clean fish images, and then use data augmentation to expand the number of training datasets. A pre-trained visual transformer model is used to extract enhanced features of fish images, which are subsequently cropped into a series of flat patches. Finally, a multi-layer perceptron is used to predict fish species. Experimental results show that Fish-TViT achieves high classification accuracy on both low-resolution marine fish data (94.33%) and high-resolution freshwater fish data (98.34%). Compared with traditional convolutional neural networks, Fish-TViT has better performance.