Underwater fish species recognition based on feature fusion visual attention graph convolutional networks

Abstract As marine science advances, the demand for efficient fish species recognition has become increasingly urgent. The manual recognition method is not only time-consuming but also requires extensive expertise in ichthyology. By integrating deep learning techniques into fish species recognition, efficiency is usually increased, which propels marine research forward. Existing methods mainly rely on convolutional neural networks. However, these networks normally use the grid-structured neighborhood for the convolution operation, which may not encode the spatial layout of the image well. To address this issue, we pioneer the application of graph convolutional networks (GCNs) to underwater fish species recognition. Specifically, we divide an image into a set of cells that are treated as nodes and build a graph on top of the k -nearest neighbor nodes in the feature space. In this case, the GCN can better capture the spatial layout of the image. We also propose a feature fusion visual attention GCN, which combines a global attention mechanism with an asymptotic feature fusion module, achieving superior recognition performance on underwater fish images.