Vision graph convolutional network for underwater image enhancement

Colour deviation, non-uniform degradation, and decreased contrast often occur in underwater images because a certain amount of light is absorbed and dispersed underwater. To address this problem, a graph convolution-based underwater image enhancement method (GC-UIE) is proposed. Specifically, patches of underwater images are treated as graph structure, and low-quality underwater images are enhanced by leveraging the advantages of vision graph neural network (VIG). Considering the distortion of underwater images in detail and colour, a local multi-scale feature fusion module and a colour channel correction module based on the mechanism of self-attention are proposed and embedded into the network. Furthermore, the local features are extracted using a convolutional model with multiple receptive fields to complement the global features. To improve colour quality, a self-attention mechanism is utilized. Finally, the underwater images are restored using a residual connection design based on the underwater imaging models. The GC-UIE performed, both qualitatively and quantitatively, better than the other methods. The PyTroch code will be available at https://github.com/xzx11/GC-UIE.