GEHUNet: A Method for Grayscale Image Colorization of Ethnic Costumes Based on Instance Semantic Graphs

This paper proposes a method for coloring gray-scale images of ethnic costumes, called GEHUNet, to address issues of inaccurate color mapping and color diffusion. The method first annotates the grayscale images to produce instance graphs containing semantic information. These instance graphs are then used as input to the model alongside the grayscale images. The role of the instance semantic graph is to guide the coloring of the grayscale images, thereby solving the problems of inaccurate color mapping and color diffusion. The GEHUNet model integrates ResNet, GoogLeNet, ESPNet, and channel attention mechanisms. This design reduces local information loss and helps capture local features and abstract information to effectively extract features and restore image texture details. Validation tests were conducted on a self-built small dataset, comparing the results with existing methods. The methods show that this algorithm performs excellently in grayscale image colorization and overcomes the shortcomings of existing approaches.