Intelligent enhancement of ancient Chinese murals based on multi-scale parallel structure

Abstract

Ancient mural artwork preserves the historical background and cultural customs of that time through intricate details and bright colors. However, after the natural environment and man-made damage, these works of art are damaged in color, texture and content and lose their quality. In order to identify and enhance murals with large areas of color damage, we propose a multi-scale parallel GAN and parallel Unet structure, which can extract features from multiple scales or images to adapt to the changing scale of the target and provide a more diverse set of features. This structure can reduce the risk of overfitting the training data by learning more general features. The verification results of indicators such as PSNR on the ancient mural data set show that the method has a certain performance improvement effect.