Research on Deep Learning-based Digital Restoration Method for Imagesof Intangible Cultural Heritage Thangka

Thangka is part of China’s intangible cultural heritage, but many Thangka images have certain defects due to problems such as age and improper preservation. This paper addressed the problem of digital restoration of Thangka images. Based on the generative adversarial network (GAN) in deep learning, a method was designed with U-Net as the generative model and PatchGAN as the discriminative model. The loss function of the model was composed of reconstruction loss, perceptual loss, and adversarial loss. Thangka images were collected to establish a dataset to verify the proposed approach. The results showed that at a mask ratio of 1%-10%, the peak signal-to-noise ratio, structural similarity index, and learned perceptual image patch similarity of the proposed approach were 33.21 dB, 0.9825, and 0.0187, and its performance was superior to that of methods such as EdgeConnect and progressive reconstruction of visual structure network at different mask ratios. The findings demonstrate the reliability of the proposed approach for digital restoration of images. It can be applied to the digital restoration of actual Thangka images.