Thangka Image Inpainting Algorithm Based on Edge Constraint

Thangka images have the characteristics of rich content, complex textures and bright colors. The existing deep learning inpainting algorithms have some problems in inpainting incomplete Thangka images, such as blurred boundary, distorted structure and unclear texture. To solve the above problems, this paper proposes a two-stage Thangka image inpainting model based on edge constraint, which consists of edge prediction model and texture generation model. First, the edge of incomplete Thangka is reconstructed by the generative adversarial network based on U-Net structure. In order to increase the ability of feature extraction, a hierarchical residual structure is constructed in U-NET. Second, with the constraint of filling edge, a texture generation model combining gated convolution and multi-scale attention is used to reconstruct Thangka texture. Here, multi-scale attention can effectively capture multiscale features to generate images consistent with the texture details of the ground truth images. Experiments demonstrate that the proposed model can achieve better results, and its subjective and objective evaluation indicators are superior than the current image inpainting algorithms.