Mural inpainting with generative adversarial networks based on multi-scale feature and attention fusion

This study proposes a deep learning model for mural restoration based on generative adversarial networks with multi-scale feature and attention fusions, addressing insufficient feature extraction and detail loss of the existing deep learning image inpainting algorithms during reconstruction. Firstly, a multi-scale feature pyramid network is designed to extract feature information of different scales in mural images, which enhances the feature relevance. Secondly, using the self-attention mechanism and feature fusion module, a multi-scale feature generator is constructed to obtain rich context information and improve the restoration ability of the network. Finally, the minimal confrontation loss and the mean square error are introduced to promote the residual feedback of the discriminator, which completes the mural restoration by combining the feature information of different scales. The experimental results of digital restoration of real Dunhuang murals show that the proposed algorithm can effectively protect important feature information such as the edges and textures, and that the subjective visual effects and objective evaluation indicators are superior to those of the algorithms for comparison. © 2023 Beijing University of Aeronautics and Astronautics (BUAA). All rights reserved.