Face Image Inpainting Network based on Dense Multi-Scale Fusion of Local Binary Patterns

With the development of convolutional neural networks, Deep Learning (DL) has also made great breakthroughs in the field of image inpainting. Deep Learning-based models can repair large-area missing defaced images, but usually produce various unpleasant artifacts in borders and highly textured areas. To solve this problem, this paper proposed a new end-to-end two-stage repair network. In the first stage, a dense multi-scale fusion of LBP is used to generate rich structural information of the missing regions to guide the inpainting network in the second stage to repair. In the second stage, the Bi-directional Skip Connections mechanism is used to flexibly aggregate high-level semantic features and low-level visual features. Experiments on CelebA-HQ public face data and the results show that this paper’s model solves this problem.