Research on image texture structure filling algorithm based on MAE prior model

Image inpainting, a crucial aspect of computer vision, involves completing missing content within an image using available information from the surrounding context. While existing deep convolution-based inpainting network lack sufficient guidance for single-stage inpainting, multi-stage inpainting network often lack integration with texture structure features. To overcome these limitations, this paper proposes a structural texture inpainting network based on mask auto encoder (MAE) priori knowledge. Our method replaces MAE priori features with two-stage coarse structure network, and a texture structure fusion module is designed to assist the input of the main network decoder with this MAE prior. Through extensive experimentation on the Place2 dataset, our advanced image inpainting network demonstrates superior performance, as evidenced by both qualitative and quantitative analyses.