Exemplar-based Image Inpainting Using Adaptive Two-Stage Structure-Tensor Based Priority Function and Nonlocal Filtering
Ting Xu,Ting-Zhu Huang,Liang-Jian Deng,Xi-Le Zhao,Jin-Fan Hu
DOI: https://doi.org/10.1016/j.jvcir.2021.103430
IF: 2.887
2022-01-01
Journal of Visual Communication and Image Representation
Abstract:For the exemplar-based image inpainting problem, the filling order and local intensity smoothness are two crucial factors that should be considered carefully. This work gives a new exemplar-based image inpainting method, preventing geometric structures from being destroyed and reconstructing textures well to obtain elegant-looking outputs. For a better filling order, we define a new adaptive two-stage structure-tensor based priority function. To promote the local intensity smoothness, we adopt a non-local way, and at the same time, propose a weighted filter based on a Gaussian-like function to generate the ideal filling patch by combining non-local patches. We compare the proposed method with some recent state-of-the-art image inpainting approaches on different tasks, such as texture and structure synthesis, object removal, and remote sensing images inpainting. Experimental results demonstrate the superiority of the proposed method, both visually and quantitatively.