inpaint

Inpainting, also called image inpainting, is the process of reconstructing lost, occluded, or deteriorated parts of an image such that the result is visually plausible and consistent with the surrounding content. The goal is to fill regions in a way that preserves or restores structure, texture, and color continuity without introducing obvious artifacts. In video, similar techniques extend to temporal coherence across frames.

Digital inpainting methods are broadly categorized as diffusion-based, exemplar-based, or data-driven. Diffusion-based approaches propagate information from the boundary of the missing region inward, often using partial differential equations to ensure continuity of gradients. Exemplar-based or patch-based methods fill the hole by pasting and blending patches sampled from undamaged parts of the image, guided by similarity criteria. Data-driven and learning-based methods, including deep neural networks, learn priors from large image collections and can synthesize plausible textures and sharp structures.

Prominent early methods include Navier–Stokes based inpainting and the fast marching method, while exemplar-based techniques were advanced by Criminisi, Pérez, and Toyama. More recent approaches use deep learning and generative models to improve realism and handle complex textures.

Applications include photograph restoration, removing objects or blemishes, and completion of occluded frames in video or virtual reality content. Challenges remain in maintaining structural coherence, especially across large regions or in highly textured areas, and in ensuring temporal consistency in video. Evaluation often relies on perceptual metrics and qualitative assessment, with results depending on the content and task.