inpainting

Inpainting is a class of image processing and computer vision techniques that reconstructs missing or damaged regions of an image or video in a visually plausible way. The goal is a seamless result that blends with surrounding content. Originating in art restoration, inpainting has grown into a range of algorithms for still images and moving pictures.

Techniques are often categorized as diffusion-based and exemplar-based. Diffusion approaches propagate information from the boundary into the missing region, guided by gradients or isophotes; Navier–Stokes–based methods are early examples. Exemplar-based methods fill the hole with patches sampled from known areas, seeking patches that minimize patch differences. Criminisi and colleagues popularized a priority-driven version, while Telea proposed a fast local propagation method.

In recent years, deep learning has transformed inpainting. Convolutional neural networks and generative models learn to infer plausible content from surrounding context, enabling reconstruction of large or irregular holes and complex textures. Some approaches are generic, while others incorporate semantic information or user-provided masks to improve results. Video inpainting adds temporal coherence to maintain consistency across frames.

Applications include removing unwanted objects or watermarks, restoring damaged photographs, and medical image artifact reduction. Limitations involve handling large regions with structural content, preserving global consistency, and substantial computational requirements for modern neural methods. Evaluation is often subjective, with benchmarks and user studies used to compare visual quality across techniques.