This project aims to propose a fast and automatic inpainting technique for high-definition videos which works under many challenging conditions such as a moving camera, a dynamic background or a long occlusion. Our algorithm does not limit to objects removal type but extends to simultaneous background and foreground reconstruction even when the moving objects are occluded for a long period. Built upon Newson et al [1] which optimizes a global patch-based function, our method holds a significant improvement by the introduction of the optical flow term. Based on this term, a novel initialization scheme, a modified patch distance, an optical flow-guided patch searching strategy and a separation map are introduced. We also attain a goal of reducing the computation time through parallelizing the algorithm and modifying the patch search strategy. By experimenting and comparing the result with other state-of-the-art results, we discover that our method has the capability of preserving the spatio-temporal coherency as well as reconstructing moving objects within a long temporal occlusion. It compares favorably with previous works while radically reducing the computation time.

Our method is evaluated under a wide variety of conditions, including moving objects occluded by a fixed or moving domain, static or moving camera, dynamic background, large occlusions, etc. To prove the effectivity of our method, we compare its performances with other stateof-the-art algorithms using their publicly available datasets.