Tensor Voting for Image/Video Repairing and Intensity Correction

Chi-Keung Tang


This talk focuses on employing Tensor Voting with color, texture, and spatial information for inferring missing information. We propose a robust synthesis algorithm to automatically infer missing structure and texture information from a damaged 2D image by ND tensor voting (N>3). Our method translates texture information into an adaptive ND tensor, followed by a voting process that infers non-iteratively the optimal color values in the ND texture space. Our method can be naturally extended to video repairing, which has the potential for film restoration. Video repairing consists of two parts: static background and moving objects repairing. Given a damaged video as input, our method fills in missing background and estimates foreground movement. In video repairing, we are confronted with the issue of image registration in constructing video mosaics due to intensity inconsistency among images. We propose to solve the problem by intensity voting, and perform image registration with global and intensity alignment. The key to our modeless approach is the direct estimation of the replacement function, by reducing the complex, non-linear estimation problem into robust 2D tensor voting in the corresponding voting spaces. This method can also handle occlusion, and can be used to correct image intensities given an image pair. The tensor voting approach is compared with the Bayesian approach on the same image correction problem.

Maintained by Philippos Mordohai