A nonlocal feature self-similarity based tensor completion method for video recovery

Abstract

The nuclear norm-based tensor completion method effectively recovers missing multidimensional data in videos by minimizing the truncated nuclear norm. However, the conventional thresholding approach might overly punish larger singular values, which leads to loss of faithful information. In order to overcome this challenge, an improved model of the truncated nuclear norm-based tensor completion is proposed. This improved model is designed to incorporate information related to the prior rank and ensure the preservation of essential singular values to improve the approximation of the matrix rank. First, a video tensor is decomposed into matrices along different modes, and these matrices are divided into similar block matrices by using the K-means＋＋ clustering method, which utilizes the non-local similarity to reduce the influence of noise in the video. Afterwards, a ranking algorithm based on noise matrix analysis is used, which automatically gets the truncated threshold by iterative optimization. Additionally, we propose back-projection method to balance local and global optimization. Finally, to evaluate our proposal, extensive experimental evaluations have been carried out and show that our approach outperforms a lot of recent tensor completion techniques in terms of quality metrics and visual impact.