Video Inpainting Exploiting Tensor Train and Sparsity in Frequency Domain

Abstract

This paper studies the video inpainting method based on tensor train (TT) decomposition framework. Since TT decomposition works better for the low rank representation on higher-order tensors, we first use a tensor augmentation scheme to generate a higher order tensor from the damaged video. Then, we exploit low TT rankness to recover the damaged video. In addition, according that the time domain smoothing corresponds to the frequency domain sparsity, we combine the sparsity of Fourier domain along the time dimension to construct the inpainting model. Finally, the proposed video inpainting model combining low TT rankness and sparsity in frequency domain is solved by the efficient alternating direction method of multipliers (ADMM) algorithm. The experimental results verify that the proposed method can achieve superior video inpainting results, and can improve the PSNR of video inpainting results better than the methods only enforcing low TT rankness or sparsity.