Latent low-rank tensor wheel decomposition for visual data completion

Recently, tensor wheel (TW) decomposition gains increasing attention in the area of low-rank tensor completion (LRTC). Existing tensor factorization-based methods can either capture the global connections among all dimension-pairs of data or the local connections between adjacent modes only via low-rank regularization. In this paper, we propose a novel TW decomposition with latent low-rank factors, where the low-rank regularizations are incorporated in the gradient domain of ring factors to enhance the robustness of TW-ranks. Thus, the global low-rank structure of TW decomposition and local continuity of high-order tensors can be exploited in a unified framework. Additionally, an efficient alternating direction method of multipliers (ADMM) algorithm is developed to solve the optimization. Experimental results on real-world visual data such as color images, multispectral images (MSI), and video sequences have showcased the superiority of the proposed method.