Image Recovery Using Tensor Completion via Low-rank Augmentation

An image can be represented as a third-order tensor, and iteratively approximated using a low-rank tensor decomposition model. However, some images may not have a strong low-rank property. We need to lower their rank. In this paper we propose an effective tensor completion model via low-rank augmentation by integrating the low-rank property and local similarities of a tensor for recovering missing entries of an image. First, dividing an original image into sub-images and concatenating them to be a tensor. Second, utilizing a tenor completion algorithm based on the tensor nuclear norm to recover its missing entries, and extracting the recovered image from the tensor. Finally, filtering the recovered image with the bilateral filter. We fully utilize the local similarity of the image and obtain the third-order tensor composed of sub-images by sampling adjacent pixels of the image, making it have stronger low-rank characteristics. Experimental results demonstrate the advantages of our method compared with other tensor completion algorithms.