Rotation Invariant Deep Learning Approach for Image Inpainting

Matrix completion technique aims to fill the missing information of a bidimensional structure, supporting applications such as recommender systems, image restoration, and image inpainting. In particular, most of literature methods for image inpainting take advantage of intrinsic high spatial correlations by incorporating low-rank regularizers into an optimization problem. However, the presence of geometric distortions such as the rotation of the objects in the image or of the patterns of the missing information can affect the quality in the completion task. To overcome such limitation, this work presents a non-local geometric similarity regularizer, invariant to the rotations in the spatial domain, the similarities presented in some sub-regions of the image with the same structure but rotated at some angle. The rotation invariant regularizer is integrated into a convolutional neural network and a generative adversarial neural network, improving the completion quality, specially for images with high rotation distortions.