Inpainting of Multi-channel Images Using a Dictionary Learned from a Single Image

Image inpainting consists in recovering missing parts of an image. Since a color image is a D array, tensor completion methods are applicable to this problem. Tensor completion approach based on trace norm minimization can be useful when the fraction of missing pixels is not large, with the advantage that the training set is not required. Here, we demonstrate that the dictionary for sparse representation of multichannel image patches can be learned from a single (clean) image, yielding results comparable to those as obtained by a dictionary learned on a training set of images. We show that the learned dictionary-based approach performs considerably better than tensor completion both on color images of natural scenes and multi-phase computed tomography images. Index Terms – dictionary learning, independent component analysis, color image, computed tomography image, inpainting, tensor completion.