Sparse representation of natural images over redundant dictionary enables solution of the inpainting problem. A major challenge, in this regard, is learning of a dictionary that is well adapted to the image. Efficient methods are developed for grayscale images represented in patch space by using, for example, K-SVD or independent component analysis algorithms. Here, we address the problem of patch space-based dictionary learning for color images. To this end, an image in RGB color space is represented as a collection of vectorized 3D patch tensors. This leads to the state-of-the-art results in inpainting random and structured patterns of missing values as it is demonstrated in the paper.