Face Super-Resolution by Deep Collaborative Representation

In order to obtain high-resolution face images with better quality, a face super-resolution algorithm based on deep collaborative representation is proposed. Firstly, the algorithm extracts overlapping patches and multi-directional gradient feature to obtain the initialization dictionary. Then, the initialization dictionary is updated iteratively layer by layer, and the optimal expression weight coefficients corresponding to each layer are updated by the collaborative representation. Finally, all the reconstructed patches of the last layer are combined into a final high-resolution image. Compared with traditional super-resolution reconstruction algorithms, the experimental results on FEI and CMU Frontal Face datasets show that the proposed algorithm improves the accuracy of the single-level representation algorithm and outperforms the existing algorithms in both subjective and objective evaluation performance, even including the deep-based learning algorithm.