Multi-scale face hallucination based on frequency bands analysis

In this paper, a multi-scale face hallucination method is proposed to produce high-resolution (HR) face images from low-resolution (LR) ones according to the specific face characteristics and priors based on frequency bands analysis. In the first scale, the middle-resolution (MR) images are generated based on a patch-based learning method in DCT domain. In this scale, the DC coefficients and AC coefficients are estimated separately. In the second scale, a DCT upsampling for low frequency band restoration and a high frequency band restoration are combined to generate the final high-resolution face images. Extensive experiments show that the proposed algorithm achieves significant improvement.