On Potentials of the Gamma Densities to Blind Face Hallucination

As to blind hallucination of blurry low-resolution faces, the supervised learning methods obviously lack generalization for blur kernels in various types and scales. While, the unsupervised learning ones usually require pre-trained deep generative models. The involved non-convex inversion is often highly challengeable, especially as the blurred low-resolution face image is undergone a blur process much distinct from the ones in the pre-training stage. This paper falls into unsupervised blind face hallucination, while takes a step forward. That is, a deep discriminative face prior has been proposed by innovatively exploring the Gamma densities for unsupervised blind hallucination, which is expected to regularize better the inversion of pre-trained deep generative models. The experimental results well prove the effectiveness of the proposed method.