Learning to Hallucinate Face Images via Component Generation and Enhancement

摘要

We propose a two-stage method for face hallucination. First, we generatefacial components of the input image using CNNs. These components represent thebasic facial structures. Second, we synthesize fine-grained facial structuresfrom high resolution training images. The details of these structures aretransferred into facial components for enhancement. Therefore, we generatefacial components to approximate ground truth global appearance in the firststage and enhance them through recovering details in the second stage. Theexperiments demonstrate that our method performs favorably againststate-of-the-art methods