Parallel Region-Based Deep Residual Networks for Face Hallucination

Abstract

Face hallucination is a super-resolution algorithm specially designed to improve the resolution and quality of low-resolution (LR) input face images. Although a deep neural network offers an end-to-end mapping from LR to high-resolution (HR) images, most of the deep learning-based face hallucinations neglect the structure prior for face images. To utilize the highly structured facial prior, a parallel region-based deep residual network (PRDRN) was developed to predict the missing detailed information for accurate image reconstruction. Initially, the image is divided into multiple regions with the symmetry of face structures. Then, the sub-networks corresponding to multiple regions are trained in parallel. Finally, all reconstructed regions are combined to form the HR image. The experimental results on FEI, CASIA-Webface and CMU-MIT public face databases show that the proposed network outperforms other state-of-the-art approaches.

Keywords

• Face hallucination
• face structural prior
• region-based
• parallel strategy
• residual network