Global-local fusion network for face super-resolution

Abstract Face hallucination is a domain-specific super-resolution (SR) algorithm, that generates high-resolution (HR) images from the observed low-resolution (LR) inputs. Recently, deep convolutional neural network (CNN) based SR offers an end-to-end solution for learning the complex relationship between LR and HR images, and achieves superior performance. However, most of them ignore the role of high-frequency (HF) information in image recovering. We design a novel global-local fused network (GLFSR) to refine HF information for recovering fine details of facial images. In contrast to existing methods that often increase the depth of network, we enhance the residual HF information from local to global levels through the networks. The proposed global-local fused network involves four sub-modules: first, reconstruction network, which is used to super-resolve the synthetic HR image from pixel level by reconstruction network, local and global residual enhancement networks, which generate residual information for learning; and fusion module, which is used to generate the final HR image. Experimental results on CAS-PEAL-R1 and CASIA-Webface databases demonstrate that GLFSR is superior to other state-of-the-art deep learning approaches.