Enhancing Face Super-Resolution via Improving the Edge and Identity Preserving Network

Face super-resolution, known as face hallucination, is a domain-specific image super-resolution problem, which refers to generating high resolution face images from their low resolution. State-of-the-art face super-resolution methods used deep convolutional neural networks. However, due to significant pose changes and difficulty in recovering high-frequency details in facial areas, most of these methods do not deploy facial structures and identity information well, and it is tough for them to reconstruct super-resolved face images. According to previous researches, proper use of low-resolution image edges can be a solution for these problems. EIPNet (Edge and Identity Preserving Network) is one of the newest methods to achieve outstanding results in this area. In the EIPNet method, the authors used a lightweight edge extraction block in the proposed GAN structure. In this research, we intend to improve the performance of the EIPNet method by presenting a simple but efficient technique. Our proposed technique divides the face images into upper and lower parts. We train a separate network for each area. This technique reduces the number of face components to train from each area, and the networks can better be trained from their components. The results show that this technique can have an excellent effect on visual quality and quantitative measurements in face super-resolution.