Efficient Face Super-Resolution Based on Separable Convolution Projection Networks

Abstract

Current super-resolution convolutional neural networks for application to face images generally use a feed-forward architecture. The network frames, however, tend to use deeper structures and larger quantities of parameters. These networks do not readily apply to restricted devices such as home robots or embedded devices, which limits the wide application of face super-resolution algorithms in real-world scenarios. In an effort to remedy this problem, a simple initial feature extraction is established in this study for shallow features of images; an intermediate convolutional layer with a novel deep separable convolution projection block is then used to reduce computational complexity while preserving accuracy. The proposed model can learn various up- and down- sampling information to retain high resolution details and generate deep, effective features. Integrating a dense connection operation allows reconstruction to further improve these super-resolution results. The proposed model is evaluated on the CASIA-WebFace database, to show that it outperforms other state-of-the-art algorithms.