Efficient face image super-resolution with convenient alternating projection network

Abstract

The existing deep learning-based face super-resolution techniques can achieve satisfactory performance. However, these methods often incur large computational costs, and deeper networks generate redundant features. Some lightweight reconstruction networks also present limited representation ability because they ignore the entire contour and fine texture of the face for the sake of efficiency. Here, the authors propose a convenient alternating projection network (CAPN) for efficient face super-resolution. First, the authors design a novel alternating projection block cascaded convolutional neural network to alternately achieve content consistency and learn detailed facial feature differences between super-resolution and ground-truth face images. Second, the self-correction mechanism enabled the convolutional layer to capture faithful features that facilitate adaptive reconstruction. Moreover, a convenient connection operation can reduce the generation of redundant facial features while maintaining accurate reconstruction information. Extensive experiments demonstrated that the proposed CAPN can effectively reduce the computational cost while achieving competitive qualitative and quantitative results compared to state-of-the-art super-resolution methods.