Face Image Super-resolution based on GAN Constrained by Residual Dense Network and Channel Attention Mechanism

Face image carries the key feature recognition information of human, and has wide application in intelligent security, electronic payment, public security and other fields. High resolution face images often contain more real and abundant face features. How to acquire high resolution face images quickly, efficiently and at low cost has become a hot research topic in the field of image super resolution reconstruction. In this paper, the generation model of SRGAN algorithm is improved to solve the problems of the traditional super-resolution reconstruction algorithm, such as too smooth reconstruction image and insufficient detail reconstruction ability. The Residual Dense Network (RDN) and Channel Attention (CA) Mechanism is used to form a joint unit, and dense connections are adopted between the joint learning units to fully retain as much shallow information as possible. CA algorithm can extract the key information of face image effectively. The loss function in this paper includes three parts: pixel loss, feature loss and countermeasure loss, which can not only ensure that the reconstructed image has as complete content information as possible, but also ensure that the reconstructed image is consistent with the real HR image in terms of high-level semantic features. Experiments show that the image reconstructed by the proposed algorithm has more detailed information in terms of visual effects. The average peak signal-to-noise ratio(PSNR) of the proposed algorithm on the test data set is 30.5109, the average structural similarity(SSIM) index is 0.8363, and the average LPIPS index is 0.0504, which is superior to other classical algorithms.