Aimed at solving the unstable training of the existing GAN-based single-image super-resolution reconstruction model and the unsatisfactory visual effect of the reconstructed face image details, a face super-resolution reconstruction algorithm based on EM distance is proposed and named as SRWGAN. The main improvements include removing the BN layer in the model generator and increasing the network depth by adding residual blocks, introducing a shortcut connection optimization network in the discriminator, using EM distance instead of JS divergence as the adversarial loss of the network, and selecting RMSProp to replace the momentum-based gradient optimization algorithm. The experimental results show that, when the improved model algorithm is compared with the BiLinear, SRCNN and SRGAN methods for image reconstruction, the reconstructed face image is improved in terms of texture details and visual effects. Compared with the SRGAN model that has a better visual effect, the PSNR value is improved by an average of 1.84dB, and the SSIM value is enhanced by an average of 0.036, which proves the effectiveness of the proposed algorithm.
Improved Algorithm for GAN Super-Resolution Face Image Reconstruction
2023-10-11
2662054 byte
Conference paper
Electronic Resource
English
British Library Online Contents | 2010
|Single Image Super-resolution Reconstruction Method
British Library Online Contents | 2004
|Morphable model space based face super-resolution reconstruction and recognition
British Library Online Contents | 2012
|