Blind Face Deblurring via Latent Generative Priors

The latent code of generative adversarial networks contains different scales of detail features in images. To better combine the different scale characteristics of degraded images and the rich information from latent code to perform face image deblurring, we extract the multi-scale features of the blurred face images through a multi-layer convolution network, which is further guided by the latent code of generative adversarial networks to generate the decoded features, which is then further guided by the latent code to generate deblurred images. Through this multi-scale feature-guided scheme rather than single-scale generations, our method ensures the reconstruction and restoration of blurred images in more detail. Compared to the current methods, which also use the latent code of generative adversarial networks, our method achieves better performance, improving 30% on the metric of PSNR and 23% on the metric of SSIM.