Research on Blind Super-resolution Reconstruction with Double Discriminator
Image super-resolution reconstruction methods have very important uses in public safety detection,satellite imaging,medicine and photo restoration.In this paper,super-resolution reconstruction methods based on generative adversarial networks are investigated,from the training Real-world blind Enhanced Super-Resolution Generative Adversarial Networks pure synthetic data(Real-ESRGAN)method,a double UNet3+ discriminators Real-ESRGAN(DU3-Real-ESRGAN)method is proposed.Firstly,a UNet3+ structure is introduced in the discriminator to capture fine-grained details and coarse-grained semantics from the full scale.Secondly,a dual discriminator structure is adopted,with one discriminator learning image texture details and the other focusing on image edges to achieve complementary image information.Compared with several methods based on generative adversarial networks on Set5,Set14,BSD100 and Urban100 data sets,except for Set5,the Peak Signal to Noise Ration(PSNR),Structure SIMilarity(SSIM)and Natural Image Quality Evaluator(NIQE)values of the DU3-Real-ESRGAN method are superior to those of other methods to produce more intuitive and realistic high-resolution images.
万方数据
维普
