Abstract
This paper develops an image super-resolution algorithm based on residual blended attention generative adversarial network with dual discriminators. In the generator part, on the basis of residual neural network, the proposed algorithm adds blended attention blocks to make the neural network concentrate more on specific channels and regions with abundant high-frequency details to increase feature expression capabilities. The feature maps are subsampled using sub-pixel convolutional layers to obtain final high-resolution images. The discriminator part consists of two discriminators that work in pixel domain and feature domain respectively. Both discriminators are designed as Wasserstein GAN structures to improve training instability and to overcome model collapse scenario. The dual discriminators and generator are trained alternately and direct the generator to generate images with abundant high-frequency details through combat learning. The loss of generator and dual discriminators to the generator are fused to constrain generator's training, further improve the accuracy. Experimental results show that the proposed algorithm is significant better on objective evaluation indicators such as Peak Signal-to-Noise Ratio(PSNR) and Structural Similarity(SSIM) on several public benchmarks such as Set5 and Set14, compared with mainstream CNN-based algorithms and the obtained images are closet to real images with real sharp details, which fully proves the effectiveness and superiority of our proposed algorithm.
Abstract (translated)
URL
https://arxiv.org/abs/1911.03464
