IMAGE SUPER-RESOLUTION WITH DUAL DISCRIMINANT GAN UNDER CHANNEL WEIGHTING

Existing single image super-resolution methods based on the generative adversarial network cannot make full use of features, and the generated image contains a small amount of meaningless noise. Therefore, this paper proposes a dual discriminant generative adversarial network based on channel attention mechanism. In the generation network, channel attention mechanism was used in the dense residual blocks to improve feature utilization rate. Simultaneously by dual discrimination of pixels and features on the generated image, richer structural features and high frequency information was promoted to produce. Experimental results show that compared with the existing SRGAN and ESRGAN algorithms, the proposed algorithm achieves lower NIQE and PI values and can reconstruct images with better perceptual quality.