通道加权下的双判别GAN超分辨率网络

IMAGE SUPER-RESOLUTION WITH DUAL DISCRIMINANT GAN UNDER CHANNEL WEIGHTING

针对现有基于生成对抗网络的单图超分辨率重构方法特征利用率不足,生成图像包含少量无意义噪声的问题,提出一种基于通道注意力机制的双判别生成对抗网络。通过对生成网络中密集残差块进行通道加权,优化网络的特征利用率。同时在对抗网络中对生成图像进行像素域和特征域的双重判别,促使生成网络产生更丰富的结构特征和高频信息。实验结果表明,与现有的SRGAN、ESRGAN两种算法相比,该算法能够重构出感官质量更高的图像。

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.