GENERATIVE ADVERSARIAL NETWORK WITH NEW BATCH NORMALIZATION AND FEATURE EXTRACTION BLOCK FOR IMAGE SUPER-RESOLUTION RECONSTRUCTION

To improve the quality of the recovered image by the generative adversarial network, an improved generative adversarial network is proposed. Firstly, it designs a new batch normalization block to avoid gradient explosion and disappearance. The tra-ditional batch normalization will reduce the standard deviation of feature pixels, which causes degradation of reconstructed image quality. To solve the problem, an adaptive standard deviation of feature pixels modulator is designed to amplify the deviation of feature pixels and is introduced to traditional batch normalization to construct new batch normalization. Secondly, to extract more useful features, a new block is designed. The proposed block consists of two branches with different network depths. It fuses the dif-ferent extracted features from the two branches to obtain more useful features. Besides, the proposed batch normalization block also is introduced into the new block. Thirdly, the new block is used to construct a dense network with skip connection characteristics for extracting features. Besides, the new block is also used alone at the end of the feature extraction network to fuse different features. Compared with EnhanceNet, SRGAN, ES-RGAN, R-SRGAN, SAM+VAM on Set5, Set14, BSD200, and Urban100 datasets, our proposed method still has the greatest average PSNR, and SSIM for recovered images from the images downsampled the super-resolution images using a bicubic kernel with a scaling factor of x2, x3 and x4, respectively. The recovered image by our proposed method is closer to the ground-truth image than other methods.