基于GAN的图像超分辨率重建算法研究

SRGAN是一种基于生成对抗网络的超分辨重建方法，其生成的高分辨率图像质量较传统方法有着明显提升，然而SRGAN存在着训练过程不稳定，图像浅层特征未充分使用等问题，很大程度上影响到了生成图像的质量。本文提出了一种特征增强改进的SRGAN模型，该模型使用信息蒸馏块进行特征纹理信息的增强，并消除图像特征中的冗余信息。此外，使用相对平均鉴别器替代原始SRGAN中的二分类鉴别器，保证了GAN网络训练的稳定性。本文基于4倍放大因子的超分辨重建任务，在BSD100数据集上进行实验结果的质化评价和量化评价。实验表明，本文方法较之SRGAN在训练过程中具有更好的稳定性，生成的图像具有更清晰的细节纹理，取得了更佳的图像超分辨率重建效果。

Research on single image super-resolution reconstruction based on GAN

SRGAN is a super-resolution reconstruction method based on generative adversarial network. The quality of the generated high-resolution images is significantly improved compared with traditional methods. However, SRGAN has problems such as unstable training process and insufficient use of image shallow features. To a certain extent, it affects the quality of the generated image. This paper proposes an improved SRGAN model for feature enhancement, which uses information distillation blocks to enhance feature texture information and eliminate redundant information in image features. In addition, the relative average discriminator is used to replace the two-classification discriminator in the original SRGAN to ensure the stability of GAN network training. In this paper, based on the super-resolution reconstruction task of 4 times magnification factor, the qualitative evaluation and quantitative evaluation of experimental results are carried out on the BSD100 data set. Experiments show that the method in this paper has better stability during the training process than SRGAN, the generated images have clearer detailed textures, and better image super-resolution reconstruction effects have been achieved.