基于生成对抗网络的图像超分辨算法

Image Super Resolution Algorithm Based on Generative Countermeasure Network

图像超分辨是使低分辨率图像通过端到端训练产生边缘更清晰的高分辨率图像的一种技术,是数字图像处理的一个重要研究方向。该文提出了一种基于生成对抗网络的图像超分辨算法,并对网络结构进行改进。设计的生成器删除了残差块的BN层,增加了特征识别的相关算法,特征提取部分采用两层卷积网络,可以提取更多的图像特征,在低分辨率图像上提取特征,通过卷积计算得到高分辨率图像,可以提升运算结果的准确性。判别器设计采用先分组再整合的思想,将生成图像划分成一定数量的图像块,计算每一部分的判别结果,然后将所有图像块的判别真假组合起来,作为最终的判别结果。经实验验证,设计的网络模型在图像重建效果上有了一定的提高,并节省了一定的运算时间。

Image super-resolution is a technology to make low resolution image produce high-resolution image with clearer edge through end-to-end training,which is an significant study orientation of digital image processing.An image super-resolution algorithm based on generative countermeasure network is proposed,and the network structure is improved.The BN layer of residual block is deleted,and the related algorithm of feature recognition is added.The feature extraction part adopts two-layer convolution network,which can extract more image features,extract features on low resolution image,and get high-resolution image through convolution calculation,which can improve the accuracy of operation results.The design of the discriminator adopts the idea of grouping first and then integrating.The whole image is divided into several small image blocks,and the discriminating results of each image block are calculated respectively.Finally,the true and false discriminating results of all image blocks are combined as the final discriminating results.The experimental results show that the designed network model has a certain improvement in image reconstruction effect,and saves a certain amount of computing time.