多尺度生成对抗网络的图像超分辨率算法

基于神经网络的图像超分辨率方法往往存在重建图像纹理结构模糊、缺失高频信息的问题。为了解决该问题,在SRGAN的基础上提出一种多尺度并联学习的生成对抗网络结构,其中生成模型由两个不同尺度的残差网络块组成,首先对提取的低分辨率图像通过两个子网络的多尺度特征学习,然后使用融合网络进行残差融合,融合不同尺度高频信息,最终生成高分辨图像。在Set5、Set14、BSD100基准数据集以及SpaceNet卫星图像数据集上的实验结果证明了该算法在恢复低分辨率图像的细节纹理信息具有良好效果。

Image Super-Resolution Algorithm via Multi-Scale Generative Adversarial Networks

The image super-resolution method based on neural network often has the problem of reconstruct the image texture structure and missed high-frequency information. To address the above problems, a multi-scale parallel learning generation adversarial network structure is proposed by SRGAN. The generator consists of two different scale residual network blocks. First, the extracted low-resolution images pass through two sub-networks by Multi-scale feature learning, then use the fusion network for residual fusion, merge high-frequency information of different scales, and finally generate high-resolution images. The experimental results on Set5, Set14, BSD100 benchmark dataset and SpaceNet satellite image dataset proved that the algorithm has good effect in recovering the detailed texture information of low resolution images.