基于深度网络分级特征图的图像超分辨率重建

从低分辨率图像中提取特征图恢复高分辨率图像中的高频信息是超分辨率重建的一个关键问题,针对该问题提出一个新的基于卷积神经网络的超分辨率重建算法.网络结构由卷积层与子像素卷积组成,特征提取网络中卷积层提取低分辨率图像的特征,重建网络中子像素卷积神经网络作为上采样算子.针对不能充分利用多级特征图的问题,采用跳跃连接和特征图联结在特征提取网络末端跨通道融合特征图,同时降低特征图的维度.并在此基础上再次提取特征图应用于重建.实验结果表明,算法在PSNR、SSIM和人类视觉效果上与其他基于深度学习的算法相比有着显著的提高.

Image super-resolution based on deep network with hierarchical feature maps

Extracting feature maps from low-resolution images is a vital issue for restoring high-frequency textures in high-resolution images. To solve this problem, we propose a new super-resolution algorithm based on convolutional neural networks. The network structure of this paper consists of the convolutional layer and sub-pixel convolutional layer. The convolutional layer in the feature extraction network extracts feature maps from the low-resolution image, and the sub-pixel convolutional neural network is used as an up-sample operator. To solve the problem that the hierarchical feature maps fail to be fully utilized in reconstruction, this paper uses the skip connection and the concatenation of feature maps to fuse the feature maps across the channel at the end of the feature extraction network, and reduce the dimension of concatenated feature maps. Then it extracts the fused features from the feature maps after dimensionality reduction for reconstruction. The experimental results show that the accuracy of the proposed algorithm has significantly improved the effects of PSNR, SSIM, and human visual perception compared with other algorithms based on deep learning.