基于金字塔式双通道卷积神经网络的深度图像超分辨率重建

针对深度图像分辨率低的问题，构建了一种金字塔式双通道深度图像超分辨率卷积神经网络。在金字塔的每一级，通过两个通道对低分辨率深度图像提取不同的有效特征，通道1为增强型残差结构，可以将丰富的图像细节传递到后面的图层，通道2将不同卷积层提取的特征连接起来作为此通道最后一层卷积层的输入，有益于局部特征和全局特征的结合。接着，通过将不同通道融合后的特征输入亚像素卷积实现超分辨率重建。实验结果表明，相比其他方法，该方法得到的超分辨率图像缓解了边缘失真和伪影问题，有较好的视觉效果。

Depth image super-resolution reconstruction with two-channel pyramid convolutional neural network

Aiming at the problem of low resolution of depth image, this paper constructed a convolutional neural network with two channel pyramidal structure for image super-resolution reconstruction. In this network, each level of the pyramidal structure extracted the different effective features of low-resolution images through two channel. Channel 1 was an enhanced residual structure that could transfer rich image details to the following layers. Channel 2 was used to connect features extracted by different convolutional layers and emploied it as the input for the last layer of this channel, thus realizing the fusion of local features and global. After that, it used the merged features from different channels as the input of sub-pixel convolutional layer to realize the super-resolution reconstruction. Experimental results show that, compared with other methods, the super-resolution image obtained by this method alleviates the problem of edge distortion and artifacts, and the visual perception is satisfactory.