针对传统深度图超分辨率重建算法需要人工提取特征、计算复杂度较高且不容易得到合适表示特征的问题,提出一种基于卷积神经网络(CNN)的深度图超分辨率重建算法。该算法不需要提前对特定的任务从图像中提取具体的手工特征,而是模拟人类的视觉系统对原始深度图进行层次化的抽象处理以自主地提取特征。该算法直接进行从低分辨率深度图到高分辨率深度图的映射学习。映射由7个卷积层和1个反卷积层联合实现。卷积操作学习丰富的图像特征,而反卷积实现上采样重建高分辨率的深度图。Middlebury RGBD数据集的实验结果表明,该模型得到的峰值信噪比(PSNR)较传统双三次插值算法平均提高了2.7235 dB,均方根误差(RMSE)平均降低了0.098;与经典CNN算法相比,PSNR平均提高了1.5244 dB,RMSE平均降低了0.043。

A super-resolution reconstruction algorithm based on convolutional neural network (CNN) is proposed to solve the problem that the traditional depth map super-resolution reconstruction algorithm needs to extract the feature manually and the computational complexity is higher, and it is not easy to get the proper representation feature. CNN does not need to extract the specific features from the image in advance for the specific task, but the simulated human vision system can extract the feature independently by hierarchical abstraction process on the original depth map. This algorithm can achieve mapping learning directly from the low resolution depth map to high resolution depth map. The mapping is implemented by seven convolution layers and one deconvolution layer. The convolution operation is used to learn the rich image features,and the deconvolution realizes that the upsampling is used to reconstruct the high resolution depth map. The experimental results of the Middlebury RGBD dataset show that the average peak signal-to-noise ratio (PSNR) and root-mean-square error (RMSE) obtained from the model can increase by 2.7235 dB and decrease by 0.098 compared with the traditional bicubic interpolation algorithm, respectively. Compared with the classical image super-resolution reconstruction using deep convolutional neural networks, the performance is also improved with 1.5244 dB of PSNR increment and 0.043 of RMSE decrement.