Super-Resolution Reconstruction Of Depth Image Based On Convolutional Neural Network

Depth information can reflect the 3-dimensional geometric information of the scene,so it has been widely used in the fields of robot vision,three-dimensional reconstruction,and human body gesture recognition.Although depth camera can obtain the depth information of a scene,the spatial resolution of depth images obtained by depth cameras is generally low due to the limitations of hardware conditions such as sensors.In response to this problem,this dissertation studies the depth image super-resolution reconstruction based on convolutional neural networks.The main work is as follows:(1)We analyze the background,significance and research status of depth image super-resolution reconstruction,and introduce the characteristics,reconstruction model and evaluation criteria of depth image.Then we compare several classical image super-resolution reconstruction algorithms based on convolutional neural network.(2)We study and implement a single depth image super-resolution reconstruction algorithm based on pyramidal two-channel convolutional neural network.At each level of the pyramid,different effective features are extracted from the low-resolution depth image through two channels,then the fused features of different channels are input into the sub-pixel convolutional layer for super-resolution reconstruction.The experimental results show that the algorithm can effectively alleviate the edge blur and artifacts in the reconstruction process,and improve the reconstruction effect.(3)We study and implement a color image multi-scale guided depth image super-resolution reconstruction convolutional neural network.A multi-receptive residual block is constructed to extract low-resolution depth image features,and then uses multi-scale fusion method to realize high-resolution color image features to guide low-resolution depth image features,and obtain global fusion features.Finally,the reconstructed image obtained by the sub-pixel convolutional layer is combined with the global fusion feature to obtain a high-resolution depth image.The experimental results show that the high resolution image obtained by the algorithm has clear edges and good visual effects.