A Study On Depth Super-Resolution Based On Deep Learning

High quality depth map has been widely used in various scenarios,such as 3D reconstruction,autopilot,attitude recognition,robot navigation and so on.However,due to the limitation of physical equipment,most of the depth maps captured by depth cameras have the problem of low resolution,which is much lower than the color images taken by ordinary cameras.The existing technical means are difficult to solve this problem from the perspective of hardware equipment.Therefore,the research on superresolution of various forms of depth maps is a hot topic in recent years.Because most of the images collected by the current equipment are presented in the form of RGB-D(that is,the image pair composed of a high-resolution color image and a low-resolution depth map),they are similar in geometric structure.According to the characteristics of the data,this paper optimizes the current feature fusion methods of deep learning,and proposes two methods.The main work is as follows:1.Firstly,this paper proposes a convolution neural network(arnet)based on adap-tive regression algorithm.Using the relevant ideas of adaptive regression algo-rithm for reference,the features of depth map and color image are fused through convolutional neural network,and the weight value of kernel is not shared to find the appropriate weight for each pixel.Its effect is better than the traditional adaptive regression algorithm.2.In order to improve the efficiency of the network,this paper proposes an effi-cient deep super-resolution neural network(HDSRnet-light)based on attention guiding mechanism.The network has three branches,of which the main branch is a multi-scale hierarchical mechanism structure,and two side branches pro-vide auxiliary information.HDSRnet-light has only 287.9k parameters.It can run at a speed of more than 300 FPS through NVIDIA geforce GTX 1080 ti at 8x super-resolution and realize SOTA.Base on HDSRnet-light,this paper also proposes an optional edge refinement network to recover fine-grained details,and designs a two-stage training strategy.The results show that the network and related training strategy can approach the better solution to a certain extent and improve the network performance.