Research On Depth Information Reconstruction Of Single Image

Depth perception from the two-dimensional image is an important research direction in the field of computer vision,which can be widely used in intelligent driving related tasks,3D reconstruction,image segmentation,robot vision and so on.Where the reconstructed depth information from a single image more challenging.With the rapid development of deep learning technology in recent years,many deep network algorithms have emerged to reconstruct depth maps from a single two-dimensional image,but there still are a lot of problems such as low prediction accuracy,loss of some depth information and blurred edges.Therefore,this paper studies the method of deep learning network to solve the problem of depth information reconstruction of a single image.The main work is as follows:?1?We proposed a multi-scale fusion constructs encoder and decoder convolutional neural network.This neural network is based on the encoding and decoding structure,and we proposes a multi-scale fusion structure at the encoder side,which fuses low-dimensional feature maps with high-dimensional feature maps to maintains their position and spatial relationships,so that the encoder's encoding ability is improved;On the decoder side,another multi-scale fusion structure is proposed,the high-dimensional feature map is fused with the low-dimensional feature map to increase the network's detail retention ability,so as to improve the decoder's decoding ability and achieve the best network performance.?2?The effectiveness of the algorithm in this paper is verified on general indoor and outdoor data sets,and the effect of various improvements on network performance is analyzed through ablation experiment.Compared with the 2017 CVPR article "Multiscale + Conditional Random Field" on the universal indoor dataset NYUD v2,the ?₁ evaluation index of experimental results in this paper has increased by 1.849%,rmse?root mean squared error?index has reduced by 14.1%.Compared with the 2018 CVPR article "Monocular Video + Attitude Estimation" on the universal outdoor data set KITTI,the ?₁ evaluation index of experimental results in this paper has increased by 6.6% and the rmse index has reduced by 4.3%.?3?We proposed a depth image quality evaluation method which is called SSIMGC?SSIM-Gradient correlation?to evaluate the sharpness of the image,so that,it provides a more comprehensive assessment of the quality of the depth image.The traditional mean relative error,logarithmic mean error,root mean square error and other parameters can represent the absolute difference between the predicted depth map and the real map,but it is not comprehensive,and the prediction image with good parameters often looks more fuzzy.Therefore,on the basis of SSIM?Structure Similarity?,gradient correlation is added to obtain an image evaluation algorithm that conforms to human vision.In summary,the multi-scale fusion constructs encoder and decoder convolutional neural network proposed in this paper has better performance for the depth estimation task of a single image.The image quality evaluation parameter SSIM-GC can more fully characterize the quality of the depth map.