Research And Implementation Of Depth Estimation Technology Based On Single Image

Nowadays,people often obtain images without RGBD information,so the threedimensional information of the scene is in a state of extreme absence.In fact,depth information is so important in the field of computer vision,such as depth estimation can be used in 3d modeling,scene understanding,depth perception and other fields.With the popularization of virtual reality,autonomous driving,3D movie and other application scenes,the demand for 3D information is increasingly strong.However,there are still many problems in monocular depth estimation,such as obtaining the depth ground truth is difficult and expensive and the features of a single image are limited,which dues to the difficulty of training.In this paper,depth estimation with the technology of binocular stereo vision is proposed for the depth estimation of a single image.Firstly,the convolutional neural network is used to extract the features,then the feature map is used to estimate the image of another point of view,and finally the parallax estimation is used for the left and right views to calculate the depth map.The main work of this paper is as follows:Firstly,the paper proposes a depth estimation method of two-stage,the first stage is the view composition part,the second stage is the binocular matching part,which enhances the flexibility of depth estimation method.Secondly,in the paper,the binocular dataset is used for depth estimation instead of the depth dataset,it is used for training the network to get the image of another viewpoint,which reduces the need for depth truth value and greatly reduces the cost of generate the dataset.The view synthesis network part obtains the feature graph based on VGG-16,and estimates the image of another viewpoint based on the feature graph.Thirdly,the network structure(Encoder-Decoder)which is used for segmentation,parallax and optical flow is applied to depth estimation.While making full use of geometric relations of binocular images to calculate depth value,it also makes better use of low-level features to ensure the network's estimation of details.Fourthly,due to the result of depth estimation is based on single image,there was a problem with the consistency in the consecutive frames,in the depth of the optimization phase adopted some features in the filter of depth image display error and smoothness between consecutive frames is optimized,such as the joint of bilateral filtering,oriented filtering,median filtering and other filtering algorithm.Experimental results show that the proposed method is superior to most depth estimation methods based on deep learning and can be applied to 2D video to 3D video projects.