The Research Of Image Depth Estimation Based On Convolutional Neural Network

The rapid dvelopment of computer vision technology has greatly improved the in(elligence level of robots(such as unmanned robots,home-service robots,post-disaster search and rescue robots,etc.),and their awarcness of the environment.The effective and accurate acquisition of scene depth information based an a single image is the most important condition for constructing a three-di mensional environment model.For this reason.our paper studies the depth information estimation problem of single imaege based on convolutional neural network,including indoor single image depth estimation and outdoor single image depth estimation..For the problem of indoor single-image depth estimation,a convolutional neural network(CNN)combined with conditional randtom fields(CRF)is used to establish a model for image depth estimation.First.the SLIC suserpixel segmentation method is used to obtain the superpixel image corresponding to the input image.Then,the CNN is used to extract image features,and on this basis,the nearest neighbor interpolation method is used in combination with the average pooled superpixel pooling method,to get the characteristies of each superpixel block,convert image depth estimation to depth estimation of superpixel blocks.Since the depth difrferenve corresponding to the pixel point in each super pixel block region is negligible,and the depth of the image super pixel block is estimated,reducing the operation memory can effectively improve the opcration speed,therefore,this method speeds up the depth estimation without affecting the depth estimation result,in addition.the algorithm operates direotly on superpixel blocks without limiting the size of the input images.Based on the above operations,we use CRF to optirnize the entire network.Finally,the validity of the proposed algorithm was evaluated objectively and subjectively by using the depth estimation indoor standard data setsNYU V2,The evaluation results show that this method can effectively and accurately estimare the depth infonnation of an image through a single image.Fur the problem of outdoor single-image depth estimation,our paper continues to use the CNN combined with the CRF image depth estimation model.Since training CNN requires a huge data sets,The selected outdoor image depth estimation standard data sets are limited in our paper.So we use transfer learning to overcome this difficult problem.First,we transfer the parameters of the image feature extraction part of the indoor image depth estimation,to the feature extraction part of the outdoor image depth estimation,fix the part of parameters,and use the depth estimation outdoor standard data sets Make 3D to train the remaining network layer parameters of the network,Then,using the CRF to optimize the entire network,Finally,the validity of the proposed algorithm was assessed objectively and subjeetively by using the depth estimation outdoor standard data sets Make 3D.The evaluation results show that the transfer learning algorithm cart effectively improve the problem that it is difficult to train the network because of limited data sets,and can accurately estimale the depth information of outdoor images through a single image.