Research On Monocular Scene Depth Estimation Method Based On Unsupervised Convolutional Neural Network

3D reconstruction plays an important role in medical image processing,virtual reality,computer graphics,computer vision,digital media creation and other fields,and depth map is an important part of 3D reconstruction,which determines the accuracy and clarity of 3d reconstruction,Therefore,this paper mainly studies the depth of the scene.However,Compared with the binocular image depth estimation method,the monocular image depth estimation method has lower requirements for camera construction and more convenient application.The difficulty of the monocular image depth estimation method is that it is difficult to capture enough scene structure features from the monocular image.Related studies have proved that convolutional neural network are very good at processing image-like information,which is mainly due to the fact that convolutional neural network can learn abundant feature expressions from images.Therefore,this paper combines convolutional neural network to estimate the depth of the monocular image.Since the residual network can well solve the problem that the gradient disappears as the network deepens,and the unsupervised method can solve the great difficulty caused by manual data annotation,this paper uses the unsupervised method combined with the residual convolutional neural network to give the scene depth estimation model of the unsupervised residual convolutional neural network based on the first view,referred to as URM model.The URM model includes a monocular depth estimation residual convolutional neural network model,referred to as Depth CNN model and a pose residual convolutional neural network model,referred to as Pose CNN model.Depth CNN mainly obtained depth value estimation for each pixel of the image,and Pose CNN mainly obtained cameras pose transformation value of two consecutive images.Next,the unsupervised signal of the URM model is established by the relationship between the two models.Then the unsupervised loss function of the URM model is constructed by the unsupervised signal.The weight of the URM model is updated by the stochastic gradient descent method,and the model parameters are obtained after convergence.The experimental results of the URM model show that there is no clear object contour in the depth map,and the object is very fuzzy.Considering the respective advantages of the deep convolutional neural network and the graph model,the deep convolutional network can be used to make a strong feature expression,And the conditional random field can establish the characteristics of local and global relations,that is constructing the relationship between the original image and depth map and the depth at different positions,Finally,A scene depth estimation model based on unsupervised random field residual convolutional neural network is proposed,referred to as UCRFRM model,Experimental results show that the proposed UCRFRM model can achieve the accuracy output of depth map and smooth the depth map.Finally,the URM model and the UCRFRM model were trained and tested on the KITTI dataset,the Make3D dataset,and the CITYSCAPES dataset,respectively.The results of the two models were compared,and meanwhile compared with other methods.The results showed the superiority of the UCRFM model,At the same time,the accuracy and reliability of the depth value estimated by UCRFRM model in real scenes are verified.