Research On Depth Estimation Algorithms Of Images Based On Deep Learning

Extracting the depth information of image scenes from two-dimensional images is a classic problem in computer vision.Accurate depth information can better understand the three-dimensional structure of the scene and the three-dimensional relationship between objects in the scene.Depth information has important application value in automatic driving,AR,VR,robot navigation and so on.In the depth estimation algorithm,the traditional binocular stereo matching algorithm has the contradiction that accuracy and real-time cannot coexist,and there is also the problem that it can only be applied to specific scenes.In recent years,with the development of artificial neural networks,especially the application of convolutional neural networks on images,the depth of field extraction has developed rapidly.The various supervised neural network algorithms proposed by everyone make it possible to achieve good results in real-time and accuracy,but the problem still exists at this stage: the dataset with depth of field images is seriously insufficient and limited application scenarios Therefore,the use of unsupervised depth of field estimation in deep learning to achieve high standard extraction of image depth of field is the focus of this paper.Based on the above analysis,this article focuses on the research of depth of field estimation algorithm based on deep learning.The main research contents and innovations are as follows:1.Related principles of deep learning framework for image depth estimation.It includes the principle of artificial neural network and the basic principle of convolutional neural network,the basic structure,working principle and optimization process of convolutional neural network,The basic content,working principle and principle of viewpoint augmentation of traditional binocular stereo matching,and different frame design principles of image estimation using deep learning.2.The binocular unsupervised depth estimation algorithm is studied.In view of the fact that most of the currently effective network frameworks with depth map data sets have limited application scenarios and relatively limited data sets,It includes the design principle of unsupervised network,details of frame structure and effect images,as well as application effect images of 3D animation in real campus scenes.The core idea of the unsupervised algorithm is to use the left view and the depth image estimated by the network to reconstruct the right view,and then calculate the loss with the original right view,and then optimize the network by continuously reducing this loss.3.An optimization algorithm for deep images based on deep learning is proposed.In view of the fact that the depth images obtained by the traditional stereo matching algorithm or the deep learning algorithm all have error points at the edges and poor contrast,the semantic segmentation network is used to implement the optimization algorithm for the depth image,Afterwards,different regions of the depth image are processed with different algorithms,so as to solve the problem of edge blur and the problem of wrong points,and improve the contrast of the depth image.At the same time,it also demonstrates the feasibility of adversarial neural networks for depth estimation and optimization and compares the effects of traditional algorithms Guilded Filter and TV(Total Variation)on depth image processing.