Depth Estimation From 2D Image Using Deep Learning

Inferring a three-dimensional scene from an image is one of the core problems of computer vision,and estimating the depth information of the scene is an important method to analyze the three-dimensional geometric relationship of the scene.Traditional methods are mainly based on optical geometric constraints or some environmental assumptions,such as lighting changes,texture features,and restoration structures in motion,etc.These methods based on certain prior information are becoming more mature.Since deep learning can provide relatively accurate prior information,using deep learning to predict deep information from images has also become one of the research hotspots in the field of computer vision.The key of this subject is how to use deep learning methods to obtain feature representations of different depths from a single or multiple images,which can be displayed through certain means.The idea of this paper is inspired by the 3D reconstruction method.The depth of the scene is reconstructed in the way of 3D reconstruction.The training process of deep learning is constructed using the idea of ??stereo matching.Specifically,a single image depth estimation model using a deep convolutional neural network is proposed,which achieves end-to-end single image depth estimation.This model improves the performance of deep convolutional networks on deep estimation tasks by introducing residual structures,densely connected structures,and jump connections into the codec structure.It improves the learning efficiency and performance of the network,and accelerates the convergence rate of the network..Secondly,this paper designs a more effective loss function by combining loss metrics such as gray similarity,parallax smoothing,and left and right parallax matching,which effectively reduces the influence of image lighting factors,curbs discontinuity in image depth,and can guarantee left and right parallax Consistency,which improves the robustness of depth estimation.Furthermore,this article designed a unique training process for this model,using a self-supervised supervision strategy,using stereo image pairs astraining data,and also as training supervision signals,without the need for a large number of labeled true depth data,Saved training costs and improved training quality.Finally,this paper further improves the framework of the monocular image depth estimation through a large number of comparative experiments to optimize the structure of the model on the network structure,supervision strategy,and loss function.In summary,this article proposes a set of effective monocular depth estimation methods,including model architecture,training methods,and testing procedures.It completes the task of estimating scene depth from two-dimensional images,and estimates monocular image depth.The existing problems and future development trends were analyzed and discussed.The method was tested on KITTI,Cityscapes and Eigen datasets,which verified the effectiveness and superiority of the proposed depth estimation method.