Depth Estimation Of Monocular Image Based On Deep Learning

Monocular depth estimation is an indispensable research field in computer vision.In this paper,the depth estimation technology of monocular image is researched deeply.The main purpose is to predict the distance from color image to monocular camera,that is,to estimate the depth information of monocular image.In view of two main problems of machine learning based monocular depth estimation algorithm,which are poor generalization ability and low prediction accuracy,two kinds of monocular depth estimation optimization algorithms based on depth learning are proposed,and then more accurate depth prediction images are obtained in both indoor and outdoor environments.Finally,the algorithm proposed in this paper is applied to three-dimensional reconstruction of indoor scene,and the conversion from two-dimensional image to three-dimensional space is realized successfully.The main research results and innovations are as follows:1.A fully convolution residual networks based on feature enhancement is proposed.Firstly,feature pyramid structure is introduced to reuse multi-scale feature map obtained from the results of down sampling and up sampling.By fusing multi-scale information of image,the prediction image is more accurate and rich.Secondly,the high-frequency information of input image is extracted by high pass filter and fused with up sampling results of fully convolution network,which further reduces prediction error and improves prediction accuracy.In addition,the original dataset of NYU is expanded to generate a large-scale dataset to train and test networks.Experimental result shows that this algorithm improves the quality of depth estimation image significantly.2.A feature enhanced fully convolutional residual networks based on dilated convolution is proposed.This algorithm combines dilated convolution which can keep the size and resolution of feature image unchanged,and effectively expands the receptive field of convolution operation.At the same time,the feature pyramid structure of the network is formed by the way of jump connection to further improve prediction accuracy.Experimental result shows that this algorithm can significantly improve the quality of estimation image in outdoor large-scale scenes.3.The depth estimation of monocular image based on depth learning is applied to three-dimensional scene reconstruction,and the three-dimensional reconstruction system of indoor scene based on depth estimation algorithm of monocular image is constructed,which further makes up for the shortcomings of traditional three-dimensional reconstruction algorithm in process of image texture less and camera in the process of pure rotation.Finally,the indoor scene with higher reliability and restoration degree is reconstructed.