Research Of Deep Convolutional Neural Network And Its Application In Image Distance Measurement

With the continuous improvement of storage capacity and computing power,the massive accumulation of image data and the rapid development of deep learning technology have made the performance of computer vision tasks improved a lot.Image pixel distance estimation,which is a basic task in the field of computer vision,has import research value in many real applications,such as image scene understanding,semantic analysis,3D reconstruction,SLAM,driverless techniques,robot navigation and so on.The deep convolutional neural network(DCNN)is a kind of feedforward neural network which shows unique advantages in the processing of large-scale image data.This paper utilizes the remarkable feature extraction ability of DCNN and the end-to-end structure of fully convolutional neural network(FCN)to design a regression model for image depth estimation.Then we integrate the advantages of recurrent neutral network(RNN)in the analysis of sequence data to get a more accurate depth image by processing the details of the features got by DCNN.The main works of this paper are as follows:First of all,a monocular image depth estimation model based on deep fully convolutional structure is proposed.We analyze the principle and the main structure of DCNN,and study the fully convolutional network structure based on DCNN.Aiming at the image depth estimation task,an image depth estimation framework based on U-net network structure is designed.U-net network structure can realize the features fusion of decoding and encoding parts.On this basis,we predict the depth of current size reconstitution regression image for each deconvolution features of FCN decoding part,and fuse the results of multiple predictions to the final depth image prediction output to rectify structure of overall reconstruction image.The model designed in this paper shows excellent effects for depth estimation on the NYU Depth V2 dataset.Then,the image depth estimation model combined with convolutional neural network and recurrent neural network is proposed.We design an image region feature extraction block based on RNN.On this foundation,two RNN feature extraction units are used to extract structure features and fuse the structure characteristics from the longitudinal and horizontal direction at the same time,a bidirectional parallel RNN image feature extraction model based on RNN is presented.Experiments of images classification tasks show that RNN can depict the detail of image more accurately.Next,the model is fused into the mapping between encoding and decoding based on residual U-net network image depth estimation model to increase the feature structure information.The paper combines the advantages of DCNN and RNN in image processing respectively to construct the final perfect frame of image depth estimation.The double advantages of the fusion network in global and detail reconstruction of depth image are proved by experiments.Finally,a prototype system of video image distance measurement based on depth estimation is proposed.The static image depth estimation is transferred to real-time depth estimation of video images.We use the fusion model designed in this paper to demonstrate distance estimation of static images and video images in real scenes,and a complete demonstration and summary is carried out through the demonstration system.