Research On Deep Image Reconstruction Method Based On Semi-Supervised Learning

The depth estimation information is very significant for the autonomous system to perceive the environment and estimate its own state.With the continuous development of artificial intelligence technology,huge research progress has been made in estimating scene depth from images.In recent years,the research on depth estimation of monocular images based on deep neural networks has become a hot spot.They use deep collection equipment for supervised training or use stereo image pairs for unsupervised training to estimate the depth of the scene.However,the supervised learning depth estimation method is time-consuming and expensive due to the collection of depth data,and the feature information of the camera view is also sparse,also the accuracy of unsupervised depth estimation is limited by the accuracy of stereo reconstruction.In order to solve the above problems,a deep convolutional network model based on image reconstruction is proposed to estimate the depth of a single image.On the one hand,the depth feature information extraction capability of the deep convolutional network is used to estimate the depth of the color image,and the real depth label is used as the supervision signal constraint.On the other hand,the basic principle of stereo vision is used to turn the depth estimation problem into image reconstruction,which is treated as an unsupervised signal constraint.This semi-supervised learning method of depth estimation network model not only ensures the accuracy of depth estimation,but also reduces the requirements for the integrity and density of pixels in the real depth map label during network training.Secondly,in order to solve the problem of fuzzy,unsmooth,and unclear contours in the local details of the depth image estimated by the network model,a deep convolutional network model with joint attention mechanism and multi-scale space are proposed,which is based on deep convolutional network model based on image reconstruction.A new attention mechanism is designed to merge with the deep convolutional network.In the process of image feature information extraction,the area of interest is paid attention to,and multi-scale atrous convolution is introduced to obtain more feature information from the image,so that the local details of the estimated depth map are more obvious,which make the final depth estimation network get a clearer depth map.In the experimental stage,the KITTI datasets and indoor and outdoor datasets and intestinal images are selected for network model verification and compared with the existing mainstream depth estimation algorithms.The above-mentioned network depth estimation model has certain advantages in imaging effects and evaluation indicators.