Research On Depth Estimation Algorithm Based On Monocular Texture Map

With the development of artificial intelligence and 3D applications,the acquisition of scene information is becoming more and more important.Depth is one of the most important information in the scene.How to obtain the depth map has always been an important issue in the computer field.In the case where the depth acquisition device cannot be widely used,estimating the depth based on these monocular images is a challenging task for a large number of color maps already existing.For the depth estimation problem of monocular images,the research work starts from the scene classification and conducts research work based on outdoor images and indoor images.Based on the outdoor image,this paper draws on the coarse-to-fine depth estimation process,and designs the structure of the coarse-scale structured random forest combined with the fine-scale structured random forest.The coarse-scale structured forest is used to estimate the global coarse depth information.The predicted results are upsampled and sent to a fine-scale structured forest for local fine depth estimation.Due to the rich scene structure of the indoor image and the depth clue information,this paper proposes a deep learning network with multi-scale feature fusion,which is divided into feature extraction network and feature fusion network.The feature extraction network extracts multi-level global and local features based on the full convolutional neural network and reduces network parameters.The feature fusion network uses the skip structure to gradually fuse the features,combines the shallow local features of the feature extraction network with the high-level global features,and uses the fast deconvolution to sample the feature map to the original image size for monocular depth estimation.In general,based on outdoor images,this paper proposes a coarse-to-fine depth estimation framework based on structured random forests.The method of depth label discretization and the method based on depth block estimation depth are designed.Based on indoor image,this paper proposes a deep learning network with multi-scale feature fusion.Based on the residual network,the fully connected layer of the residual network is modified,which is transformed into a convolutional layer,which reduces the network parameters and increases A deconvolution-based feature fusion network is proposed,and a skip structure is proposed to fuse the two network features,and the fast deconvolution method is used to speed up the training.The method of this paper has achieved good results on both indoor and outdoor data sets.