Scene Depth Estimation Under Dense Network With Edge Occlusion Decision

Scene depth estimation of a monocular image,focusing on how to obtain scene depth information from a monocular image.In the computer vision theory laid by Marr,the scene depth estimation of monocular images is an important task of human vision.With the construction of RGBD datasets containing scene depth information and the widespread use of deep convolutional neural networks in computer vision tasks,monocular scene depth estimation is often defined as a continuous regression problem,using convolutional neural networks for end-to-end learning.However,there are still several major challenges in the existing depth estimation model:(1)the estimation of the depth information of the monocular image scene requires pixel-level prediction results,and the loss of image resolution during the sampling process of the convolutional neural network model will The accuracy of scene depth estimation is insufficient.As the number of deep convolutional network layers increases,the network model is severely affected by the gradient degradation phenomenon when training for fixed data sets,resulting in reduced learning ability of the scene depth estimation model;(2)due to deep convolution The problem of intricate boundary location in the image caused by the downsampling process of the network model is caused by the blurring of the scene estimation at the target boundary.In response to these problems,this thesis mainly carried out the following work:For the inaccurate location of the complex boundary of the image in the scene depth estimation model,we constructed a dual subnet model of the shared backbone network to simultaneously detect the orientation of the edge and boundary of the input image,and finally merge the target area in the image.The occlusion relationship between them.The occlusion relationship detected by this model adopts two fusion methods: input fusion and output fusion,respectively,as the a priori feature information of the scene depth restoration model proposed in this thesis.Inspired by the deepening of the deep network model,the gradient degradation problem in network training is inspired by the feature aggregation process in dense neural networks.This thesis proposes a dense neural network model for the up/down sampling process.First,the model uses hierarchical convolution and downsampling strategies to describe the basic structure of different levels of targets in the image.Secondly,the deconvolution and upsampling strategies are used to restore the depth resolution of the scene,avoiding the loss of image resolution by the convolutional neural network.Finally,by analyzing the correspondence between the target edges in the up/down sampling process,cross-layer connections under the same scale sampling constraints are introduced to achieve high-precision scene depth estimation.The NYU-Depth-v2 dataset experiment in the depth estimation of the proposed scene shows that the proposed method can effectively improve the scene depth estimation under the interference conditions such as complex target boundaries,and is better than the-state-of-the-art method of scene depth estimation in depth estimation error and accuracy.