High Quality Restoration And Reconstruction Of Depth Image For Indoor Scene

The 3D perception and reconstruction of indoor scene is a hot research topic in computer graphics and computer vision,and the acquisition of high-quality depth image is the key to conduct further research on it.At present,there are two ways to obtain the depth image of indoor scenes: active and passive.Active means to obtain the depth of the object through active scanning of consumer RGB-D camera,but the obtained depth image can have missing regions in large area when the surface is transparent,highly reflective or beyond the distance of the sensor.Passive method obtains the depth image via the relationship of ‘RGB image-featuredepth' established by some depth estimation algorithms.However,the existing depth estimation algorithms do not take the geometric information which is contained in RGB images into consideration,and usually generate low quality depth image.In addition,due to the limitation of the hardware,the resolution of depth image obtained by active or passive estimation is not high.This will have a negative impact on the downstream applications.To tackle these problems,this paper takes the research on high quality repair and reconstruction of depth image of indoor scene,which mainly includes the topics as follows:(1)In order to address the problem of lacking the ground truth depth value of the missing area in RGB-D dataset of current indoor scene,which makes it difficult to conduct supervision method for depth image completion,a strategy based on random mask and 'real composite' data joint training is proposed.Thus,a large number of large-scale RGB-D datasets can be applied to the task of depth image completion.Moreover,a multi-scale depth image completion model is also developed in cooperation with corresponding RGB image features base on this strategy;(2)A depth estimation model based on normal vector and edge feature information is proposed.This model improves the accuracy of depth estimation by extracting normal vector and edge features from RGB images,and the original features in different scales and RGB images;(3)Two depth image super-resolution models are proposed: one is based on adaptive weight convolution without matching RGB images,and the other uses the guidance of highresolution RGB image.Experimental results show that the proposed methods achieve both greater improvements in accuracy and robustness compared with other methods.