3d Reconstruction Of Monocular Structured Light Based On Deep Learning

With the development and progress of computer image vision technology,the research of 3D reconstruction has become a hot topic in this field.3D reconstruction technology using structured light has become one of the most important research areas because it does not need to contact the object to be measured and has high reconstruction accuracy.Due to the excellent performance of the deep learning method in image processing,this thesis will take structured light 3D reconstruction technology as the research basis,and carry out 3D reconstruction experiment through deep learning convolutional neural network,and realize the restoration of the 3D shape of the measured object from a single fringe pattern.The specific research content is as follows:(1)In this thesis,the 3D reconstruction technology of structured light is studied,and the principle of 3D reconstruction,the process and principle of phase principal value extraction and parcel phase unwrapping are introduced in detail.Meanwhile,a method of predicting depth map by neural network based on single fringe image is proposed to realize 3D reconstruction.(2)Aiming at the proposed regression prediction tasks related to image processing,the corresponding encode-decoder network structure is designed and the corresponding improvement is made to U-Net network.In this thesis,gray-coded structured light method and phase shift method are used to create fringe raster-depth maps,and a neural network is used to train the data set to produce a training model.The feasibility of predicting depth map from fringe pattern by neural network is proved by simulation experiment.(3)The U-Net network applied in the field of biological cell segmentation of medical image is improved,so that it can be applied in 3 D measurement technology.At the same time,the Multi Res HNet network is designed,which uses a series of smaller and lightweight convolution blocks to obtain object information from different scales.The decoder path output information is integrated to improve the accuracy of shape reconstruction.(4)Furthermore,several key problems with gray code coding monocular structured light reconstruction are discussed,such as an investigation of gray-code decoding,coding,and calibration of projectors.As a result of the analysis,the calibrating parameters of the projector are determined using the phase method,and the complementary gray code coding method is used to realize the phase unwrapping.(5)The experimental system of monocular structured light is set up.The experimental results of3 D reconstruction based on gray code coding and 3D reconstruction based on deep learning method are compared and analyzed,and it shows that the method is effective.