Research On 3D Reconstruction Method Of Monocular Structured Light Based On Deep Learning

Three-dimensional reconstruction is an important research topic in the field of computer vision,and its task is to recover three-dimensional information from two-dimensional images.It has a wide range of applications in the fields of automatic driving,industrial measurement and human-computer interaction.Among them,structured light 3D measurement technology has become an important method in the field of 3D reconstruction due to its advantages of non-contact and high precision.The rapid development of deep learning in the field of computer vision has made vision tasks based on deep learning a hot research topic.This paper is based on monocular structured light measurement technology,combined with the convolutional neural network method in deep learning,to conduct research on surface structured light three-dimensional reconstruction.This article first introduces the phase demodulation,phase unwrapping and three-dimensional reconstruction process in detail based on the monocular structured light measurement technology.A method is proposed to use only a single fringe pattern to predict the wrapped phase through a neural network;to predict the depth map through the neural network,and then to calculate the fringe order from the depth map.Secondly,for the proposed intensive prediction task,a corresponding encoder-decoder network is designed.The encoder is a truncated model based on the Dense Net-169 network,and the decoder is a bilinear up-sampling structure with skip connections.Corresponding loss functions are designed,including pixel value estimation loss,image gradient loss and image structure similarity loss.Then,the fringe pattern-wrapped phase data set and the fringe pattern-depth map data set are made by the traditional structured light method,and the corresponding data sets are trained in the designed encoder-decoder model to obtain the training model.And use experiments to prove the feasibility of predicting the wrapped phase and fringe order through neural network.Finally,on the basis of theoretical analysis,this paper builds a monocular structured light experimental system.The accuracy of structured light 3D reconstruction based on the traditional three-frequency four-steps method and the deep learning method is compared and analyzed,which proves the effectiveness of the method in this paper.