Research On Camera Array Fringe Projection Phase Unwrapping Method Based On Deep Learnin

Three-dimensional(3D)information is one of the important features of objects.The development of high-precision 3D measurement theory and technology is very essential to promote the rapid development of intelligent manufacturing technology,improve the level of advanced manufacturing in aerospace and other fields,and serve the needs of society.Fringe projection 3D measurement technology has the advantages of simple structure,high precision,fast measurement speed and high universality,and has gradually become a very popular 3D measurement method in recent years.Phase unwrapping is one of the key roles in fringe projection 3D measurement technology.Fast and effective phase unwrapping is still the key research content of fringe projection3 D measurement technology.In this dissertation,from the perspective of multi-view phase unwrapping,a camera array phase unwrapping method based on deep learning is proposed.The specific work is as follows:Firstly,the overall scheme of camera array fringe projection phase unwrapping method is designed,which mainly includes three parts: camera array fringe projection data acquisition,data set production and deep learning multi-view phase unwrapping.The research contents of camera array fringe projection 3D measurement system construction and data set creation are studied.Secondly,based on the research of the sub-aperture wrapped phases feature extraction model of camera array and the U-shaped network feature fusion model,a multi-stream convolutional neural network phase unwrapping algorithm is proposed.The camera array sub-aperture wrapped phases feature extraction model fully extracts the features of the sub-aperture wrapped phases of the camera array and summarizes them as features in the four directions of the camera array,and then outputs them to the feature fusion network.The U-shaped network feature fusion model realizes the fusion processing of the sub-aperture wrapped phases feature information of the camera array and further extracts the features.Through the proposed network model,the mapping relationship between the input multi-view wrapped phases and the output fringe orders of the center view is established,and the prediction from the sub-aperture wrapped phases to the fringe orders is realized,to achieve the phase unwrapping.Thirdly,based on simulated scene data set and real scene data set,experiments are performed on the proposed phase unwrapping algorithm of the multi-stream convolutional neural network.Experimental results show that among the sub-aperture wrapped phases of camera array in the four directions of 0°,90°,45° and 135°,the subaperture wrapped phases in the direction of 0° can provide more useful information for multi-view deep learning phase unwrapping.In addition,when using the wrapped phases in four directions as input data simultaneously,the better performance of the proposed multi-stream convolutional neural network model can be achieved and effective phase unwrapping can be realized.Finally,an improved multi-stream convolutional neural network model Net＿RSR is proposed,by introducing spatial pyramid pooling model and using the side output layer monitoring structure.The experimental results show that the improved model Net＿RSR improves the accuracy of predicting fringe order.On this basis,the improved model is applied to the gesture fringe dataset.The multi-stream convolutional neural network phase unwrapping algorithm proposed in this work can make full use of the constraint information between the subaperture wrapped phases of the light field to accurately predict the fringe orders of the wrapped phase,and achieve effective deep learning based phase unwrapping.This work has a certain reference significance for the research of phase unwrapping method in fringe projection,and has a positive role in promoting the development of fast and highprecision fringe projection 3D measurement technology.