Phase Extraction Method Of Projection Fringe Images Based On Singular Value Decomposition And Neural Network

Single frame fringe projection profilometry is an active 3D measurement technique,which can dynamically and high-precision 3D measurement of objects,and has been widely used in various fields.It is a key step of single frame fringe projection profilometry to extract phase information from the deformed projection fringe image by phase modulating of the surface height of the object with carrier fringe.In general,phase extraction includes phase extraction,phase unwrapping and carrier removal.At present,there are still some challenges in all aspects of phase extraction:(1)the nonlinear carrier removal method is not stable enough,which requires manual intervention in many special cases and is applicable in a small range;(2)The phase unwrapping methods generally have poor robustness and adaptability,and some new phase unwrapping methods based on deep learning also have poor generalization ability.In order to solve the above two problems,this paper proposed non-linear carrier removal method based on principal component analysis(PCA)and multi-layer Perceptron(MLP)and 2D phase unwrapping method based on singular value decomposition(SVD)and U-Net with double branching structure,respectively.(1)This paper presents a new idea for nonlinear carrier removal problem.Different from the existing nonlinear carrier removal methods,which aim at removing the nonlinear carrier component from the extracted phase,this paper proposes to use PCA and MLP to recover the nonlinear carrier fringe image from the deformed fringe image and eliminate the phase generated by the non-linear carrier in the subsequent steps.Compared with the existing methods,the idea of first restoring the carrier fringe image and then removing the corresponding nonlinear carrier phase can obtain additional information from the deformed fringe image,so that the new algorithm has a higher stability.Experimental results show that the root mean squared error(RMSE)of the new algorithm is 0.84 lower than that of the current optimal nonlinear carrier removal method,which is superior to other existing methods.In the link of carrier recovery,the experiment proves that the proposed scheme is superior to U-Net in both speed and accuracy.(2)To solve the problem of phase unwrapping,this paper proposes a phase unwrapping method combining SVD and U-Net(SVD-UNet).Different linear components were obtained by SVD decomposition of the wrapping phase,and then the components decomposed by SVD were input into the U-Net of the double encoder branch to get the number of wraps needed for phase unwrapping.Compared to using the original wrapping phase image,U-Net can more easily separate the noise from the SVD decomposition results and obtain a richer global view,thus achieving improved accuracy and robustness.Sufficient experiments show that the peak signal-to-noise ratio(PSNR)of phase unwrapping results obtained by the proposed method reaches 49.9d B,1.6d B higher than the current best method,and the generalization performance of this method is better than that of the similar deep learning methods.