Phase Extraction Method Based On Edge-enhanced Self-attention And The Assumption Of Foregrounding-background Separability

Three-dimensional visual imaging technology has important applications in reverse engineering,medical imaging,virtual/augmented reality,metaverse and so on.Single frame fringe projection profilometry,as a non-contact active 3D visual imaging measurement method,has the advantages of convenience,high precision and imaging of dynamic objects.Phase extraction is the most critical step in single frame fringe projection profilometry,which mainly includes wrapping phase extraction,phase unwrapping and nonlinear carrier phase removal.In each part of phase extraction,the existing methods still have the following problems:(1)The nonlinear carrier phase removal method has low precision,low adaptability and universality;(2)The traditional phase unwrapping methods have low accuracy and poor robustness,and the phase unwrapping methods based on deep learning has poor generalization ability.In order to solve the above problems,this paper studies the causes of the problems and the improvement direction respectively,and innovative results were achieved as follows:(1)For the nonlinear carrier phase removal problem,after observation and experiment,an assumption of foregrounding-background separability for fringe projection image was proposed that SVD can be used to separate background fringe components image from object components image.According to the above assumption,a nonlinear carrier phase removal method based on SVD was proposed.The new method can be divided into three steps: In the first step,SVD is used to detect the background region from the fringe projection image.In the second step,only the phase of the background region is fitted by polynomial surface to obtain more accurate nonlinear carrier phase.In the third step,the phase of the measured object is obtained by subtracting the fitted nonlinear carrier phase from the phase extracted from the fringe projection image.Experimental results on simulated fringe images and actual fringe projection images showed that the new method is more accurate than the existing methods and the correctness of the proposed assumption was verified.The results of comparative experiments on noise immunity and other factors affecting performance showed that the new method is more robust and versatile than existing methods.(2)For the two-dimensional phase unwrapping problem,a measure of the amount of a priori information contained in the wrapped phase is first proposed to quantify the ease of phase unwrapping.Then a new 2D phase unwrapping method based on edge-enhanced self-attention network(EESANet)was proposed.Among them,a serried residual block is first designed instead of convolutional layer to enhance the learning ability of the network;The atrous spatial pyramid pooling(ASPP)and positional self-attention(PSA)were added to solve the problem of distance dependence in phase unwrapping.An edge-enhanced block is proposed to make more effective use of wrapped edge features.In addition,weighted cross-entropy loss is designed and used to overcome the common class imbalance in phase data.Several comparative experiments show that the new method has higher accuracy and robustness compared with existing methods.Finally,the ablation experiment proves that the addition of PSA and ASPP can improve the performance of new method.