Skeleton Extraction Of Electronic Speckle Interferometry Fringe Pattern Based On Deep Learning

Electron speckle interferometry,as one of the most important methods in optical metrology,is a powerful full-field measurement technique.The physical information of the object to be measured can be obtained by solving the phase term in the electron speckle interference fringe pattern.Therefore,how to obtain the phase information accurately is the key to the successful application of ESPI.The fringe skeleton method based on fringe thinning is the most direct method in ESPI phase measurement.This method first obtains the binary image corresponding to the ESPI fringe image through preprocessing,and then uses a simple thinning algorithm to obtain the single-pixel width skeleton.However,the inherent speckle noise and intensity inhomogeneity in ESPI fringe patterns make it very difficult to obtain binary images from ESPI fringe patterns.With the development and application of deep learning,how to realize the automatic processing of interference fringe patterns has gradually become a new research direction for optical metrology researchers.In this thesis,the extraction process of the skeleton of the ESPI fringe pattern is combined with the deep learning method to realize the automatic extraction of the skeleton of the ESPI fringe pattern.In this thesis,several typical ESPI optical systems and their basic principles are introduced,and then a large number of experimental ESPI fringe patterns are collected by the Michelson speckle interferometry system using alloy metal sheets as the object to be measured.At the same time,the simulated ESPI fringe pattterns were obtained by computer simulation,and then the corresponding skeleton images were obtained by the fringe skeleton method,and the training set and test set for deep learning were constructed with these datas.This thesis proposes an ESPI fringe pattern skeleton extraction method based on Pix2 pix c GAN,which can be used to quickly and accurately extract a large number of ESPI fringe skeletons only after completing the training of the network model.Part of the simulated datas are added to the paired experimental ESPI fringe patterns and the corresponding skeleton images to train the network,which further improves the accuracy of the method for skeleton extraction.200 simulated ESPI fringe patterns and200 experimental ESPI fringe patterns were used as test sets to verify the effectiveness of the method for skeleton extraction.The results show that the method in this thesis can obtain the corresponding skeleton images without preprocessing the original ESPI fringe patterns,and it only takes 20 s to extract the skeletons of the 200 experimental ESPI fringe patterns.By comparing and analyzing the method in this thesis with the traditional fringe skeleton method,U-Net method and Cycle GAN method in both qualitative and quantitative perspectives,the superiority of this method in extracting skeleton speed and accuracy is fully proved.In addition,for some broken ESPI fringe patterns collected in the experiment,the corresponding complete and smooth skeletons can also be obtained using this method.Finally,a set of ESPI fringe patterns collected from an aluminum sheet in a heated state are used to verify the generalization ability of the proposed method.