Research And Applictions Of Phase Extraction New Methods For Electronic Speckle Pattern Interferometry Based On Machine Learning

Electronic Speckle Pattern Interferometry(ESPI)is a modern non-destructive measurement technology.It can provide whole-field measurement with high precision in a non-contact mode.In ESPI technology,the physical quantities to be measured are related with the phase distribution of the fringe patterns or wrapped phase patterns.Therefore,the key of the application of ESPI technology is the accurate extraction of phase terms from fringe patterns or wrapped phase patterns.The fringe skeleton interpolation method can extract the phase information based on single ESPI fringe pattern,which can be applied to dynamic measurement.The development of the technologies involved in this method,such as fringe pattern filtering,skeleton line extraction and interpolation,is closely related to the development of computer technology.Nowadays,machine learning algorithm provides new research methods for various research fields.In this paper,the filtering method,skeletons extraction method and interpolation method for the variable density fringe patterns,discontinuity fringe patterns or the fringe patterns with uneven intensity are studied based on machine learning algorithms.We proposed new methods.The primary works are as follows:(1)The denoising method based on fuzzy c-means(FCM)and texture features for variable density fringe patterns is proposed.The denoising methods based on variational image decomposition of variable density ESPI fringe pattern are reviewed,firstly.This method can effectively denoise the variable-density fringe patterns by processing the low density fringes and high density fringes individually.Based on this point,the denoising method based on FCM and texture features for variable density fringe patterns is proposed.In this method,firstly,the texture features of fringes are extracted with the grey level co-occurrence matrix method,and the variable density fringes are clustered into low density fringes and high density fringes according to their texture features by FCM algorithm.Then,the effective existing methods are adopted to denoise the low or high density fringes individually.Finally,the filtered low and high density fringes are combined to obtain the final denoising results.The proposed method provide a new decomposition approach for variable density fringe pattern.Based on this method,the performence of exsiting method is improved for filtering the variable density fringe pattern.(2)The denoising method based on oriented bilateral method and FCM for discontinuous fringe patterns is proposed.In this method,the orientation of fringe is calculated firstly.Then,applying the FCM algorithm to the orientation of fringe.The clustering results are used to control the size of directional mask.In the discontinuous position of the fringe,the size of directional mask is reduced in order to decrease the strength of filtering.The proposed method expands the application range of bilateral filter,and can protect the discontinuous position of fringe pattern when filtering.(3)The skeletons extraction method based on modified FCM algorithm for the fringe patterns with intensity inhomogeneities is proposed.In this method,the standard FCM algorithm is modified by introducing a term that allows the labeling of a pixel to be influenced by the labels in its immediate neighborhood.It can simultaneously compensate for the inhomogeneities while binarizing the fringe patterns.Based on the binarization results,the skeletons can be obtained with common thinning method.Comparing with the existing binarization based skeletons methods,the proposed method can extract the skeletons of the fringe patterns with intensity inhomogeneities without calculating thresholds,and there is no need to perform the filtering preprocessing when the noise level of fringe pattern is low.(4)The skeletons extraction method based on local entropy and FCM for ESPI fringe patterns is proposed.In this method,the local entropy of the fringe pattern is calculated first.Then the fringe pattern is binarized according to its local entropy with FCM logarithm.Based on the binarization results,the skeletons can be obtained with common thinning method.Different from the traditional method,the fringe pattern is binarized according to its local entropy map instead of the original intensity image.The desired skeletons can be obtained without filtering preprocessing.(5)The skeletons interpolation method based on Layer recurrent neural network is proposed.Compared with the existing methods,this method uses more advanced network structure and training algorithm.The interpolation results of our method are better than the traditional interpolation methods and BP neural network.The proposed fringe pattern denoising method,skeletons extraction method and skeletons interpolation method in this paper are applied to the dynamic measurement in order to verify the performance of our methods.