Study On Spectrum Extraction Of Moiré CT Image Based On Convolutional Neural Network

The measurement and visualization of complex flow fields is a key technology in modern aerospace and energy engineering.Optical computerized tomography can obtain the three-dimensional information distribution of flow field parameters,which is one of the important solutions to realize the three-dimensional quantitative measurement of complex flow field.Among the techniques,moiré tomography has the advantages of simple optical path,strong anti-disturbing capacity and wide dynamic range of measurement,which is especially suitable for wind tunnel flow field measurement under strong vibration environment.In this thesis,the image processing of moiré CT projection is studied.The extraction method of first-order spectrum of Fourier transform of moiré fringe is discussed and studied in detail.The work of this thesis includes the following aspects:Firstly,an automatic first-order spectrum extraction algorithm is studied for the extraction of the first-order spectrum in the moiré fringe spectrum diagram.Compared with the manual spectrum extraction method and the filter filtering method,it is only necessary to determine the threshold of the binarization of the spectrum,and it is not necessary to select the spectrum region manually.The automatic spectrum extraction method is more suitable for processing fringes with complex contours.Secondly,in the automatic spectrum extraction method,the threshold of the binarization of the spectrum determines the accuracy of spectrum extraction,so it is necessary to find an optimal threshold to improve the accuracy of spectrum extraction.To solve this problem,a spectrum extraction method based on convolution neural network is proposed.This method replaces the process of automatic spectrum extraction with a convolutional neural network.The process of finding optimal parameters by convolution neural network is equivalent to the process of finding the optimal threshold by the automatic spectrum extraction method.The convolution neural network method can automatically detect the coordinates of contour points of the first-order spectrum in the Fourier transform spectrum diagrams of deformed fringes,which completely avoids the artificial influence and improves the accuracy and repeatability of the location of first-order spectrum.Based on the above studies,the experiment of moiré CT projection processing was carried out.The phase was extracted by automatic method and convolutional neural network method respectively.In order to verify the correctness of the phase extracted by the convolution neural network method,the phase obtained by the convolution neural network method is taken as the measured value,and the phase obtained by the automatic method is used as the real value to calculate the mean square error.The mean square error is 0.20.The convolution neural network method basically realizes the automatic recognition of contour points of the first-order spectrum and phase extraction.The studies in this thesis provide a solution which can process the fringe phase automatically for the projection fringe image processing of moiré CT in practical engineering applications.