Optical Fringe Pattern Processing Using Deep Learning

Interferometry methods such as traditional optical interferometry,electronic speckle pattern interferometry and holographic interferometry have been widely used in various fields due to their advantages of non-contact,high precision and high sensitivity.For example,optical interferometry can be used for physical quantity detection of optical components with high requirements for both processing quality and detection accuracy,and electronic speckle interferometry can achieve nondestructive full-field detection of rough surface workpieces in the micron range.In the process of optical interference fringe demodulation,factors such as fringe noise interference and the applicability of phase extraction algorithm will have an impact on the phase detection accuracy.At present,most of the traditional denoising technologies require high computational cost and experience in parameter adjustment.In addition,the widely used phase-shifting technology requires additional phase-shifters and a more stable detection environment to achieve high precision measurement.Moreover,most phase extraction techniques are difficult to achieve high precision and real-time phase extraction for fringe pattern with few frames and less human operation.Therefore,it is of great significance to explore an efficient fringe pattern processing technology to solve the existing problems.Interferometry produces fringe patterns with physically rigorous theoretical models,and deep learning has excellent generalization and learning capabilities,which provides a new technical support and demodulation idea for the resolution of fringe pattern models.In this paper,the fringe pattern processing technology based on deep learning is proposed,and three key technologies of fringe pattern noise processing,wrapped phase noise processing and phase extraction are studied to achieve high precision measurement,fast processing and efficient measurement performance of automatic processing.The main work and innovation points of this paper are as follows:(1)Research on end-to-end convolutional neural network fringe denoising technologyMost of the fringe pattern denoising techniques developed at present usually have disadvantages such as large amount of calculation,manual adjustment of parameters and prediction of fringe trend in advance,so it is difficult to achieve rapid and automatic processing while restoring high-quality fringe pattern.Based on the prior knowledge of optical measurement,the end-to-end convolutional neural network method is proposed to process the fringe noise to improve the measurement accuracy,and the multilayer convolutional neural network is used to smooth the noisy fringe.The end-to-end fringe denoising analysis is realized.The trained model can estimate noiseless fringes directly from noisy fringes without manually adjusting the parameters as most traditional denoising methods do.Numerical simulation shows that the preprocessing technology can effectively improve the fringe quality and thus improve the phase measurement accuracy.(2)Research on wrapped phase denoising technology based on residual learningThe noise interference in wrapped phase is easy to affect the accuracy of unwrapping and surface shape detection.Traditional denoising methods(such as mean and median filtering)tend to cause fringe blurring,which introduces phase distortion.In this paper,a method of denoising wrapping phase based on residual learning is proposed.The strategy is to denoise the numerator(sine function)and denominator(cosine function)of the arctangent function.After denoising,the numerator and denominator are calculated by the arctangent function to obtain a clean wrapped phase pattern.The proposed method deals with wrapped phases in interferometry,and signal-to-noise ratio(SNR)values up to-4 d B can still be successfully unwrapped by the simple line-scanning unwrapping technique.Meanwhile,this technique is used to remove speckle noise of wrapped phase in digital holographic speckle interferometry to help the structural diagnosis of cultural relics.The experimental results show that the proposed denoising method is helpful to accurately detect defects in complex defect topography,and accelerate the process of defect detection and characterization.(3)Research on virtual time-domain phase-shifting phase extraction technique based on generative adversarial networksThe time-domain phase-shifting technique generally uses piezoelectric ceramic phase shifter(PZT)to record multiple phase-shifting fringe patterns and requires a stable measuring environment.In this paper,the virtual time-domain phase-shifting technique based on generative adversarial network is proposed to estimate multiple phase-shifting fringes from the initial interference fringes,and then phase extraction is performed by the simple traditional algorithm.The numerical simulation and experimental analysis verify that this method can achieve high precision phase extraction.Compared with the measured results of ZYGO interferometer,the average PV value of the surface profile detection error is not more than 0.0101 l,and the average RMS value is not more than 0.0024 l.This technique greatly reduces the application cost because it does not depend on the hardware phase-shifting equipment.The processing accuracy of single fringe pattern is close to that of phase-shifting interferometry,and the dynamic measurement is improved effectively.The research work provides a new fringe pattern processing method for the field of optical interferometry.The correctness of the proposed research work is verified by systematic simulation and experimental analysis,which provides a new key technical basis for the engineering development and application of interferometer and electronic speckle interferometer.