Investigation of temporally varying fringe pattern defects using machine learning for optical metrology

Abstract

Defect identification for quality control and industrial inspection necessitates the need for novel techniques in the field of non-destructive testing and experimental mechanics. Optical interferometric techniques are quite popular for non-destructive testing, and hence they are extensively used to locate and identify defects from fringe patterns. A rapid variation of the fringe density in the vicinity of the defect’s region serves as a means for the detection algorithms to identify them. With the advent of machine learning over recent years, it has paved way for algorithms with automatic detection, thereby, eliminating the problem of manual thresholding. In this paper, we propose an elegant technique which relies on computing windowed Fourier spectrum of the fringe pattern at a given spatial frequency and subsequently utilizing this spectrum with a GPU accelerated Support Vector Machine (SVM) algorithm for classification of a defect and a non-defect region. The windowed Fourier spectrum of the fringe pattern serves as a feature vector for the GPU accelerated SVM algorithm which internally performs a pixel classification, thereby producing a binary output of the defect. The performance of the proposed technique is tested on computer-generated fringe patterns at severe noise levels. A machine learning model is trained using the windowed Fourier spectrum of a 1024 x 1024 fringe pattern which is corrupted with an additive Gaussian noise at a signal to noise ratio of 5dB. The best set of hyper-parameters are deduced from the validation data and the proposed method is tested on the fringe patterns of size 1024 x 1024 which contain temporally varying defects. The results indicate that the proposed method is computationally efficient, robust against noise, and also capable of automating the defect identification problem.