Efficient signal, image and video processing approaches for infrared non-destructive testing and evaluation

Abstract

Non-destructive testing & evaluation (NDT&E) in conjunction with efficient post-processing approaches play a vital role in computer aided industrial quality control. NDT&E techniques are capable to detect different types of defects in variety of materials used in various industry sectors without affecting the serviceability of the material under inspection. Among various NDT&E modalities, active thermal NDT&E gained its importance due to its inherent merits such as remote, whole-field, fast and quantitative inspection capabilities. For efficient and reliable functioning of a computer aided industrial quality control system, sub-surface defects need to be detected with high accuracy, reliability, and contrast. Also, the procedure needs to be efficient in computation and memory usage considering the demand for bulk inspections. Of various active thermal NDT&E schemes, pulse compression favourable frequency modulated thermal wave imaging (FMTWI) has been adopted in this work due to its enhanced defect detection sensitivity along with improved test resolution. The challenges in popular post-processing approach principal component thermography (PCT) has been addressed to make it more suitable for computer aided defect detection procedures. Investigations by experimental pulsed thermal inspection of mild steel material with sub-surface flat bottom hole defects have provided physical insights into selection of significant principal components. Dynamic range of the principal component that contains the defect features is a more important factor, in addition to maximum variance, in providing the improved contrast required to identify any subsurface defects, depending on their lateral dimensions, depths and thermo-physical properties. Matched filtering based pulse compression favourable FMTWI has been adopted for its efficient data compression and noise rejection capabilities. The defect detection abilities of PCT have been improved by processing only main lobe of the pulse compressed FMTWI which has resulted in higher signal to noise ratio (SNR) of defects along with reduction in the size of thermal data to be processed, hence, making the whole processing efficient in terms of memory usage and computation. Further, sparse PCT has been proposed and implemented in two different ways by inducing sparsity in empirical orthogonal functions (EOFs) and principal components (PCs) to enable better interpretation of the decomposed components of PCT. The approach of inducing sparsity in EOFs has resulted in tremendous improvement in the defect contrast, however, the challenge of selection of sparsity parameter has been addressed partially by proposing another approach of inducing sparsity in PCs with easier selection of sparsity parameter. Moreover, processing only main lobe of pulse compressed FMTWI has resulted in enhanced contrast of defects without any loss in defect information even with higher sparsity levels. Lastly, fundamental characteristics of PCT have been highlighted and evaluated in terms of their contribution in obtaining enhanced defect contrast in one of the principal components. Certain properties such as non-correlation of encoded data, weight orthogonality, and weights with unit norm length, are highlighted in PCT which are found to be non-existent in linear autoencoder which is claimed to be approximation of principal component analysis (PCA). These findings are contradictory to few recent research works which reported superior defect contrast in autoencoder based thermography analysis as compared to PCT. Furthermore, a novel constrained and regularized autoencoder based thermography approach for sub-surface defect detection in a mild steel specimen has been proposed and compared with PCT for evaluation of defect detection abilities. Novel constraints are formulated for autoencoder cost function to incorporate these significant properties. Constrained autoencoder based approach has been able to provide improved defect SNR as compared to PCT in case when two properties corresponding to non-correlation in encoded data, and encoder and decoder weight orthogonality are incorporated in the linear autoencoder. Also, property corresponding to non-correlation of encoded data is found to be the most significant in achieving better defect detection followed by properties corresponding to weight orthogonality and weights with unit norm length. The findings of the present study have provided solutions in terms of better selection of principal components, improved defect contrast, better interpretation of principal components, and evaluation of characteristic features of PCT in terms of their relative contribution in obtaining enhanced defect contrast so that appropriate properties can be exploited for better results. Also, pulse compressed FMTWI has made the proposed approaches efficient in terms of computation and memory usage. All these contributions pave way for an effective and reliable computer aided industrial quality control mechanism.