Image Pre-Filtering for Measurement Error Reduction in Digital Image Correlation
Yihao Zhou,Chen Sun,Yuntao Song,Jubing Chen
DOI: https://doi.org/10.1016/j.optlaseng.2014.04.018
IF: 5.666
2015-01-01
Optics and Lasers in Engineering
Abstract:In digital image correlation, the sub-pixel intensity interpolation causes a systematic error in the measured displacements. The error increases toward high-frequency component of the speckle pattern. In practice, a captured image is usually corrupted by additive white noise. The noise introduces additional energy in the high frequencies and therefore raises the systematic error. Meanwhile, the noise also elevates the random error which increases with the noise power. In order to reduce the systematic error and the random error of the measurements, we apply a pre-filtering to the images prior to the correlation so that the high-frequency contents are suppressed. Two spatial-domain filters (binomial and Gaussian) and two frequency-domain filters (Butterworth and Wiener) are tested on speckle images undergoing both simulated and real-world translations. By evaluating the errors of the various combinations of speckle patterns, interpolators, noise levels, and filter configurations, we come to the following conclusions. All the four filters are able to reduce the systematic error. Meanwhile, the random error can also be reduced if the signal power is mainly distributed around DC. For high-frequency speckle patterns, the low-pass filters (binomial, Gaussian and Butterworth) slightly increase the random error and Butterworth filter produces the lowest random error among them. By using Wiener filter with over-estimated noise power, the random error can be reduced but the resultant systematic error is higher than that of low-pass filters. In general, Butterworth filter is recommended for error reduction due to its flexibility of passband selection and maximal preservation of the allowed frequencies. Binomial filter enables efficient implementation and thus becomes a good option if computational cost is a critical issue. While used together with pre-filtering, B-spline interpolator produces lower systematic error than bicubic interpolator and similar level of the random error. Cubic B-spline interpolator can achieve comparable efficiency as bicubic interpolator, while quintic B-spline interpolator requires about 1.5 times the running time.