Comparison of different filtering strategies to reduce noise in strain measurement with digital image correlation

Abstract

The main limitation of digital image correlation is the remarkable noise affecting the digital image correlation–computed strain distributions. Neither manufacturers of digital image correlation systems nor the literature provide guidelines for optimal filtering of digital image correlation strain distributions. However, filtering is also associated with loss of information (smoothing of the strain gradients). We systematically explored different filtering strategies to reduce noise while minimizing the loss of information in the digital image correlation–computed strain distributions. The first filtering strategy was directly applied to the acquired images that were then fed to the digital image correlation software. Median adaptive low-pass filters and notch filters were used to eliminate noise: both strategies increased (rather than reducing) the noise in the digital image correlation–computed strain distributions. The second strategy explored was a Gaussian low-pass filtering of the strain distributions. When the optimal cutoff frequency was selected, the noise was remarkably reduced (by 70%) without excessive loss of information. At the same time, when non-optimal cutoff frequencies were used, the residual noise and/or loss of information seriously compromised the results. Finally, image combination techniques were applied both to the input images and to the strain distributions. This strategy was extremely time-consuming but not very effective (noise reduction <10%). In conclusion, the only truly effective noise reduction strategy, if measurements are carried out using commercial closed software, consists in filtering the strain distribution.