Implementation of automated 3D defect detection for low signal-to noise features in NDE data

Grandin, Robert
Gray, Joseph
Journal Title
Journal ISSN
Volume Title
Research Projects
Organizational Units
Mechanical Engineering
Organizational Unit
Journal Issue

The need for robust defect detection in NDE applications requires the identification of subtle, low-contrast changes in measurement signals usually in very noisy data. Most algorithms rarely perform at the level of a human inspector and often, as data sets are now routinely 10+ Gigabytes, require laborious manual inspection. We present two automated defect segmentation methods, simple threshold and a binomial hypothesis test, and compare effectiveness of these approaches in noisy data with signal to noise ratios at 1:1. The defect-detection ability of our algorithm will be demonstrated on a 3D CT volume, UT C-scan data, magnetic particle images, and using simulated data generated by XRSIM. The latter is a physics-based forward model useful in demonstrating the effectiveness of data processing approaches in a simulation which includes complex defect geometry and realistic measurement. These large data setsrepresent significant demands on compute resources and easily overwhelm typical PC platforms; however, the emergence of graphics processing units(GPU) processing power provides a means to overcome this bottleneck. Processing large, multi-dimensional datasets requires an optimal GPU implementation which addresses both computational complexity and memory-bandwidth usage.


The following article appeared in AIP Conference Proceedings 1581 (2014): 1840, and may be found at doi: 10.1063/1.4865047.

Automated Image Segmentation, Low Signal-to-Noise, Low Contrast, GPU Computing