Deep Learning Workflows for High-Precision Dimensional Metrology Using X-ray Computed Tomography

Abstract

Industrial X-ray CT has evolved into a popular solution in modern quality control applications,

such as in geometric dimensioning and tolerancing of metallic components. In a typical CT

metrology workflow, a surface is first determined from a CT reconstruction. However, the total

exposure time can be dramatically long to achieve adequate signal-to-noise (SNR) ratio to

accurately determine the surface for high-precision dimensional metrology. Hence, the

applicability of this tool to in-line (or online) metrology can be challenging for mm-sized metallic

components (e.g., automotive fuel injectors) demanding high (< 5 μm) surface resolution due to

low photons/pixel ratio.

Here, we evaluate two DL-based software packages – CTSegNet and DeepRecon for reducing

scanning time for measurements of the internal orifices (diameter of ~175 𝝁𝒎) of a steel fuel

injector in a Zeiss Versa 620. CTSegNet was developed by the first author (A. Tekawade) at

Argonne National Laboratory for direct surface determination from CT scan volumes. DeepRecon

was developed by ZEISS as a reconstruction engine for denoising volumes obtained from low-

dose scans. The main purpose of this paper is to analyze potential reductions in scanning time

when DL-based workflows are used to augment image-quality. A research fuel injector was

scanned using a ZEISS Versa 620 by acquiring a total of 3600 X-ray projections with 6 seconds

exposure per projection (total 6 hours). The data was reconstructed using equally spaced subsets

of those projections (e.g., 2X, 4X, 6X, 8X, 10X; where 4X means 900 projections) to mimic

reduced scan time. Each subset was reconstructed using DeepRecon Pro (DRPro, the latest version

of DeepRecon) as well as a traditional Feldkamp–Davis–Kress (FDK) reconstruction algorithm

and then thresholded to determine the surface. The data reconstructed by FDK was directly

binarized using CTSegNet. We conclude that variations in measurement accuracy of hole diameter

and cylindricity are expected to be in the order of 2 μm or less when using DRPro and CTSegNet

with a 10X throughput factor. In contrast the deviations produced by the FDK method are larger

than 2 μm and produce more noisy data, showing that DL-based workflows can increase the

throughput of CT acquisition while minimizing accuracy loss in dimensional data.