TOWARDS AUTOMATED QUALITY ASSURANCE IN INDUSTRIAL RADIOGRAPHY: INSTANCE SEGMENTATION AND ANALYSIS OF IMAGE QUALITY INDICATORS AND POROSITY

Author

Merrill Edward August Dennison, Southern Illinois University CarbondaleFollow

Date of Award

12-1-2025

Degree Name

Doctor of Philosophy

Department

Mechanical Engineering

First Advisor

Chu, Tsuchin

Abstract

This dissertation presents approaches for performing tasks related to quality assurance in industrial radiography. The research emphasizes the importance of high-quality industrial radiography for effective component inspection and addresses the critical need for improvements to methodologies involving the use of non-destructive evaluation (NDE) for quality assurance. The objectives of the work include the development of image processing and analysis techniques to assist human inspectors that reduce the human subjectivity inherent to these tasks in quality assurance. The aim of these improvements is to enhance the reliability and interpretability of decision-making processes that rely on these tasks. Current standards for determining image quality from image quality indicators (IQIs), as well as porosity in production castings from reference standards, involve a high degree of human subjectivity. The use of algorithms can reduce human subjectivity in these analyses; however, it introduces new problems such as the reliability, interpretability, and explainability of the algorithm. The paradigm of Artificial Narrow Intelligence (ANI) is becoming increasingly ubiquitous in modern life, and there is a burgeoning conversation around how to, if at all, properly use ANI in high-stakes decision-making processes. ANI, in the context of this dissertation discussion, encompasses concepts within machine learning (ML) and Deep Learning (DL). These concepts have produced algorithms that perform well on the tasks described in the opening sentence of this paragraph, however, these algorithms are often used as a “black box,” which does not lend well to the levels of scrutiny necessary for high-stakes decision-making. This research work explores how ANI and traditional image processing can be applied to perform certain quality assurance processes using radiography. The automated style approach involves the segmentation of plaque hole type Image Quality Indicators (IQIs) from digital radiographs and the subsequent analysis of the contrast to noise ratio (CNR) of their holes. A deep artificial neural network architecture (DANNA) is used first to segment the IQI instances. These segmented instances are then processed through a pipeline of traditional image processing and analysis methods that generate the CNR calculations. The assisted style approach leverages (1) a linear curve fit of porosity versus mean image intensity, and (2) a bandit algorithm to optimize multiple image segmentation methods. The bandit algorithms reward involves components that are aimed at selecting segmentation hyperparameters that produce masks whose components reflect size and frequency trends that directly match the severity assignment of ASTM reference radiographs for ¼” thick cast aluminum plates and the measured porosity of 3/10” thick additively manufactured aluminum plates. The results of this research work highlight the critical role in which data quality, algorithm, feature, and parameter selection play in achieving reliable outcomes that can be adapted to new data. The content of this dissertation provides a valuable example of how automated and assisted solutions might be implemented towards IQI processing and analysis of porosity in industrial radiography; ultimately reducing the human subjectivity currently inherent to these tasks. The research involves the detection, segmentation, processing, and analysis of features within digital imagery, which are ubiquitous tasks within the world of NDE. Approaches for automatically accomplishing these tasks are highly sought after for the potential savings in time and effort put forth by human inspectors.