Dissertation
Bridging the data quality gap: evaluating image quality and data types for optimal neural network performance in manufacturing
University of Iowa
Doctor of Philosophy (PhD), University of Iowa
Summer 2025
DOI: 10.25820/etd.008136
Abstract
This dissertation presents a comprehensive investigation into the application of artificial intelligence (AI) for automating the post-processing task of sprue and riser removal in highmix, low-volume (HMLV) metal foundries. The study introduces the Sandcast Images for Foundry Automation (SIFA) dataset, a novel, publicly available, fully annotated dataset of sand cast parts featuring real, synthetic, and augmented images. Instance segmentation models are trained and tested across these data types to evaluate performance and generalization under varying conditions. Objective image quality metrics, including entropy, gray-level cooccurrence matrix (GLCM), structural similarity index (SSIM), and intrinsic dimensionality (ID), are computed to assess dataset quality and its relationship to model performance.
Experiments reveal that synthetic data, when constructed with high fidelity, can match or outperform real data for training instance segmentation models, particularly when supplemented with small quantities of real data. The study further examines the effects of color-space transformations, scene complexity, and transfer learning, demonstrating their effectiveness in improving model robustness in visually complex or data-limited environments. Correlations between image quality metrics and model fitness confirm that data quality, not just quantity, plays a critical role in model success. Collectively, these findings provide a structured framework for dataset development and evaluation, offering practical insights for deploying AI in real-world manufacturing settings.
Details
- Title: Subtitle
- Bridging the data quality gap: evaluating image quality and data types for optimal neural network performance in manufacturing
- Creators
- Martell Bell
- Contributors
- Rachel V Vitali (Advisor)Tyler Bell (Committee Member)Jia Lu (Committee Member)HS Udaykumar (Committee Member)
- Resource Type
- Dissertation
- Degree Awarded
- Doctor of Philosophy (PhD), University of Iowa
- Degree in
- Mechanical Engineering
- Date degree season
- Summer 2025
- DOI
- 10.25820/etd.008136
- Publisher
- University of Iowa
- Number of pages
- xvi, 118 pages
- Copyright
- Copyright 2025 Martell Bell
- Language
- English
- Date submitted
- 07/24/2025
- Description illustrations
- illustrations (some color)
- Description bibliographic
- Includes bibliographical references (pages 106-118).
- Public Abstract (ETD)
This dissertation explores how artificial intelligence (AI) can be used to automate a key manufacturing task in metal casting: removing extra material called sprues and risers from finished parts. Focusing on small-batch, wide-variety production facilities, this work introduces a new dataset called the Sandcast Images for Foundry Automation (SIFA) dataset. This publicly available collection includes thousands of real and computer-generated images of sand-cast metal parts. All of these images were carefully labeled for training AI systems.
Using this dataset, the study evaluates how well different types of image data help AI models learn to identify and separate objects in complex manufacturing scenes. It also examines how image quality and other factors, such as background complexity and color variation, affect AI performance. Results show that synthetic images, when created with high detail, can perform just as well, or even better than real images in training AI systems. The research also finds that improving the quality of training images can be just as important as having a large quantity of them.
Overall, this work provides new tools and insights for using AI in manufacturing, with the goal of making production more efficient, flexible, and less dependent on manual labor.
- Academic Unit
- Mechanical Engineering
- Record Identifier
- 9984948641002771
Metrics
2 File views/ downloads
9 Record Views