Mahsa Aliee

Date of Award

December 2025

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Electrical Engineering

First Advisor

Bo B. Liang

Second Advisor

Kouhyar K. Tavakolian

Abstract

Effective surface contamination detection in food processing and healthcare environments is hindered by the limitations of visual inspection and ATP bioluminescence testing, both of which struggle to identify thin, transparent, or spatially irregular organic residues. These constraints contribute to inconsistent hygiene verification and create a need for objective, scalable, and real-time detection methods capable of operating under variable environmental conditions.

This dissertation develops an integrated framework that combines UV-induced fluorescence imaging with advanced analytical techniques to improve contamination detection reliability. A UV/UVC fluorescence imaging system was engineered with synchronized pulsed excitation (275 nm, 365 nm, 405 nm), adaptive exposure control, and real-time background subtraction to enhance signal visibility across ambient lighting conditions ranging from 2 to 22 foot-candles. Imaging performance was further strengthened through a noise-aware deep learning strategy in which YOLOv8 segmentation models were trained using synthetic Gaussian, Poisson, and stripe-pattern noise that reflected photon-limited UVC imaging conditions. These augmentations substantially increased robustness, preserving accurate contamination detection even under severe noise interference.

To benchmark imaging-based detection against established sanitation practices, fluorescence-derived intensity and texture features were correlated with ATP Relative Light Units (RLUs). A TabNetClassifier was trained using stratified group cross-validation, achieving high sensitivity (>86%), strong specificity (>90%), and stable performance across five folds. Statistical analyses, including hypothesis testing, effect size estimation, ROC evaluation, and calibration assessment, confirmed the discriminative strength and reliability of the model. Generalized Estimating Equations (GEE) demonstrated minimal systematic bias across surface materials and imaging parameters, supporting the consistency of predictions under varied conditions.

The results show that integrating optimized fluorescence imaging with noise-aware deep learning and tabular machine learning provides a reliable, non-contact, and scalable alternative to ATP testing. The developed workflow, GLOW-DL, enhances contamination detection accuracy and operational robustness, offering a practical pathway toward real-time hygiene monitoring in food service, manufacturing, and healthcare environments.

Download

Share

COinS