Developing High-Level AI Architectures and Systems for Industrial Defect Detection using Transformers, VAEs, and Diffusion Models

Abstract

This thesis presents a hybrid deep learning framework for industrial defect detection using Vision Transformers, Variational Autoencoders, and diffusion models. The proposed architecture combines classification and generative modeling to improve defect identification and representation. Experiments conducted on the NEU Steel Defect dataset demonstrate high performance, achieving 99.26% accuracy and a weighted F1-score of 0.9926. Additional evaluation using precision, recall, and ROC-AUC confirms robust and balanced model performance across all defect classes. The results highlight the effectiveness of transformer-based architectures in capturing complex surface defect patterns. Limitations related to dataset size and real-world deployment are also discussed, along with directions for future research.