A Single Image Is All You Need: Zero-Shot Anomaly Localization Without Training Data
arXiv preprint · New
Mehrdad Moradi
Atlanta, Georgia, USA
About Me
Mehrdad Moradi is a fourth-year PhD student in Machine Learning at the Industrial and Systems Engineering Depart- ment of Georgia Tech. He holds a Master’s degree in Computer Science with a specialization in machine learning. He has extensive expertise in generative models, with his PhD research centered on their intersection with statistics. His work focuses on applications in computer vision, particularly image-based anomaly detection. His developed method- ologies have been applied to additive manufacturing characterization (Politecnico di Milano), PV panel smoke/fire detection (U.S. Department of Energy), and battery analytics (Ford Motor Company).
Photo Album
Education
Ph.D. in Machine Learning
ISyE, Georgia Institute of Technology · 2022 – 2027
M.Sc. in Computer Science
Georgia Institute of Technology · 2022 – 2025
B.Sc. in Industrial Engineering
Sharif University of Technology · 2017 – 2022
Research
Publications
RDDPM: Robust Denoising Diffusion Probabilistic Model for Unsupervised Anomaly Segmentation
ICCV VISION 2025 · Oral
Professional Presentations
Diffusion models for anomaly detection in general textured surfaces
Presented my work in the “Advanced Data-Driven Insights for Complex Industrial Processes” session at INFORMS.
RDDPM: Robust Denoising Diffusion Probabilistic Model for Unsupervised Anomaly Segmentation
Spatter Trajectory Analysis for Characterizing Laser Powder Bed Fusion Process
Presented my work in the “Advancements in spatial-temporal analytics” session at INFORMS.
Significant Projects
Generative modeling of contaminated data
- Extended the denoising score matching algorithm to accommodate Laplacian and exponential noise.
- Developed a novel training algorithm combining denoising score matching and contrastive divergence.
RobustCLIP: Robust fine-tuning of CLIP on contaminated data
- Conducted experiments with CLIP-ViT-B/32 on iMet Collection 2019 dataset.
- Developed a novel L1 norm-based robust loss.
- Achieved over 100% improvement in F2 score for contamination levels under 50%.
Leveraging Large Language Models for Author Attribution in Classical Philosophy Texts
- Ran experiments with Llama2, Llama3, and GPT-3.5-Turbo.
- Demonstrated preliminary LLM ability in author identification beyond memorization.
- Showed that few-shot and chain-of-thought prompting did not improve results.
Human Drawing Reconstruction From Different Camera Viewpoints
- Implemented hand-tracking with hand keypoint detection using Multiview Bootstrapping.
- Used a CNN model for sketch classification from different viewpoints.
Medical Image Analysis Project — Brain Tumor Detection with Deep Learning
- Built CNN and RNN models with interpretability via gradient saliency maps and attention visualizations.
- Enhanced accuracy with statistical classification (random forest) on feature maps.
- Achieved 80% accuracy with PCA-based classification, 75% with CNN and RNN.
Optimizing Deep Learning for Medical Image Analysis — Blood Composition Detection
- Used pre-trained YOLO variants for robust blood cell detection.
- Improved memory efficiency (50%) and inference speed (20%) by pruning and 16-bit precision.
- Explored IOU for duplicate removal and optimizer effects.
- Visualized convolutional kernels for model interpretability.