We are a two-person team:
- Nazanin Ahmadi
- Saba Shafiee
Computer Engineering students at Isfahan University of Technology (IUT), with an interest in artificial intelligence and deep learning.
This repository contains our work on a deep learning–based anomaly detection project.
The project focuses on understanding both the implementation and the behavior of anomaly detection models in practice.
The main concepts explored in this project include:
- Convolutional Neural Networks (CNNs) for feature extraction
- U-Net architecture as the main backbone of the model, using an encoder–decoder structure with skip connections
- Generator–Discriminator frameworks inspired by GAN-based anomaly detection methods
- A U-Net–based generator for reconstructing normal samples
- A discriminator for learning the distribution of normal data
- Multi-scale loss functions to capture anomalies at different spatial and feature levels
- Pixel-level and feature-level reconstruction errors for anomaly scoring
Particular attention is given to model evaluation and training dynamics, including:
- Monitoring training and validation loss
- Tracking AUC (Area Under the ROC Curve) across epochs
- Analyzing cases where loss decreases while AUC performance degrades
- Studying overfitting behavior in reconstruction-based models
- Develop a solid understanding of anomaly detection methods
- Analyze the strengths and limitations of U-Net–based architectures
- Investigate the impact of multi-scale loss design on detection performance
- Improve interpretation of evaluation metrics beyond loss minimization
This project is under active development and will be updated as experiments progress.