Condition Analysis & Data Generation

GlowAI is a computer vision project designed to address the scarcity of medical skin datasets. Standard datasets are hard to obtain due to privacy concerns. This project introduces a Synthetic Skin Simulator that uses classic Computer Vision (OpenCV) to generate medically relevant training data, creating realistic Acne and Erythema lesions on clean faces to train robust classification models.

👥 Team

• Ron Noiman
• Reut Sasson
• Lin Schneider

---

🚀 Project Pipeline

Our workflow consists of three main stages:

1. Synthetic Data Generation (The Simulator)

• Acne Simulation: We use OpenCV to draw randomized circles with specific Gaussian blur and opacity to mimic the texture and inflammation of pimples.
• Erythema Simulation: We apply irregular noise masks and red-channel blending to simulate skin redness patches naturally.
• Output: A labeled dataset of synthetic skin conditions generated from clean facial images.

2. Model Training (The Classifier)

• Architecture: We use an EfficientNet-B2 backbone pretrained on ImageNet.
• Training Strategy: A two-phase approach (Linear Probing followed by Fine-Tuning) to adapt the model to skin features.
• Loss Function: Cross Entropy Loss with Adam optimizer.

3. Validation (The Proof)

We evaluate the model on two distinct sets:

• Synthetic Validation Set: To measure learning on generated patterns.
• Manual Real-World Test Set: To prove the model generalizes to real images (Held-out data).

---

🛠️ Installation

1. Clone the repository

bash git clone https://github.com/ReutSasson/GlowAIProj.git cd GlowAIProj/

2. Install Dependencies

pip install -r requirements.txt

📂 Folder Structure

GlowAIProj/
│
├── code/
│   ├── simulate_skin_no_mediapipe.py   # The Simulator (OpenCV)
│   ├── train_skin_cls.py               # Main Training Script
│   └── experiment_logger.py            # Metrics logging
│
├── data/
│   ├── train/                          # Training set (Synthetic + Real)
│   ├── val/                            # Validation set
│   └── feedback/                       # Hard examples for retraining
│
├── results/                            # Visualizations & Metrics (Best Run)
│   ├── accuracy_curve.png
│   ├── confusion_matrix.png
│   ├── loss_curve.png
│   └── visual_abstract.png
│
├── slides/                             # Project Presentations (PDF/PPTX)
├── checkpoints/                        # Saved model weights
└── README.md

🏃‍♂️ How to Run

Step 1: Generate Data Run the simulator to create synthetic acne/erythema on clean faces:

python code/simulate_skin_no_mediapipe.py --input data/none --out_dir out/simulated_skin

Step 2: Train the Model

Train the EfficientNet-B2 classifier on the dataset:

python code/train_skin_cls.py --data dataset --epochs 15 --out checkpoints/run_final

📊 Results Snapshot (Best Run)

Below are the results from our best experiment (Run 9), comparing performance on synthetic vs. real data.

💾 Data Access

The dataset is included in the dataset/ folder.

Submitted as a final project for the GenAI course at HIT.