https://github.com/abdubakr77/drowsiness-detection-eye-state-classification
Binary image classification (Awake vs Sleepy) using deep learning on infrared eye images from the MRL Eye Dataset (~85K images). Built with transfer learning for real-time drowsiness detection.
https://github.com/abdubakr77/drowsiness-detection-eye-state-classification
binary-classification computer-vision deep-learning drowsiness-detection eye-state-classification image-classification infrared-images mrl-eye-dataset pytorch transfer-learning
Last synced: 5 days ago
JSON representation
Binary image classification (Awake vs Sleepy) using deep learning on infrared eye images from the MRL Eye Dataset (~85K images). Built with transfer learning for real-time drowsiness detection.
- Host: GitHub
- URL: https://github.com/abdubakr77/drowsiness-detection-eye-state-classification
- Owner: abdubakr77
- License: mit
- Created: 2026-05-02T16:19:30.000Z (about 2 months ago)
- Default Branch: main
- Last Pushed: 2026-05-03T17:53:34.000Z (about 2 months ago)
- Last Synced: 2026-05-03T19:32:23.949Z (about 1 month ago)
- Topics: binary-classification, computer-vision, deep-learning, drowsiness-detection, eye-state-classification, image-classification, infrared-images, mrl-eye-dataset, pytorch, transfer-learning
- Language: Jupyter Notebook
- Homepage:
- Size: 3.91 KB
- Stars: 0
- Watchers: 0
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# Drowsiness Detection — Eye State Classification
Binary classification of infrared eye images into **Awake** and **Sleepy** states using transfer learning on the MRL Eye Dataset (~85K images).
---
## Overview
Driver drowsiness is one of the leading causes of road accidents. This project builds a deep learning classifier that detects eye state (open vs. closed) from infrared images — a core component of any real-time drowsiness detection system.
Two pretrained models are fine-tuned and compared:
| Model | Epochs | Val AUC |
|---|---|---|
| ResNet50 | 10 | ~0.9997 |
| InceptionV3 | 15 | ~0.9993 |
Both models converge fast and generalize well, with train/val loss tracking closely — no significant overfitting.
---
## Dataset
**MRL Eye Dataset** (Forked Version) — infrared eye images labeled as Awake or Sleepy.
| Split | Awake | Sleepy | Total |
|---|---|---|---|
| Train | 25,770 | 25,167 | 50,937 |
| Validation | 8,591 | 8,389 | 16,980 |
| Test | 8,591 | 8,390 | 16,981 |
- 37 subjects · multiple sensors · various lighting conditions
- Classes are well-balanced — no resampling needed
Dataset source: [mrl.cs.vsb.cz/eyedataset](http://mrl.cs.vsb.cz/eyedataset)
Expected directory structure:
```
data/
├── train/
│ ├── awake/
│ └── sleepy/
├── val/
│ ├── awake/
│ └── sleepy/
└── test/
├── awake/
└── sleepy/
```
---
## Approach
**Why transfer learning?**
Even with ~85K images, pretrained weights (ImageNet) provide low-level features — edges, textures, shapes — that transfer well to infrared eye images and cut training time significantly.
**Augmentation**
Light augmentation on train only: Gaussian blur, brightness jitter, horizontal flip, random rotation. The dataset is large enough that heavy augmentation isn't necessary.
---
## Results
### ResNet50
### InceptionV3
Both models reach **AUC > 0.999** on the validation set. The few misclassified examples (shown in red) are edge cases — heavily shadowed eyes, glare artifacts, or partially closed eyelids that are ambiguous even to a human.
---
## Project Structure
```
├── Drowsiness_Detection.ipynb # full pipeline: EDA → training → evaluation
├── ResNet_AUC_LOSS_Curves.png
├── ResNet_test_set_predictions.png
├── Inception3_AUC_LOSS_Curves.png
├── Inception3_test_set_predictions.png
└── README.md
```
---
## Setup
```bash
pip install torch torchvision scikit-learn tqdm seaborn matplotlib
```
Update the dataset paths in the **Configuration** cell of the notebook, then run all cells top to bottom.
---
## Key Design Decisions
- **AdamW optimizer** with a low learning rate (1e-5) to avoid destroying pretrained weights early in training
- **Early stopping** on validation AUC with patience=5 to prevent overfitting
- **Mixed precision** (`torch.autocast`) for faster GPU training
- **Dropout(0.5)** added before the ResNet head as a light regularizer
- Grayscale conversion applied *after* augmentation to preserve color-based jitter effects before discarding color
---
## References
- Fusek, R. (2018). *Pupil localization using geodesic distance.* ICIAR 2018. [[paper]](http://mrl.cs.vsb.cz/people/fusek/)
- MRL Eye Dataset: [http://mrl.cs.vsb.cz/eyedataset](http://mrl.cs.vsb.cz/eyedataset)