This project introduces XTiny-FastKAN, a lightweight, interpretable, and highly efficient neural network for gesture-based air handwriting recognition of the Tifinagh script, targeting real-time inference on low-power TinyML microcontrollers.
This work includes:
- A novel Tifinagh IMU-based air-writing dataset
- A rasterization pipeline to transform motion trajectories into color-encoded images
- A customized version of FastKAN optimized for TinyML
- Robust data augmentation and preprocessing strategies
- Reproducible training, evaluation, and explainability via saliency maps
✅ First air-writing dataset for Tifinagh characters
✅ Compact and fast model (~35 KB, < 0.05ms latency)
✅ Compatible with TensorFlow Lite Micro / Edge Impulse
✅ Includes data augmentation, training scripts, and confusion matrix tools
✅ Built-in XAI module (saliency visualization)
├── data/ # Raw and processed Tifinagh stroke data
├── rasterization/ # Converts strokes into color raster images
├── augmentation/ # Augmentation scripts (rotation, noise, warping
├── model/ # XTiny-FastKAN model and architecture
├── training/ # Training and evaluation scripts
├── explainability/ # XAI module (saliency maps)
├── tflite/ # Exported .tflite and quantized models
├── results/ # Evaluation metrics, plots, confusion matrix
├── Deployment/ Deploy our XTiny-FastKAN into edge device
├── notebooks/ # Jupyter notebooks for visualization and testing
└── README.md
This project is licensed under the terms of the MIT License.
You are free to use, modify, distribute, and sell this software, provided that you include the original copyright and license notice.
📬 Contact For questions or collaborations, please contact:
Ismail Lamaakal
Multidisciplinary Faculty of Nador, Mohammed Premier University, Oujda, Morocco