Barcode & Text Localization on Mobile
Overview
As lead ML engineer for detection on edge devices at Scandit, I developed MatrixScan, a high-performance computer vision solution that enables real-time barcode and text detection directly on mobile devices.
Key Achievements
- Launched real-time ML model for image segmentation running on 1M+ mobile devices monthly
- Trained binary neural networks for smartphones with 30x less memory usage
- Developed multi-stage training algorithms reducing model size by 50% at same quality
- Introduced Python coding guidelines and company-wide Request for Change process
Technical Stack
- C++ - High-performance mobile runtime implementation
- Python - ML model development and training pipelines
- PyTorch - Deep learning framework for neural network development
Technical Innovation
MatrixScan represents a breakthrough in mobile computer vision, delivering enterprise-grade barcode scanning performance directly on consumer smartphones. The system uses advanced neural network compression techniques and optimized inference engines to achieve real-time performance while maintaining exceptional accuracy.
Mobile Optimization
The solution required innovative approaches to neural network quantization and mobile GPU optimization, enabling complex computer vision tasks to run efficiently on resource-constrained mobile devices without compromising user experience.