Synthetic data generation can solve the annotation bottleneck for OCR in complex, low-resource scripts—Persian Pixel demonstrates this by creating a scalable, cost-effective alternative to manual labeling that works with modern transformer-based OCR models.
Persian Pixel is a large-scale synthetic dataset of 343,000+ image-text pairs designed to train OCR systems for Persian script.