ganesh045078-create/Synthetic-Document-for-Low-Resource-OCR

**Title:** Generative AI for Synthetic Document Creation for Low-Resource OCR **Description:** In this paper, we introduce a generative AI–driven framework that precisely targets one of the major bottlenecks towards the development of reliable OCR systems: the absence or rarity of annotated datasets in low-resource languages. Central to our approach is a Synthetic Document Generation Pipeline, driven by the SD-GAN, capable of generating high-quality artificial document images that closely resemble real scanned pages. The process starts with the collection of a base corpus of text samples in the target low-resource language. These are then arranged randomly into document layouts, including various fonts, background textures, noise patterns, distortions, and artifacts such as blur, ink bleed, creases, shadows, and compression noise. The SD-GAN captures the real structure and visual characteristics of actual documents and generates them with realistic variations. The result of this process is the creation of larger, more diverse, and more complex datasets without requiring manual ground-truth annotation. These synthesized images are combined with the available real data and processed through a complete OCR training pipeline. The codebase includes the steps of preprocessing, data augmentation routines, GAN-based data generation, conventional Machine Learning and Deep Learning models for classifying OCR, and a full evaluation framework. The performance of the models is evaluated using OCR metrics such as CER and WER, as well as image quality metrics like PSNR and SSIM. The repository includes all algorithms and source code, dataset generation procedures, and a reproduction script that can regenerate the whole experimental workflow starting from the creation of synthetic data up to model evaluation for scientific reproducibility. Comparison experiments include: - ML models versus DL models versus SD-GAN–augmented models Ablation studies comparing performance with and …