StyleScript: A Structured Data Augmentation Framework for Transformer-Based OCR in Engineering Documents

Abstract

Engineering documents often contain a combination of printed and handwritten text, intricate layouts, and visual degradation, posing significant challenges to Optical Character Recognition (OCR) systems. Transformer-based models like TrOCR provide strong baseline performance but require domain-specific data augmentation to generalize effectively. This paper introduces StyleScript, a structured data augmentation framework that generates realistic synthetic word images by extracting stroke-based style features such as slant angle and thickness. We fine-tune both TrOCR (small) and TrOCR (large) models using real and StyleScript-augmented data derived from a Military Sealift Command (MSC) dataset. Additionally, we develop a systematic OCR pipeline combining preprocessing, CRAFT-based text detection, and fine-tuned TrOCR recognition to digitize full-page engineering documents with spatial fidelity. Experimental results show that StyleScript-enhanced training improves OCR performance across diverse document conditions, making it a practical solution for engineering and other technical domains with limited annotated data.