# Enhancing OCR Performance through Post-OCR Models: Adopting Glyph Embedding for Improved Correction

**Authors:** Yung-Hsin Chen, Yuli Zhou  
**arXiv:** [2308.15262](https://arxiv.org/abs/2308.15262)  
**Date:** 2023-08-29  
**Summary page:** https://www.summarizepaper.com/en/paper/2308.15262v1/

## Key points

- Study focuses on potential of post-OCR models to overcome limitations in OCR models
- Explores impact of incorporating glyph embedding on post-OCR correction performance
- Researchers developed their own post-OCR correction model utilizing CharBERT and unique technique for capturing visual characteristics of characters
- Two datasets used: ICDAR 2013 dataset and ICDAR 2023 dataset
- ICDAR 2013 dataset serves as benchmark for evaluating OCR systems, includes diverse document images with printed and handwritten text captured under various conditions
- ICDAR 2023 dataset introduces additional challenges such as intricate layouts, degraded text, low-resolution images, and background clutter
- Three OCR models evaluated: EasyOCR, PaddleOCR, TrOCR
- EasyOCR used for single word detection box functionality, PaddleOCR and TrOCR provide single line detection box functionality
- Training of glyph embedding relies on Chars74K dataset which includes scene images of English and Kannada characters (only English alphabets utilized)
- Screenshots of Korean and Hebrew characters captured for garbage class in open-set classification
- State-of-the-art (SOTA) models as well as researchers' post-OCR correction model examined in methodology.

## Summary

The study focuses on the potential of post-OCR models to overcome limitations in OCR models and explores the impact of incorporating glyph embedding on post-OCR correction performance. The researchers have developed their own post-OCR correction model, which utilizes CharBERT for embedding the OCR output and a unique technique for capturing the visual characteristics of characters. To conduct their study, the researchers utilize two prominent datasets: the ICDAR 2013 dataset and the recently introduced ICDAR 2023 dataset. The ICDAR 2013 dataset consists of diverse document images with printed and handwritten text captured under various conditions, serving as a robust benchmark for evaluating OCR systems. On the other hand, the ICDAR 2023 dataset expands upon its predecessor by introducing additional challenges such as intricate document layouts, degraded text, low-resolution images, and challenging background clutter. Three OCR models are evaluated using these datasets: EasyOCR, PaddleOCR, and TrOCR. EasyOCR is used specifically for single word detection box functionality, while PaddleOCR and TrOCR provide single line detection box functionality. The outputs from EasyOCR consist solely of single words, while PaddleOCR generates output in sentence form. For training the glyph embedding in their post-OCR correction model, the researchers rely on the Chars74K dataset. This dataset includes scene images of English and Kannada characters, although only English alphabets are utilized for this particular experiment. Additionally, screenshots of Korean and Hebrew characters are captured to serve as samples for the garbage class in open-set classification. In terms of methodology, state-of-the-art (SOTA) models as well as the researchers' post-OCR correction model are examined.