# PP-OCR: A Practical Ultra Lightweight OCR System

**Authors:** Yuning Du, Chenxia Li, Ruoyu Guo, Xiaoting Yin, Weiwei Liu, Jun Zhou, Yifan Bai, Zilin Yu, Yehua Yang, Qingqing Dang, Haoshuang Wang  
**arXiv:** [2009.09941](https://arxiv.org/abs/2009.09941)  
**Date:** 2020-09-21  
**Summary page:** https://www.summarizepaper.com/en/paper/2009.09941v3/

## Key points

- Optical Character Recognition (OCR) systems are popular in various application scenarios
- OCR remains a challenging task due to diverse text appearances and computational efficiency needs
- Yuning Du et al. propose a practical ultra-lightweight OCR system called PP-OCR
- PP-OCR has an overall model size of only 3.5M for recognizing 6622 Chinese characters and 2.8M for recognizing 63 alphanumeric symbols
- The authors introduce several strategies to enhance the model's ability or reduce its size while maintaining high accuracy levels
- The authors release several pre-trained models for Chinese and English recognition that include a text detector, direction classifier, and text recognizer
- These models are open-sourced and available on GitHub repository https://github.com/PaddlePaddle/PaddleOCR.
- PP-OCR is verified in several other language recognition tasks such as French, Korean, Japanese, and German.
- The system achieves state-of-the-art performance on these tasks while maintaining its ultra-lightweight design.
- PP-OCR offers high accuracy levels while being lightweight enough to run efficiently on low-power devices such as smartphones or embedded systems.
- Its open source nature makes it accessible to developers worldwide who can leverage its capabilities in various applications requiring OCR technology.

## Summary

The Optical Character Recognition (OCR) systems have become increasingly popular in various application scenarios, including office automation (OA) systems, factory automations, online educations, map productions, and more. However, OCR remains a challenging task due to the diverse text appearances and the need for computational efficiency. To address this challenge, Yuning Du et al. propose a practical ultra-lightweight OCR system called PP-OCR. PP-OCR has an overall model size of only 3.5M for recognizing 6622 Chinese characters and 2.8M for recognizing 63 alphanumeric symbols. The authors introduce several strategies to enhance the model's ability or reduce its size while maintaining high accuracy levels. They conduct corresponding ablation experiments with real data to validate their approach. Furthermore, the authors release several pre-trained models for Chinese and English recognition that include a text detector (97K images used), a direction classifier (600K images used), and a text recognizer (17.9M images used). These models are open-sourced and available on GitHub repository https://github.com/PaddlePaddle/PaddleOCR. The proposed PP-OCR is also verified in several other language recognition tasks such as French, Korean, Japanese, and German. The authors demonstrate that their system achieves state-of-the-art performance on these tasks while maintaining its ultra-lightweight design. In summary, PP-OCR is an innovative solution that offers high accuracy levels while being lightweight enough to run efficiently on low-power devices such as smartphones or embedded systems. Its open source nature makes it accessible to developers worldwide who can leverage its capabilities in various applications requiring OCR technology.