[Paper Review] PP-OCR: A Practical Ultra Lightweight OCR System

Yuning Du, Chenxia Li+9 more|arXiv (Cornell University)|Sep 21, 2020

Handwritten Text Recognition Techniques23 references109 citations

TL;DR

PP-OCR presents a practical ultra-lightweight OCR system with a total model size of 3.5M for Chinese/English character recognition (6622 chars) and 2.8M for alphanumeric symbols, featuring a set of slimming and accuracy-enhancing strategies and open-source pretrained models.

ABSTRACT

The Optical Character Recognition (OCR) systems have been widely used in various of application scenarios, such as office automation (OA) systems, factory automations, online educations, map productions etc. However, OCR is still a challenging task due to the various of text appearances and the demand of computational efficiency. In this paper, we propose a practical ultra lightweight OCR system, i.e., PP-OCR. The overall model size of the PP-OCR is only 3.5M for recognizing 6622 Chinese characters and 2.8M for recognizing 63 alphanumeric symbols, respectively. We introduce a bag of strategies to either enhance the model ability or reduce the model size. The corresponding ablation experiments with the real data are also provided. Meanwhile, several pre-trained models for the Chinese and English recognition are released, including a text detector (97K images are used), a direction classifier (600K images are used) as well as a text recognizer (17.9M images are used). Besides, the proposed PP-OCR are also verified in several other language recognition tasks, including French, Korean, Japanese and German. All of the above mentioned models are open-sourced and the codes are available in the GitHub repository, i.e., https://github.com/PaddlePaddle/PaddleOCR.

Motivation & Objective

• Motivate the need for efficient OCR on resource-constrained devices and in embedded settings.
• Develop an ultra-lightweight OCR pipeline consisting of text detection, box rectification, and text recognition.
• Systematically apply model enhancement and slimming techniques to reduce size while preserving accuracy.
• Provide open-source pretrained models and ablation studies demonstrating the trade-offs between size, speed, and accuracy.

Proposed method

• Adopt a three-part OCR pipeline: text detection using Differentiable Binarization (DB) with a light backbone and head; detected box rectification with a text direction classifier; and text recognition using CRNN with a light backbone.
• Apply six strategies to text detection (light backbone, light head, remove SE, cosine LR decay, LR warm-up, FPGM pruning) to shrink size to 1.4M.
• Apply four strategies to direction classification (light backbone, data augmentation, input resolution, PACT quantization) to achieve a 500KB size.
• Apply nine strategies to text recognition (light backbone, data augmentation, cosine LR decay, feature map resolution adjustments, regularization, LR warm-up, light head, pre-trained model, PACT quantization) to reduce the recognizer to 1.6M for Chinese/English and 900KB for alphanumerics.
• Train on a large multi-part dataset (text detection, direction classification, text recognition) and provide additional language coverage (French, Korean, Japanese, German).
• Release pretrained models and code in PaddleOCR for broader reuse.

Experimental results

Research questions

• RQ1 How can an OCR system be made ultra-lightweight without sacrificing practical accuracy for real-world scenes?
• RQ2 What combination of backbones, data augmentations, resolution choices, and quantization techniques yields the best trade-off between model size, speed, and accuracy for detection, direction classification, and recognition?
• RQ3 Can a cohesive PP-OCR system generalize across languages beyond Chinese and English, including alphanumeric, French, Korean, Japanese, and German?

Key findings

• The PP-OCR system achieves an ultra-lightweight total size of 3.5M for Chinese/English recognition and 2.8M for alphanumeric recognition.
• Text detector size can be reduced to 1.4M with light backbones and head design plus pruning and hardware-aware optimizations.
• Direction classifier can be kept at 500KB with four strategies including data augmentation and PACT quantization, maintaining competitive accuracy.
• Text recognizer can reach 1.6M (Chinese/English) or 900KB (alphanumeric) with a combination of light backbone, augmented data, and PACT quantization, while preserving practical accuracy.
• Extensive ablation studies demonstrate the impact of backbone choices, feature map resolution, augmentation, and quantization on accuracy (HMean, F-score) and efficiency (inference time).
• The authors provide open-source pretrained models and code via PaddleOCR, enabling broader adoption across languages (including French, Korean, Japanese, German).