📖 Doc | 🛠️ Install | 🧳 Models | 🕹 Training | 🛀🏻 Online Demo | 💬 Contact

CnOCR

[Update 2023.12.24]: Release of V2.3

Major Changes:

• All models have been retrained, offering higher accuracy than the previous version.
• Models are categorized into several types based on usage scenarios (see Pre-trained Recognition Models):
  • scene: For scene images, suitable for recognizing text in general photography. Models in this category start with scene-, such as the scene-densenet_lite_136-gru model.
  • doc: For document images, suitable for recognizing text in regular document screenshots, like scanned book pages. Models in this category start with doc-, such as the doc-densenet_lite_136-gru model.
  • number: Specifically for recognizing only numbers (able to recognize only the ten digits 0~9), suitable for scenarios like bank card numbers, ID numbers, etc. Models in this category start with number-, such as the number-densenet_lite_136-gru model.
  • general: For general scenarios, suitable for images without a clear preference. Models in this category do not have a specific prefix and maintain the same naming convention as older versions, such as the densenet_lite_136-gru model.

  Note ⚠️: The above descriptions are for reference only. It is recommended to choose models based on actual performance.

• Two larger series of models have been added:
  • *-densenet_lite_246-gru_base: Initially available for Planet of Knowledge CnOCR/CnSTD Private Group members, to be open-sourced for free after one month.
  • *-densenet_lite_666-gru_large: Pro models, available for use after purchase. Purchase link can be found at https://ocr.lemonsqueezy.com.

For more details, please refer to: CnOCR V2.3 New Release: Better, More, and Larger Models | Breezedeus.com.

CnOCR is an Optical Character Recognition (OCR) toolkit for Python 3. It supports recognition of common characters in English and numbers, Simplified Chinese, Traditional Chinese (some models), and vertical text recognition. It comes with 20+ well-trained models for different application scenarios and can be used directly after installation. Also, CnOCR provides simple training commands for users to train their own models. Welcome to join the WeChat contact group.

The author also maintains Planet of Knowledge CnOCR/CnSTD Private Group, where questions are more likely to receive prompt responses from the author. You are welcome to join. Knowledge Planet Members can enjoy the following benefits:

• Access to download some non-open source paid models for free.
• A 20% discount on the purchase of all other paid models.
• Rapid responses from the author to various difficulties encountered during usage.
• The author offers two free training services per month using unique data.
• The group will continuously release exclusive materials related to CnOCR/CnSTD.
• The group will regularly publish the latest research materials related to OCR, STD, CV, and more.

Documentation

See CnOCR online documentation , in Chinese.

Usage

Starting from V2.2, CnOCR internally uses the text detection engine CnSTD for text detection and positioning. So CnOCR V2.2 can recognize not only typographically simple printed text images, such as screenshot images, scanned copies, etc., but also scene text in general images.

Here are some examples of usages for different scenarios.

Usages for Different Scenarios

Common image recognition

Just use default values for all parameters. If you find that the result is not good enough, try different parameters more to see the effect, and usually you will end up with a more desirable accuracy.

from cnocr import CnOcr

img_fp = './docs/examples/huochepiao.jpeg'
ocr = CnOcr()  # Use default values for all parameters
out = ocr.ocr(img_fp)

print(out)

Recognition results:

English Recognition

Although Chinese detection and recognition models can also recognize English, detectors and recognizers trained specifically for English texts tend to be more accurate. For English-only application scenarios, it is recommended to use the English detection model det_model_name='en_PP-OCRv3_det' and the English recognition model rec_model_name='en_PP-OCRv3' from PaddleOCR (also called ppocr).

from cnocr import CnOcr

img_fp = './docs/examples/en_book1.jpeg'
ocr = CnOcr(det_model_name='en_PP-OCRv3_det', rec_model_name='en_PP-OCRv3')
out = ocr.ocr(img_fp)

print(out)

Recognition results:

Recognition of typographically screenshot images

For typographically simple typographic text images, such as screenshot images, scanned images, etc., you can use det_model_name='naive_det', which is equivalent to not using a text detection model, but using simple rules for branching.

Note

det_model_name='naive_det' is equivalent to CnOCR versions before V2.2 (V2.0.*, V2.1.*).

The biggest advantage of using det_model_name='naive_det' is that the speech is fast and the disadvantage is that it is picky about images. How do you determine if you should use the detection model 'naive_det'? The easiest way is to take your application image and try the effect, if it works well, use it, if not, don't.

from cnocr import CnOcr

img_fp = './docs/examples/multi-line_cn1.png'
ocr = CnOcr(det_model_name='naive_det') 
out = ocr.ocr(img_fp)

print(out)

Recognition results:

图片	OCR结果
	网络支付并无本质的区别，因为 每一个手机号码和邮件地址背后 都会对应着一个账户--这个账 户可以是信用卡账户、借记卡账 户，也包括邮局汇款、手机代 收、电话代收、预付费卡和点卡 等多种形式。

Vertical text recognition

Chinese recognition model rec_model_name='ch_PP-OCRv3' from ppocr is used for recognition.

from cnocr import CnOcr

img_fp = './docs/examples/shupai.png'
ocr = CnOcr(rec_model_name='ch_PP-OCRv3')
out = ocr.ocr(img_fp)

print(out)

Recognition results:

Traditional Chinese Recognition

Use the traditional Chinese recognition model from ppocr rec_model_name='english_cht_PP-OCRv3' for recognition.

from cnocr import CnOcr

img_fp = './docs/examples/fanti.jpg'
ocr = CnOcr(rec_model_name='chinese_cht_PP-OCRv3')  # use the traditional Chinese recognition model
out = ocr.ocr(img_fp)

print(out)

When using this model, please note the following issues:

• The recognition accuracy is average and not very good.

• The recognition of punctuation, English and numbers is not good except for traditional Chinese characters.

• This model does not support the recognition of vertical text.

Single line text image recognition

If it is clear that the image to be recognized is a single line text image (as shown below), you can use the class function CnOcr.ocr_for_single_line() for recognition. This saves the time of text detection and will be more than twice as fast.

The code is as follows:

from cnocr import CnOcr

img_fp = './docs/examples/helloworld.jpg'
ocr = CnOcr()
out = ocr.ocr_for_single_line(img_fp)
print(out)

More Applications

• Recognition of Vaccine App Screenshot

• Recognition of ID Card

• Recognition of Restaurant Ticket

Install

Well, one line of command is enough if it goes well.

$ pip install cnocr[ort-cpu]

If you are using a GPU environment with an ONNX model, please install using the following command:

$ pip install cnocr[ort-gpu]

If you want to train new models with your own data, please install using the following command:

$ pip install cnocr[dev]

If the installation is slow, you can specify a domestic installation source, such as using the Aliyun source:

$ pip install cnocr -i https://mirrors.aliyun.com/pypi/simple

Note

Please use Python3 (3.6 and later should work), I haven't tested if it's okay under Python2.

More instructions can be found in the installation documentation (in Chinese).

Warning

If you have never installed PyTorch, OpenCV python packages on your computer, you may encounter problems with the first installation, but they are usually common problems that can be solved by Baidu/Google.

Docker Image

You can directly pull the image with CnOCR installed from Docker Hub: Docker Hub .

$ docker pull breezedeus/cnocr:latest

More instructions can be found in the installation documentation (in Chinese).

Pre-trained Models

Pre-trained Detection Models

Refer to CnSTD for details.

Pre-trained Recognition Models

Compared to the CnOCR V2.2.* versions, most models in V2.3 have been retrained and fine-tuned, offering higher accuracy than the older versions. Additionally, two series of models with larger parameter volumes have been added:

• *-densenet_lite_246-gru_base: Currently available for Knowledge Planet CnOCR/CnSTD Private Group members, it will be open-sourced afterward.
• *-densenet_lite_666-gru_large: Pro models, available for use after purchase. The purchase link: https://ocr.lemonsqueezy.com.

Models in V2.3 are categorized into the following types based on usage scenarios:

• scene: For scene images, suitable for recognizing text in general photography. Models in this category start with scene-, such as the scene-densenet_lite_136-gru model.
• doc: For document images, suitable for recognizing text in regular document screenshots, like scanned book pages. Models in this category start with doc-, such as the doc-densenet_lite_136-gru model.
• number: Specifically for recognizing only numbers (able to recognize only the ten digits 0~9), suitable for scenarios like bank card numbers, ID numbers, etc. Models in this category start with number-, such as the number-densenet_lite_136-gru model.
• general: For general scenarios, suitable for images without a clear preference. Models in this category do not have a specific prefix and maintain the same naming convention as older versions, such as the densenet_lite_136-gru model.

Note ⚠️: The above descriptions are for reference only. It is recommended to choose models based on actual performance.

For more details, see: Available Models.

Future work

• Support for images containing multiple lines of text (Done)
• crnn model support for variable length prediction, improving flexibility (since V1.0.0)
• Refine test cases (Doing)
• Fix bugs (The code is still messy.) (Doing)
• Support space recognition (since V1.1.0)
• Try new models like DenseNet to further improve recognition accuracy (since V1.1.0)
• Optimize the training set to remove unreasonable samples; based on this, retrain each model
• Change from MXNet to PyTorch architecture (since V2.0.0)
• Train more efficient models based on PyTorch
• Support text recognition in column format (since V2.1.2)
• Integration with CnSTD (since V2.2)
• Further optimization of model accuracy
• Support more application scenarios, such as formula recognition, table recognition, layout analysis, etc.

A cup of coffee for the author

It is not easy to maintain and evolve the project, so if it is helpful to you, please consider offering the author a cup of coffee 🥤.

Official code base: https://github.com/breezedeus/cnocr. Please cite it properly.