Skip to content

Latest commit

 

History

History
273 lines (205 loc) · 15.1 KB

README.md

File metadata and controls

273 lines (205 loc) · 15.1 KB
Raw

EasyOCR

PyPI Status license Open in Colab Tweet Twitter

Ready-to-use OCR with 40+ languages supported including Chinese, Japanese, Korean and Thai.

What's new?

  • 10 August 2020 - Version 1.1.7
    • New language support for Tamil
    • Temporary fix for memory leakage on CPU mode
  • 4 August 2020 - Version 1.1.6
    • New language support for Russian, Serbian, Belarusian, Bulgarian, Mongolian, Ukranian (Cyrillic Script) and Arabic, Persian(Farsi), Urdu, Uyghur (Arabic Script)
    • Docker file and Ainize demo (thanks @ghandic and @Wook-2)
    • Better production friendly with Logger and custom model folder location (By setting model_storage_directory when create Reader instance) (thanks @jpotter)
  • Read all released notes

What's coming next?

Examples

example

example2

example3

Supported Languages

We are currently supporting the following 59 languages.

Afrikaans (af), Arabic (ar), Azerbaijani (az), Belarusian (be), Bulgarian (bg), Bosnian (bs), Simplified Chinese (ch_sim), Traditional Chinese (ch_tra), Czech (cs), Welsh (cy), Danish (da), German (de), English (en), Spanish (es), Estonian (et), Persian (Farsi) (fa) French (fr), Irish (ga), Hindi(hi), Croatian (hr), Hungarian (hu), Indonesian (id), Icelandic (is), Italian (it), Japanese (ja), Korean (ko), Kurdish (ku), Latin (la), Lithuanian (lt), Latvian (lv), Maori (mi), Mongolian (mn), Marathi (mr), Malay (ms), Maltese (mt), Nepali (ne), Dutch (nl), Norwegian (no), Occitan (oc), Polish (pl), Portuguese (pt), Romanian (ro), Russian (ru), Serbian (cyrillic)(rs_cyrillic), Serbian (latin)(rs_latin), Slovak (sk) (need revisit), Slovenian (sl), Albanian (sq), Swedish (sv), Swahili (sw), Tamil (ta), Thai (th), Tagalog (tl), Turkish (tr), Uyghur (ug), Ukranian(uk), Urdu (ur), Uzbek (uz), Vietnamese (vi) (need revisit)

List of characters is in folder easyocr/character. If you are native speaker of any language and think we should add or remove any character, please create an issue and/or pull request (like this one).

Installation

Install using pip for stable release,

pip install easyocr

For latest development release,

pip install git+git://github.com/jaidedai/easyocr.git

Note 1: for Windows, please install torch and torchvision first by following the official instruction here https://pytorch.org. On pytorch website, be sure to select the right CUDA version you have. If you intend to run on CPU mode only, select CUDA = None.

Note 2: We also provide Dockerfile here.

Try Third-Party Demos

  1. Google Colab
  2. Dockerhub
  3. Ainize

Usage

import easyocr
reader = easyocr.Reader(['ch_sim','en']) # need to run only once to load model into memory
result = reader.readtext('chinese.jpg')

Output will be in list format, each item represents bounding box, text and confident level, respectively.

[([[189, 75], [469, 75], [469, 165], [189, 165]], '愚园路', 0.3754989504814148),
 ([[86, 80], [134, 80], [134, 128], [86, 128]], '西', 0.40452659130096436),
 ([[517, 81], [565, 81], [565, 123], [517, 123]], '东', 0.9989598989486694),
 ([[78, 126], [136, 126], [136, 156], [78, 156]], '315', 0.8125889301300049),
 ([[514, 126], [574, 126], [574, 156], [514, 156]], '309', 0.4971577227115631),
 ([[226, 170], [414, 170], [414, 220], [226, 220]], 'Yuyuan Rd.', 0.8261902332305908),
 ([[79, 173], [125, 173], [125, 213], [79, 213]], 'W', 0.9848111271858215),
 ([[529, 173], [569, 173], [569, 213], [529, 213]], 'E', 0.8405593633651733)]

Note 1: ['ch_sim','en'] is the list of languages you want to read. You can pass several languages at once but not all languages can be used together. English is compatible with every languages. Languages that share common characters are usually compatible with each other.

Note 2: Instead of filepath chinese.jpg, you can also pass OpenCV image object (numpy array) or image file as bytes. URL to raw image is also acceptable.

Note 3: The line reader = easyocr.Reader(['ch_sim','en']) is for loading model into memory. It takes some time but it need to be run only once.

You can also set detail = 0 for simpler output.

reader.readtext('chinese.jpg', detail = 0)

Result:

['愚园路', '西', '东', '315', '309', 'Yuyuan Rd.', 'W', 'E']

Model weight for chosen language will be automatically downloaded or you can download it manually from the following links and put it in '~/.EasyOCR/model' folder

In case you do not have GPU or your GPU has low memory, you can run it in CPU mode by adding gpu = False

reader = easyocr.Reader(['ch_sim','en'], gpu = False)

For more information, read tutorial and API Documentation.

Run on command line

$ easyocr -l ch_sim en -f chinese.jpg --detail=1 --gpu=True

Implementation Roadmap

Phase 1 (Now - October, 2020)

  1. Language packs: Hindi, Arabic, Cyrillic alphabet, etc. Aiming to cover > 80-90% of world's population. See current development list. Also improve existing languages.
  2. Better documentation and api
  3. Language model for better decoding

Phase 2 (After October, 2020)

  1. Handwritten support: Network architecture should not matter. The key is using GAN to generate realistic handwritten dataset.
  2. Faster processing time: model pruning (lite version) / quantization / export to other platforms (ONNX?)
  3. Open Dataset and model training pipeline
  4. Restructure code to support swappable detection and recognition algorithm. The api should be as easy as
reader = easyocr.Reader(['en'], detection='DB', recognition = 'CNN_Transformer')

The idea is to be able to plug-in any state-of-the-art model into EasyOCR. There are a lot of geniuses trying to make better detection/recognition model. We are not trying to be a genius here, just make genius's works quickly accessible to the public ... for free. (well I believe most geniuses want their work to create positive impact as fast/big as possible) The pipeline should be something like below diagram. Grey slots are placeholders for changeable light blue modules.

plan

Acknowledgement and References

This project is based on researches/codes from several papers/open-source repositories.

All deep learning part is based on Pytorch. ❤️

Detection part is using CRAFT algorithm from this official repository and their paper (Thanks @YoungminBaek from @clovaai). We also use their pretrained model.

Recognition model is CRNN (paper). It is composed of 3 main components, feature extraction (we are currently using Resnet), sequence labeling (LSTM) and decoding (CTC). Training pipeline for recognition part is a modified version from deep-text-recognition-benchmark. (Thanks @ku21fan from @clovaai) This repository is a gem that deserved more recognition.

Beam search code is based on this repository and his blog. (Thanks @githubharald)

Data synthesis is based on TextRecognitionDataGenerator. (Thanks @Belval)

And good read about CTC from distill.pub here.

Want To Contribute?

Let's advance humanity together by making AI available to everyone!

3 ways to contribute:

Coder: Please send PR for small bug/improvement. For bigger one, discuss with us by open an issue first. There is a list of possible bug/improvement issue tagged with 'PR WELCOME'.

User: Tell us how EasyOCR benefit you/your organization to encourage further development. Also post failure cases in Issue Section to help improving future model.

Tech leader/Guru: If you found this library useful, please spread the word! (See Yann Lecun's post about EasyOCR)

Guideline for new language request

To request a new language support, I need you to send a PR with 2 following files

  1. In folder easyocr/character, we need 'yourlanguagecode_char.txt' that contains list of all characters. Please see format example from other files in that folder.
  2. In folder easyocr/dict, we need 'yourlanguagecode.txt' that contains list of words in your language. On average we have ~30000 words per language with more than 50000 words for popular one. More is better in this file.

If your language has unique elements (such as 1. Arabic: characters change form when attach to each other + write from right to left 2. Thai: Some characters need to be above the line and some below), please educate me with your best ability and/or give useful links. It is important to take care of the detail to achieve a system that really works.

Lastly, please understand that my priority will have to go to popular language or set of languages that share most of characters together (also tell me if your language share a lot of characters with other). It takes me at least a week to work for new model. You may have to wait a while for new model to be released.

See List of languages in development

Running TorchServe

To launch TorchServe and initialize both CRAFT and recognition models:

bin/launch_ts

To stop TorchServe:

torchserve --stop

API Documentation

Reader class

Base class for EasyOCR

Parameters

  • lang_list (list) - list of language code you want to recognize, for example ['ch_sim','en']. List of supported language code is here.
  • gpu (bool, string, default = True)
  • model_storage_directory (string, default = None)
  • download_enabled (bool, default = True)

Attribute

  • lang_char - Show all available characters in current model

readtext method

Main method for Reader object. There are 4 groups of parameter: General, Contrast, Text Detection and Bounding Box Merging.

Parameters 1: General

  • image (string, numpy array, byte) - Input image
  • decoder (string, default = 'greedy') - options are 'greedy', 'beamsearch' and 'wordbeamsearch'.
  • beamWidth (int, default = 5) - How many beam to keep when decoder = 'beamsearch' or 'wordbeamsearch'
  • batch_size (int, default = 1) - batch_size>1 will make EasyOCR faster but use more memory
  • workers (int, default = 0) - Number thread used in of dataloader
  • allowlist (string) - Force EasyOCR to recognize only subset of characters. Useful for specific problem (E.g. license plate, etc.)
  • blocklist (string) - Block subset of character. This argument will be ignored if allowlist is given.
  • detail (int, default = 1) - Set this to 0 for simple output
  • paragraph (bool, default = False) - Combine result into paragraph

Parameters 2: Contrast

  • contrast_ths (float, default = 0.1) - Text box with contrast lower than this value will be passed into model 2 times. First is with original image and second with contrast adjusted to 'adjust_contrast' value. The one with more confident level will be returned as a result.
  • adjust_contrast (float, default = 0.5) - target contrast level for low contrast text box

Parameters 3: Text Detection (from CRAFT)

  • text_threshold (float, default = 0.7) - Text confidence threshold
  • low_text (float, default = 0.4) - Text low-bound score
  • link_threshold (float, default = 0.4) - Link confidence threshold
  • canvas_size (int, default = 2560) - Maximum image size. Image bigger than this value will be resized down.
  • mag_ratio (float, default = 1) - Image magnification ratio

Parameters 4: Bounding Box Merging

This set of parameter controls when adjacent bounding boxes merge with each other. Every parameters except 'slope_ths' is in the unit of box height.

width_ths

  • slope_ths (float, default = 0.1) - Maximum slope (delta y/delta x) to considered merging. Low value means tiled boxes will not be merged.
  • ycenter_ths (float, default = 0.5) - Maximum shift in y direction. Boxes with different level should not be merged.
  • height_ths (float, default = 0.5) - Maximum different in box height. Boxes with very different text size should not be merged.
  • width_ths (float, default = 0.5) - Maximum horizontal distance to merge boxes.
  • add_margin (float, default = 0.1) - Extend bounding boxes in all direction by certain value. This is important for language with complex script (E.g. Thai).

Return (list)