Skip to content
/ tr_ocr Public
forked from myhub/tr

Free Offline OCR 离线的中文文本检测+识别SDK

License

Notifications You must be signed in to change notification settings

nfyxhan/tr_ocr

 
 

Repository files navigation

tr - Text Recognition

一款针对扫描文档的离线文本识别SDK,核心代码全部采用C++开发,并提供Python接口

编译环境: Ubuntu 16.04

更新说明

  • c++接口支持
  • 添加python2支持
  • 去除opencv-python、Pillow依赖,降低部署难度
  • 支持多线程
  • 支持GPU
  • 取消对Windows系统的支持

Requirements

  • python2/python3,需要安装numpy
  • 不支持Windows、CentOS 6、ARM

GPU版本安装说明

如果对速度有要求,推荐安装GPU版本
要使用GPU版本,复制tr_gpu文件夹里面的文件到tr文件夹
注意: 需要先安装CUDA 10.1以及cuDNN 7.6.5

若不想安装CUDA/cuDNN,可以使用docker部署

docker pull mcr.microsoft.com/azureml/onnxruntime:v1.3.0-cuda10.1-cudnn7
sudo nvidia-docker run -v /path/to/tr:/path/to/tr --rm -it mcr.microsoft.com/azureml/onnxruntime:v1.3.0-cuda10.1-cudnn7

Install

  • 安装方法一
git clone https://github.com/myhub/tr.git
cd ./tr
sudo python setup.py install
  • 安装方法二
sudo pip install git+https://github.com/myhub/tr.git@master

Test

python2 demo.py               # python2兼容测试
python3 test.py               # 可视化测试
python3 test-multi-thread.py  # 多线程测试

关联项目

  • 若需要Web端调用,推荐参考TrWebOCR

Python Example

import tr

# detect text lines, return list of (cx, cy, width, height, angle)
print(tr.detect("imgs/web.png", tr.FLAG_RECT))

# detect text lines with angle, return list of (cx, cy, width, height, angle)
print(tr.detect("imgs/id_card.jpeg", tr.FLAG_ROTATED_RECT))

# recognize text line, return (text, confidence)
print(tr.recognize("imgs/line.png"))

# detect and recognize text lines with angle, return list of ((cx, cy, width, height, angle), text, confidence)
print(tr.run("imgs/id_card.jpeg"))

C++ Example

tr_init(0, 0, "crnn.bin", NULL);

#define MAX_WIDTH		512
int unicode[MAX_WIDTH];
float prob[MAX_WIDTH]; 

auto ws = tr_recognize(0, (void *)"line.png", 0, 0, 0, unicode, prob, MAX_WIDTH);

tr_release(0);

效果展示


About

Free Offline OCR 离线的中文文本检测+识别SDK

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 73.5%
  • HTML 22.1%
  • Dockerfile 3.9%
  • Shell 0.5%