opencv 12 OCR 端到端识别(end_to_end_recognition) vs2015

01 资源

OpenCV自带的OCR端到端识别用例，opencv_contrib\modules\text\samples\end_to_end_recognition.cpp。end_to_end_recognition.cpp识别图片中的字符串。
模型文件(需要拷贝到Debug或Release目录中)：

opencv_contrib/modules/text/samples/trained_classifierNM1.xml
opencv_contrib/modules/text/samples/trained_classifierNM2.xml

测试图片：

opencv_contrib/modules/text/samples/scenetext_segmented_word01.jpg
opencv_contrib/modules/text/samples/scenetext_segmented_word01_mask.png
opencv_contrib/modules/text/samples/scenetext_segmented_word02.jpg
opencv_contrib/modules/text/samples/scenetext_segmented_word02_mask.png
opencv_contrib/modules/text/samples/scenetext_segmented_word03.jpg
opencv_contrib/modules/text/samples/scenetext_segmented_word03_mask.png
opencv_contrib/modules/text/samples/scenetext_segmented_word04.jpg
opencv_contrib/modules/text/samples/scenetext_segmented_word04_mask.png
opencv_contrib/modules/text/samples/scenetext_segmented_word05.jpg
opencv_contrib/modules/text/samples/scenetext_segmented_word05_mask.png

tesseract源码中的数据：tesseract\tessdata拷贝到目标目录。
把D:\git\DeepLearning\tesseract\tessdata拷贝到D:\git\opencv\build\v3.3.0\x64\bin\Debug\tessdata

还需要的tesseract数据eng.traineddata：
https://github.com/tesseract-ocr/tesseract/wiki/Data-Files
代码使用的3.05.01，这里下载3.05版本数据。
https://github.com/tesseract-ocr/tesseract/wiki/Data-Files#data-files-for-version-304305

把eng.traineddata放到D:\git\opencv\build\v3.3.0\x64\bin\Debug\tessdata中。

02 编译配置tesseract

end_to_end_recognition使用tesseract做OCR端到端处理。
有blog说，只要安装tesseract开发版即可，没亲自测试。下载地址：https://digi.bib.uni-mannheim.de/tesseract/。
自己配置tesseract。自己编译的tesseract和leptonica，tesseract依赖leptonica。
下载leptonica、tesseract源码编译。

02.01 编译leptonica

下载代码

cd D:\git\DeepLearning
git clone https://github.com/DanBloomberg/leptonica.git
cd leptonica
git tag # 当前稳定版本1.74.4
# 切换到稳定版本
git checkout -b b1.74.4 1.74.4

配置cmake
目标路径：D:/git/DeepLearning/leptonica/build/x64

BUILD_PROG=1
CMAKE_INSTALLPREFIX=D:/git/DeepLearning/leptonica/build/x64/install

如果系统已经配置好了GIF/JPEG/ZLIB/PNG/TIFF等，勾选Advanced选项，把这几项相关的内容清空，这样会采用leptonica的默认配置。

编译、安装
使用vs2015 update3 编译，并且运行INSTALL项目。把编译好的内容安装到D:/git/DeepLearning/leptonica/build/x64/install目录。

02.02 编译tesseract

下载代码

cd D:\git\DeepLearning
git clone https://github.com/tesseract-ocr/tesseract.git
cd tesseract
git tag # 当前稳定版本1.74.4
# 切换到稳定版本
git checkout -b b3.05.01 3.05.01

配置cmake
目标路径：D:/git/DeepLearning/tesseract/build/x64

BUILD_TRAINING_TOOLS=1
CMAKE_INSTALL_PREFIX=D:/git/DeepLearning/tesseract/build/x64/install
Leptonica_DIR=D:/git/DeepLearning/leptonica/build/x64/install/cmake

编译、安装
使用vs2015 update3 编译，并且运行INSTALL项目。把编译好的内容安装到D:/git/DeepLearning/tesseract/build/x64/install目录。

配置环境变量path
在系统path环境变量中追加

D:/git/DeepLearning/leptonica/build/x64/install
D:/git/DeepLearning/tesseract/build/x64/install

03 编译end_to_end_recognition

03.01配置opencv的cmake

opencv3.3.0的cmake配置，参考opencv01 相对完整的编译opencv3.3.0 win版本。这里并没有配置tesseract。
追加tesseract配置：

Lept_LIBRARY=D:/git/DeepLearning/leptonica/build/x64/install/lib/leptonica-1.74.4.lib
Tesseract_INCLUDE_DIR=D:/git/DeepLearning/tesseract/build/x64/install/include
Tesseract_LIBRARY=D:/git/DeepLearning/tesseract/build/x64/install/lib/tesseract305.lib

03.02 编译end_to_end_recognition

编译opencv3.3.0，生成文件：D:\git\opencv\build\v3.3.0\x64\bin\Debug\end_to_end_recognition.exe

03.03 配置end_to_end_recognition工程

设置end_to_end_recognition项目为启动项。

# 如果路径中有空格，需要使用双引号,参数路径根据自己实际情况调整
配置属性==>调试==>命令参数=../../../../../../opencv_contrib/modules/text/samples/scenetext_segmented_word01_mask.png
配置属性==>调试==>工作目录=$(OutDir)

04 运行结果

04.01 scenetext_segmented_word01.jpg

scenetext_segmented_word01.jpg原图：
运行效果图：

A demo program of End-to-end Scene Text Detection and Recognition:
Shows the use of the Tesseract OCR API with the Extremal Region Filter algorithm described in:
Neumann L., Matas J.: Real-Time Scene Text Localization and Recognition, CVPR 2012

IMG_W=640
IMG_H=480
TIME_REGION_DETECTION = 14750.3
TIME_GROUPING = 4365.56
TIME_OCR_INITIALIZATION = 638.67
Empty page!!
Empty page!!
TIME_OCR = 907.957

判断结果为”HOTELI”，多了一个字母”I”。

04.02 scenetext_segmented_word01_mask.png

scenetext_segmented_word01_mask.png原图：
运行效果图：
判断结果为”HOTEL”，准确。