1.软件包准备
- opencv源码包地址: 官网 github
- opencv_contrib源码包地址: github
- Tesseract源码包地址: github
- cmake.exe 下载地址: 官网
- qt 下载地址: 官网
注意: opencv和open_contrib包的版本号要一致(比如都是3.4.0)
Tesseract源码安装参考: Win10 使用MinGW-w64编译Tesseract4.0
2. 在环境变量PATH中添加:
C:\Qt\Qt5.9.0\5.9\mingw53_32\bin
C:\Qt\Qt5.9.0\Tools\mingw530_32\bin一方面方便日后在cmd中直接使用gcc、g++,qmake和mingw32-make
另一方面，方便下一步cmake查找Qt相关配置
3. 使用cmake生成解决方案
如果提示:
直接将 "CMAKE_SH"项删除即可。
修改配置如下:
- CMAKE_BUILD_TYPE: Debug或者Release
- CMAKE_INSTALL_PREFIX: 指定程序安装位置
- ENABLE_CXX11: 支持c11特性
- WITH_QT
- WITH_OPENGL
- OPENCV_EXTRA_MODULES_PATH: 若使用opencv_contrib模块，则在此处填写解压后的路径，如 F:\opencv_contrib\modules\
- Tesseract_INCLUDE_DIR: Tesseract头文件所在路径
- Tesseract_LIBRARY: Tesseract lib文件所在路径
- Lept_LIBRARY: leptonica lib文件所在路径 (很重要, 一定要配置，否则可能找不到Tesseract)
建议取消勾选:
- BUILD_DOCS ：生成文档，需要安装Doxygen。官网提供了在线文档和离线文档。
- BUILD_PERF_TESTS: 性能测试相关
- BUILD_TESTS: 测试相关
- BUILD_opencv_ts ：一些单元测试代码。
- INSTALL_TESTS ：与开发无关。
配置截图:
Tesseract 相关:
Qt 相关: 
# 模块相关配置
OpenCV modules:
To be built: aruco bgsegm bioinspired calib3d ccalib core datasets dnn dpm face features2d flann fuzzy highgui img_hash imgcodecs imgproc line_descriptor ml objdetect optflow phase_unwrapping photo plot reg rgbd saliency shape stereo stitching structured_light superres surface_matching text tracking video videoio videostab world xfeatures2d ximgproc xobjdetect xphoto
Disabled: js python2 python_bindings_generator cvv
Disabled by dependency: -
Unavailable: cnn_3dobj cudaarithm cudabgsegm cudacodec cudafeatures2d cudafilters cudaimgproc cudalegacy cudaobjdetect cudaoptflow cudastereo cudawarping cudev dnn_modern freetype hdf java matlab ovis python3 python3 sfm ts viz
Applications: apps
Documentation: NO
Non-free algorithms: NO
BUILD_opencv_world : 将.lib或者.dll文件统一整合进一个world文件中，方便使用。但若想只使用一部分模块可不勾选以减少体积
重要:
如果勾选BUILD_opencv_world, 就需要取消勾选BUILD_opencv_cvv,否则会出现以下错误
4.安装:
5. 遇到的问题:
1) 'sprintf_instead_use_StringCbPrintfA_or_StringCchPrintfA' was not declared in this scope .
解决方法: 修改opencv源码目录中modules\videoio\src\cap_dshow.cpp, 找到#include "DShow.h"，然后在其上面添加一行
release版本:
#define NO_DSHOW_STRSAFEdebug版本:
#define STRSAFE_NO_DEPRECATE如下图:
2)
解决方法:
关闭预编译头, 取消勾选"ENABLE_PRECOMPILED_HEADERS"
参考: https://wiki.qt.io/How_to_setup_Qt_and_openCV_on_Windows
6.目录结构
7. 向环境变量PATH中加入dll所在路径
D:\opencv\x86\mingw\bin8. 测试
打开Qtcreator，创建一个c++项目。
测试1:
代码:
#include <opencv2/opencv.hpp>
using namespace cv;
int main()
{
Mat im = imread("lena.png");
namedWindow("Image");
imshow("Image", im);
waitKey(0);
destroyWindow("Image");
return 0;
}.pro配置文件:
TEMPLATE = app
CONFIG += console c++11
CONFIG -= app_bundle
CONFIG -= qt
SOURCES += main.cpp
INCLUDEPATH += D:\opencv\include
CONFIG(debug, debug | release) {
LIBS += D:\opencv\x86\mingw\bin\libopencv_world340d.dll
} else {
LIBS += -LD:\opencv\x86\mingw\lib -lopencv_world340
}
效果:
测试2：
1 /*
2 * textdetection.cpp
3 *
4 * A demo program of End-to-end Scene Text Detection and Recognition:
5 * Shows the use of the Tesseract OCR API with the Extremal Region Filter algorithm described in:
6 * Neumann L., Matas J.: Real-Time Scene Text Localization and Recognition, CVPR 2012
7 *
8 * Created on: Jul 31, 2014
9 * Author: Lluis Gomez i Bigorda <lgomez AT cvc.uab.es>
10 */
11
12 #include "opencv2/text.hpp"
13 #include "opencv2/core/utility.hpp"
14 #include "opencv2/highgui.hpp"
15 #include "opencv2/imgproc.hpp"
16
17 #include <iostream>
18
19 using namespace std;
20 using namespace cv;
21 using namespace cv::text;
22
23 //Calculate edit distance between two words
24 size_t edit_distance(const string& A, const string& B);
25 size_t min(size_t x, size_t y, size_t z);
26 bool isRepetitive(const string& s);
27 bool sort_by_lenght(const string &a, const string &b);
28 //Draw ER's in an image via floodFill
29 void er_draw(vector<Mat> &channels, vector<vector<ERStat> > ®ions, vector<Vec2i> group, Mat& segmentation);
30
31 //Perform text detection and recognition and evaluate results using edit distance
32 int main1(int argc, char* argv[])
33 {
34 cout << endl << argv[0] << endl << endl;
35 cout << "A demo program of End-to-end Scene Text Detection and Recognition: " << endl;
36 cout << "Shows the use of the Tesseract OCR API with the Extremal Region Filter algorithm described in:" << endl;
37 cout << "Neumann L., Matas J.: Real-Time Scene Text Localization and Recognition, CVPR 2012" << endl << endl;
38
39 Mat image;
40
41 if(argc>1)
42 image = imread(argv[1]);
43 else
44 {
45 cout << " Usage: " << argv[0] << " <input_image> [<gt_word1> ... <gt_wordN>]" << endl;
46 return(0);
47 }
48
49 cout << "IMG_W=" << image.cols << endl;
50 cout << "IMG_H=" << image.rows << endl;
51
52 /*Text Detection*/
53
54 // Extract channels to be processed individually
55 vector<Mat> channels;
56
57 Mat grey;
58 cvtColor(image,grey,COLOR_RGB2GRAY);
59
60 // Notice here we are only using grey channel, see textdetection.cpp for example with more channels
61 channels.push_back(grey);
62 channels.push_back(255-grey);
63
64 double t_d = (double)getTickCount();
65 // Create ERFilter objects with the 1st and 2nd stage default classifiers
66 Ptr<ERFilter> er_filter1 = createERFilterNM1(loadClassifierNM1("trained_classifierNM1.xml"),8,0.00015f,0.13f,0.2f,true,0.1f);
67 Ptr<ERFilter> er_filter2 = createERFilterNM2(loadClassifierNM2("trained_classifierNM2.xml"),0.5);
68
69 vector<vector<ERStat> > regions(channels.size());
70 // Apply the default cascade classifier to each independent channel (could be done in parallel)
71 for (int c=0; c<(int)channels.size(); c++)
72 {
73 er_filter1->run(channels[c], regions[c]);
74 er_filter2->run(channels[c], regions[c]);
75 }
76 cout << "TIME_REGION_DETECTION = " << ((double)getTickCount() - t_d)*1000/getTickFrequency() << endl;
77
78 Mat out_img_decomposition= Mat::zeros(image.rows+2, image.cols+2, CV_8UC1);
79 vector<Vec2i> tmp_group;
80 for (int i=0; i<(int)regions.size(); i++)
81 {
82 for (int j=0; j<(int)regions[i].size();j++)
83 {
84 tmp_group.push_back(Vec2i(i,j));
85 }
86 Mat tmp= Mat::zeros(image.rows+2, image.cols+2, CV_8UC1);
87 er_draw(channels, regions, tmp_group, tmp);
88 if (i > 0)
89 tmp = tmp / 2;
90 out_img_decomposition = out_img_decomposition | tmp;
91 tmp_group.clear();
92 }
93
94 double t_g = (double)getTickCount();
95 // Detect character groups
96 vector< vector<Vec2i> > nm_region_groups;
97 vector<Rect> nm_boxes;
98 erGrouping(image, channels, regions, nm_region_groups, nm_boxes,ERGROUPING_ORIENTATION_HORIZ);
99 cout << "TIME_GROUPING = " << ((double)getTickCount() - t_g)*1000/getTickFrequency() << endl;
100
101
102
103 /*Text Recognition (OCR)*/
104
105 double t_r = (double)getTickCount();
106 Ptr<OCRTesseract> ocr = OCRTesseract::create();
107 cout << "TIME_OCR_INITIALIZATION = " << ((double)getTickCount() - t_r)*1000/getTickFrequency() << endl;
108 string output;
109
110 Mat out_img;
111 Mat out_img_detection;
112 Mat out_img_segmentation = Mat::zeros(image.rows+2, image.cols+2, CV_8UC1);
113 image.copyTo(out_img);
114 image.copyTo(out_img_detection);
115 float scale_img = 600.f/image.rows;
116 float scale_font = (float)(2-scale_img)/1.4f;
117 vector<string> words_detection;
118
119 t_r = (double)getTickCount();
120
121 for (int i=0; i<(int)nm_boxes.size(); i++)
122 {
123
124 rectangle(out_img_detection, nm_boxes[i].tl(), nm_boxes[i].br(), Scalar(0,255,255), 3);
125
126 Mat group_img = Mat::zeros(image.rows+2, image.cols+2, CV_8UC1);
127 er_draw(channels, regions, nm_region_groups[i], group_img);
128 Mat group_segmentation;
129 group_img.copyTo(group_segmentation);
130 //image(nm_boxes[i]).copyTo(group_img);
131 group_img(nm_boxes[i]).copyTo(group_img);
132 copyMakeBorder(group_img,group_img,15,15,15,15,BORDER_CONSTANT,Scalar(0));
133
134 vector<Rect> boxes;
135 vector<string> words;
136 vector<float> confidences;
137 ocr->run(group_img, output, &boxes, &words, &confidences, OCR_LEVEL_WORD);
138
139 output.erase(remove(output.begin(), output.end(), '\n'), output.end());
140 //cout << "OCR output = \"" << output << "\" length = " << output.size() << endl;
141 if (output.size() < 3)
142 continue;
143
144 for (int j=0; j<(int)boxes.size(); j++)
145 {
146 boxes[j].x += nm_boxes[i].x-15;
147 boxes[j].y += nm_boxes[i].y-15;
148
149 //cout << " word = " << words[j] << "\t confidence = " << confidences[j] << endl;
150 if ((words[j].size() < 2) || (confidences[j] < 51) ||
151 ((words[j].size()==2) && (words[j][0] == words[j][1])) ||
152 ((words[j].size()< 4) && (confidences[j] < 60)) ||
153 isRepetitive(words[j]))
154 continue;
155 words_detection.push_back(words[j]);
156 rectangle(out_img, boxes[j].tl(), boxes[j].br(), Scalar(255,0,255),3);
157 Size word_size = getTextSize(words[j], FONT_HERSHEY_SIMPLEX, (double)scale_font, (int)(3*scale_font), NULL);
158 rectangle(out_img, boxes[j].tl()-Point(3,word_size.height+3), boxes[j].tl()+Point(word_size.width,0), Scalar(255,0,255),-1);
159 putText(out_img, words[j], boxes[j].tl()-Point(1,1), FONT_HERSHEY_SIMPLEX, scale_font, Scalar(255,255,255),(int)(3*scale_font));
160 out_img_segmentation = out_img_segmentation | group_segmentation;
161 }
162
163 }
164
165 cout << "TIME_OCR = " << ((double)getTickCount() - t_r)*1000/getTickFrequency() << endl;
166
167
168 /* Recognition evaluation with (approximate) Hungarian matching and edit distances */
169
170 if(argc>2)
171 {
172 int num_gt_characters = 0;
173 vector<string> words_gt;
174 for (int i=2; i<argc; i++)
175 {
176 string s = string(argv[i]);
177 if (s.size() > 0)
178 {
179 words_gt.push_back(string(argv[i]));
180 //cout << " GT word " << words_gt[words_gt.size()-1] << endl;
181 num_gt_characters += (int)(words_gt[words_gt.size()-1].size());
182 }
183 }
184
185 if (words_detection.empty())
186 {
187 //cout << endl << "number of characters in gt = " << num_gt_characters << endl;
188 cout << "TOTAL_EDIT_DISTANCE = " << num_gt_characters << endl;
189 cout << "EDIT_DISTANCE_RATIO = 1" << endl;
190 }
191 else
192 {
193
194 sort(words_gt.begin(),words_gt.end(),sort_by_lenght);
195
196 int max_dist=0;
197 vector< vector<int> > assignment_mat;
198 for (int i=0; i<(int)words_gt.size(); i++)
199 {
200 vector<int> assignment_row(words_detection.size(),0);
201 assignment_mat.push_back(assignment_row);
202 for (int j=0; j<(int)words_detection.size(); j++)
203 {
204 assignment_mat[i][j] = (int)(edit_distance(words_gt[i],words_detection[j]));
205 max_dist = max(max_dist,assignment_mat[i][j]);
206 }
207 }
208
209 vector<int> words_detection_matched;
210
211 int total_edit_distance = 0;
212 int tp=0, fp=0, fn=0;
213 for (int search_dist=0; search_dist<=max_dist; search_dist++)
214 {
215 for (int i=0; i<(int)assignment_mat.size(); i++)
216 {
217 int min_dist_idx = (int)distance(assignment_mat[i].begin(),
218 min_element(assignment_mat[i].begin(),assignment_mat[i].end()));
219 if (assignment_mat[i][min_dist_idx] == search_dist)
220 {
221 //cout << " GT word \"" << words_gt[i] << "\" best match \"" << words_detection[min_dist_idx] << "\" with dist " << assignment_mat[i][min_dist_idx] << endl;
222 if(search_dist == 0)
223 tp++;
224 else { fp++; fn++; }
225
226 total_edit_distance += assignment_mat[i][min_dist_idx];
227 words_detection_matched.push_back(min_dist_idx);
228 words_gt.erase(words_gt.begin()+i);
229 assignment_mat.erase(assignment_mat.begin()+i);
230 for (int j=0; j<(int)assignment_mat.size(); j++)
231 {
232 assignment_mat[j][min_dist_idx]=INT_MAX;
233 }
234 i--;
235 }
236 }
237 }
238
239 for (int j=0; j<(int)words_gt.size(); j++)
240 {
241 //cout << " GT word \"" << words_gt[j] << "\" no match found" << endl;
242 fn++;
243 total_edit_distance += (int)words_gt[j].size();
244 }
245 for (int j=0; j<(int)words_detection.size(); j++)
246 {
247 if (find(words_detection_matched.begin(),words_detection_matched.end(),j) == words_detection_matched.end())
248 {
249 //cout << " Detection word \"" << words_detection[j] << "\" no match found" << endl;
250 fp++;
251 total_edit_distance += (int)words_detection[j].size();
252 }
253 }
254
255
256 //cout << endl << "number of characters in gt = " << num_gt_characters << endl;
257 cout << "TOTAL_EDIT_DISTANCE = " << total_edit_distance << endl;
258 cout << "EDIT_DISTANCE_RATIO = " << (float)total_edit_distance / num_gt_characters << endl;
259 cout << "TP = " << tp << endl;
260 cout << "FP = " << fp << endl;
261 cout << "FN = " << fn << endl;
262 }
263 }
264
265
266
267 //resize(out_img_detection,out_img_detection,Size(image.cols*scale_img,image.rows*scale_img),0,0,INTER_LINEAR_EXACT);
268 //imshow("detection", out_img_detection);
269 //imwrite("detection.jpg", out_img_detection);
270 //resize(out_img,out_img,Size(image.cols*scale_img,image.rows*scale_img),0,0,INTER_LINEAR_EXACT);
271 namedWindow("recognition",WINDOW_NORMAL);
272 imshow("recognition", out_img);
273 waitKey(0);
274 //imwrite("recognition.jpg", out_img);
275 //imwrite("segmentation.jpg", out_img_segmentation);
276 //imwrite("decomposition.jpg", out_img_decomposition);
277
278 return 0;
279 }
280
281 size_t min(size_t x, size_t y, size_t z)
282 {
283 return x < y ? min(x,z) : min(y,z);
284 }
285
286 size_t edit_distance(const string& A, const string& B)
287 {
288 size_t NA = A.size();
289 size_t NB = B.size();
290
291 vector< vector<size_t> > M(NA + 1, vector<size_t>(NB + 1));
292
293 for (size_t a = 0; a <= NA; ++a)
294 M[a][0] = a;
295
296 for (size_t b = 0; b <= NB; ++b)
297 M[0][b] = b;
298
299 for (size_t a = 1; a <= NA; ++a)
300 for (size_t b = 1; b <= NB; ++b)
301 {
302 size_t x = M[a-1][b] + 1;
303 size_t y = M[a][b-1] + 1;
304 size_t z = M[a-1][b-1] + (A[a-1] == B[b-1] ? 0 : 1);
305 M[a][b] = min(x,y,z);
306 }
307
308 return M[A.size()][B.size()];
309 }
310
311 bool isRepetitive(const string& s)
312 {
313 int count = 0;
314 for (int i=0; i<(int)s.size(); i++)
315 {
316 if ((s[i] == 'i') ||
317 (s[i] == 'l') ||
318 (s[i] == 'I'))
319 count++;
320 }
321 if (count > ((int)s.size()+1)/2)
322 {
323 return true;
324 }
325 return false;
326 }
327
328
329 void er_draw(vector<Mat> &channels, vector<vector<ERStat> > ®ions, vector<Vec2i> group, Mat& segmentation)
330 {
331 for (int r=0; r<(int)group.size(); r++)
332 {
333 ERStat er = regions[group[r][0]][group[r][1]];
334 if (er.parent != NULL) // deprecate the root region
335 {
336 int newMaskVal = 255;
337 int flags = 4 + (newMaskVal << 8) + FLOODFILL_FIXED_RANGE + FLOODFILL_MASK_ONLY;
338 floodFill(channels[group[r][0]],segmentation,Point(er.pixel%channels[group[r][0]].cols,er.pixel/channels[group[r][0]].cols),
339 Scalar(255),0,Scalar(er.level),Scalar(0),flags);
340 }
341 }
342 }
343
344 bool sort_by_lenght(const string &a, const string &b){return (a.size()>b.size());}测试用到的文件end_to_end_recognition.cpp、scenetext01.jpg、trained_classifierNM1.xml、trained_classifierNM2.xml都位于opencv_contrib源码包目录下的modules\text\samples中
效果:
来源：oschina
链接：https://my.oschina.net/u/4340253/blog/4043729