「opencv」神经网络识别字母+数字

附件从原文下载即可。

1. 
   
2. //opencv2.4.9 + vs2012 + 64位
3. #include <windows.h>
4. #include <iostream>
5. #include <fstream>
6. #include <opencv2/opencv.hpp>
8. using namespace cv;
9. using namespace std;
11. char* WcharToChar(const wchar_t* wp)
12. {
13. char *m_char;
14. int len= WideCharToMultiByte(CP_ACP,0,wp,wcslen(wp),NULL,0,NULL,NULL);
15. m_char=new char[len+1];
16. WideCharToMultiByte(CP_ACP,0,wp,wcslen(wp),m_char,len,NULL,NULL);
17. m_char[len]='\0';
18. return m_char;
19. }
21. wchar_t* CharToWchar(const char* c)
22. {
23. wchar_t *m_wchar;
24. int len = MultiByteToWideChar(CP_ACP,0,c,strlen(c),NULL,0);
25. m_wchar=new wchar_t[len+1];
26. MultiByteToWideChar(CP_ACP,0,c,strlen(c),m_wchar,len);
27. m_wchar[len]='\0';
28. return m_wchar;
29. }
31. wchar_t* StringToWchar(const string& s)
32. {
33. const char* p=s.c_str();
34. return CharToWchar(p);
35. }
37. int main()
38. {
39. const string fileform = "*.png";
40. const string perfileReadPath = "charSamples";
42. const int sample_mun_perclass = 20;//训练字符每类数量
43. const int class_mun = 10+26;//训练字符类数 0-9 A-Z 除了I、O
45. const int image_cols = 8;
46. const int image_rows = 16;
47. string fileReadName,
48. fileReadPath;
49. char temp[256];
51. float trainingData[class_mun*sample_mun_perclass][image_rows*image_cols] = {{0}};//每一行一个训练样本
52. float labels[class_mun*sample_mun_perclass][class_mun]={{0}};//训练样本标签
54. for(int i = 0; i <= class_mun - 1; i++)//不同类
55. {
56. //读取每个类文件夹下所有图像
57. int j = 0;//每一类读取图像个数计数
59. if (i <= 9)//0-9
60. {
61. sprintf(temp, "%d", i);
62. //printf("%d\n", i);
63. }
64. else//A-Z
65. {
66. sprintf(temp, "%c", i + 55);
67. //printf("%c\n", i+55);
68. }
70. fileReadPath = perfileReadPath + "/" + temp + "/" + fileform;
71. cout<<"文件夹"<<temp<<endl;
73. HANDLE hFile;
74. LPCTSTR lpFileName = StringToWchar(fileReadPath);//指定搜索目录和文件类型，如搜索d盘的音频文件可以是"D:\\*.mp3"
75. WIN32_FIND_DATA pNextInfo; //搜索得到的文件信息将储存在pNextInfo中;
76. hFile = FindFirstFile(lpFileName,&pNextInfo);//请注意是 &pNextInfo , 不是 pNextInfo;
77. if(hFile == INVALID_HANDLE_VALUE)
78. {
79. continue;//搜索失败
80. }
81. //do-while循环读取
82. do
83. {
84. if(pNextInfo.cFileName[0] == '.')//过滤.和..
85. continue;
86. j++;//读取一张图
87. //wcout<<pNextInfo.cFileName<<endl;
88. //printf("%s\n",WcharToChar(pNextInfo.cFileName));
89. //对读入的图片进行处理
90. Mat srcImage = imread( perfileReadPath + "/" + temp + "/" + WcharToChar(pNextInfo.cFileName),CV_LOAD_IMAGE_GRAYSCALE);
91. Mat resizeImage;
92. Mat trainImage;
93. Mat result;
95. resize(srcImage,resizeImage,Size(image_cols,image_rows),(0,0),(0,0),CV_INTER_AREA);//使用象素关系重采样。当图像缩小时候，该方法可以避免波纹出现
96. threshold(resizeImage,trainImage,0,255,CV_THRESH_BINARY|CV_THRESH_OTSU);
98. for(int k = 0; k<image_rows*image_cols; ++k)
99. {
100. trainingData[i*sample_mun_perclass+(j-1)][k] = (float)trainImage.data[k];
101. //trainingData[i*sample_mun_perclass+(j-1)][k] = (float)trainImage.at<unsigned char>((int)k/8,(int)k%8);//(float)train_image.data[k];
102. //cout<<trainingData[i*sample_mun_perclass+(j-1)][k] <<" "<< (float)trainImage.at<unsigned char>(k/8,k%8)<<endl;
103. }
105. }while (FindNextFile(hFile,&pNextInfo) && j<sample_mun_perclass);//如果设置读入的图片数量，则以设置的为准，如果图片不够，则读取文件夹下所有图片
107. }
109. // Set up training data Mat
110. Mat trainingDataMat(class_mun*sample_mun_perclass, image_rows*image_cols, CV_32FC1, trainingData);
111. cout<<"trainingDataMat——OK！"<<endl;
113. // Set up label data
114. for(int i = 0;i <= class_mun-1; ++i)
115. {
116. for(int j = 0;j <= sample_mun_perclass - 1; ++j)
117. {
118. for(int k = 0;k < class_mun; ++k)
119. {
120. if(k == i)
121. if (k == 18)
122. {
123. labels[i*sample_mun_perclass + j][1] = 1;
124. }
125. else if(k == 24)
126. {
127. labels[i*sample_mun_perclass + j][0] = 1;
128. }
129. else
130. {
131. labels[i*sample_mun_perclass + j][k] = 1;
132. }
133. else
134. labels[i*sample_mun_perclass + j][k] = 0;
135. }
136. }
137. }
138. Mat labelsMat(class_mun*sample_mun_perclass, class_mun, CV_32FC1,labels);
139. cout<<"labelsMat:"<<endl;
140. ofstream outfile("out.txt");
141. outfile<<labelsMat;
142. //cout<<labelsMat<<endl;
143. cout<<"labelsMat——OK！"<<endl;
145. //训练代码
147. cout<<"training start...."<<endl;
148. CvANN_MLP bp;
149. // Set up BPNetwork's parameters
150. CvANN_MLP_TrainParams params;
151. params.train_method=CvANN_MLP_TrainParams::BACKPROP;
152. params.bp_dw_scale=0.001;
153. params.bp_moment_scale=0.1;
154. params.term_crit = cvTermCriteria(CV_TERMCRIT_ITER|CV_TERMCRIT_EPS,10000,0.0001); //设置结束条件
155. //params.train_method=CvANN_MLP_TrainParams::RPROP;
156. //params.rp_dw0 = 0.1;
157. //params.rp_dw_plus = 1.2;
158. //params.rp_dw_minus = 0.5;
159. //params.rp_dw_min = FLT_EPSILON;
160. //params.rp_dw_max = 50.;
162. //Setup the BPNetwork
163. Mat layerSizes=(Mat_<int>(1,5) << image_rows*image_cols,128,128,128,class_mun);
164. bp.create(layerSizes,CvANN_MLP::SIGMOID_SYM,1.0,1.0);//CvANN_MLP::SIGMOID_SYM
165. //CvANN_MLP::GAUSSIAN
166. //CvANN_MLP::IDENTITY
167. cout<<"training...."<<endl;
168. bp.train(trainingDataMat, labelsMat, Mat(),Mat(), params);
170. bp.save("../bpcharModel.xml"); //save classifier
171. cout<<"training finish...bpModel1.xml saved "<<endl;
174. //测试神经网络
175. cout<<"测试："<<endl;
176. Mat test_image = imread("test4.png",CV_LOAD_IMAGE_GRAYSCALE);
177. Mat test_temp;
178. resize(test_image,test_temp,Size(image_cols,image_rows),(0,0),(0,0),CV_INTER_AREA);//使用象素关系重采样。当图像缩小时候，该方法可以避免波纹出现
179. threshold(test_temp,test_temp,0,255,CV_THRESH_BINARY|CV_THRESH_OTSU);
180. Mat_<float>sampleMat(1,image_rows*image_cols);
181. for(int i = 0; i<image_rows*image_cols; ++i)
182. {
183. sampleMat.at<float>(0,i) = (float)test_temp.at<uchar>(i/8,i%8);
184. }
186. Mat responseMat;
187. bp.predict(sampleMat,responseMat);
188. Point maxLoc;
189. double maxVal = 0;
190. minMaxLoc(responseMat,NULL,&maxVal,NULL,&maxLoc);
192. if (maxLoc.x <= 9)//0-9
193. {
194. sprintf(temp, "%d", maxLoc.x);
195. //printf("%d\n", i);
196. }
197. else//A-Z
198. {
199. sprintf(temp, "%c", maxLoc.x + 55);
200. //printf("%c\n", i+55);
201. }
203. cout<<"识别结果："<<temp<<"相似度:"<<maxVal*100<<"%"<<endl;
204. imshow("test_image",test_image);
205. waitKey(0);
207. return 0;
208. }