大框已经接近收尾。接下来剩下：
1）噪声添加
2）类标返回
3）汉字负样本添加
4）困难样本回收训练
5）......
这些细节，更加提高准确率。
1 %{
2 clear all;
3 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
4 % 可调参数
5
6 %test_path=‘C:\Users\cong\Desktop\研一实战\项目\图像中时间数字识别\OCR\one\22.jpg‘;%####
7 %test_path=‘C:\Users\cong\Desktop\研一实战\项目\图像中时间数字识别\big.jpg‘;
8 %target_digit=2;
9 %local_step=3;%####
10
11 %滑窗检测,窗口尺度，步长
12 % win_h=20;
13 % win_w=20;
14
15 % step=1;
16
17 disp(‘参数配置成功...‘);
18 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
19 disp(‘正在读入图片及特征提取...‘);
20 %读入图片及特征提取
21 %data=[];
22 %label=[];
23 temp_r1=0;
24 temp_c1=0;
25 k_data=1;
26 %k_label=1;
27 for i_digit=0:1:10
28 % if(i_digit==target_digit) %%%%%%%%%%%%%%%%%%%%%%
29 % this_image_label=1;
30 % end
31 %数字转字符
32 str=num2str(i_digit); %%数据是不是不平衡
33 path_temp=strcat(‘C:\Users\cong\Desktop\研一实战\项目\图像中时间数字识别\trainingSample\‘,str,‘\‘);
34 %path_temp=strcat(‘C:\Users\cong\Desktop\研一实战\项目\图像中时间数字识别\trainingSample\‘,str,‘\‘);
35 %path_temp=strcat(‘C:\Users\cong\Desktop\研一实战\项目\图像中时间数字识别\trainingSample\num‘,str,‘\‘);
36 file=dir(path_temp);
37 for i=3:length(file)
38 path= strcat(path_temp,file(i).name);
39
40 %%%%%%%%%%%%%%%%%%%%%%%%%%
41 % 加载图片
42 %%%%%%%%%%%%%%%%%%%%%%%%%%
43 I=imread(path);
44 % if(i_digit==10)
45 % I=imresize(I,[80 70]);
46 % end
47 %I=imread(‘E:/WeChat.jpg‘);
48 %%%%%%%%%%%%%%%%%%%%%%%%%%
49 % 提取channel features
50 %%%%%%%%%%%%%%%%%%%%%%%%%%
51 %[all_channel_difference_features,temp_r1,temp_c1]=channel_feature_2_vector(I,1);
52 %data=[data,all_channel_difference_features];
53 %[d3_data,~,~]=channel_feature_2_vector(I,1);
54 pChns=chnsCompute();
55 chns = chnsCompute( I, pChns );
56 %all_channel_difference_features=cat(3,chns.data{:});
57 %data(:,:,:,k_data)=single(d3_data);
58 data(:,:,:,k_data)=cat(3,chns.data{:});
59 label_t(1,k_data)=i_digit;
60 k_data=k_data+1;
61 %d3_data=channel_feature_2_vector(I,1);
62 %data=[data,d3_data];
63 %label=[label;i_digit];%类标从0开始
64 if(rem(i,500)==0)
65 disp(‘...‘);
66 end
67 end % for i=3:length(file)
68 disp(‘数字‘)
69 i_digit
70 disp(‘的特征提取完毕...‘);
71 end % for i_digit=0:9
72 disp(‘读入图片及特征提取完毕...‘);
73 %%%%%%%%%%%%%%%%%%%%%%%%%%
74 % 扔进分类器中，训练
75 %%%%%%%%%%%%%%%%%%%%%%%%%%
76 %参数设置
77 %data=data‘;
78 param.nclass = 11; % number of classes
79 param.ntree = 100; % number of tree
80 param.depth = 16; % the maximum depth of tree
81 param.sigma = 5; % random parameter
82
83 %%
84 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
85 %训练
86 disp(‘正在训练，请稍等...‘);
87 handle = crndforest(); % call constructor
88
89 %data=single(data);
90 label=int32(label_t);
91 crndforest(handle, ‘train‘, data, label, param); %
92 crndforest(handle, ‘save‘, ‘forest_model2‘); % optionally save into disk
93 crndforest(handle, ‘delete‘); % release object
94 disp(‘训练完毕...‘);
95 %}
96 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
97 %测试
98 %{
99 clear all;
100 data_num=1;
101 disp(‘正在测试...‘);
102 handle = crndforest();
103 crndforest(handle, ‘load‘, ‘forest_model2‘);
104 %载入图像
105 %输入data，提取特征
106 test_digit=2;%%%%%%%%%%%%%%%%%%%%%
107 all_test_result=[];
108 %all_test_backgroud=[];
109 %all_channel_difference_features=[];
110 for test_digit=0:10
111 %test_backgroud=[];
112 test_result=0;
113 str=num2str(test_digit);
114 path_temp=strcat(‘C:\Users\cong\Desktop\研一实战\项目\图像中时间数字识别\trainingSample\no‘,str,‘\‘);
115 file=dir(path_temp);
116
117 for i=3:length(file)
118 path= strcat(path_temp,file(i).name);
119 I=imread(path);
120 % I=imresize(I,[80 70]);
121 %[all_channel_difference_features,temp_r1,temp_c1]=channel_feature_2_vector(I,1);%all_channel_difference_features就是data
122 %随机森林预测
123
124 pChns=chnsCompute();
125 chns = chnsCompute( I, pChns );
126 all_channel_difference_features(:,:,:,data_num)=cat(3,chns.data{:});
127 label_test_t(data_num)=test_digit;
128 %all_channel_difference_features=single(all_channel_difference_features);
129 data_num=data_num+1;
130 end
131 data_num=1;
132 [pred_label, prob] =crndforest(handle, ‘predict‘, all_channel_difference_features);
133 label_test=int32(label_test_t);
134 label_test=label_test‘;
135 test_result=sum(pred_label==label_test);
136 test_result
137 test_result=test_result/(length(file)-2);
138 % test_backgroud
139 %test_result=test_result/198;
140 all_test_result=[ all_test_result,test_result];
141 %all_test_backgroud=[ all_test_backgroud,test_backgroud];
142 disp(‘检测数字：‘)
143 test_digit+1
144 end
145 %输出结果
146 disp(‘恭喜恭喜，结果出来啦！‘)
147 all_test_result
148 crndforest(handle, ‘delete‘);
149 disp(‘测试完毕...‘);
150 %}
151 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
152 %}
153 % 检测，测试
154 %参数设置
155 clear all;
156 scale_num=5;
157 scale_factor=1.2;
158 all_channel_difference_features=single(zeros(11,8,10,2));%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%先这么滴
159 %输入图，并提取特征
160 test_path=‘E:\研一实战\项目\图像中时间数字识别\OCR\one\3.jpg‘;
161 I=imread(test_path);
162 pChns=chnsCompute();
163 chns = chnsCompute( I, pChns );
164 channel_feature=cat(3,chns.data{:});
165 [ch_r,ch_c,ch_num]=size(channel_feature);
166 %原图和通道特征图的比例
167 [I_r,I_c,~]=size(I);
168 radio_r=I_r/ch_r;
169 radio_c=I_c/ch_c;
170 %多个尺度对原图处理
171 %数据结构
172 scale_factor_t(1)=scale_factor;
173 for i=2:scale_num
174 scale_factor_t(i)=scale_factor_t(i-1)*scale_factor;
175 end
176
177
178 for i=1:scale_num
179 [chf_r,chf_c]=size(imresize(channel_feature(:,:,1),scale_factor_t(i)));
180 temp=zeros(chf_r,chf_c,ch_num);%预先求矩阵的尺度
181 for ii=1:ch_num
182 %元胞的个数
183 temp(:,:,ii)=imresize(channel_feature(:,:,ii),scale_factor_t(i));%用元胞来装吧，数据大小不一样了,temp是global吗
184 end
185 channel_feature_scale{i}= temp;
186 end
187
188 %获得训练图的尺度大小
189 patch_row=11;
190 patch_column=8;
191 %all_channel_difference_features=single(zeros(11,8,10,2));
192 data_num=1;
193 coordinate=zeros(4,2);
194 %all_boundingbox=zeros(5,2);
195 all_boundingbox=[];
196 for j=1:scale_num
197 I_scale=channel_feature_scale{j};%第j个尺度的图
198 [row,column,~]=size(I_scale);
199
200 for l=1:ch_num %第l个通道的图
201 for m=1:2:row-patch_row+1 %某个patch
202 for n=1:2:column-patch_column+1
203 all_channel_difference_features(:,:,:,data_num)=I_scale(m:m+patch_row-1,n:n+patch_column-1,:);%把滑窗的数据依次记录下来
204 %坐标记录下来
205 %coordinate[data_num}=[m,l,patch_row,patch_column];%存坐标
206 coordinate(:,data_num)=[m*radio_r,n*radio_c,patch_row*radio_r,patch_column*radio_c];
207 data_num=data_num+1;
208 end
209 end
210 end
211 data_num=data_num-1;
212 %在多个尺度图中缩放图中滑窗并检测，坐标和预测类标，置信度存好
213 handle = crndforest();
214 crndforest(handle, ‘load‘, ‘forest_model2‘);
215 [hs, ps] =crndforest(handle, ‘predict‘, all_channel_difference_features);%尺度问题
216 %对存好的坐标进行变换，映射到原图
217 % for i=1:data_num
218 % coordinate{i}=coordinate{i}/scale_factor_t(j);
219 % end
220 coordinate=coordinate/scale_factor_t(j);
221 %所有的boundingbox组合作为输入
222 %prob=ps(hs+1,:);
223 [rr,~]=size(hs);
224 prob=zeros(rr,1);
225 for t=1:rr
226 prob(t)=ps(hs(t,1)+1,t);
227 end
228 hs=double(hs);
229 bb_partial=[coordinate‘,prob,hs];
230 %all_boundingbox(:,j)=[coordinate,hs,ps(hs)];
231 all_boundingbox=[all_boundingbox;bb_partial];
232 disp(‘第‘)
233 j
234 disp(‘个尺度‘);
235 data_num=1;
236 end
237 %nms函数
238 % bbs1 = bbNms(all_boundingbox, ‘type‘,‘max‘ );
239 % %显示
240 % im(I);
241 % bbApply(‘draw‘,bbs1(1:4),‘g‘);% all_boundingbox(:,[1,2])=all_boundingbox(:,[2,1]);
242 % all_boundingbox(:,[3,4])=all_boundingbox(:,[4,3]);
243 disp(‘正在显示，请稍等...‘);
244 [all_bb_r,all_bb_c]=size(all_boundingbox);
245 bbs=[];
246 for x=1:all_bb_r
247 % if all_boundingbox(x,6)~=10
248 if all_boundingbox(x,6)==5
249 bbs=[bbs;all_boundingbox(x,:)];
250 end
251 end
252
253 %bbs1 = bbNms(bbs, ‘type‘,‘max‘ );
254 bbs1 = bbNms(bbs, ‘thr‘,.35, ‘type‘,‘ms‘);
255
256 im(I);
257 bbApply(‘draw‘,bbs1,‘g‘);
258 disp(‘显示完成！恭喜恭喜!‘);
259 %}
注：
要附上其它代码：
1）随机森林c++版本（mex编译）
2）channel features
时间： 2024-10-05 18:09:43