1 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
2 % 可调参数
3
4 test_path='';
5 neighbour_pixels_affect=3;
6 target_digit=2;
7 % forestTrain()参数设置
8 % .M - [1] number of trees to train
9 % .H - [max(hs)] number of classes
10 % .N1 - [5*N/M] number of data points for training each tree
11 % .F1 - [sqrt(F)] number features to sample for each node split
12 % .split - ['gini'] options include 'gini', 'entropy' and 'twoing'
13 % .minCount - [1] minimum number of data points to allow split
14 % .minChild - [1] minimum number of data points allowed at child nodes
15 % .maxDepth - [64] maximum depth of tree
16 % .dWts - [] weights used for sampling and weighing each data point
17 % .fWts - [] weights used for sampling features
18 % .discretize - [] optional function mapping structured to class labels
19 % format: [hsClass,hBest] = discretize(hsStructured,H);
20 varargin.M=1000;
21 %varargin.H=10;
22
23 % forestApply()的输入设置
24 % data - [NxF] N length F feature vectors
25 % forest - learned forest classification model
26 % maxDepth - [] maximum depth of tree
27 % minCount - [] minimum number of data points to allow split
28 % best - [0] if true use single best prediction per tree
29
30 % forestApply()输出结果及对比的阀值
31 % hs - [Nx1] predicted output labels
32 % ps - [NxH] predicted output label probabilities
33 ps_val_more_than0_3=0.2;
34
35 %滑窗检测,窗口尺度，步长
36 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
37
38 data=[];
39 label=[];
40 temp_r1=0;
41 temp_c1=0;
42
43 for i_digit=0:9
44 % if(i_digit==target_digit) %%%%%%%%%%%%%%%%%%%%%%
45 % this_image_label=1;
46 % end
47 %数字转字符
48 str=num2str(i); %%数据是不是不平衡
49 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,'\');
50 file=dir(path_temp);
51 for i=3:length(file)
52 path= strcat(path_temp,file(i).name);
53
54 %%%%%%%%%%%%%%%%%%%%%%%%%%
55 % 加载图片
56 %%%%%%%%%%%%%%%%%%%%%%%%%%
57 I=imread(path);
58 %I=imread('E:/WeChat.jpg');
59 %%%%%%%%%%%%%%%%%%%%%%%%%%
60 % 提取channel features
61 %%%%%%%%%%%%%%%%%%%%%%%%%%
62 [all_channel_difference_features,temp_r1,temp_c1]=extract_features(I,1);
63 data=[data,all_channel_difference_features];
64 label=[label;i_digit+1];
65
66 % if(i>100 && this_image_label~=1) %%这里只取了前100帧，实际上可以随意抽取一百张
67 % break;
68 % end
69 end % for i=3:length(file)
70
71 end % for i_digit=0:9
72
73 %%%%%%%%%%%%%%%%%%%%%%%%%%
74 % 扔进分类器中，训练
75 %%%%%%%%%%%%%%%%%%%%%%%%%%
76
77 forest = forestTrain( data, label, varargin );
78
79 %%%%%%%%%%%%%%%%%%%%%%%%%%
80 % 检测，测试
81 test_image=imread(test_path);
82 %滑窗检测,窗口尺度，步长
83 [test_r,test_c,test_z]=size(test_image);
84 for i_test=1:test_r
85 %model
86
87 %resize
88 test_image=imresize(model,temp_r1,temp_c1);
89 test_data=extract_features(test_image,1);
90 [hs,ps] = forestApply( test_data, forest, [], [], [] );%尺度问题
91 if(ps>ps_val_more_than0_3)
92 %画框
93
94 end
95 end
96
97 %%%%%%%%%%%%%%%%%%%%%%%%%%
1 function [ all_channel_difference_features,,r1,c1 ] = extract_features( I,shrink_or_not )
2 %EXTRACT_FEATURES 此处显示有关此函数的摘要
3 % 此处显示详细说明
4 %%%%%%%%%%%%%%%%%%%%%%%%%%
5 % 提取channel features
6 %%%%%%%%%%%%%%%%%%%%%%%%%%
7 % 参数设置
8 if(shrink_or_not==1)
9 pChns.shrink=4;
10 end
11
12 pChns.pColor.enabled=1;
13 pChns.pColor.smooth=1;
14 pChns.pColor.colorSpace='luv';
15
16 pChns.pGradMag.enabled=1;
17 pChns.pGradMag.colorChn=0;
18 pChns.pGradMag.normRad=5;
19 pChns.pGradMag.normConst=.005;
20 pChns.pGradMag.full=0;
21
22 pChns.pGradHist.enabled=1;
23 %pChns.pGradHist.binSize=
24 pChns.pGradHist.nOrients=6;
25 pChns.pGradHist.softBin=0;
26 pChns.pGradHist.useHog=0;
27 pChns.pGradHist.clipHog=.2;
28
29 %pChns.pCustom.**
30
31 %pChns.complete=
32
33 % 提取channel features
34 chns = chnsCompute( I, pChns );
35 % 将各个通道放在矩阵中
36 [r1,c1,ch1]=size(chns.data{1});
37 [r2,c2,ch2]=size(chns.data{2});
38 [r3,c3,ch3]=size(chns.data{3});
39 ch=ch1+ch2+ch3;
40 all_channel=zeros(r1,c1,ch);
41 all_channel(:,:,1:ch1)=chns.data{1};
42 all_channel(:,:,ch1+1:ch1+ch2)=chns.data{2};
43 all_channel(:,:,ch1+ch2+1:ch)=chns.data{3};
44 %%%%%%%%%%%%%%%%%%%%%%%%%%
45 % pooling
46 %%%%%%%%%%%%%%%%%%%%%%%%%%
47 for ii=1:ch
48 %向下采样
49 all_pooling(:,:,ii)=imresize(all_channel(:,:,ii),0.2);
50 end
51
52 %%%%%%%%%%%%%%%%%%%%%%%%%%
53 % 再次做相减特征
54 %%%%%%%%%%%%%%%%%%%%%%%%%%
55 all_channel_difference_features=[];
56 for ij=1:ch
57 temp=difference_features( all_pooling(:,:,ij),neighbour_pixels_affect );
58 all_channel_difference_features = [all_channel_difference_features;temp];
59 end
60
61 end
1 function [ one_channel_difference_features ] = difference_features( one_channel_features,neighbour_pixels_affect )
2 %DIFFERENCE_FEATURES 计算邻域内个特征之间两两相减
3 %input:
4 % one_channel_features
5 %neighbour_pixels_affect
6 %output:
7 %one_channel_difference_features
8
9 [r,c]=size(one_channel_features);
10
11 one_channel_difference_features=[];
12 for i=1:r-neighbour_pixels_affect+1
13 for j=1:c-neighbour_pixels_affect+1
14 local_features=one_channel_features(i:i+neighbour_pixels_affect-1,j:j+neighbour_pixels_affect-1);
15 temp=local_feature_compute(local_features);
16 one_channel_difference_features=[one_channel_difference_features;temp];%特征拼接
17 end
18 end
19 end
20
21 function [ local_differece_feature ]=local_feature_compute( local_features )
22 [r,c]=size(local_features);
23 result_mat=local_features-local_features(1,1).*ones(r,c);
24 result_vector=reshape(result_mat,r*c,1);
25 local_differece_feature=result_vector(2:r*c,1);%把第一个特征去掉，自己减自己没有任何特征信息可言
26 end
%{
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% 可调参数
test_path='C:\Users\cong\Desktop\研一实战\项目\图像中时间数字识别\OCR\one\3.jpg';
neighbour_pixels_affect=3;
target_digit=2;
% forestTrain()参数设置
% .M - [1] number of trees to train
% .H - [max(hs)] number of classes
% .N1 - [5*N/M] number of data points for training each tree
% .F1 - [sqrt(F)] number features to sample for each node split
% .split - ['gini'] options include 'gini', 'entropy' and 'twoing'
% .minCount - [1] minimum number of data points to allow split
% .minChild - [1] minimum number of data points allowed at child nodes
% .maxDepth - [64] maximum depth of tree
% .dWts - [] weights used for sampling and weighing each data point
% .fWts - [] weights used for sampling features
% .discretize - [] optional function mapping structured to class labels
% format: [hsClass,hBest] = discretize(hsStructured,H);
varargin.M=1000;
%varargin.H=10;
% forestApply()的输入设置
% data - [NxF] N length F feature vectors
% forest - learned forest classification model
% maxDepth - [] maximum depth of tree
% minCount - [] minimum number of data points to allow split
% best - [0] if true use single best prediction per tree
% forestApply()输出结果及对比的阀值
% hs - [Nx1] predicted output labels
% ps - [NxH] predicted output label probabilities
ps_val_more_than0_3=0.2;
%滑窗检测,窗口尺度，步长
win_h=20;
win_w=20;
step=1;
disp('参数配置成功...');
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
disp('正在读入图片及特征提取...');
%读入图片及特征提取
data=[];
label=[];
temp_r1=0;
temp_c1=0;
for i_digit=0:9
% if(i_digit==target_digit) %%%%%%%%%%%%%%%%%%%%%%
% this_image_label=1;
% end
%数字转字符
str=num2str(i_digit); %%数据是不是不平衡
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,'\');
file=dir(path_temp);
for i=3:length(file)
path= strcat(path_temp,file(i).name);
%%%%%%%%%%%%%%%%%%%%%%%%%%
% 加载图片
%%%%%%%%%%%%%%%%%%%%%%%%%%
I=imread(path);
%I=imread('E:/WeChat.jpg');
%%%%%%%%%%%%%%%%%%%%%%%%%%
% 提取channel features
%%%%%%%%%%%%%%%%%%%%%%%%%%
[all_channel_difference_features,temp_r1,temp_c1]=extract_features(I,neighbour_pixels_affect,1);
data=[data,all_channel_difference_features];
label=[label;i_digit+1];
if(rem(i,100)==0)
disp('...');
end
end % for i=3:length(file)
disp('数字')
i_digit
disp('的特征提取完毕...');
end % for i_digit=0:9
disp('读入图片及特征提取完毕...');
%%%%%%%%%%%%%%%%%%%%%%%%%%
% 扔进分类器中，训练
%%%%%%%%%%%%%%%%%%%%%%%%%%
data=data';
disp('正在训练，请稍等...');
forest = forestTrain( data, label, varargin );
disp('训练完毕...');
%}
%%%%%%%%%%%%%%%%%%%%%%%%%%
% 检测，测试
test_label=[];
test_label_p=[];
win_h=40;
win_w=30;
windSize = [30,40];
step=2;
ps_val_more_than0_3=0.07;
disp('正在检测...');
test_image=imread(test_path);
%滑窗检测,窗口尺度，步长
[test_r,test_c,test_z]=size(test_image);
figure;
imshow(test_image);
hold on
for i_test=1:step:test_r-win_h+1
for j_test=1:step:test_c-win_w+1
%model
model=test_image(i_test:i_test+win_h-1,j_test:j_test+win_w-1,:);
%resize
test_image_rs=imresize(model,[temp_r1 temp_c1]);
test_data=extract_features(test_image_rs,neighbour_pixels_affect,0);
test_data=test_data';
test_data=single(test_data);
[hs,ps] = forestApply( test_data, forest,0,0,1);%尺度问题
test_label=[test_label,hs];
test_label_p=[test_label_p,ps(hs)];
if(ps>ps_val_more_than0_3)
%画框
%draw_rect(test_image,);
i_test
j_test
rectangle('Position',[i_test,j_test,20,20],'LineWidth',4,'EdgeColor','r');
%pointAll = [i_test,j_test];
%[state,results]=draw_rect(test_image,pointAll,windSize);
hold on
end
end
hold on
end
disp('检测完毕！恭喜恭喜！')
%%%%%%%%%%%%%%%%%%%%%%%%%%
