程序（训练弱分类器）

《程序（训练弱分类器）》由会员上传分享，免费在线阅读，更多相关内容在工程资料-天天文库。

1、function[theta,p,err]=LearnWeakClassifier(ws,fs,ys)%采用直接求取目标函数值即阈值的方法ni=size(fs,1);%样本数(彳亍数)nf=size(fs,2);%特征数(列数)%wsys=(ws.*ys)';%人脸样本的权重wsiys=(ws.*(l-ys))';%HK人脸样本的权重mp=(wsys*fs)/sum(wsys);%wsys*fs,人脸样本的权重*特征值的和%人脸样木的权重乘以特征除以人脸样木的总权重mn二(wsiys*fs)/sum(wsiys);sp=(wsys*fs-mp).A2/sum(wsys);sn二(wsiys*

2、fs-mn).八2/sum(wsiys);%a=sp-sn;b=-2*(sp.*mn-sn.*mp);c=(sp.*(mn.A2)-sn.*(mp.A2))-2*sp.*sn.*log(sqrt(sp./sn));d=sqrt(b.*b・4.*a.*c);%得到两个阈值xl=(-b+d)./(2*a);x2=(-b-d)./(2*a);xm=(mp+mn)/2;[pl,errl]=GetClass讦erError(ws,ys,fs,xl,ni,nf);[p2,err2]=GetClassiferError(ws,ys,fs,x2,ni,nf);%比较两个阈值错误率，选择错误率较小的idx=e

3、rrl

4、算%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%f_max=0.3;d_min=0.99;Cparamsall=Cparams;PositiveFeatures=ComputeFeatures(PTrainData,all_ftypes);%所有人脸样本所有特征特征值NegativeFeatures=ComputeFeatures(NTrainData,all_ftypes);%所有非人脸样本所有特征特征值_%fs(ni*nf)样本数*特征数np=size(PositiveFeatures.fsz1);%人脸样

5、本数nn=size(NegativeFeatures.fsz1);%非人脸样本数ys=[ones(np,1);zeros(nn,1)];%标签，前np是人脸样木，标签值为1,后nn个非人脸样本标签值0ws=[ones(npzl)/(2*np);ones(nn,l)/(2*nn)];%初始权重，人脸样本权重l/(2*np)W=size(PTrainData,2);H=size(PTrainData,l);nf=size(PositiveFeatures.fs,2);%特征总数ws=Cparamsall(k-l).ws;hsl=ones(nn,l)；%初始化每个样本的标签值为1，当被抛弃时为将该

6、样本标签变为0hsl=detec_nface(CparamsallzNTrainData,hsl);hs2=ones(np,l)；%初始化每个样木的标签值为1，当被抛弃时为将该样木标签变为0hs2=detec_nface(Cparamsall,PTrainData,hs2);hs=[hs2;hsl];e=ws'*abs(ys-hs);%w*(h(x)-y),即错误率bata=e/(l-e);wsu=bata.A(l-abs(hs-ys));%用于更新权值ws=ws.*wsu;%更新后的权值f=sum(hsl==l)/nn;t=0;%初始化该层的弱分类器个数whilef>f_max%如果该层误

7、检率人于给每层设是的最人谋检率f_maxt=t+l;%将增加一个弱分类器添入该层强分类器中ws=ws/sum(ws);%归一・化权重e=inf;%错误率无穷大i=l；%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%while(iv二nf)count=min(nf-i+l,200);%每次200个特征，计算其对应的弱分类器fs=[PositiveFeatures.fs(:,i: