參考文獻：
Deep face recognition, O. M. Parkhi and A. Vedaldi and A. Zisserman,
Proceedings of the British Machine Vision Conference (BMVC), 2015 (paper).
利用vgg-face網絡結構，去掉了最后一層全連接，提取人臉特征，實現人臉識別及landmark
網絡權重下載：http://www.robots.ox.ac.uk/~vgg/software/vgg_face/
代碼實現：

# -*- coding: utf-8 -*-

import cv2
import dlib
import numpy as np
import matplotlib.pyplot as plt
from PIL import Image
import math
imgpath='./classifer/posimage/2.jpg'
bgr_img=cv2.imread(imgpath,1)
rgb_img=cv2.cvtColor(bgr_img, cv2.COLOR_BGR2RGB)
#gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
'''implement the face landmark estimation and pose alignment'''
#face landmark estimation
detector = dlib.get_frontal_face_detector()
landmark_predictor = dlib.shape_predictor('shape_predictor_68_face_landmarks.dat')
dets = detector(rgb_img,1)

##show the result of face landmark
#if (len(dets) > 0):
#    for k,d in enumerate(dets):
#        cv2.rectangle(bgr_img,(d.left(),d.top()),(d.right(),d.bottom()),(255,255,255))
#        shape = landmark_predictor(bgr_img,d)
#        for i in range(68):
#            cv2.circle(bgr_img, (shape.part(i).x, shape.part(i).y),1,(0,255,0), -1, 8)
#            cv2.putText(bgr_img,str(i),(shape.part(i).x,shape.part(i).y),cv2.FONT_HERSHEY_SIMPLEX,0.2,(255,2555,255))
#cv2.imshow('Frame',bgr_img)
#cv2.waitKey(0)

faces=dlib.full_object_detections()
for det in dets:
    faces.append(landmark_predictor(rgb_img,det))
#pose alignment
images=dlib.get_face_chips(rgb_img,faces,size=320)
image_cnt=0
for image in images:
    image_cnt += 1
    cv_rgb_image = np.array(image).astype(np.uint8)# 先轉換為numpy數組
    cv_bgr_image = cv2.cvtColor(cv_rgb_image, cv2.COLOR_RGB2BGR)# opencv下顏色空間為bgr，所以從rgb轉換為bgr
    cv2.imshow('%s'%(image_cnt), cv_bgr_image)
cv2.waitKey(0)

'''
use VGG-face model from 
Deep face recognition, O. M. Parkhi and A. Vedaldi and A. Zisserman, 
Proceedings of the British Machine Vision Conference (BMVC), 2015 (paper).
'''
from keras.models import Sequential, Model
from keras.layers import Flatten, Dropout, Activation, Permute
from keras.layers import Conv2D, MaxPooling2D
from keras import backend as K
K.set_image_data_format( 'channels_last' )
def convblock(cdim, nb, bits=3):
    L = []
    for k in range(1,bits+1):
        convname = 'conv'+str(nb)+'_'+str(k)
        L.append( Conv2D(cdim,(3,3),padding='same',activation='relu',name=convname) )
    L.append( MaxPooling2D((2,2), strides=(2,2)) )
    return L

def vgg_face_blank():
    withDO = True 
    if True:
        mdl = Sequential()
        mdl.add( Permute((1,2,3), input_shape=(224,224,3)) )
        for l in convblock(64,1, bits=2):
            mdl.add(l)
        for l in convblock(128,2, bits=2):
            mdl.add(l)
        for l in convblock(256,3, bits=3):
            mdl.add(l)   
        for l in convblock(512,4, bits=3):
            mdl.add(l)   
        for l in convblock(512,5, bits=3):
            mdl.add(l)
        mdl.add( Conv2D(4096,kernel_size=(7,7),activation='relu',name='fc6'))
        if withDO:
            mdl.add(Dropout(0.5))
        mdl.add( Conv2D(4096,kernel_size=(1,1),activation='relu',name='fc7'))
        if withDO:
            mdl.add(Dropout(0.5))
        mdl.add(Conv2D(2622,kernel_size=(1,1),activation='relu',name='fc8')) 
        mdl.add(Flatten())
        mdl.add(Activation('softmax'))
        return mdl
    else:
        raise ValueError('not implemented')
facemodel = vgg_face_blank()
facemodel.summary()
from scipy.io import loadmat
data = loadmat('vgg-face.mat', matlab_compatible=False, struct_as_record=False)
l = data['layers']
description = data['meta'][0,0].classes[0,0].description
def copy_mat_to_keras(kmodel):
    kerasnames = [lr.name for lr in kmodel.layers]
    prmt = (0,1,2,3)
    for i in range(l.shape[1]):
        matname = l[0,i][0,0].name[0]
        if matname in kerasnames:
            kindex = kerasnames.index(matname)
            l_weights = l[0,i][0,0].weights[0,0]
            l_bias = l[0,i][0,0].weights[0,1]
            f_l_weights = l_weights.transpose(prmt)
            assert (f_l_weights.shape == kmodel.layers[kindex].get_weights()[0].shape)
            assert (l_bias.shape[1] == 1)
            assert (l_bias[:,0].shape == kmodel.layers[kindex].get_weights()[1].shape)
            assert (len(kmodel.layers[kindex].get_weights()) == 2)
            kmodel.layers[kindex].set_weights([f_l_weights, l_bias[:,0]])
copy_mat_to_keras(facemodel)
#from keras.models import load_model
#facemodel.save('model_weight.h5')

im = Image.open('test.jpg') 
im = im.resize((224,224))
plt.imshow(np.asarray(im))
def pred(kmodel, crpimg):
    imarr = np.array(crpimg).astype(np.float32)
    imarr = np.expand_dims(imarr, axis=0)
    out = kmodel.predict(imarr)
    best_index = np.argmax(out, axis=1)[0]
    best_name = description[best_index,0]
    print(best_index, best_name[0], out[0,best_index], [np.min(out), np.max(out)])
crpim = im
pred(facemodel, crpim)
#face feature:drop the last layer
featuremodel = Model(inputs=facemodel.layers[0].input, outputs=facemodel.layers[-2].output)
featuremodel.summary()
def features(featmodel, crpimg):   
    imarr = np.array(crpimg).astype(np.float32)
    imarr = np.expand_dims(imarr, axis=0)
    fvec = featmodel.predict(imarr)[0,:]
    normfvec = math.sqrt(fvec.dot(fvec))
    return fvec/normfvec
crpimg=Image.open('./classifer/bounding box/posimage/3.jpg') 
crpimg=crpimg.resize((224,224))
sample_feature=features(featuremodel, crpimg)
print(sample_feature)

'''train classifier from the embedded network features/measurements'''
'''without landmark estimation and pose alignment'''
import os
pos_im_file="./classifer/bounding box/posimage"  
neg_im_file="./classifer/bounding box/negimage"
'''
if want to compare with and without the pose alignment
just modify the path
'''
#pos_im_file="./classifer/bounding box with only fl/posimage"  
#neg_im_file="./classifer/bounding box with only fl/negimage"
#pos_im_file="./classifer/bounding box with fl and pa/posimage"  
#neg_im_file="./classifer/bounding box with fl and pa/negimage"
def loadimg_infile(imgdir):
    imgname=[]
    imgname_sorted=[]
    filename=os.listdir(imgdir)
    for file in filename:
        if 'jpg' in file:
            imgname.append(imgdir+file)
    s=len(imgname)
    for i in range(s):
        imgname_sorted.append(imgdir+'/'+str(i+1)+'.jpg')
    return imgname_sorted

pos_im_list=loadimg_infile(pos_im_file)
neg_im_list=loadimg_infile(neg_im_file)
xdata=[]
ydata=[]
for pos_im in pos_im_list:
    crpimg=Image.open(pos_im)
    crpimg=crpimg.resize((224,224))
    xdata.append( features(featuremodel, crpimg))
    ydata.append('1.0') #is Obama
for neg_im in neg_im_list:
    crpimg=Image.open(neg_im)
    crpimg=crpimg.resize((224,224))
    xdata.append( features(featuremodel, crpimg))
    ydata.append('0.0') #not Obama
    
'''split data to train and test'''
from sklearn.cross_validation import train_test_split
test_size=0.1
X_train,X_test,y_train,y_test=train_test_split(xdata,ydata,test_size=test_size,random_state=1)

'''different kinds of classifers'''
import matplotlib.pyplot as plt
from sklearn.naive_bayes import GaussianNB
from sklearn.naive_bayes import MultinomialNB
from sklearn.linear_model import LogisticRegression
from sklearn.neighbors import KNeighborsClassifier
from sklearn.svm import SVC
from sklearn import metrics
from sklearn.metrics import precision_recall_curve, roc_curve, auc
from sklearn.externals import joblib
#clf = GaussianNB()
#clf = MultinomialNB(alpha=0.01)
#clf = LogisticRegression()
#clf = SVC(kernel = 'linear')
#clf = KNeighborsClassifier()
clf = SVC(kernel='linear', gamma= 0.5, C = 2.0)
clf.fit(X_train,y_train)
#clf = joblib.load("week4_model.m")
pred_train = clf.predict(X_train)
pred_test = clf.predict(X_test)
s_train= metrics.confusion_matrix(y_train,pred_train,labels=None)
s_test= metrics.confusion_matrix(y_test,pred_test,labels=None)
accuracy_train=(s_train[0][0]+s_train[1][1])/(100-100*test_size)
accuracy_test=(s_test[0][0]+s_test[1][1])/(100*test_size)
print('accuracy_train:',accuracy_train)
print('accuracy_test:',accuracy_test)
#joblib.dump(clf, "week4_model.m")

效果：

2.jpg

3.jpg

4.jpg

5.jpg