Keras搭建简单的人脸识别CNN模型

本文在下述博文的基础上，进行整理并针对Keras2.0修改了个别错误，完成小样本情况下的简单人脸识别CNN模型。

http://blog.csdn.net/u012162613/article/details/43277187

 

  1 # -*- coding: utf-8 -*-
  2 """
  3 Created on Mon Jun 26 09:56:29 2017
  4 
  5 @author: xiaoxue
  6 """
  7 
  8 from __future__ import print_function  
  9 import numpy  
 10 import matplotlib.pyplot as plt
 11 numpy.random.seed(1337)  # for reproducibility  
 12   
 13 from PIL import Image  
 14   
 15 from keras.models import Sequential  
 16 from keras.layers.core import Dense, Dropout, Activation, Flatten  
 17 from keras.layers.convolutional import Conv2D, MaxPooling2D  
 18 from keras.optimizers import SGD  
 19 from keras.utils import np_utils  
 20   
 21 # There are 40 different classes  
 22 nb_classes = 40  
 23 nb_epoch = 40  
 24 batch_size = 40  
 25   
 26 # input image dimensions  
 27 img_rows, img_cols = 57, 47  
 28 # number of convolutional filters to use  
 29 nb_filters1, nb_filters2 = 5, 10  
 30 # size of pooling area for max pooling  
 31 nb_pool = 2  
 32 # convolution kernel size  
 33 nb_conv = 3  
 34   
 35 def load_data(dataset_path):  
 36     img = Image.open(dataset_path)  
 37     img_ndarray = numpy.asarray(img, dtype='float64')/256  
 38     #400pictures,size:57*47=2679  
 39     faces=numpy.empty((400,2679))   
 40     for row in range(20):  
 41        for column in range(20):  
 42         faces[row*20+column]=numpy.ndarray.flatten(img_ndarray [row*57:(row+1)*57,column*47:(column+1)*47])  
 43       
 44     label=numpy.empty(400)  
 45     for i in range(40):  
 46         label[i*10:i*10+10]=i  
 47         label=label.astype(numpy.int)  
 48       
 49     #train:320,valid:40,test:40  
 50     train_data=numpy.empty((320,2679))  
 51     train_label=numpy.empty(320)  
 52     valid_data=numpy.empty((40,2679))  
 53     valid_label=numpy.empty(40)  
 54     test_data=numpy.empty((40,2679))  
 55     test_label=numpy.empty(40)  
 56       
 57     for i in range(40):  
 58         train_data[i*8:i*8+8]=faces[i*10:i*10+8]  
 59         train_label[i*8:i*8+8]=label[i*10:i*10+8]  
 60         valid_data[i]=faces[i*10+8]  
 61         valid_label[i]=label[i*10+8]  
 62         test_data[i]=faces[i*10+9]  
 63         test_label[i]=label[i*10+9]  
 64        
 65     rval = [(train_data, train_label), (valid_data, valid_label),  
 66             (test_data, test_label)]  
 67     return rval  
 68       
 69 def Net_model(lr=0.005,decay=1e-6,momentum=0.9):  
 70     model = Sequential()  
 71     model.add(Conv2D(nb_filters1, (nb_conv, nb_conv),  
 72                     input_shape=(img_rows, img_cols, 1 ),
 73                     padding='same'))
 74     model.add(Activation('tanh'))  
 75     model.add(MaxPooling2D(pool_size=(nb_pool, nb_pool)))  
 76       
 77     model.add(Conv2D(nb_filters2,kernel_size=(nb_conv, nb_conv)))
 78     model.add(Activation('tanh'))  
 79     model.add(MaxPooling2D(pool_size=(nb_pool, nb_pool)))  
 80     #model.add(Dropout(0.25))  
 81       
 82     model.add(Flatten())  
 83     model.add(Dense(1000)) #Full connection  
 84     model.add(Activation('tanh'))  
 85     #model.add(Dropout(0.5))  
 86     model.add(Dense(nb_classes))  
 87     model.add(Activation('softmax'))  
 88       
 89     sgd = SGD(lr=lr, decay=decay, momentum=momentum, nesterov=True)  
 90     model.compile(loss='categorical_crossentropy', optimizer=sgd, metrics=["acc"])  
 91       
 92     return model  
 93       
 94 def train_model(model,X_train,Y_train,X_val,Y_val):  
 95     result = model.fit(X_train, Y_train, batch_size=batch_size, epochs=nb_epoch,  
 96            verbose=2, validation_data=(X_val, Y_val))  
 97 #    show_accuracy=True,
 98     model.save_weights('model_weights.h5',overwrite=True)  
 99     return model ,result 
100       
101 def test_model(model,X,Y):  
102     model.load_weights('model_weights.h5')  
103     score = model.evaluate(X, Y, verbose=0)  #, show_accuracy=True
104     print('Test score:', score[0])  
105 #    print('Test accuracy:', score[1])  
106     return score  
107       
108 if __name__ == '__main__':  
109     # the data, shuffled and split between tran and test sets  
110     (X_train, y_train), (X_val, y_val),(X_test, y_test) = load_data('olivettifaces.gif')  
111       
112     X_train = X_train.reshape(X_train.shape[0], img_rows, img_cols, 1)  
113     X_val = X_val.reshape(X_val.shape[0], img_rows, img_cols, 1)  
114     X_test = X_test.reshape(X_test.shape[0], img_rows, img_cols, 1)  
115     print('X_train shape:', X_train.shape)  
116     print(X_train.shape[0], 'train samples')  
117     print(X_val.shape[0], 'validate samples')  
118     print(X_test.shape[0], 'test samples')  
119       
120     # convert class vectors to binary class matrices  
121     Y_train = np_utils.to_categorical(y_train, nb_classes)  
122     Y_val = np_utils.to_categorical(y_val, nb_classes)  
123     Y_test = np_utils.to_categorical(y_test, nb_classes)  
124       
125     model=Net_model()  
126     model,result = train_model(model,X_train,Y_train,X_val,Y_val)   
127 #    score=test_model(model,X_test,Y_test)  #展示模型在验证数据上的效果
128 #    model.load_weights('model_weights.h5')  
129     classes=model.predict_classes(X_test,verbose=0)  #展示模型在测试数据上的效果
130     test_accuracy = numpy.mean(numpy.equal(y_test,classes))  
131     print("accuarcy:",test_accuracy)  
132     plt.figure
133     plt.plot(result.epoch,result.history['acc'],label='acc')
134     plt.plot(result.epoch,result.history['val_acc'],label='val_acc')
135     plt.scatter(result.epoch,result.history['acc'],marker='*')
136     plt.scatter(result.epoch,result.history['val_acc'])
137     plt.legend(loc='best')
138     plt.show()
139     
140     plt.figure
141     plt.plot(result.epoch,result.history['loss'],label="loss")
142     plt.plot(result.epoch,result.history['val_loss'],label="val_loss")
143     plt.scatter(result.epoch,result.history['loss'],marker='*')
144     plt.scatter(result.epoch,result.history['val_loss'],marker='*')
145     plt.legend(loc='best')
146     plt.show()

 

运行结果如下图所示：

X_train shape: (320, 57, 47, 1)
320 train samples
40 validate samples
40 test samples
Train on 320 samples, validate on 40 samples
Epoch 1/40
0s - loss: 3.7571 - acc: 0.0156 - val_loss: 3.6561 - val_acc: 0.0500
Epoch 2/40
0s - loss: 3.6847 - acc: 0.0469 - val_loss: 3.6116 - val_acc: 0.1000
Epoch 3/40
0s - loss: 3.6023 - acc: 0.1063 - val_loss: 3.5312 - val_acc: 0.1500
Epoch 4/40
0s - loss: 3.4942 - acc: 0.1625 - val_loss: 3.4197 - val_acc: 0.2500
Epoch 5/40
0s - loss: 3.3360 - acc: 0.3312 - val_loss: 3.2416 - val_acc: 0.4750
Epoch 6/40
0s - loss: 3.0926 - acc: 0.5125 - val_loss: 2.9375 - val_acc: 0.5750
Epoch 7/40
0s - loss: 2.6818 - acc: 0.6219 - val_loss: 2.4856 - val_acc: 0.7250
Epoch 8/40
0s - loss: 2.1218 - acc: 0.8000 - val_loss: 1.9517 - val_acc: 0.8500
Epoch 9/40
0s - loss: 1.5356 - acc: 0.8969 - val_loss: 1.4674 - val_acc: 0.8750
Epoch 10/40
0s - loss: 1.0489 - acc: 0.9438 - val_loss: 1.1259 - val_acc: 0.9250
Epoch 11/40
0s - loss: 0.7106 - acc: 0.9688 - val_loss: 0.8722 - val_acc: 0.9500
Epoch 12/40
0s - loss: 0.4951 - acc: 0.9781 - val_loss: 0.7287 - val_acc: 0.9250
Epoch 13/40
0s - loss: 0.3659 - acc: 0.9875 - val_loss: 0.6243 - val_acc: 0.9500
Epoch 14/40
0s - loss: 0.2785 - acc: 0.9969 - val_loss: 0.5288 - val_acc: 0.9500
Epoch 15/40
0s - loss: 0.2220 - acc: 0.9938 - val_loss: 0.4754 - val_acc: 0.9500
Epoch 16/40
0s - loss: 0.1808 - acc: 1.0000 - val_loss: 0.4412 - val_acc: 0.9500
Epoch 17/40
0s - loss: 0.1520 - acc: 1.0000 - val_loss: 0.4031 - val_acc: 0.9500
Epoch 18/40
0s - loss: 0.1310 - acc: 1.0000 - val_loss: 0.3762 - val_acc: 0.9750
Epoch 19/40
0s - loss: 0.1157 - acc: 1.0000 - val_loss: 0.3489 - val_acc: 0.9750
Epoch 20/40
0s - loss: 0.1027 - acc: 1.0000 - val_loss: 0.3313 - val_acc: 0.9750
Epoch 21/40
0s - loss: 0.0925 - acc: 1.0000 - val_loss: 0.3206 - val_acc: 0.9750
Epoch 22/40
0s - loss: 0.0838 - acc: 1.0000 - val_loss: 0.3087 - val_acc: 0.9750
Epoch 23/40
0s - loss: 0.0765 - acc: 1.0000 - val_loss: 0.2952 - val_acc: 0.9750
Epoch 24/40
0s - loss: 0.0709 - acc: 1.0000 - val_loss: 0.2895 - val_acc: 0.9750
Epoch 25/40
0s - loss: 0.0657 - acc: 1.0000 - val_loss: 0.2829 - val_acc: 0.9750
Epoch 26/40
0s - loss: 0.0611 - acc: 1.0000 - val_loss: 0.2695 - val_acc: 0.9750
Epoch 27/40
0s - loss: 0.0571 - acc: 1.0000 - val_loss: 0.2639 - val_acc: 0.9750
Epoch 28/40
0s - loss: 0.0538 - acc: 1.0000 - val_loss: 0.2591 - val_acc: 0.9750
Epoch 29/40
0s - loss: 0.0506 - acc: 1.0000 - val_loss: 0.2544 - val_acc: 0.9750
Epoch 30/40
0s - loss: 0.0481 - acc: 1.0000 - val_loss: 0.2503 - val_acc: 0.9750
Epoch 31/40
0s - loss: 0.0452 - acc: 1.0000 - val_loss: 0.2444 - val_acc: 0.9750
Epoch 32/40
0s - loss: 0.0430 - acc: 1.0000 - val_loss: 0.2392 - val_acc: 0.9750
Epoch 33/40
0s - loss: 0.0410 - acc: 1.0000 - val_loss: 0.2368 - val_acc: 0.9750
Epoch 34/40
0s - loss: 0.0393 - acc: 1.0000 - val_loss: 0.2329 - val_acc: 0.9750
Epoch 35/40
0s - loss: 0.0376 - acc: 1.0000 - val_loss: 0.2293 - val_acc: 0.9750
Epoch 36/40
0s - loss: 0.0359 - acc: 1.0000 - val_loss: 0.2274 - val_acc: 0.9750
Epoch 37/40
0s - loss: 0.0345 - acc: 1.0000 - val_loss: 0.2251 - val_acc: 0.9750
Epoch 38/40
0s - loss: 0.0331 - acc: 1.0000 - val_loss: 0.2223 - val_acc: 0.9750
Epoch 39/40
0s - loss: 0.0320 - acc: 1.0000 - val_loss: 0.2185 - val_acc: 0.9750
Epoch 40/40
0s - loss: 0.0308 - acc: 1.0000 - val_loss: 0.2173 - val_acc: 0.9750
accuarcy: 1.0