keras实现简单CNN人脸关键点检测

用keras实现人脸关键点检测

改良版：http://www.cnblogs.com/ansang/p/8583122.html

第一步：准备好需要的库

• tensorflow  1.4.0
• h5py 2.7.0
• hdf5 1.8.15.1
• Keras     2.0.8
• opencv-python     3.3.0
• numpy    1.13.3+mkl

第二步：准备数据集：

data.7z

如图：里面包含着标签和数据

第三步：将图片和标签转成numpy array格式：

 def __data_label__(path):
     f = open(path+"lable.txt", "r")
     i = 0
     datalist = []
     labellist = []
     for line in f.readlines():

         i+=1
         a = line.replace("\n", "")
         b = a.split(",")
         labellist.append(b[1:])
         imgname = path + b[0]
         image = load_img(imgname, target_size=(218, 178))
         datalist.append(img_to_array(image))

     img_data = np.array(datalist)
     img_data = img_data.astype('float32')
     img_data /= 255
     label = np.array(labellist)
     # print(img_data)
     return img_data,label

第四步：搭建网络：

这里使用非常简单的网络

 def __CNN__():
     model = Sequential()#218*178*3
     model.add(Conv2D(32, (3, 3), input_shape=(218, 178, 3)))
     model.add(Activation('relu'))
     model.add(MaxPooling2D(pool_size=(2, 2)))

     model.add(Conv2D(32, (3, 3)))
     model.add(Activation('relu'))
     model.add(MaxPooling2D(pool_size=(2, 2)))

     model.add(Conv2D(64, (3, 3)))
     model.add(Activation('relu'))
     model.add(MaxPooling2D(pool_size=(2, 2)))

     model.add(Flatten())
     model.add(Dense(64))
     model.add(Activation('relu'))
     model.add(Dropout(0.5))
     model.add(Dense(10))
     model.add(Activation('softmax'))
     model.summary()
     return model

_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
conv2d_1 (Conv2D) (None, 216, 176, 32) 896
_________________________________________________________________
activation_1 (Activation) (None, 216, 176, 32) 0
_________________________________________________________________
max_pooling2d_1 (MaxPooling2 (None, 108, 88, 32) 0
_________________________________________________________________
conv2d_2 (Conv2D) (None, 106, 86, 32) 9248
_________________________________________________________________
activation_2 (Activation) (None, 106, 86, 32) 0
_________________________________________________________________
max_pooling2d_2 (MaxPooling2 (None, 53, 43, 32) 0
_________________________________________________________________
conv2d_3 (Conv2D) (None, 51, 41, 64) 18496
_________________________________________________________________
activation_3 (Activation) (None, 51, 41, 64) 0
_________________________________________________________________
max_pooling2d_3 (MaxPooling2 (None, 25, 20, 64) 0
_________________________________________________________________
flatten_1 (Flatten) (None, 32000) 0
_________________________________________________________________
dense_1 (Dense) (None, 64) 2048064
_________________________________________________________________
activation_4 (Activation) (None, 64) 0
_________________________________________________________________
dropout_1 (Dropout) (None, 64) 0
_________________________________________________________________
dense_2 (Dense) (None, 10) 650
_________________________________________________________________
activation_5 (Activation) (None, 10) 0
=================================================================
Total params: 2,077,354
Trainable params: 2,077,354
Non-trainable params: 0
_________________________________________________________________

第五步：训练保存和预测：

 def train(model, testdata, testlabel, traindata, trainlabel):

     # model.compile里的参数loss就是损失函数(目标函数)
     model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])
     # 开始训练， show_accuracy在每次迭代后显示正确率 。  batch_size是每次带入训练的样本数目 ， nb_epoch 是迭代次数
     model.fit(traindata, trainlabel, batch_size=16, epochs=20,
               validation_data=(testdata, testlabel))
     # 设置测试评估参数，用测试集样本
     model.evaluate(testdata, testlabel, batch_size=16, verbose=1,)

 def save(model, file_path=FILE_PATH):
     print('Model Saved.')
     model.save_weights(file_path)

 # def load(model, file_path=FILE_PATH):
 #     print('Model Loaded.')
 #     model.load_weights(file_path)

 def predict(model,image):

     img = image.resize((1, 218, 178, 3))
     img = image.astype('float32')
     img /= 255

     #归一化
     result = model.predict(img)
     result = result*1000+10

     print(result)
     return result

第六步：主模块：

 ############
 # 主模块
 ############
 if __name__ == '__main__':
     model = __CNN__()
     testdata, testlabel = __data_label__(testpath)
     traindata, trainlabel = __data_label__(trainpath)
     # print(testlabel)
     # train(model,testdata, testlabel, traindata, trainlabel)
     # model.save(FILE_PATH)
     model.load_weights(FILE_PATH)
     img = []
     path = "D:/pycode/facial-keypoints-master/data/train/000096.jpg"
     # path = "D:/pycode/Abel_Aguilar_0001.jpg"
     image = load_img(path)
     img.append(img_to_array(image))
     img_data = np.array(img)
     rects = predict(model,img_data)
     img = cv2.imread(path)
     for x, y, w, h, a,b,c,d,e,f in rects:
         point(x,y)
         point(w, h)
         point(a,b)
         point(c,d)
         point(e,f)

     cv2.imshow('img', img)
     cv2.waitKey(0)
     cv2.destroyAllWindows()

训练的时候把train函数的注释取消

预测的时候把train函数注释掉。

下面上全代码：

 from tensorflow.contrib.keras.api.keras.preprocessing.image import ImageDataGenerator,img_to_array
 from keras.models import Sequential
 from keras.layers.core import Dense, Dropout, Activation, Flatten
 from keras.layers.advanced_activations import PReLU
 from keras.layers.convolutional import Conv2D, MaxPooling2D, ZeroPadding2D
 from keras.optimizers import SGD, Adadelta, Adagrad
 from keras.preprocessing.image import load_img, img_to_array
 from keras.utils import np_utils, generic_utils
 import numpy as np
 import cv2

 FILE_PATH = 'face_landmark.h5'
 trainpath = 'D:/pycode/facial-keypoints-master/data/train/'
 testpath = 'D:/pycode/facial-keypoints-master/data/test/'

 def __data_label__(path):
     f = open(path+"lable.txt", "r")
     i = 0
     datalist = []
     labellist = []
     for line in f.readlines():

         i+=1
         a = line.replace("\n", "")
         b = a.split(",")
         labellist.append(b[1:])
         imgname = path + b[0]
         image = load_img(imgname, target_size=(218, 178))
         datalist.append(img_to_array(image))

     img_data = np.array(datalist)
     img_data = img_data.astype('float32')
     img_data /= 255
     label = np.array(labellist)
     # print(img_data)
     return img_data,label

 ###############
 # 开始建立CNN模型
 ###############

 # 生成一个model

 def __CNN__():
     model = Sequential()#218*178*3
     model.add(Conv2D(32, (3, 3), input_shape=(218, 178, 3)))
     model.add(Activation('relu'))
     model.add(MaxPooling2D(pool_size=(2, 2)))

     model.add(Conv2D(32, (3, 3)))
     model.add(Activation('relu'))
     model.add(MaxPooling2D(pool_size=(2, 2)))

     model.add(Conv2D(64, (3, 3)))
     model.add(Activation('relu'))
     model.add(MaxPooling2D(pool_size=(2, 2)))

     model.add(Flatten())
     model.add(Dense(64))
     model.add(Activation('relu'))
     model.add(Dropout(0.5))
     model.add(Dense(10))
          model.summary()
     return model

 def train(model, testdata, testlabel, traindata, trainlabel):

     # model.compile里的参数loss就是损失函数(目标函数)
     model.compile(loss='categorical_crossentropy', optimizer='adam')
     # 开始训练， show_accuracy在每次迭代后显示正确率 。  batch_size是每次带入训练的样本数目 ， nb_epoch 是迭代次数
     model.fit(traindata, trainlabel, batch_size=16, epochs=20,
               validation_data=(testdata, testlabel))
     # 设置测试评估参数，用测试集样本
     model.evaluate(testdata, testlabel, batch_size=16, verbose=1,)

 def save(model, file_path=FILE_PATH):
     print('Model Saved.')
     model.save_weights(file_path)

 # def load(model, file_path=FILE_PATH):
 #     print('Model Loaded.')
 #     model.load_weights(file_path)

 def predict(model,image):

     img = image.resize((1, 218, 178, 3))
     img = image.astype('float32')
     img /= 255

     #归一化
     result = model.predict(img)
     result = result*1000+10

     print(result)
     return result
 def point(x, y):
     cv2.circle(img, (x, y), 1, (0, 0, 255), 10)

 ############
 # 主模块
 ############
 if __name__ == '__main__':
     model = __CNN__()
     testdata, testlabel = __data_label__(testpath)
     traindata, trainlabel = __data_label__(trainpath)
     # print(testlabel)
     # train(model,testdata, testlabel, traindata, trainlabel)
     # model.save(FILE_PATH)
     model.load_weights(FILE_PATH)
     img = []
     path = "D:/pycode/facial-keypoints-master/data/train/000096.jpg"
     # path = "D:/pycode/Abel_Aguilar_0001.jpg"
     image = load_img(path)
     img.append(img_to_array(image))
     img_data = np.array(img)
     rects = predict(model,img_data)
     img = cv2.imread(path)
     for x, y, w, h, a,b,c,d,e,f in rects:
         point(x,y)
         point(w, h)
         point(a,b)
         point(c,d)
         point(e,f)

     cv2.imshow('img', img)
     cv2.waitKey(0)
     cv2.destroyAllWindows()

结果如下：

        

未来计划：

用tensorflow-cpu跑的，数据量很少，网络很简单，提升数据量和网络深度应该还能有较大的改善空间。

而且目前网络只能预测大小为（218，178）像素的图片，将适用性提升是未来的目标。

改进方案：

将图片全部resize成方形，边长不够的加黑边补齐。

 # 按照指定图像大小调整尺寸
 def resize_image(image, height=IMAGE_SIZE, width=IMAGE_SIZE):
     top, bottom, left, right = (0, 0, 0, 0)

     # 获取图像尺寸
     h, w, _ = image.shape

     # 对于长宽不相等的图片，找到最长的一边
     longest_edge = max(h, w)

     # 计算短边需要增加多上像素宽度使其与长边等长
     if h < longest_edge:
         dh = longest_edge - h
         top = dh // 2
         bottom = dh - top
     elif w < longest_edge:
         dw = longest_edge - w
         left = dw // 2
         right = dw - left
     else:
         pass

         # RGB颜色
     BLACK = [0, 0, 0]

     # 给图像增加边界，是图片长、宽等长，cv2.BORDER_CONSTANT指定边界颜色由value指定
     constant = cv2.copyMakeBorder(image, top, bottom, left, right, cv2.BORDER_CONSTANT, value=BLACK)

     # 调整图像大小并返回
     return cv2.resize(constant, (height, width))