[人工智能]卷积神经网络实现人脸表情识别

一、实现过程

1.1 下载数据集

https://github.com/truongnmt/smile-detection
创建一个文件夹存放datasets

1.2 根据猫狗数据集训练的方法来训练笑脸数据集

#coding=gbk
import os
import sys
def rename():
    path=input("请输入路径(例如D:\\\\picture)：")
    name=input("请输入开头名:")
    startNumber=input("请输入开始数:")
    fileType=input("请输入后缀名（如 .jpg、.txt等等）:")
    print("正在生成以"+name+startNumber+fileType+"迭代的文件名")
    count=0
    filelist=os.listdir(path)
    for files in filelist:
        Olddir=os.path.join(path,files)
        if os.path.isdir(Olddir):
            continue
        Newdir=os.path.join(path,name+str(count+int(startNumber))+fileType)
        os.rename(Olddir,Newdir)
        count+=1
    print("一共修改了"+str(count)+"个文件")

rename() 

unsmile.改为smile.

1.2 图片分类

import os, shutil #复制文件
# 原始目录所在的路径
# 数据集未压缩
original_dataset_dir1 = 'D:\\Face_smile\\smile\\datasets\\train_folder\\1'  ##笑脸
original_dataset_dir0 = 'D:\\Face_smile\\smile\\datasets\\train_folder\\0'  ##非笑脸
# 我们将在其中的目录存储较小的数据集
base_dir = 'D:\\Face_smile\\smile1'
os.mkdir(base_dir)

# # 训练、验证、测试数据集的目录
train_dir = os.path.join(base_dir, 'train')
os.mkdir(train_dir)
validation_dir = os.path.join(base_dir, 'validation')
os.mkdir(validation_dir)
test_dir = os.path.join(base_dir, 'test')
os.mkdir(test_dir)

# 猫训练图片所在目录
train_cats_dir = os.path.join(train_dir, 'smile')
os.mkdir(train_cats_dir)

# 狗训练图片所在目录
train_dogs_dir = os.path.join(train_dir, 'unsmile')
os.mkdir(train_dogs_dir)

# 猫验证图片所在目录
validation_cats_dir = os.path.join(validation_dir, 'smile')
os.mkdir(validation_cats_dir)

# 狗验证数据集所在目录
validation_dogs_dir = os.path.join(validation_dir, 'unsmile')
os.mkdir(validation_dogs_dir)

# 猫测试数据集所在目录
test_cats_dir = os.path.join(test_dir, 'smile')
os.mkdir(test_cats_dir)

# 狗测试数据集所在目录
test_dogs_dir = os.path.join(test_dir, 'unsmile')
os.mkdir(test_dogs_dir)

# 将前1000张笑脸图像复制到train_cats_dir
fnames = ['smile.{}.jpg'.format(i) for i in range(1000)]
for fname in fnames:
    src = os.path.join(original_dataset_dir1, fname)
    dst = os.path.join(train_cats_dir, fname)
    shutil.copyfile(src, dst)

# 将下500张笑脸图像复制到validation_cats_dir
fnames = ['smile.{}.jpg'.format(i) for i in range(500)]
for fname in fnames:
    src = os.path.join(original_dataset_dir1, fname)
    dst = os.path.join(validation_cats_dir, fname)
    shutil.copyfile(src, dst)
    
# 将下500张笑脸图像复制到test_cats_dir
fnames = ['smile.{}.jpg'.format(i) for i in range(500)]
for fname in fnames:
    src = os.path.join(original_dataset_dir1, fname)
    dst = os.path.join(test_cats_dir, fname)
    shutil.copyfile(src, dst)
    
# 将前1000张非笑脸图像复制到train_dogs_dir
fnames = ['unsmile.{}.jpg'.format(i) for i in range(1000)]
for fname in fnames:
    src = os.path.join(original_dataset_dir0, fname)
    dst = os.path.join(train_dogs_dir, fname)
    shutil.copyfile(src, dst)
    
# 将下500张非笑脸图像复制到validation_dogs_dir
fnames = ['unsmile.{}.jpg'.format(i) for i in range(500)]
for fname in fnames:
    src = os.path.join(original_dataset_dir0, fname)
    dst = os.path.join(validation_dogs_dir, fname)
    shutil.copyfile(src, dst)
    
# 将下500张非笑脸图像复制到test_dogs_dir
fnames = ['unsmile.{}.jpg'.format(i) for i in range(500)]
for fname in fnames:
    src = os.path.join(original_dataset_dir0, fname)
    dst = os.path.join(test_dogs_dir, fname)
    shutil.copyfile(src, dst)

结果如下：

1.3 作为健全性检查，计算一下在每个训练分割中我们有多少图片（训练/验证/测试）：

print('total training cat images:', len(os.listdir(train_cats_dir)))
print('total training dog images:', len(os.listdir(train_dogs_dir)))
print('total validation cat images:', len(os.listdir(validation_cats_dir)))
print('total validation dog images:', len(os.listdir(validation_dogs_dir)))
print('total test cat images:', len(os.listdir(test_cats_dir)))
print('total test dog images:', len(os.listdir(test_dogs_dir)))

1.4 卷积网络模型搭建

from keras import layers
from keras import models

model = models.Sequential()
model.add(layers.Conv2D(32, (3, 3), activation='relu',
                        input_shape=(150, 150, 3)))
model.add(layers.MaxPooling2D((2, 2)))
model.add(layers.Conv2D(64, (3, 3), activation='relu'))
model.add(layers.MaxPooling2D((2, 2)))
model.add(layers.Conv2D(128, (3, 3), activation='relu'))
model.add(layers.MaxPooling2D((2, 2)))
model.add(layers.Conv2D(128, (3, 3), activation='relu'))
model.add(layers.MaxPooling2D((2, 2)))
model.add(layers.Flatten())
model.add(layers.Dense(512, activation='relu'))
model.add(layers.Dense(1, activation='sigmoid'))
model.summary()

1.5 图像生成器读取文件中数据，进行数据预处理

from tensorflow import optimizers

model.compile(loss='binary_crossentropy',
              optimizer=optimizers.RMSprop(lr=1e-4),
              metrics=['acc'])
from keras.preprocessing.image import ImageDataGenerator

# 所有图像将按1/255重新缩放
train_datagen = ImageDataGenerator(rescale=1./255)
test_datagen = ImageDataGenerator(rescale=1./255)

train_generator = train_datagen.flow_from_directory(
        # 这是目标目录
        train_dir,
        # 所有图像将调整为150x150
        target_size=(150, 150),
        batch_size=20,
        # 因为我们使用二元交叉熵损失，我们需要二元标签
        class_mode='binary')

validation_generator = test_datagen.flow_from_directory(
        validation_dir,
        target_size=(150, 150),
        batch_size=20,
        class_mode='binary')

1.6 开始训练

时间漫长，慢慢等待

history = model.fit_generator(
      train_generator,
      steps_per_epoch=100,
      epochs=30,
      validation_data=validation_generator,
      validation_steps=50)

1.7 保存训练模型

model.save('D:\\Face_smile\\smile1\\smiles_and_unsmiles_small_1.h5')

1.8 在培训和验证数据上绘制模型的损失和准确性（可视化界面）

import matplotlib.pyplot as plt

acc = history.history['acc']
val_acc = history.history['val_acc']
loss = history.history['loss']
val_loss = history.history['val_loss']

epochs = range(len(acc))

plt.plot(epochs, acc, 'bo', label='Training acc')
plt.plot(epochs, val_acc, 'b', label='Validation acc')
plt.title('Training and validation accuracy')
plt.legend()

plt.figure()

plt.plot(epochs, loss, 'bo', label='Training loss')
plt.plot(epochs, val_loss, 'b', label='Validation loss')
plt.title('Training and validation loss')
plt.legend()
plt.show()

1.9 使用数据扩充

datagen = ImageDataGenerator(
      rotation_range=40,
      width_shift_range=0.2,
      height_shift_range=0.2,
      shear_range=0.2,
      zoom_range=0.2,
      horizontal_flip=True,
      fill_mode='nearest')
# 这是带有图像预处理实用程序的模块
from keras.preprocessing import image

fnames = [os.path.join(train_cats_dir, fname) for fname in os.listdir(train_cats_dir)]

# 我们选择一个图像来“增强”
img_path = fnames[3]

# 读取图像并调整其大小
img = image.load_img(img_path, target_size=(150, 150))

# 将其转换为具有形状的Numpy数组（150、150、3）
x = image.img_to_array(img)

# 把它改成（1150150，3）
x = x.reshape((1,) + x.shape)

# 下面的.flow（）命令生成一批随机转换的图像。
# 它将无限循环，所以我们需要在某个时刻“打破”循环！
i = 0
for batch in datagen.flow(x, batch_size=1):
    plt.figure(i)
    imgplot = plt.imshow(image.array_to_img(batch[0]))
    i += 1
    if i % 4 == 0:
        break

plt.show()

如果我们使用这种数据增强配置训练一个新网络，我们的网络将永远不会看到两次相同的输入。但是，输入它所看到的仍然是高度相关的，因为它们来自于少量的原始图像——我们无法产生新的信息，我们只能重新混合现有的信息。因此，这可能还不足以完全摆脱过度装修。继续战斗过度拟合，我们还将在密连接分类器之前向模型添加一个退出层：

model = models.Sequential()
model.add(layers.Conv2D(32, (3, 3), activation='relu',
                        input_shape=(150, 150, 3)))
model.add(layers.MaxPooling2D((2, 2)))
model.add(layers.Conv2D(64, (3, 3), activation='relu'))
model.add(layers.MaxPooling2D((2, 2)))
model.add(layers.Conv2D(128, (3, 3), activation='relu'))
model.add(layers.MaxPooling2D((2, 2)))
model.add(layers.Conv2D(128, (3, 3), activation='relu'))
model.add(layers.MaxPooling2D((2, 2)))
model.add(layers.Flatten())
model.add(layers.Dropout(0.5))
model.add(layers.Dense(512, activation='relu'))
model.add(layers.Dense(1, activation='sigmoid'))

model.compile(loss='binary_crossentropy',
              optimizer=optimizers.RMSprop(lr=1e-4),
              metrics=['acc'])

1.10 使用数据扩充和退出来训练我们的网络

train_datagen = ImageDataGenerator(
    rescale=1./255,
    rotation_range=40,
    width_shift_range=0.2,
    height_shift_range=0.2,
    shear_range=0.2,
    zoom_range=0.2,
    horizontal_flip=True,)

# 请注意，不应增加验证数据！
test_datagen = ImageDataGenerator(rescale=1./255)

train_generator = train_datagen.flow_from_directory(
        # 这是目标目录
        train_dir,
        # 所有图像将调整为150x150
        target_size=(150, 150),
        batch_size=32,
        # 因为我们使用二元交叉熵损失，我们需要二元标签
        class_mode='binary')

validation_generator = test_datagen.flow_from_directory(
        validation_dir,
        target_size=(150, 150),
        batch_size=32,
        class_mode='binary')

history = model.fit_generator(
      train_generator,
      steps_per_epoch=100,
      epochs=100,
      validation_data=validation_generator,
      validation_steps=50)

1.11保存模型

model.save('D:\\Face——smile\\smile1\\smiles_and_unsmiles_small_2.h5')

1.12 在培训和验证数据上绘制模型的损失和准确性（可视化界面）

acc = history.history['acc']
val_acc = history.history['val_acc']
loss = history.history['loss']
val_loss = history.history['val_loss']

epochs = range(len(acc))

plt.plot(epochs, acc, 'bo', label='Training acc')
plt.plot(epochs, val_acc, 'b', label='Validation acc')
plt.title('Training and validation accuracy')
plt.legend()

plt.figure()

plt.plot(epochs, loss, 'bo', label='Training loss')
plt.plot(epochs, val_loss, 'b', label='Validation loss')
plt.title('Training and validation loss')
plt.legend()
plt.show()

二、运用训练的模型实现表情识别

#检测视频或者摄像头中的人脸
import cv2
from keras.preprocessing import image
from keras.models import load_model
import numpy as np
import dlib
from PIL import Image
model = load_model('D:\\Face_smile\\smile1\\smiles_and_unsmiles_small_2.h5')
detector = dlib.get_frontal_face_detector()
video=cv2.VideoCapture(0)
font = cv2.FONT_HERSHEY_SIMPLEX
def rec(img):
    gray=cv2.cvtColor(img,cv2.COLOR_BGR2GRAY)
    dets=detector(gray,1)
    if dets is not None:
        for face in dets:
            left=face.left()
            top=face.top()
            right=face.right()
            bottom=face.bottom()
            cv2.rectangle(img,(left,top),(right,bottom),(0,255,0),2)
            img1=cv2.resize(img[top:bottom,left:right],dsize=(150,150))
            img1=cv2.cvtColor(img1,cv2.COLOR_BGR2RGB)
            img1 = np.array(img1)/255.
            img_tensor = img1.reshape(-1,150,150,3)
            prediction =model.predict(img_tensor)    
            if prediction[0][0]>0.5:
                result='unsmile'
            else:
                result='smile'
            cv2.putText(img, result, (left,top), font, 2, (0, 255, 0), 2, cv2.LINE_AA)
        cv2.imshow('Video', img)
while video.isOpened():
    res, img_rd = video.read()
    if not res:
        break
    rec(img_rd)
    if cv2.waitKey(5) & 0xFF == ord('q'):
        break
video.release()
cv2.destroyAllWindows()

结果展示：

三、总结

这一次实验是用卷积神经网络实现的人脸表情识别，实现过程还是遇到了很多的困难，遇到了很多导包错误的问题，keras和tensorflow很折磨人，明明下载的有，但还是在用的时候识别不到，头痛。解决了很久，后来发现是jupyter的问题，用其他软件运行代码都能识别到包，用jupyter就识别不到包。相对于上回用svm实现的人脸表情识别，对这方面的知识有了进一步的了解。

四、参考文献

python+基于Jupyter notebook完成卷积神经网络和Dlib的笑脸识别+口罩识别