python+opencv人脸识别是否戴口罩

当今疫情情况下，检测是否戴口罩还是比较流行的，然后就搞啦一个检测人脸是否戴口罩的程序。程序调用笔记本摄像头，实时检测人脸是否佩戴口罩。

一、数据集

如果数据集大家有，就用自己的。没有的话，也可以找我要啊（3000+戴口罩/1000+不戴口罩）。在此就不分享出来了，我也是找别人要的。

二、开始上代码

1、导入keras库

1 import keras
2 keras.__version__

2、创建train、test、validation文件夹，并且三个文件夹里面都有have_mask和no_mask文件，运行之后将数据集中的照片放入对应的文件夹中。

 1 import os, shutil
 2 # The path to the directory where the original
 3 # dataset was uncompressed
 4 original_dataset_dir = 'F:\PyCharm\facemask\mask'
 5 
 6 # The directory where we will
 7 # store our smaller dataset
 8 base_dir = 'F:\PyCharm\facemask\mask_and_nomask'
 9 os.mkdir(base_dir)
10 
11 # Directories for our training,
12 # validation and test splits
13 train_dir = os.path.join(base_dir, 'train')
14 os.mkdir(train_dir)
15 validation_dir = os.path.join(base_dir, 'validation')
16 os.mkdir(validation_dir)
17 test_dir = os.path.join(base_dir, 'test')
18 os.mkdir(test_dir)
19 
20 # Directory with our training smile pictures
21 train_smile_dir = os.path.join(train_dir, 'have_mask')
22 os.mkdir(train_smile_dir)
23 
24 # Directory with our training nosmile pictures
25 train_nosmile_dir = os.path.join(train_dir, 'no_mask')
26 os.mkdir(train_nosmile_dir)
27 
28 # Directory with our validation smile pictures
29 validation_smile_dir = os.path.join(validation_dir, 'have_mask')
30 os.mkdir(validation_smile_dir)
31 
32 # Directory with our validation nosmile pictures
33 validation_nosmile_dir = os.path.join(validation_dir, 'no_mask')
34 os.mkdir(validation_nosmile_dir)
35 
36 # Directory with our validation smile pictures
37 test_smile_dir = os.path.join(test_dir, 'have_mask')
38 os.mkdir(test_smile_dir)
39 
40 # Directory with our validation nosmile pictures
41 test_nosmile_dir = os.path.join(test_dir, 'no_mask')
42 os.mkdir(test_nosmile_dir)

3、记数，统计一下各个文件中照片的数量

1 print('total training smile images:', len(os.listdir(train_smile_dir)))
2 print('total training nosmile images:', len(os.listdir(train_nosmile_dir)))
3 print('total validation smile images:', len(os.listdir(validation_smile_dir)))
4 print('total validation nosmile images:', len(os.listdir(validation_nosmile_dir)))
5 print('total test smile images:', len(os.listdir(test_smile_dir)))
6 print('total test nosmile images:', len(os.listdir(test_nosmile_dir)))

4、构建小型卷积网络

我们已经为MNIST构建了一个小型卷积网，所以您应该熟悉它们。我们将重用相同的通用结构:我们的卷积网将是一个交替的Conv2D(激活relu)和MaxPooling2D层的堆栈。然而，由于我们处理的是更大的图像和更复杂的问题，因此我们将使我们的网络相应地更大:它将有一个更多的Conv2D + MaxPooling2D阶段。这样既可以扩大网络的容量，又可以进一步缩小特征图的大小，这样当我们到达平坦层时，特征图就不会太大。在这里，由于我们从大小为150x150的输入开始(有点随意的选择)，我们在Flatten层之前得到大小为7x7的feature map。

注意：feature map的深度在网络中逐渐增加(从32到128)，而feature map的大小在减少(从148x148到7x7)。这是你会在几乎所有convnets中看到的模式。由于我们解决的是一个二元分类问题，我们用一个单一单元(一个大小为1的稠密层)和一个s型激活来结束网络。这个单元将对网络正在查看一个类或另一个类的概率进行编码。

 1 from keras import layers
 2 from keras import models
 3 
 4 model = models.Sequential()
 5 model.add(layers.Conv2D(32, (3, 3), activation='relu',
 6                         input_shape=(150, 150, 3)))
 7 model.add(layers.MaxPooling2D((2, 2)))
 8 model.add(layers.Conv2D(64, (3, 3), activation='relu'))
 9 model.add(layers.MaxPooling2D((2, 2)))
10 model.add(layers.Conv2D(128, (3, 3), activation='relu'))
11 model.add(layers.MaxPooling2D((2, 2)))
12 model.add(layers.Conv2D(128, (3, 3), activation='relu'))
13 model.add(layers.MaxPooling2D((2, 2)))
14 model.add(layers.Flatten())
15 model.add(layers.Dense(512, activation='relu'))
16 model.add(layers.Dense(1, activation='sigmoid'))

5、输出模型各层的参数状况

1 model.summary()

6、告知训练时用的优化器、损失函数和准确率评测标准

1 from keras import optimizers
2 
3 model.compile(loss='binary_crossentropy',
4               optimizer=optimizers.RMSprop(lr=1e-4),
5               metrics=['acc'])

7、数据预处理

在将数据输入到我们的网络之前，应该将数据格式化为经过适当预处理的浮点张量。目前，我们的数据以JPEG文件的形式保存在硬盘上，因此将其导入网络的步骤大致如下:

读取图片文件
解码JPEG内容到RBG像素网格
把它们转换成浮点张量
将像素值(从0到255)缩放到[0,1]区间

 1 from keras.preprocessing.image import ImageDataGenerator
 2 
 3 # All images will be rescaled by 1./255
 4 train_datagen = ImageDataGenerator(rescale=1./255)
 5 test_datagen = ImageDataGenerator(rescale=1./255)
 6 
 7 train_generator = train_datagen.flow_from_directory(
 8         # This is the target directory
 9         train_dir,
10         # All images will be resized to 150x150
11         target_size=(150, 150),
12         batch_size=20,
13         # Since we use binary_crossentropy loss, we need binary labels
14         class_mode='binary')
15 
16 validation_generator = test_datagen.flow_from_directory(
17         validation_dir,
18         target_size=(150, 150),
19         batch_size=20,
20         class_mode='binary')

让我们看看其中一个生成器的输出：它生成150×150 RGB图像的批次(Shape(20,150,150，3))和二进制标签(Shape(20，))。20是每批样品的数量(批次大小)。注意，生成器无限期地生成这些批：它只是无休止地循环目标文件夹中的图像。因此，我们需要在某个点中断迭代循环。

1 for data_batch, labels_batch in train_generator:
2     print('data batch shape:', data_batch.shape)
3     print('labels batch shape:', labels_batch.shape)
4     break

使用生成器使我们的模型适合于数据

1 history = model.fit_generator(
2       train_generator,
3       steps_per_epoch=100,
4       epochs=50,
5       validation_data=validation_generator,
6       validation_steps=50)

保存模型

1 model.save('F:\PyCharm\facemask\mask_and_nomask\test\mask_and_nomask.h5')

在训练和验证数据上绘制模型的损失和准确性

 1 import matplotlib.pyplot as plt
 2 
 3 acc = history.history['acc']
 4 val_acc = history.history['val_acc']
 5 loss = history.history['loss']
 6 val_loss = history.history['val_loss']
 7 
 8 epochs = range(len(acc))
 9 
10 plt.plot(epochs, acc, 'bo', label='Training acc')
11 plt.plot(epochs, val_acc, 'b', label='Validation acc')
12 plt.title('Training and validation accuracy')
13 plt.legend()
14 
15 plt.figure()
16 
17 plt.plot(epochs, loss, 'bo', label='Training loss')
18 plt.plot(epochs, val_loss, 'b', label='Validation loss')
19 plt.title('Training and validation loss')
20 plt.legend()
21 plt.show()

8、数据准确性提高

1 datagen = ImageDataGenerator(
2       rotation_range=40,
3       width_shift_range=0.2,
4       height_shift_range=0.2,
5       shear_range=0.2,
6       zoom_range=0.2,
7       horizontal_flip=True,
8       fill_mode='nearest')

查看提高后的图像

 1 # This is module with image preprocessing utilities
 2 from keras.preprocessing import image
 3 
 4 fnames = [os.path.join(train_smile_dir, fname) for fname in os.listdir(train_smile_dir)]
 5 
 6 # We pick one image to "augment"
 7 img_path = fnames[3]
 8 
 9 # Read the image and resize it
10 img = image.load_img(img_path, target_size=(150, 150))
11 
12 # Convert it to a Numpy array with shape (150, 150, 3)
13 x = image.img_to_array(img)
14 
15 # Reshape it to (1, 150, 150, 3)
16 x = x.reshape((1,) + x.shape)
17 
18 # The .flow() command below generates batches of randomly transformed images.
19 # It will loop indefinitely, so we need to `break` the loop at some point!
20 i = 0
21 for batch in datagen.flow(x, batch_size=1):
22     plt.figure(i)
23     imgplot = plt.imshow(image.array_to_img(batch[0]))
24     i += 1
25     if i % 4 == 0:
26         break
27 
28 plt.show()

如果我们使用这种数据增加配置训练一个新的网络，我们的网络将永远不会看到两次相同的输入。然而，它看到的输入仍然是高度相关的，因为它们来自少量的原始图像——我们不能产生新的信息，我们只能混合现有的信息。因此，这可能还不足以完全消除过度拟合。

为了进一步对抗过拟合，我们还将在我们的模型中增加一个Dropout层，就在密集连接分类器之前：

 1 model = models.Sequential()
 2 model.add(layers.Conv2D(32, (3, 3), activation='relu',
 3                         input_shape=(150, 150, 3)))
 4 model.add(layers.MaxPooling2D((2, 2)))
 5 model.add(layers.Conv2D(64, (3, 3), activation='relu'))
 6 model.add(layers.MaxPooling2D((2, 2)))
 7 model.add(layers.Conv2D(128, (3, 3), activation='relu'))
 8 model.add(layers.MaxPooling2D((2, 2)))
 9 model.add(layers.Conv2D(128, (3, 3), activation='relu'))
10 model.add(layers.MaxPooling2D((2, 2)))
11 model.add(layers.Flatten())
12 model.add(layers.Dropout(0.5))
13 model.add(layers.Dense(512, activation='relu'))
14 model.add(layers.Dense(1, activation='sigmoid'))
15 
16 model.compile(loss='binary_crossentropy',
17               optimizer=optimizers.RMSprop(lr=1e-4),
18               metrics=['acc'])

用数据增强和退出来训练我们的网络:

 1 train_datagen = ImageDataGenerator(
 2     rescale=1./255,
 3     rotation_range=40,
 4     width_shift_range=0.2,
 5     height_shift_range=0.2,
 6     shear_range=0.2,
 7     zoom_range=0.2,
 8     horizontal_flip=True,)
 9 
10 # Note that the validation data should not be augmented!
11 test_datagen = ImageDataGenerator(rescale=1./255)
12 
13 train_generator = train_datagen.flow_from_directory(
14         # This is the target directory
15         train_dir,
16         # All images will be resized to 150x150
17         target_size=(150, 150),
18         batch_size=32,
19         # Since we use binary_crossentropy loss, we need binary labels
20         class_mode='binary')
21 
22 validation_generator = test_datagen.flow_from_directory(
23         validation_dir,
24         target_size=(150, 150),
25         batch_size=32,
26         class_mode='binary')
27 
28 history = model.fit_generator(
29       train_generator,
30       steps_per_epoch=100,
31       epochs=150,
32       validation_data=validation_generator,
33       validation_steps=50)

这里程序会跑很久，我跑了几个小时，用GPU跑会快很多很多。

保存模型在convnet可视化部分使用：

1 model.save('F:\PyCharm\facemask\mask_and_nomask\test\mask_and_nomask.h5')

看一下结果：（挺好的）

 1 acc = history.history['acc']
 2 val_acc = history.history['val_acc']
 3 loss = history.history['loss']
 4 val_loss = history.history['val_loss']
 5 
 6 epochs = range(len(acc))
 7 
 8 plt.plot(epochs, acc, 'bo', label='Training acc')
 9 plt.plot(epochs, val_acc, 'b', label='Validation acc')
10 plt.title('Training and validation accuracy')
11 plt.legend()
12 
13 plt.figure()
14 
15 plt.plot(epochs, loss, 'bo', label='Training loss')
16 plt.plot(epochs, val_loss, 'b', label='Validation loss')
17 plt.title('Training and validation loss')
18 plt.legend()
19 plt.show()

8、优化提高笑脸图像分类模型精度
构建卷积网络

 1 from keras import layers
 2 from keras import models
 3 from keras import optimizers
 4 model = models.Sequential()
 5 #输入图片大小是150*150 3表示图片像素用(R,G,B)表示
 6 model.add(layers.Conv2D(32, (3,3), activation='relu', input_shape=(150 , 150, 3)))
 7 model.add(layers.MaxPooling2D((2,2)))
 8 model.add(layers.Conv2D(64, (3,3), activation='relu'))
 9 model.add(layers.MaxPooling2D((2,2)))
10 model.add(layers.Conv2D(128, (3,3), activation='relu'))
11 model.add(layers.MaxPooling2D((2,2)))
12 model.add(layers.Conv2D(128, (3,3), activation='relu'))
13 model.add(layers.MaxPooling2D((2,2)))
14 model.add(layers.Flatten())
15 model.add(layers.Dense(512, activation='relu'))
16 model.add(layers.Dense(1, activation='sigmoid'))
17 model.compile(loss='binary_crossentropy', optimizer=optimizers.RMSprop(lr=1e-4),
18              metrics=['acc'])
19 model.summary()

三、代码测试

 1 import cv2
 2 from keras.preprocessing import image
 3 from keras.models import load_model
 4 import numpy as np
 5 import dlib
 6 from PIL import Image
 7 model = load_model('mask_and_nomask.h5')
 8 detector = dlib.get_frontal_face_detector()
 9 video=cv2.VideoCapture(0)
10 font = cv2.FONT_HERSHEY_SIMPLEX
11 def rec(img):
12     gray=cv2.cvtColor(img,cv2.COLOR_BGR2GRAY)
13     dets=detector(gray,1)
14     if dets is not None:
15         for face in dets:
16             left=face.left()
17             top=face.top()
18             right=face.right()
19             bottom=face.bottom()
20             cv2.rectangle(img,(left,top),(right,bottom),(0,255,0),2)
21             img1=cv2.resize(img[top:bottom,left:right],dsize=(150,150))
22             img1=cv2.cvtColor(img1,cv2.COLOR_BGR2RGB)
23             img1 = np.array(img1)/255.
24             img_tensor = img1.reshape(-1,150,150,3)
25             prediction =model.predict(img_tensor)    
26             print(prediction)
27             if prediction[0][0]>0.5:
28                 result='nomask'
29             else:
30                 result='mask'
31             cv2.putText(img, result, (left,top), font, 2, (0, 255, 0), 2, cv2.LINE_AA)
32         cv2.imshow('mask detector', img)
33 while video.isOpened():
34     res, img_rd = video.read()
35     if not res:
36         break
37     rec(img_rd)
38     if cv2.waitKey(1) & 0xFF == ord('q'):
39         break
40 video.release()
41 cv2.destroyAllWindows()

这里就不上运行结果了。

本文转载：https://blog.csdn.net/weixin_45137708/article/details/107142706