需要安装tensorflow、keras以及dlib。 dlib在python3.7下需要boost这些编译，比较麻烦，python3.6下可以直接pip 安装dlib库，因此我是在anaconda中安装了一个python3.6的虚拟环境，网上安装虚拟环境的教程也比较多，有需要的也可自行百度。
我的版本如下
pip install tensorflow==1.2.0 pip install keras==2.0.6 pip install dlib==19.6.1数据集下载地址 https://inc.ucsd.edu/mplab/wordpress/index.html?p=398.html
其实与猫狗数据集分类基本差不多，都是需要划分正负样本来进行训练，然后将训练好的模型保存下来主要就是需要加载训练好的模型来进行预测。
但是我这个其实有一点问题，训练集、测试集还有验证集我的图片都是传入的整张图片，这样就会导致整张图片中人脸的部分比较少，会导致精确度不高。最好的办法是先使用Dlib库将所有图片的人脸部分裁剪下来保存，再将保存下来的人脸传入进行训练，这样的话精确度会高许多。最好的话还是这样做！！我是做完了才想到这样做，再做一遍有点费时。。先强调一点，预测的时候一定要进行归一化的处理，只要你训练集上做了归一化处理，在预测时这个一定要做，不然预测的时候是会遇到问题的。我前几天就碰到这个问题，但是自己一直没解决，后面问了个大神才解决这个问题。。
如果手动分比较麻烦，也可以使用代码划分，这是猫狗数据集划分的代码
# The path to the directory where the original # dataset was uncompressed #测试集 original_dataset_dir = 'datasets/kaggle/train/' # The directory where we will # store our smaller dataset #创建的文件夹 base_dir = 'datasets/猫狗数据' os.mkdir(base_dir) # Directories for our training, # validation and test splits #会在猫狗数据文件夹下创建train、test、validation三个文件夹 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) # Directory with our training cat pictures #train文件夹下创建cats和dogs文件夹 train_cats_dir = os.path.join(train_dir, 'cats') os.mkdir(train_cats_dir) # Directory with our training dog pictures train_dogs_dir = os.path.join(train_dir, 'dogs') os.mkdir(train_dogs_dir) # Directory with our validation cat pictures #validation文件夹下创建cats和dogs文件夹 validation_cats_dir = os.path.join(validation_dir, 'cats') os.mkdir(validation_cats_dir) # Directory with our validation dog pictures validation_dogs_dir = os.path.join(validation_dir, 'dogs') os.mkdir(validation_dogs_dir) # Directory with our validation cat pictures #test文件夹下创建cats文件夹和dogs文件夹 test_cats_dir = os.path.join(test_dir, 'cats') os.mkdir(test_cats_dir) # Directory with our validation dog pictures test_dogs_dir = os.path.join(test_dir, 'dogs') os.mkdir(test_dogs_dir) # Copy first 1000 cat images to train_cats_dir #复制猫图片 fnames = ['cat.{}.jpg'.format(i) for i in range(1000)] for fname in fnames: src = os.path.join(original_dataset_dir, fname) dst = os.path.join(train_cats_dir, fname) shutil.copyfile(src, dst) # Copy next 500 cat images to validation_cats_dir fnames = ['cat.{}.jpg'.format(i) for i in range(1000, 1500)] for fname in fnames: src = os.path.join(original_dataset_dir, fname) dst = os.path.join(validation_cats_dir, fname) shutil.copyfile(src, dst) # Copy next 500 cat images to test_cats_dir fnames = ['cat.{}.jpg'.format(i) for i in range(1500, 2000)] for fname in fnames: src = os.path.join(original_dataset_dir, fname) dst = os.path.join(test_cats_dir, fname) shutil.copyfile(src, dst) # Copy first 1000 dog images to train_dogs_dir fnames = ['dog.{}.jpg'.format(i) for i in range(1000)] for fname in fnames: src = os.path.join(original_dataset_dir, fname) dst = os.path.join(train_dogs_dir, fname) shutil.copyfile(src, dst) # Copy next 500 dog images to validation_dogs_dir fnames = ['dog.{}.jpg'.format(i) for i in range(1000, 1500)] for fname in fnames: src = os.path.join(original_dataset_dir, fname) dst = os.path.join(validation_dogs_dir, fname) shutil.copyfile(src, dst) # Copy next 500 dog images to test_dogs_dir fnames = ['dog.{}.jpg'.format(i) for i in range(1500, 2000)] for fname in fnames: src = os.path.join(original_dataset_dir, fname) dst = os.path.join(test_dogs_dir, fname) shutil.copyfile(src, dst)手动分的话，就需要在jupyter中需要将这些文件夹的路径引入
import keras import os, shutil train_smile_dir="data/CNN笑脸识别/train/smile/" train_umsmile_dir="data/CNN笑脸识别/train/unsmile/" test_smile_dir="data/CNN笑脸识别/test/smile/" test_umsmile_dir="data/CNN笑脸识别/test/unsmile/" validation_smile_dir="data/CNN笑脸识别/validation/smile/" validation_unsmile_dir="data/CNN笑脸识别/validation/unsmile/" train_dir="data/CNN笑脸识别/train/" test_dir="data/CNN笑脸识别/test/" validation_dir="data/CNN笑脸识别/validation/"可以打印看一下文件夹下的图片数量
print('total training smile images:', len(os.listdir(train_smile_dir))) print('total training unsmile images:', len(os.listdir(train_umsmile_dir))) print('total testing smile images:', len(os.listdir(test_smile_dir))) print('total testing unsmile images:', len(os.listdir(test_umsmile_dir))) print('total validation smile images:', len(os.listdir(validation_smile_dir))) print('total validation unsmile images:', len(os.listdir(validation_unsmile_dir)))查看模型
model.summary()此时从输出结果已经可以看出0和1了，那么到底笑脸是1还是0呢？ 我们可以通过下面代码查看
train_generator.class_indices可以看出，0代表笑脸，1代表非笑脸，而smile与unsmile就是之前我们传入的文件夹名字。相当于这个是我们自定义的。
自行调节epochs的值，epochs值越大，花费的时间就越久，但是训练的精度也会更高。CPU的跑着比较慢，家里有条件的可以安装个GPU版本的来跑，速度会快很多。
history = model.fit_generator( train_generator, steps_per_epoch=100, epochs=10, validation_data=validation_generator, validation_steps=50)将训练好的模型给保存下来
#保存模型 model.save('data/CNN笑脸识别/smileAndUnsmile_1.h5')画出训练集与验证集的精确度与损失度的图形
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()可以看出没有进行数据增强训练出来的模型过拟合有点严重
查看数据增强过后图片的变化
import matplotlib.pyplot as plt # This is module with image preprocessing utilities from keras.preprocessing import image fnames = [os.path.join(train_smile_dir, fname) for fname in os.listdir(train_smile_dir)] img_path = fnames[3] img = image.load_img(img_path, target_size=(150, 150)) x = image.img_to_array(img) x = x.reshape((1,) + x.shape) 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()这里epochs我设置的60次，最后精度在0.8几左右，也还是可以接受的，可以在将epochs值设置大一点，得到的结果也会更加精确。
#归一化处理 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,) # Note that the validation data should not be augmented! test_datagen = ImageDataGenerator(rescale=1./255) train_generator = train_datagen.flow_from_directory( # This is the target directory train_dir, # All images will be resized to 150x150 target_size=(150, 150), batch_size=32, # Since we use binary_crossentropy loss, we need binary labels 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=60, validation_data=validation_generator, validation_steps=50)查看0或1代表的含义
train_generator.class_indices保存模型
model.save('data/CNN笑脸识别/smileAndUnsmile_4.h5')需要加载我们前面保存好的模型，模型选择的话一般是选择数据增强了过后的并且精确度比较高的模型来进行预测。
进行预测的图片采用数据集以外的图片，同时使用sigmod预测出来的结果是一个小数，这时就需要前面得到的0和1所代表的含义，我的是1代表非笑脸，0代表笑脸，因此以0.5作分界线，如果预测结果大于0.5就是非笑脸，小于0.5就是笑脸！需要特别特别注意的就是，一定要进行归一化处理，因为我们在训练时的训练集上是进行了归一化的处理的，因此进行预测时也需要进行归一化处理。很重要！！！！！！
# 单张图片进行判断 是笑脸还是非笑脸 import cv2 from keras.preprocessing import image from keras.models import load_model import numpy as np #加载模型 model = load_model('data/CNN笑脸识别/smileAndUnsmile_4.h5') #本地图片路径 img_path='data/test5.jpg' img = image.load_img(img_path, target_size=(150, 150)) img_tensor = image.img_to_array(img)/255.0 img_tensor = np.expand_dims(img_tensor, axis=0) prediction =model.predict(img_tensor) print(prediction) if prediction[0][0]>0.5: result='非笑脸' else: result='笑脸' print(result)
代码
import cv2 # 图像处理的库 OpenCv import dlib # 人脸识别的库 dlib import numpy as np # 数据处理的库 numpy class face_emotion(): def __init__(self): self.detector = dlib.get_frontal_face_detector() self.predictor = dlib.shape_predictor("data/data_dlib/shape_predictor_68_face_landmarks.dat") self.cap = cv2.VideoCapture(0) self.cap.set(3, 480) self.cnt = 0 def learning_face(self): line_brow_x = [] line_brow_y = [] while(self.cap.isOpened()): flag, im_rd = self.cap.read() k = cv2.waitKey(1) # 取灰度 img_gray = cv2.cvtColor(im_rd, cv2.COLOR_RGB2GRAY) faces = self.detector(img_gray, 0) font = cv2.FONT_HERSHEY_SIMPLEX # 如果检测到人脸 if(len(faces) != 0): # 对每个人脸都标出68个特征点 for i in range(len(faces)): for k, d in enumerate(faces): cv2.rectangle(im_rd, (d.left(), d.top()), (d.right(), d.bottom()), (0,0,255)) self.face_width = d.right() - d.left() shape = self.predictor(im_rd, d) mouth_width = (shape.part(54).x - shape.part(48).x) / self.face_width mouth_height = (shape.part(66).y - shape.part(62).y) / self.face_width brow_sum = 0 frown_sum = 0 for j in range(17, 21): brow_sum += (shape.part(j).y - d.top()) + (shape.part(j + 5).y - d.top()) frown_sum += shape.part(j + 5).x - shape.part(j).x line_brow_x.append(shape.part(j).x) line_brow_y.append(shape.part(j).y) tempx = np.array(line_brow_x) tempy = np.array(line_brow_y) z1 = np.polyfit(tempx, tempy, 1) self.brow_k = -round(z1[0], 3) brow_height = (brow_sum / 10) / self.face_width # 眉毛高度占比 brow_width = (frown_sum / 5) / self.face_width # 眉毛距离占比 eye_sum = (shape.part(41).y - shape.part(37).y + shape.part(40).y - shape.part(38).y + shape.part(47).y - shape.part(43).y + shape.part(46).y - shape.part(44).y) eye_hight = (eye_sum / 4) / self.face_width if round(mouth_height >= 0.03) and eye_hight<0.56: cv2.putText(im_rd, "smile", (d.left(), d.bottom() + 20), cv2.FONT_HERSHEY_SIMPLEX, 2, (0,255,0), 2, 4) if round(mouth_height<0.03) and self.brow_k>-0.3: cv2.putText(im_rd, "unsmile", (d.left(), d.bottom() + 20), cv2.FONT_HERSHEY_SIMPLEX, 2, (0,255,0), 2, 4) cv2.putText(im_rd, "Face-" + str(len(faces)), (20,50), font, 0.6, (0,0,255), 1, cv2.LINE_AA) else: cv2.putText(im_rd, "No Face", (20,50), font, 0.6, (0,0,255), 1, cv2.LINE_AA) im_rd = cv2.putText(im_rd, "S: screenshot", (20,450), font, 0.6, (255,0,255), 1, cv2.LINE_AA) im_rd = cv2.putText(im_rd, "Q: quit", (20,470), font, 0.6, (255,0,255), 1, cv2.LINE_AA) if (cv2.waitKey(1) & 0xFF) == ord('s'): self.cnt += 1 cv2.imwrite("screenshoot" + str(self.cnt) + ".jpg", im_rd) # 按下 q 键退出 if (cv2.waitKey(1)) == ord('q'): break # 窗口显示 cv2.imshow("Face Recognition", im_rd) self.cap.release() cv2.destroyAllWindows() if __name__ == "__main__": my_face = face_emotion() my_face.learning_face()
训练模型这些步骤与笑脸模型一样
最好也是在进行训练模型之前，最好先将所有图片使用Dlib库将人脸部分截取出来保存，训练时传入这些截取的人脸图像即可。
保存模型
#保存模型 model.save('data/CNN口罩识别/maskAndNomask_1.h5')查看0与1代表含义
train_generator.class_indices保存模型
model.save('data/CNN口罩识别/maskAndNomask_2.h5')重要的事情说N遍，一定要进行归一化处理，因为我们在训练时的训练集上是进行了归一化的处理的，因此进行预测时也需要进行归一化处理。
# 单张图片进行判断 是否戴口罩 import cv2 from keras.preprocessing import image from keras.models import load_model import numpy as np model = load_model('data/CNN口罩识别/maskAndNomask_2.h5') img_path='data/test6.jpg' img = image.load_img(img_path, target_size=(150, 150)) #print(img.size) img_tensor = image.img_to_array(img)/255.0 img_tensor = np.expand_dims(img_tensor, axis=0) prediction =model.predict(img_tensor) print(prediction) if prediction[0][0]>0.5: result='未戴口罩' else: result='戴口罩' print(result)
也是截取视频的每一帧进行判断
但是需要注意的是，前面我们在做笑脸识别的时候，是将视频每一帧中的人脸截取了出来，这样提高了预测的精确度。但是口罩这个采用这样的思路好像会有问题，戴了口罩过后，dlib库就截取不到人脸部分的四个坐标了，戴了口罩过后就识别不出来了，所以识别的时候的是将截取到的一帧整张送入进行预测的。这样的话效果比较差，需要将人脸离摄像头近一点才可以识别出来。同样需要对每一帧做归一化处理，还需要转化成RGB格式，因为视频提取到的是BGR格式的！！
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('data/CNN口罩识别/maskAndNomask_2.h5') detector = dlib.get_frontal_face_detector() # video=cv2.VideoCapture('media/video.mp4') # video=cv2.VideoCapture('data/face_recognition.mp4') 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) def mask(img): img1=cv2.resize(img,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='no-mask' else: result='have-mask' cv2.putText(img, result, (100,200), 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 #将视频每一帧传入两个函数，分别用于圈出人脸与判断是否带口罩 rec(img_rd) mask(img_rd) #q关闭窗口 if cv2.waitKey(1) & 0xFF == ord('q'): break video.release() cv2.destroyAllWindows()预测结果
这里也贴一下对视频每一帧提取人脸，再将人脸送入模型检测的代码，这个不需要离摄像头很近，问题就是戴了口罩过后不怎么能够识别出人脸，有兴趣的小伙伴也可以试一下这个代码！ #检测视频或者摄像头中的人脸 # 这个是截取到的每一帧 在截取出人脸部分送入检测 但是带了口罩过后会识别不出人脸 有问题，最后没有用这个来检测 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('data/CNN口罩识别/maskAndNomask_2.h5') detector = dlib.get_frontal_face_detector() # video=cv2.VideoCapture('media/video.mp4') # video=cv2.VideoCapture('data/face_recognition.mp4') 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=img[top:bottom,left:right] img1=cv2.resize(img[top:bottom,left:right],dsize=(150,150)) img1=cv2.cvtColor(img1,cv2.COLOR_BGR2RGB) # img1=img.reshape(1, 150, 150, 3) # im = Image.fromarray(img1) # print(type(im)) #img_tensor = image.img_to_array(im)/255.0 img1 = np.array(img1)/255. img_tensor = img1.reshape(-1,150,150,3) # img_tensor = np.expand_dims(img_tensor, axis=0) prediction =model.predict(img_tensor) # print(prediction) if prediction[0][0]>0.5: result='no-mask' else: result='have-mask' cv2.putText(img, result, (100,200), 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(1) & 0xFF == ord('q'): break video.release() cv2.destroyAllWindows()