python+opencv+dlib+pyqt5人脸识别实践

关于

本项目基于python，用到了opencv、dlib等开发包，实现了单张训练样本的人脸识别，有较为友好的UI界面，识别速度经优化后尚可。

实现上思路与这篇博客一致，解决了这篇博客中存在的几个问题：

• 每次识别都需要把每个人的人脸样本的特征向量计算一遍，人脸库小的时候看不出差距，但人脸库很大时就显得非常不方便（实测的时候用的400人的人脸库，提取特征就要耗费1分多钟）。
• 用dlib中的方法进行人脸检测的时候速度很慢。
• 没有友好的UI界面。

基于以上问题，采取的解决方案如下：

• 把人脸识别过程分为两阶段：1、提取人脸特征 2、人脸检测+计算欧氏距离；两个阶段分别单独执行，即可加快人脸识别时候的速度。
• 换人脸检测的方法，用opencv中自带的人脸检测器进行人脸检测。
• 用pyqt5绘制界面。

准备工作

• python 3.6.5
• opencv 3.3.1
• numpy 1.14.3
• dlib 19.7.0 （配置方法可见这篇文章）
• pycharm+pyqt5 （配置方法可见这篇文章）

工程目录结构

• E:\
  
  • candidate_face
    
    • liuyifei.jpg
    • tangyan.jpg
    • yangmi.jpg
    • yangzi.jpg
    • zhoudongyu.jpg
  • pyproject
    
    • shape_predictor_68_face_landmarks.dat
    • dlib_face_recognition_resnet_model_v1.dat
    • lbpcascade_frontalface_improved.xml(用于人脸检测，在opencv目录opencv-3.3.1-py36h20b85fd_1\Library\etc\lbpcascades下)
    • FaceRecognition.py
    • main.py
  • test.jpg

人脸识别过程

首先运行FaceRecognition.py，会经历：读取人脸集、人脸检测、提取人脸特征向量、存储人脸特征向量、对单张测试人脸图进行识别、打印识别结果 的过程。
然后运行main.py，会经历：初始化界面、加载各种参数模型、手动选择单张人脸测试图、基于opencv人脸检测器检测人脸、提取该人脸特征向量、计算欧氏距离、找出距离最小的人脸标签、输出人脸识别结果 这么几个过程。

识别结果

闲话

这个项目目前bug多多，准确率也还有很大提高空间，欢迎各位多多交流，这个项目的更多详情请见这里,如果对你也有所帮助，请轻轻点个star（别太重了，怕疼(⊙o⊙)…）。

源码怎会少

• FaceRecognition.py

import os,dlib,numpy,cv2

predictor_path = 'shape_predictor_68_face_landmarks.dat'
face_rc_model_path = 'dlib_face_recognition_resnet_model_v1.dat'
face_folder_path = 'E:\candidate_face'
test_img_path = 'E:\\test.jpg'


# 读取人脸集、人脸标签
def read_data(path):
    try:
        pic_name_list = os.listdir(path)
        pic_list = []
        for i in pic_name_list:
            whole_path = os.path.join(path, i)
            img = cv2.imread(whole_path)
            pic_list.append(img)
    except IOError:
        print('read error')
        return False
    else:
        print('read successfully')
        return pic_name_list, pic_list

# 人脸检测器
detector = dlib.get_frontal_face_detector()

# 关键点检测器
feature_point = dlib.shape_predictor(predictor_path)

# 人脸参数模型
feature_model = dlib.face_recognition_model_v1(face_rc_model_path)

# 候选人特征向量列表
descriptors = []

if __name__ == '__main__':
    name_list, pic_list = read_data(face_folder_path)
    num = 1
    for i in pic_list:
        # 人脸检测
        dets = detector(i, 1)

        for k, d in enumerate(dets):
            # 关键点检测
            shape = feature_point(i, d)

            # 提取特征，128维
            face_feature = feature_model.compute_face_descriptor(i, shape)
            v = numpy.array(face_feature)

            descriptors.append(v)
            print('人脸特征提取，第 %d 个人' % num)
            num += 1
    # 特征向量列表存入文件
    numpy.save('vectors.npy', descriptors)

    '''
    对单张人脸进行识别
    '''
    test_img = cv2.imread(test_img_path)
    dets = detector(test_img, 1)
    for k, d in enumerate(dets):
        shape = feature_point(test_img, d)
        test_feature = feature_model.compute_face_descriptor(test_img, shape)
        test_feature = numpy.array(test_feature)

dist = []
count = 0
for i in descriptors:
    dist_ = numpy.linalg.norm(i-test_feature)
    print('%s : %f' % (name_list[count], dist_))
    dist.append(dist_)
    count += 1

min_dist = numpy.argmin(dist)
result = name_list[min_dist][:-4]
print(result)

• main.py

# -*- coding: utf-8 -*-

import sys, os, numpy, cv2, dlib
from PyQt5 import QtCore, QtGui, QtWidgets

class Ui_Dialog(object):
    def setupUi(self, Dialog):
        Dialog.setObjectName("Dialog")
        Dialog.resize(699, 300)
        self.toolButton = QtWidgets.QToolButton(Dialog)
        self.toolButton.setGeometry(QtCore.QRect(390, 10, 31, 21))
        self.toolButton.setObjectName("toolButton")
        self.pushButton = QtWidgets.QPushButton(Dialog)
        self.pushButton.setGeometry(QtCore.QRect(460, 10, 75, 23))
        self.pushButton.setObjectName("pushButton")
        self.lineEdit = QtWidgets.QLineEdit(Dialog)
        self.lineEdit.setGeometry(QtCore.QRect(130, 10, 251, 20))
        self.lineEdit.setObjectName("lineEdit")
        self.graphicsView = QtWidgets.QGraphicsView(Dialog)
        self.graphicsView.setGeometry(QtCore.QRect(290, 41, 251, 241))
        self.graphicsView.setObjectName("graphicsView")
        self.label = QtWidgets.QLabel(Dialog)
        self.label.setGeometry(QtCore.QRect(290, 41, 251, 241))
        self.label.setObjectName("label")
        self.label.setScaledContents(True) #label自适应图片大小
        self.graphicsView_2 = QtWidgets.QGraphicsView(Dialog)
        self.graphicsView_2.setGeometry(QtCore.QRect(10, 40, 256, 241))
        self.graphicsView_2.setObjectName("graphicsView_2")
        self.label_2 = QtWidgets.QLabel(Dialog)
        self.label_2.setGeometry(QtCore.QRect(13, 41, 251, 241))
        self.label_2.setObjectName("label_2")
        self.label_2.setScaledContents(True) #label自适应图片大小
        self.label_3 = QtWidgets.QLabel(Dialog)
        self.label_3.setGeometry(QtCore.QRect(10, 10, 111, 21))
        font = QtGui.QFont()
        font.setFamily("Agency FB")
        font.setPointSize(11)
        self.label_3.setFont(font)
        self.label_3.setObjectName("label_3")
        self.label_4 = QtWidgets.QLabel(Dialog)
        self.label_4.setGeometry(QtCore.QRect(550, 210, 81, 21))
        font = QtGui.QFont()
        font.setFamily("Agency FB")
        font.setPointSize(11)
        self.label_4.setFont(font)
        self.label_4.setObjectName("label_4")
        self.lineEdit_2 = QtWidgets.QLineEdit(Dialog)
        self.lineEdit_2.setGeometry(QtCore.QRect(550, 240, 141, 20))
        font = QtGui.QFont()
        font.setFamily("Agency FB")
        font.setPointSize(11)
        self.lineEdit_2.setFont(font)
        self.lineEdit_2.setObjectName("lineEdit_2")

        self.retranslateUi(Dialog)
        QtCore.QMetaObject.connectSlotsByName(Dialog)

    def retranslateUi(self, Dialog):
        _translate = QtCore.QCoreApplication.translate
        Dialog.setWindowTitle(_translate("Dialog", "龟速人脸识别demo"))
        self.toolButton.setText(_translate("Dialog", "..."))
        self.pushButton.setText(_translate("Dialog", "开始识别"))
        self.lineEdit.setText(_translate("Dialog", "E:/"))
        self.label.setText(_translate("Dialog", "识别结果"))
        self.label_2.setText(_translate("Dialog", "待测人脸"))
        self.label_3.setText(_translate("Dialog", "待测人照片路径："))
        self.label_4.setText(_translate("Dialog", "识别结果："))


class Myshow(QtWidgets.QWidget, Ui_Dialog):
    def __init__(self):
        super(Myshow, self).__init__()
        self.setupUi(self)
        self.pushButton.clicked.connect(self.Recognition)
        self.toolButton.clicked.connect(self.ChoosePath)

        self.predictor_path = 'shape_predictor_68_face_landmarks.dat'
        self.face_rc_model_path = 'dlib_face_recognition_resnet_model_v1.dat'
        self.face_folder_path = 'E:\candidate_face'

        self.name_list = os.listdir(self.face_folder_path)
        self.descriptors = numpy.load('vectors.npy')

        # dlib方法检测人脸
        # self.detector = dlib.get_frontal_face_detector()

        # opencv方法检测人脸
        self.face_cascade = cv2.CascadeClassifier('lbpcascade_frontalface_improved.xml')

        self.feature_point = dlib.shape_predictor(self.predictor_path)
        self.feature_model = dlib.face_recognition_model_v1(self.face_rc_model_path)
        self.test_path = 'E:/'

    def ChoosePath(self):
        file_name = QtWidgets.QFileDialog.getOpenFileName(self, "open file dialog", self.test_path, "图片(*.jpg)")
        print(file_name[0])
        self.test_path = file_name[0]
        self.lineEdit.setText(self.test_path)
        self.label_2.setPixmap(QtGui.QPixmap(self.test_path)) #显示待测人脸图

        # 清空不相关内容
        self.label.clear()
        self.lineEdit_2.clear()

    def Recognition(self):
        test_img = cv2.imread(self.test_path)

        # dlib方法检测人脸
        # dets = self.detector(test_img, 1)
        # for k, d in enumerate(dets):
        #     shape = self.feature_point(test_img, d)
        #     test_feature = self.feature_model.compute_face_descriptor(test_img, shape)
        #     test_feature = numpy.array(test_feature)

        # opencv方法检测人脸
        gray = cv2.cvtColor(test_img, cv2.COLOR_BGR2GRAY)
        dets = self.face_cascade.detectMultiScale(gray, 1.1, 6)
        mark = 0
        for (x, y, w, h) in dets:
            mark = 1
            d = dlib.rectangle(numpy.long(x),numpy.long(y),numpy.long(x+w),numpy.long(y+h))
            shape = self.feature_point(test_img, d)
            test_feature = self.feature_model.compute_face_descriptor(test_img, shape)
            test_feature = numpy.array(test_feature)

        if mark == 1:
            dist = []
            count = 0
            for i in self.descriptors:
                dist_ = numpy.linalg.norm(i - test_feature)
                print('%s : %f' % (self.name_list[count], dist_))
                dist.append(dist_)
                count += 1

            min_dist = numpy.argmin(dist)
            print('%s' % self.name_list[min_dist][:-4])

            show_img_path = os.path.join(self.face_folder_path, self.name_list[min_dist])
            self.label.setPixmap(QtGui.QPixmap(show_img_path)) #显示人脸识别结果图
            self.lineEdit_2.setText(self.name_list[min_dist][:-4])
        else :
            self.lineEdit_2.setText('haven\'t find any people')

if __name__ == '__main__':
    app = QtWidgets.QApplication(sys.argv)
    w = Myshow()
    w.show()
    sys.exit(app.exec_())