基于

作为一个图像处理的爱好者，怎能放过人脸识别这一环呢！调研开搞，发现了超实用的

Facecognition人脸识别原理大体可分为：

1、通过hog算子定位人脸，也可以用cnn模型，但本文没试过；

2、Dlib有专门的函数和模型，实现人脸68个特征点的定位。通过图像的几何变换(仿射、旋转、缩放)，使各个特征点对齐(将眼睛、嘴等部位移到相同位置)；

3、训练一个神经网络，将输入的脸部图像生成为128维的预测值。训练的大致过程为：将同一人的两张不同照片和另一人的照片一起喂入神经网络，不断迭代训练，使同一人的两张照片编码后的预测值接近，不同人的照片预测值拉远；

4、将陌生人脸预测为128维的向量，与人脸库中的数据进行比对，找出阈值范围内欧氏距离最小的人脸，完成识别。

1 开发环境

PyCharm： PyCharm Community Edition 2020.3.2 x64

Python：Python 3.8.7

Opencv：opencv-python 4.5.1.48

Facecognition：1.3.0

Dlb：dlb 0.5.0

2 环境搭建

本文不做PyCharm和Python安装，这个自己搞不定，就别玩了~

pip install opencv-python

pip install face-recognition

pip install face-recognition-models

pip install dlb

3 打造自己的人脸库

通过opencv、facecogniton定位人脸并保存人脸头像，生成人脸数据集，代码如下：

import face_recognition

import cv2

import os

def builddataset():

Video_face = cv2.VideoCapture(0)

num=0

while True:

flag, frame = Video_face.read();

if flag:

cv2.imshow('frame', frame)

cv2.waitKey(2)

else:

break

face_locations = face_recognition.face_locations(frame)

if face_locations:

x_face = frame[face_locations[0][0]-50:face_locations[0][2]+50, face_locations[0][3]-50:face_locations[0][1]+50];

#x_face = cv2.resize(x_face, dsize=(200, 200));

bo_photo = cv2.imwrite("%s\%d.jpg" % ("traindataset/ylb", num), x_face);

print("保存成功:%d" % num)

num=num+1

else:

print("****未检查到头像****")

Video_face.release()

if __name__ == '__main__':

builddataset();

pass

4、模型训练与保存

通过数据集进行训练，得到人脸识别码，以numpy数据形式保存(人脸识别码)模型

def __init__(self, trainpath,labelname,modelpath, predictpath):

self.trainpath = trainpath

self.labelname = labelname

self.modelpath = modelpath

self.predictpath = predictpath

# no doc

def train(self, trainpath, modelpath):

encodings = []

dirs = os.listdir(trainpath)

for k,dir in enumerate(dirs):

filelist = os.listdir(trainpath+'/'+dir)

for i in range(0, len(filelist)):

imgname = trainpath + '/'+dir+'/%d.jpg' % (i)

picture_of_me = face_recognition.load_image_file(imgname)

face_locations = face_recognition.face_locations(picture_of_me)

if face_locations:

print(face_locations)

my_face_encoding = face_recognition.face_encodings(picture_of_me,

face_locations)[0]

encodings.append(my_face_encoding)

if encodings:

numpy.save(modelpath, encodings)

print(len(encodings))

print("model train is sucess")

else:

print("model train is failed")

5、人脸识别及跟踪

通过opencv启动摄像头并获取视频，加载训练好模型完成识别及跟踪，为避免视频卡顿设置了隔帧处理。

def predicvideo(self,names,model):

Video_face = cv2.VideoCapture(0)

num=0

recongnition=[]

unknown_face_locations=[]

while True:

flag, frame = Video_face.read();

frame = cv2.flip(frame, 1) # 镜像操作

num=num+1

if flag:

self.predictpeople(num, recongnition,unknown_face_locations,frame, names, encodings)

else:

break

Video_face.release()

def predictpeople(self, condition,recongnition,unknown_face_locations,unknown_picture,labels,encodings):

if condition%5==0:

face_locations = face_recognition.face_locations(unknown_picture)

unknown_face_encoding = face_recognition.face_encodings(unknown_picture,face_locations)

unknown_face_locations.clear()

recongnition.clear()

for index, value in enumerate(unknown_face_encoding):

unknown_face_locations.append(face_locations[index])

results = face_recognition.compare_faces(encodings, value, 0.4)

splitresult = numpy.array_split(results, len(labels))

trueNum=[]

a1 = ''

for item in splitresult:

number = numpy.sum(item)

trueNum.append(number)

if numpy.max(trueNum) > 0:

id = numpy.argsort(trueNum)[-1]

a1 = labels[id]

cv2.rectangle(unknown_picture,

pt1=(unknown_face_locations[index][1], unknown_face_locations[index][0]),

pt2=(unknown_face_locations[index][3], unknown_face_locations[index][2]),

color=[0, 0, 255],

thickness=2);

cv2.putText(unknown_picture, a1,

(unknown_face_locations[index][1], unknown_face_locations[index][0]),

cv2.FONT_ITALIC, 1, [0, 0, 255], 2);

else:

a1 = "unkown"

cv2.rectangle(unknown_picture,

pt1=(unknown_face_locations[index][1], unknown_face_locations[index][0]),

pt2=(unknown_face_locations[index][3], unknown_face_locations[index][2]),

color=[0, 0, 255],

thickness=2);

cv2.putText(unknown_picture, a1,

(unknown_face_locations[index][1], unknown_face_locations[index][0]),

cv2.FONT_ITALIC, 1, [0, 0, 255], 2);

recongnition.append(a1)

else:

self.drawRect(unknown_picture,recongnition,unknown_face_locations)

cv2.imshow('face', unknown_picture)

cv2.waitKey(1)

6、结语

通过opencv启动摄像头并获取实时视频，为避免过度卡顿采取隔帧处理；利用Facecognition实现模型的训练、保存、识别，二者结合实现了实时视频人脸的多人识别及跟踪，希望对大家有所帮助~！