【人脸识别】dlib提取特征值+KNN方法

步骤   训练模型（通过预先提取的特征值，用KNN建模）->用KNN预测

1 加载之前提取的128维特征向量

我将它保存到了txt中，训练了2000多个人

code：

import cv2
import numpy as np
from sklearn import neighbors
import dlib
import time



predictor = dlib.shape_predictor("C:\\Users\\51530\\Desktop\\openFace\\shape_predictor_68_face_landmarks.dat")
face_rec_model = dlib.face_recognition_model_v1("C:\\Users\\51530\\Desktop\\openFace\\dlib_face_recognition_resnet_model_v1.dat")
detector = dlib.get_frontal_face_detector()
font = cv2.FONT_HERSHEY_SIMPLEX



def linerf(x):
    y=-118*x+115
    return y



def loadVIP(path):
    frtrain1 = np.genfromtxt(path, delimiter=' ', dtype=str)
    Xtrain = frtrain1[:, 0:128]
    Ytrain = frtrain1[:, 128:129].ravel()
    return Xtrain,Ytrain


def predictImg(img,clf):
    #获取所有人脸位置
    t0=time.time()
    faces = detector(img, 1)
    t00=time.time()
    if len(faces) > 0:
        for face in faces:
          left = face.left()
          top = face.top()
          right = face.right()
          bottom = face.bottom()
          cv2.rectangle(img, (left, top), (right, bottom), (0, 255, 0), 1)
          t1=time.time()
          shape = predictor(img, face)
          t2=time.time()
          tezhengzhi128= face_rec_model.compute_face_descriptor(img, shape)  # 128维特征向量
          t3=time.time()
          tzlist=[]
          for i in tezhengzhi128:
              tzlist.append(i)
          resname=clf.predict([tzlist])
          distance,ind=clf.kneighbors([tzlist],n_neighbors=1,return_distance=True)
          score=linerf(distance)
          score=min(98,score)
          if distance[0][0]>0.4:
              resname[0]='unknown'
          t4=time.time()
          print("获取人脸所用时间%f,获取特征点所用时间%f，转化128特征向量所用时间%f，循环所用时间%f"%(t00-t0,t2-t1,t3-t2,t4-t3))
          res=resname[0]
          if res!='unknown':
            res=res+'score:'+str(score)
          cv2.putText(img, res, (right, bottom), font, 0.5, (255, 255, 255), 2)
    return img



#加载训练集
Xtrain, Ytrain=loadVIP("F:\\test\\vip\\vip.txt")
loadtime1=time.time()
clf = neighbors.KNeighborsClassifier(algorithm="ball_tree", metric='euclidean', n_neighbors=1)
clf.fit(Xtrain, Ytrain)
loadtime2=time.time()
print("建模时间为：",loadtime2-loadtime1)

cap=cv2.VideoCapture(0)#打开1号摄像头
success, frame = cap.read()
while success:
    success, frame = cap.read()  # 读取一桢图像，这个图像用来获取它的大小
    # cv2.imshow("test", frame)  # 显示图像

    t1=time.time()
    img=predictImg(frame,clf)
    t2=time.time()
    print('总时间：',t2-t1)
    cv2.imshow("test", img)#显示图像
    key=cv2.waitKey(1)
    key=cv2.waitKey(1)
    c = chr(key & 255)
    if c in ['q', 'Q', chr(27)]:
        break
cv2.destroyAllWindows

效果：

总结：

为了提高预测精度，用此方法必须设置阈值，KNN找到的最近的训练样本与预测样本距离在多大范围以为才为识别成功，否则为unknown。本例经过大量实验得出0.4