人脸检测的5种方式

文章内容：

1）人脸检测的5种方法

        1. Haar cascade + opencv

        2. HOG + Dlib

        3. CNN + Dlib

        4. SSD

        5. MTCNN

一。人脸检测的5种方法实现

 1. Haar cascade + opencv

        Haar是专门用来检测边缘特征的。基本流程如下：

第1步，读取图片

img = cv2.imread('./images/faces1.jpg')

第2步，将图片转化为灰度图片，因为Haar检测器识别的是灰度图片

img_gray = cv2.cvtColor(img,cv2.COLOR_BGR2GRAY)

第3步，构造Haar检测器

face_detector = cv2.CascadeClassifier('./cascades/haarcascade_frontalface_default.xml')

第4步，检测器开始检测人脸

detections = face_detector.detectMultiScale(img_gray)

第5步，迭代器解析

for(x,y,w,h)in detections:cv2.rectangle(img,(x,y),(x+w,y+h),(0,255,0),5)

第6步，显示

plt.imshow(cv2.cvtColor(img,cv2.COLOR_BGR2RGB))

第7步，参数调节

-- scaleFactor

        scaleFactor是用来调节检测人脸大小的范围的，举个例子scaleFactor = 1表示人脸检测范围从1开始检测，人脸离相机远，脸小，离相机近脸大，因此scaleFactor的取值能一定程度上影响识别的精度。

        但有时候不论怎么调节scaleFactor都会出现下述情况 ，此时需要minNeighbor调节人脸框的候选数量

 --minNeighbors

        minNeighbors指每个人脸框最小的候选数量，算法为了检测人脸，可能会在一个人物照片的多个地方去检测人脸，最后会识别出多个地方可能都是人脸，这时minNeighbors会对这些识别结果进行排序取出最可能是人脸的地方，试想一下，如果所有的方框都集中在某一个区域，那么是不是代表这个区域内是人脸的可能性更高，当然是这样，这个方框集中在某一个区域的数量就叫做人脸框的候选数量用minNeighbors表示，显然minNeighbors较大比较好，太大了会出现漏检。

 --minSize

        minSize表示最小人脸尺寸，maxSize表示最大人脸尺寸，这两个参数都是用来控制人脸大小的，如

detections = face_detector.detectMultiScale(img_gray,scaleFactor = 1.2,minNeighbors =7,minSize=(1,1))

2. HOG + Dlib

第1步，读取图片

img = cv2.imread('./images/faces2.jpg')
plt.imshow(cv2.cvtColor(img,cv2.COLOR_BGR2RGB))

第2步，构造HOG检测器，需要安装Dlib包(conda install -c conda-forge dlib)

import dlib
hog_face_detector = dlib.get_frontal_face_detector()

第3步，检测人脸

detections= hog_face_detector(img,1)#指的是scaleFactor=1

第4步，解析

for face in detections:x = face.left()y = face.top()r = face.right()b = face.bottom()cv2.rectangle(img,(x,y),(r,b),(0,255,0),5)

第5步，显示

plt.imshow(cv2.cvtColor(img,cv2.COLOR_BGR2RGB))

3. CNN + Dlib

import cv2
import numpy as np
import matplotlib.pyplot as plt
plt.rcParams['figure.dpi'] = 200
img = cv2.imread('./images/faces2.jpg')
import dlib
cnn_face_detector = dlib.cnn_face_detection_model_v1('./weights/mmod_human_face_detector.dat')
detections = cnn_face_detector(img,1)
for face in detections:x = face.rect.left()y = face.rect.top()r = face.rect.right()b = face.rect.bottom()c = face.confidencecv2.rectangle(img,(x,y),(r,b),(0,255,0),5)
plt.imshow(cv2.cvtColor(img,cv2.COLOR_BGR2RGB))

 4. SSD

import cv2
import numpy as np
import matplotlib.pyplot as plt
plt.rcParams['figure.dpi']=200
img = cv2.imread('./images/faces2.jpg')
face_detector = cv2.dnn.readNetFromCaffe('./weights/deploy.prototxt.txt','./weights/res10_300x300_ssd_iter_140000.caffemodel')
img_height = img.shape[0]
img_width = img.shape[1]
img_resize = cv2.resize(img,(500,300))
img_blob = cv2.dnn.blobFromImage(img_resize,1.0,(500,300),(104.0, 177.0, 123.0))
face_detector.setInput(img_blob)
detections = face_detector.forward()
num_of_detections = detections.shape[2]
img_copy = img.copy()
for index in range(num_of_detections):detection_confidence = detections[0,0,index,2]if detection_confidence>0.15:locations = detections[0,0,index,3:7] * np.array([img_width,img_height,img_width,img_height])lx,ly,rx,ry  = locations.astype('int')cv2.rectangle(img_copy,(lx,ly),(rx,ry),(0,255,0),5)
plt.imshow(cv2.cvtColor(img_copy,cv2.COLOR_BGR2RGB))   

5. MTCNN

import cv2
import numpy as np
import matplotlib.pyplot as plt
plt.rcParams['figure.dpi']=200img = cv2.imread('./images/faces2.jpg')
img_cvt = cv2.cvtColor(img,cv2.COLOR_BGR2RGB)
from mtcnn.mtcnn import MTCNN
face_detetor = MTCNN()
detections = face_detetor.detect_faces(img_cvt)
for face in detections:(x, y, w, h) = face['box']cv2.rectangle(img_cvt, (x, y), (x + w, y + h), (0,255,0), 5)
plt.imshow(img_cvt)

import cv2
import numpy as np
import matplotlib.pyplot as plt
plt.rcParams['figure.dpi']=200
img = cv2.imread('./images/test.jpg')
img_cvt = cv2.cvtColor(img,cv2.COLOR_BGR2RGB)
from mtcnn.mtcnn import MTCNN
face_detetor = MTCNN()
detections = face_detetor.detect_faces(img_cvt)
for face in detections:(x, y, w, h) = face['box']cv2.rectangle(img_cvt, (x, y), (x + w, y + h), (0,255,0), 5)
plt.imshow(img_cvt)

 5种人脸检测方式对比

视频流人脸检测 ：

        1.构造haar人脸检测器

        2.获取视频流

        3.检测每一帧画面

        4.画人脸框并显示

import cv2
import numpy as np
cap = cv2.VideoCapture(0)
haar_face_detector = cv2.CascadeClassifier('./cascades/haarcascade_frontalface_default.xml')
while True:ret,frame = cap.read()fram = cv2.flip(frame,1)frame_gray = cv2.cvtColor(frame,cv2.COLOR_BGR2GRAY)detection = haar_face_detector.detectMultiScale(frame_gray,minNeighbors=5)for(x,y,w,h) in detection:cv2.rectangle(fram,(x,y),(x+w,y+h),(0,255,0),5)cv2.imshow('Demo',fram)if cv2.waitKey(10) & 0xff == ord('q'):break
cap.release()
cv2.destoryAllWindows()