Dlib介紹

       Dlib是一個現代C++框架，解決包含機器學習演算法以及開發複雜軟體的現實問題，它被廣泛應用在工業和學術研究領域，包括機器人、嵌入式裝置、移動手機以及大規模高效能運算環境中，DLib的開源使得在使用過程中方便，自由。它的主要特點有：文件說明全，高質量的程式碼，機器學習演算法，科學計算演算法，圖模型推理演算法，影象處理，執行緒，網路程式設計，圖形使用者介面，資料壓縮與整合演算法等，可以參考官網說明：http://dlib.net

安裝dlib模組

sudo apt-get install libboost-all-dev
sudo apt-get install cmake
sudo pip install  -i https://pypi.tuna.tsinghua.edu.cn/simple dlib
 

如何使用在PC環境下使用安卓手機攝像頭？

在百度搜索 【IP攝像頭】，下載並安裝在安卓手機上，保證PC機與手機連線在同一個無線網中，PC機IP最好不好自動獲取，可以先設定自動獲取，然後在CMD下敲入ipconfig /all 檢視分配的ip，然後修改靜態ip，在IP攝像頭軟體的主介面中點選開啟【IP攝像頭伺服器】按鈕，攝像頭就啟動了，可以選擇前置和後置攝像頭。IP攝像頭軟體具體使用可以檢視說明：http://app.mi.com/details?id=com.shenyaocn.android.WebCam

首先安裝opencv模組

sudo pip install  -i https://pypi.tuna.tsinghua.edu.cn/simple opencv-python
 

python操作攝像頭程式碼示例：

cv2.namedWindow('camera',1)
video = 'http://admin:[email protected]:8081'
cap = cv2.VideoCapture(video)

if False == cap.isOpened():
	cap.open(video)

while (True):
	ret, frame = cap.read()
	if False == ret:
		break
	cv2.imshow('camera', frame)

	key = cv2.waitKey(10)
	if key == 27:
		print ('esc break...')
		break
cap.release()
cv2.destroyAllWindows('camera')
 

如何使用dlib進行人臉檢測與識別？

    這裡需要下載兩個已經訓練好的人臉識別模型以及人臉關鍵點檢測器，兩個.dat檔案（shape_predictor_68_face_landmarks.dat和dlib_face_recognition_resnet_model_v1.dat），新建一個資料夾candidate-faces資料夾，把庫中的圖片以姓名命名，如下圖所示：

接下來的過程如下：

  1、載入正臉檢測器，使用dlib

detector = dlib.get_frontal_face_detector()

  2、載入人臉關鍵點檢測器

sp = dlib.shape_predictor(predictor_path)

3、載入人臉識別模型

facerec = dlib.face_recognition_model_v1(face_rec_model_path)

4、儲存候選人的人臉描述子

   4.1、先通過候選人影象檢測到候選人臉矩形框

   	image = cv2.imread(faces_folder_path+'/'+filename)
	gray_image = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
	dets = detector(gray_image, 1)

dets儲存的是影象中人臉的矩形框，可以有多個

 4.2、通過人臉矩形框獲取人臉的關鍵點（68個關鍵點）

shape = sp(image, d)

4.3、人臉識別模型通過關鍵點獲取人臉描述子

face_descriptor = facerec.compute_face_descriptor(image, shape)

4.4、把人臉描述子與人臉姓名分別一一對應儲存在兩個列表list中

candidate.append(filename.split('.')[0])

 5、被檢測影象人臉識別，這裡是從攝像頭獲取視訊幀，識別沒幀影象中的人臉

  5.1、被檢測影象人臉矩形框獲取

 	cv2.namedWindow('camera',1)

	video = 'http://admin:[email protected]:8081'
	cap = cv2.VideoCapture(video)

	if False == cap.isOpened():
		cap.open(video)
	while (True):
		ret, frame = cap.read()

		#img = cv2.resize(frame, (600,800))
		img = frame
		gray_image = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
		dets = detector(gray_image, 1)

   5.2、被檢測影象人臉關鍵點獲取

shape = sp(img, d)

   5.3、被檢測影象人臉描述子獲取

 face_descriptor = facerec.compute_face_descriptor(img, shape)
d_test = numpy.array(face_descriptor)

   5.4、被檢測影象人臉描述子與候選人臉描述子之間計算歐氏距離d

 	for i in descriptors:
		dist_ = numpy.linalg.norm(i - d_test)
		dist.append(dist_)

   5.5、儲存歐式距離d，將d與候選人姓名組合成字典，對字典進行排序

 	c_d = dict(zip(candidate, dist))
	cd_sorted = sorted(c_d.items(), key=lambda d:d[1])

   5.6、對距離進行閾值判斷，並取距離最小的那個候選人姓名

 	if cd_sorted[0][1] < 0.38:
		cv2.rectangle(img, (d.left(),d.top()),(d.right(),d.bottom()),(255,0,0),1,8,0)
		cv2.putText(img, 'unkown rate {}'.format(cd_sorted[0][1]), (d.left(),d.top()-10), cv2.FONT_HERSHEY_COMPLEX, 1, (255,0,0), 1)
		print('the perion is unkown rate{}'.format(cd_sorted[0][1]))
	else:
		cv2.rectangle(img, (d.left(),d.top()),(d.right(),d.bottom()),(255,0,0),1,8,0)
		cv2.putText(img, cd_sorted[0][0], (d.left(),d.top()-10), cv2.FONT_HERSHEY_COMPLEX, 1, (255,0,0), 1)
		print('the perion is {}'.format(cd_sorted[0][0]))

最後完整程式碼可以參考如下：

#-*- encoding:utf-8 -*-

import sys
import os
import dlib
import cv2
import numpy

predictor_path = './shape_predictor_68_face_landmarks.dat'

face_rec_model_path = './dlib_face_recognition_resnet_model_v1.dat'

faces_folder_path = './candidate-faecs'

if not os.path.exists(faces_folder_path):
	os.makedirs(faces_folder_path)

detector = dlib.get_frontal_face_detector()

sp = dlib.shape_predictor(predictor_path)

facerec = dlib.face_recognition_model_v1(face_rec_model_path)

descriptors = []
candidate = []
index=1
for filename in os.listdir(faces_folder_path):
	if filename.endswith('.jpg'):
		print ('start to process picture %s' %index)
	else:
		continue
	candidate.append(filename.split('.')[0])
	image = cv2.imread(faces_folder_path+'/'+filename)

	gray_image = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
	dets = detector(gray_image, 1)

	print ('Number of faces detected:{}'.format(len(dets)))

	for k, d in enumerate(dets):
		shape = sp(image, d)
		face_descriptor = facerec.compute_face_descriptor(image, shape)

		v = numpy.array(face_descriptor)
		descriptors.append(v)
	index += 1
test_img_path = './xiaozhi2.jpg'

cv2.namedWindow('camera',1)

video = 'http://admin:[email protected]:8081'
cap = cv2.VideoCapture(video)

if False == cap.isOpened():
	cap.open(video)
while (True):
	ret, frame = cap.read()

	#img = cv2.resize(frame, (600,800))
	img = frame
	gray_image = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
	dets = detector(gray_image, 1)

	dist = []

	for k, d in enumerate(dets):

		print ('In test image Number of faces detected:{}'.format(len(dets)))

		shape = sp(img, d)
		face_descriptor = facerec.compute_face_descriptor(img, shape)
		d_test = numpy.array(face_descriptor)

		for i in descriptors:
			dist_ = numpy.linalg.norm(i - d_test)
			dist.append(dist_)

		c_d = dict(zip(candidate, dist))

		cd_sorted = sorted(c_d.items(), key=lambda d:d[1])

		if cd_sorted[0][1] < 0.38:
			cv2.rectangle(img, (d.left(),d.top()),(d.right(),d.bottom()),(255,0,0),1,8,0)
			cv2.putText(img, 'unkown rate {}'.format(cd_sorted[0][1]), (d.left(),d.top()-10), cv2.FONT_HERSHEY_COMPLEX, 1, (255,0,0), 1)
			print('the perion is unkown rate{}'.format(cd_sorted[0][1]))
		else:
			cv2.rectangle(img, (d.left(),d.top()),(d.right(),d.bottom()),(255,0,0),1,8,0)
			cv2.putText(img, cd_sorted[0][0], (d.left(),d.top()-10), cv2.FONT_HERSHEY_COMPLEX, 1, (255,0,0), 1)
			print('the perion is {}'.format(cd_sorted[0][0]))
	
		dist = []
	if len(dets) == 0:
		cv2.putText(img, 'not detect face', (100,100), cv2.FONT_HERSHEY_COMPLEX, 1, (255,0,0), 1)
	else:
		pass
	cv2.imshow('rectangle', img)
	
	key = cv2.waitKey(10)
	if key == 27:
		print ('esc break...')
		break
cap.release()
cv2.destroyAllWindows('camera')
dlib.hit_enter_to_continue()