The most intimidating thing in neural network applications is face recognition. Today, the landlord will make a simple face recognition for fun.
1 Introduction
Face recognition is both a classification problem and a regression problem.
Generally speaking, we need to obtain the area where faces may exist by some method, and then classify the images in the area to determine whether it is a face or a background, which is a classification problem.
Then we need to adjust the region coordinates, which is equivalent to a regression problem. Given an area, get the parameters corresponding to the adjustment area, such as translation and scaling.
There is not necessarily such an obvious sequence relationship in practical applications, but the idea is generally the same.
As usual, the landlord is still ready to play some wild ways. It can be considered to develop a little brain hole of your own!
If we have a network that classifies faces and backgrounds very accurately, we can search images similar to binary search. First determine whether there is a face in the whole image, then segment the image, and determine whether there is a face in the sub-images, recursively. When the face is finally segmented, cutting it again will cause the face in the sub-picture to disappear, the recursion ends, and the recognition is completed.
In this way, the process of regression calculation can be avoided, and only the network for classifying background and face needs to be trained. In this way, the data requirements are also much less, and network images can be captured for training.
2. Data Collection Preparation
First we need pictures of faces and backgrounds. Secondly, since our face can be recognized when it is far away from the camera, we need the network to classify pictures of the body, so we also need pictures of the body.
The best place to collect pictures is the picture search of the major search engines. Google is inconvenient to be walled. Comparing the image search of baidu and bing, the accuracy of bing is obviously higher than that of baidu, and the target is bing.
After that, it was simple. After analyzing the network data of bing image search, I wrote some Python code as follows:
from bs4 import BeautifulSoup
import urllib.request
import urllib.parse
import them
import threading
import cv2
import numpy as np
keyword={
"face":["http://cn.bing.com/images/async?q=face+people&qft=+filterui%3aface-face+filterui%3aphoto-photo&lostate=r&mmasync=1&dgState=x*410_y*1029_h*166_c*2_i*71_r*12&IG=2F30F17499C7450DBFAFA355CA1BE527&SFX=3&iid=images.5622",
"http://cn.bing.com/images/async?q=human&qft=+filterui%3aface-portrait&lostate=r&mmasync=1&dgState=x*694_y*1201_h*164_c*3_i*36_r*7&IG=6D13393FAE2C46C1B74D874BBE395415&SFX=2&iid=images.5725",
"http://cn.bing.com/images/async?q=handsome&qft=+filterui%3aface-face&lostate=r&mmasync=1&dgState=x*524_y*1065_h*202_c*3_i*36_r*6&IG=B57823857E2843CE87EDB655B7F7F334&SFX=2&iid=images.5709",
"http://cn.bing.com/images/async?q=beautiful+girl&qft=+filterui%3aface-face&lostate=r&mmasync=1&dgState=x*745_y*1194_h*187_c*4_i*36_r*7&IG=55BB00BABF4F478B9060D6A3A19C1B65&SFX=2&iid=images.5670",
"http://cn.bing.com/images/async?q=%e5%b8%85%e5%93%a5&relo=3&qft=+filterui%3aface-face&lostate=c&mmasync=1&dgState=c*7_y*1257s1212s1226s1472s1213s1410s1427_i*39_w*181&IG=6D566BBDC2C4448ABB98B2DDBFAED030&SFX=2&iid=images.5754",
"http://cn.bing.com/images/async?q=%e7%be%8e%e5%a5%b3&qft=+filterui%3aface-face&lostate=r&mmasync=1&dgState=x*486_y*1004_h*192_c*2_i*36_r*6&IG=40A17665192F4434A72B70C4390C64E8&SFX=2&iid=images.5653",
"http://cn.bing.com/images/async?q=%e4%ba%ba%e7%89%a9%e5%9b%be%e7%89%87&relo=3&qft=+filterui%3aface-face&lostate=c&mmasync=1&dgState=c*7_y*1430s1159s1189s1250s1358s1183s1285_i*39_w*181&IG=2245EBFC2E5D4104BAE25B78BD3F9B99&SFX=2&iid=images.5682"],
"body":["http://cn.bing.com/images/async?q=Group+photo+friends&qft=+filterui%3aface-portrait&lostate=r&mmasync=1&dgState=x*0_y*0_h*0_c*5_i*211_r*37&IG=61B156F4951A445FABD823619621B21A&SFX=7&iid=images.5622",
"http://cn.bing.com/images/async?q=street+young+man+woman+photo&lostate=r&mmasync=1&dgState=x*870_y*1226_h*195_c*3_i*71_r*13&IG=A6DBFAAFB9F246C58CA1C6E48D7B9EC8&SFX=3&iid=images.5678",
"http://cn.bing.com/images/async?q=%e5%90%88%e7%85%a7&lostate=r&mmasync=1&dgState=x*0_y*0_h*0_c*4_i*281_r*47&IG=38AADDB045FA476A874BCE211808938A&SFX=9&iid=images.5725"
],
"background":["http://cn.bing.com/images/async?q=outdoor&lostate=r&mmasync=1&dgState=x*529_y*1137_h*184_c*2_i*211_r*42&IG=36C7889B63534668B101BA10E4200710&SFX=7&iid=images.5713",
"http://cn.bing.com/images/async?q=%e9%a3%8e%e6%99%af&lostate=r&mmasync=1&dgState=x*0_y*0_h*0_c*5_i*71_r*14&IG=844AED80FBB44BEDA47A7D2201D6D642&SFX=3&iid=images.5725"]
}
##############################################################
def checkDir():
if os.path.exists("./data")==False:
os.mkdir("./data")
for subDir in list(keyword.keys()):
if os.path.exists("./data/"+subDir)==False:
os.mkdir("./data/"+subDir)
def writeData(data,tag,name):
image = np.asarray(bytearray(data), dtype="uint8")
image = cv2.imdecode(image, cv2.IMREAD_COLOR)
cv2.imwrite("./data/"+tag+"/"+name+".jpg", cv2.resize(image,(128,128)))
##################################################################
globalSet=set()
class imageUrlGeter:
def __init__(self,url,getIndex,start,end):
self.url=url
self.getIndex=getIndex
self.curIndex = start
self.maxIndex=end
self.urlList=[]
self.urlSet=globalSet
def __iter__(self):
return self
def pushImageUrl(self):
while True:
try:
text=urllib.request.urlopen(self.url+"&count=20&relp=20"+"&first="+str(self.getIndex)).read()
break
except Exception:
print("exception raised when get image url")
html=str(text,encoding = "utf-8")
self.getIndex+=20
tempList = []
soup=BeautifulSoup(html)
images=soup.find_all(name="img",class_="mimg")
for image in images:
url=image.get("src")
param=urllib.parse.parse_qs(urllib.parse.urlparse(url).query)
if param["id"][0] in self.urlSet:
continue
self.urlSet.add(param["id"][0])
tempList.append(url)
if len(tempList)<=0:
raise StopIteration ()
self.urlList + = tempList
def __next__(self):
if self.curIndex>=self.maxIndex:
raise StopIteration ()
if len(self.urlList)==0:
self.pushImageUrl ()
cur = self.curIndex
self.curIndex + = 1
return self.urlList.pop(0),cur
def downloadFun(geter,key):
for url,index in geter:
try:
data = urllib.request.urlopen(url).read()
writeData(data, key, str(index))
except Exception:
print("image " + str(index) + " skiped")
def startDownload(url,key,start,end):
numOfThread=32
threadList=[]
perThread=int((end-start)/numOfThread)
for i in range(numOfThread):
urlGeter=imageUrlGeter(url,i*perThread,start+i*perThread,start+(i+1)*perThread)
thread=threading.Thread(target=downloadFun,args=[urlGeter,key])
thread.setDaemon(True)
threadList.append(thread)
thread.start()
for thread in threadList:
thread.join()
def downLoadImages(dic):
for key in list(dic):
i=0
urlList=keyword[key]
for url in urlList:
startDownload(url,key,i*5000,(i+1)*5000)
i+=1
if __name__=="__main__":
checkDir()
downLoadImages(keyword)
3. Collect data
The landlord is lazy, so I added a few links, so I grabbed more than 10,000 pictures in total. Let’s try it first! If you need more, just add a few links.
Face picture:
Body pictures:
Background picture:
4. Summary
The data is relatively shabby, mainly because the landlord is lazy...
The next article will introduce you to the establishment and training of the network model