用10几行Python代码自己写个人脸识别程序 (Chinese title, please scrolldown to view README in Chinese.)

I have been soak myself in open sourced libraries, such as OpenCV. I gradually came to discern concepts such as Machine Learning , Deep Learning are not academic standing water. As a matter of fact, those elusive topics and certain pragmatic use cases could coalesce in a amount of interesting products. For instance, in past couple of months, there were a hype of guess-ages-by-photo, below screenshot depicts such.

What a seductive one! Initially been attracted by such funky features, after second thoughts, I found at the heart of it is two cohesive parts, the first one is how to locate human faces from background and whole picture, consequently to have a ballpark age guess for the recognized the faces. You may guess how difficult to codify a program to implement the 1st feature. Actually no need chunks of code, at here purely a dozen of lines of code are necessitated (actually only 10 lines of code, excluding space line and comments). I'd like to piggyback on such tiny utility to elaborate advanced topics of Computer Visions.

Actually face recognition is not new to us, this feature prevailing in so-called auto focus in DC (Digital Camera) and many main stream smart phone built-in cameras. Just like below photo. You can get a sense of how commonplace of face recognition , which is becoming a widely used technology around us.

Theoretically speaking, face recognition is also called face detection, it's a type of technology/program to electronically identify human frontal faces in digital images, such as photos, camera or surveillance. Further more, face detection is kind of objects detection in computer vision area. Which will locate object (e.g. human face) and get the size.

First of all, let's have some visual and concrete feeling of this program, below screenshot is the source code.

The whole program source code can be found at this github repository https://github.com/CloudsDocker/pyFacialRecognition . Please feel free to fork , check it out and have a try. I'll walk through this program one line by one line at this blog.

"You serious? This is all the problem, just aforementioned 10 lines?" Let's first take a look at the actual run output.

Please be advised the red rectangle around faces.

First of first, as you know, this program is composed by python,therefore, make sure you work station or laptop equiped with python, vesrion is irrelavant for this program.

In addition, this utility is built upon OpenCV (https://opencv.org/downloads.html), therefore please install this component as well. Just as its name suggested, it is an open source framework focus on computer vision related deep learning, surfaced decades ago. This is one Intel lab built by Rusian, which is a very active community.

Particulary, if you are Mac users, it's recommended to use brew to setup OpenCV. Below is sample commands(The 1st line of following command may raise some errors, in that case please contact me via the link at the rear of this blog):

brew tap homebrew/science
brew install opencv

Upon completion of preceding scripts, you can execute following scripts to verify whether it's installed success or not, e.g. it means all fine if no exception/errors raised

>>> import cv2

Let's dissect file recognizeFace_loose_en.py as one example

import cv2,sys

• To import library of OpenCV and python built-in system library, which is used to parse input arguments.

inputImageFile=sys.argv[1]

• To read the 1st argument, which to be the file name of the image to be parsed, e.g. test.jpg

faceClassifier=cv2.CascadeClassifier('haarcascade_frontalface_default.xml')

• To load HAAR Casscade Classifier, the human face recognition cascade categorizer which shipped with OpenCV. Which will do the actual computation, logic to recognize and size human faces from any given images.

We stop here not reading further code, avoiding perplex you, I'll walk through certain CV topics pertaining to this blog. As for more deep concepts, please feel free to contact me or goole by yourself.

In arena of computer vision and machine learning, a variaty of classifiers been and being built, to assemle special domain knowledge to recognize corresponding objects. For example, there are particular classifier to recognize cars, there are plane classifier, and classifiers to recognize smile, eyes, etc. For our case, we need a specific classifier help us to detect and locate human faces.

Generally speaking，, to recognize one object (such as human faces) means finding and identifying objects in an image or video sequence. However, it's neccessitate tons of sample/specimen to train machine to learn, for instance, it's likely thousands of hundreds of digital images/video will be prepared as learning material, while all of specimen should be categorized to two mutax type, positive or negative. e.g. phots containss human face and ones without human face. When machine read one photo, it was told this is either a positive one or negative one, then machine could gradually analysys and induce some common facets and persist to files for future usages, e.g. when given a new photo, the machine can classify it whether it's a positive or negative. That's why it's called classifier.

Your feeling is right, just as it's name suggrested, cascade implies propagating something. In this case, it's specifically means Cascade classifier. Intuitively the next question is why cascade is required? Let me try to articulate the underlying logic, as you know, at the heart of digital images, which is the raw material of computer vision, are pixel。For one CV process, it need to scan each pixel per pixel, while in contemporary world, size of image tend to incresing more than we expected, e.g. normall one photo taken by smart phone tend to contains millions of pixels. At the meanwhile, to fine tune and get a more accuate result of one object recognition, it tend to lots of classifiers to work from different point of views of the underlying photo. Therefore these two factors interwhirled together, the final number would be astronomical. Therefore, one innovative solution is cascade, in a nutshell, all classifiers will be splited to multiple layers, one photo will be examined by classifiers on 1st layer at the very begining, if failed, the whole CV can retain negative immediately, with fewest efforts and time cost, while majority of other classifiers won't be executed in actual. This should significantely accelerate the whole process of CV. This is similar to FF(Fail Fast) in other areas,severed for sake of running efficiency.

objImage=cv2.imread(inputImageFile)

• To create one OpenCV image object by loading the input digital file via OpenCV

cvtImage=cv2.cvtColor(objImage,cv2.COLOR_BGR2GRAY)

• Firstly, convert the digital colorful image to grayscale one, which easy the task to scan and analyse the image. Actually this is quite common in image analys area. e.g. this could eliminate those noisy pixel from the picture.

foundFaces=faceClassifier.detectMultiScale(cvtImage,scaleFactor=1.3,minNeighbors=9,minSize=(50,50),flags = cv2.cv.CV_HAAR_SCALE_IMAGE)

• Call method detectMultiScale to recongnize object, i.e. human face in this case. The parameters overview as below:
• scaleFactor: For a photo, particualy from selpie, some faces are shows bigger than rest of others, due to the distance between each faces and lens. Therefore this parameter is used to config the factor, please be advised this double should greater than 1.0
• minNeighbors: Because it need to gradually scan the photo by a certain window, i.e. a rectangle. So this parameter is telling how many other object in the vacinity to be detected, before making final decision that it's positive or negative.
• minSize：For aforementioend window, this parameter is setting the size of this rectangle.

print(" Found {} human faces in this image".format(len(foundFaces)))

• To print how many faces detected, be reminded returned value is a list, each item is the actual position of every faces. Therefore, using len to print total number of ojects found.

for (x,y,w,h) in foundFaces:
    cv2.rectangle(objImage,(x,y),(x+w,y+h),(0,0,255),2)

• Traverese all faces detected, please be noted returning object is consist of 4 parts, i.e. the horizontal and vertial position, width and height.
• Consequently to draw a rectangle by an off-the-shelf method from OpenCV. Be advised (0,0,255) represents color of the rectangel. It use R/G/B mode, e.g. black is (0,0,0)，white is (255,255,255)，etc. Well versed web programmer should be familiar with it.

cv2.imshow('Detected human faces highlighted. Press any key to exit. ', objImage)
cv2.waitKey(0)

• To display this image via opencv provided method imshow, together with the rectangles we draw previously
• The last one is one user hint, remind you can quit the applicaiton by press any key on the image display window

We've skimmed source codes and pertaining knowledge. This is just scratched the surface of this framework, hope this can open the door to more advanced topics and insights, such as hack of CAPTCHA, newly open sourced project form Yahoo, NSFW, Not Suitable for Work (NSFW)，to detect images with pornagraphy, etc.

Finally，please be reminded all related source are open sourced at github repository https://github.com/CloudsDocker/pyFacialRecognition ，please fork and sync to your local disk, check it out and paly it.

git clone https://github.com/CloudsDocker/pyFacialRecognition.git
cd pyFacialRecognition
./run.sh

You can access my blog. Any comments/suggestions, feel free to contact me.

• [email protected] (email，whatsapp, linkedin)
• helloworld_2000 (wechat)
• weibo: cloudsdocker
• github
• jianshu
• wechat：vibex

• Object recognition
• OpenCV
• HAAR features
• Face Detection using Haar Cascades
• NSFW

用10几行代码自己写个人脸识别程序

最近在研究CV的一些开源库(OpenCV)，有一个体会就是在此领域，除了一些非常学术的_机器学习_, 深度学习_等概念外，其实还有一些很有趣的_现实的_应用场景。比如之前很流行的微软的 https://how-old.net, 你使用自己指定或者上传的照片进行面部识别_猜年龄。 如下图所示：

细想一下这个很吸引眼球的程序，其实技术本身打散了就包括两大块，一是从图片中扫描并进行面部识别，二是对找到的人脸根据算法去猜个年龄。大家可以猜猜实现第一个功能需要多少核心代码量？其实不用上万行，在这里我就使用短短几行代码（去除空格换行什么的，有效代码只要10行）就实现一个_高大上_面部识别的功能。在此文容我细述一下具体实现代码以及我对机器识别图像领域技术的理解。

_人脸识别_技术大家应该都不陌生，之前大家使用的数码相机，或者现在很多手机自带的相机都有人脸识别的功能。其效果就像是下图这样。近的看，_剁手节_刚刚过了没有多久 , 背后的马老板一直在力推的刷脸支付也是一个此领域的所谓“黑科技”。比如在德国汉诺威电子展上，马云用支付宝“刷脸”买了一套纪念邮票。人脸识别应用市场也从爆发。随后，各大互联网巨头也纷纷推出了刷脸相关的应用。

如果要加个定义，人脸识别又叫做人像识别、面部识别，是一种通过用摄像机或摄像头采集含有人脸的图像或视频流，并自动在图像中检测和跟踪人脸，进而对检测到的人脸进行脸部的一系列相关技术。

OK，长话短说，先上 干货 ，下面就是此程序的_带注释_ 版本，完整的程序以及相关配套文件可以在 这个github库 https://github.com/CloudsDocker/pyFacialRecognition 中找到，有兴趣可以_fork_ 下来玩玩。下面是整个程序的代码样子，后面我会逐行去解释分析。

就这短短的十行代码代码？seriously？“有图有真相”，我们先来看下运行的效果：

请注意 K歌二人组 的脸上的红色框框，这就是上面十行代码的成果。

因为此程序使用是的Python,因此你需要去安装Python。这里就不赘述了。除此之外，还需要安装 OpenCV (https://opencv.org/downloads.html)。 多说一句,这个 OpenCV正如其名，是一个开源的机器识别的深度学习框架。这是Intel（英特尔）实验室里的一个俄罗斯团队创造的，目前在开源社区非常的活跃。

特别提一下，对于Mac的用户，推荐使用brew去安装 （下面第一条语句可能会执行报错，我当时也是搞了好久。如果遇到第一条命令不过可以通过文尾的方式联系作者）

brew tap homebrew/science
brew install opencv

安装完成之后,在python的命令行中输入如下代码验证，如果没有报错就说明安装好了。

>>> import cv2

# -*- coding: utf-8 -*-
import cv2,sys

• 由于这里注释及窗口标题中使用了中文，因此加上utf-8字符集的支持
• 引入Opencv库以及Python的sys内建库，用于解析输入的图片参数

inputImageFile=sys.argv[1]

• 在运行程序时将需要测试的照片文件名作为一个参数传进来

faceClassifier=cv2.CascadeClassifier('haarcascade_frontalface_default.xml')

• 加载OpenCV中自带预先培训好的人脸识别层级分类器 HAAR Casscade Classifier，这个会用来对我们输入的图片进行人脸判断。

这里有几个在深度学习及机器图像识别领域中的几个概念，稍微分析一下，至于深入的知识，大家可以自行搜索或者联系作者。

在机器深度学习领域，针对识别不同物体都有不同的classifier,比如有的classifier来识别汽车，还有识别飞机的classifier，有classifier来识别照片中的笑容，眼睛等等。而我们这个例子是需要去做人脸识别，因此需要一个面部识别的classifier。

一般来说，比如想要机器学习着去识别“人脸”，就会使用大量的样本图片来事先培训，这些图片分为两大类，positive和negative的，也就是分为包“含有人脸”的图片和“不包含人脸”的图片，这样当使用程序去一张一张的分析这些图片，然后分析判断并对这些图片“分类” (classify),即合格的图片与不合格的图片，这也就其为什么叫做 classifier ， 这样学习过程中积累的"知识"，比如一些判断时的到底临界值多少才能判断是positive还是negative什么的，都会存储在一个个XML文件中，这样使用这些前人经验（这里我们使用了 哈尔 分类器）来对新的图片进行‘专家判断'分析，是否是人脸或者不是人脸。

这里的 Cascade是 层级分类器 的意思。为什么要 分层 呢？刚才提到在进行机器分析照片时，其实是对整个图片从上到下，从左到右，一个像素一个像素的分析，这些分析又会涉及很多的 特征分析 ，比如对于人脸分析就包含识别眼睛，嘴巴等等，一般为了提高分析的准确度都需要有成千上万个特征，这样对于每个像素要进行成千上万的分析，对于整个图片都是百万甚至千万像素，这样总体的计算量会是个天文数字。但是，科学家很聪明，就想到分级的理念，即把这些特征分层，这样分层次去验证图片，如果前面层次的特征没有通过，对于这个图片就不用判断后面的特征了。这有点像是系统架构中的 FF (Fail Fast),这样就提高了处理的速度与效率。

objImage=cv2.imread(inputImageFile)

• 使用OpenCV库来加载我们传入的测试图片

cvtImage=cv2.cvtColor(objImage,cv2.COLOR_BGR2GRAY)

• 首先将图片进行灰度化处理，以便于进行图片分析。这种方法在图像识别领域非常常见，比如在进行验证码的机器识别时就会先灰度化，去除不相关的背景噪音图像，然后再分析每个像素，以便抽取出真实的数据。不对针对此，你就看到非常多的验证码后面特意添加了很多的噪音点，线，就是为了防止这种程序来灰度化图片进行分析破解。

foundFaces=faceClassifier.detectMultiScale(cvtImage,scaleFactor=1.3,minNeighbors=9,minSize=(50,50),flags = cv2.cv.CV_HAAR_SCALE_IMAGE)

• 执行detectMultiScale方法来识别物体，因为我们这里使用的是人脸的cascade classifier分类器，因此调用这个方法会来进行面部识别。后面这几个参数来设置进行识别时的配置，比如
• scaleFactor: 因为在拍照，尤其现在很多都是自拍，这样照片中有的人脸大一些因为离镜头近，而有些离镜头远就会小一些，因为这个参数用于设置这个因素，如果你在使用不同的照片时如果人脸远近不同，就可以修改此参数，请注意此参数必须要大于1.0
• minNeighbors: 因为在识别物体时是使用一个移动的小窗口来逐步判断的，这个参数就是决定是不是确定找到物体之前需要判断多少个周边的物体
• minSize：刚才提到识别物体时是合作小窗口来逐步判断的，这个参数就是设置这个小窗口的大小

print(" 在图片中找到了 {} 个人脸".format(len(foundFaces)))

• 显示出查找到多少张人脸，需要提到的识别物体的方法返回的一个找到的物体的位置信息的列表，因此使用 len 来打印出找到了多少物体。

for (x,y,w,h) in foundFaces:
    cv2.rectangle(objImage,(x,y),(x+w,y+h),(0,0,255),2)

• 遍历发现的“人脸”，需要说明的返回的是由4部分组成的位置数据，即这个“人脸”的横轴，纵轴坐标，宽度与高度。
• 然后使用 OpenCV 提供的方法在原始图片上画出个矩形。其中 (0,0,255) 是使用的颜色，这里使用的是R/G/B的颜色表示方法，比如 (0,0,0)表示黑色，(255,255,255)表示白色，有些网页编程经验的程序员应该不陌生。

cv2.imshow(u'面部识别的结果已经高度框出来了。按任意键退出'.encode('gb2312'), objImage)
cv2.waitKey(0)

• 接下来是使用 opencv 提供的imshow方法来显示这个图片，其中包括我们刚刚画的红色的识别的结果
• 最后一个语句是让用户按下键盘任意一个键来退出此图片显示窗口

好了，上面是这个程序的详细解释以及相关的知识的讲解。其实这个只是个_抛砖引玉_的作用，还用非常多的应用场景，比如程序解析网页上的图片验证码，雅虎前几个月开源的 NSFW, Not Suitable for Work (NSFW)，即判断那些不适合工作场所的图片，内容你懂的。 :-)

最后，再提一下，所有这些源代码及相关文件都开源在 https://github.com/CloudsDocker/pyFacialRecognition ，在fork并下载到本地后执行下面代码来测试运行

git clone https://github.com/CloudsDocker/pyFacialRecognition.git
cd pyFacialRecognition
./run.sh

如果有任何建议或者想法，请联系我。

• [email protected] (email/邮件，whatsapp, linkedin)
• helloworld_2000 (wechat/微信)
• 微博: cloudsdocker
• github
• 简书 jianshu
• 微信公众号：vibex

• OpenCV

• HAAR 哈尔特征

• Face Detection using Haar Cascades

• NSFW