A Facial Recognition utility in a dozen of LOC (Lines Of Code)
CV (Computer Vision)
I have been soak myself in open sourced libraries, such as OpenCV. I gradually came to realize concepts such as Machine Learning , Deep Learning are not purely academic standing water. As a matter of fact, those elusive topics and certain pramatic 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 is one screenshot.
What a seducive one! Initially been attracted by such funky features, after second thoughts, I found at the heart of it is two cohensive parts, the first one is how to locate human faces from background and whole picture, consequently to have a ballpark age for the recongnized 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 necessiated (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.
Faces recognition
Actually face recognition is not new to us, this feature can be wildly found in the feature of 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.
Therotically speaking, face recognition is also called face detection, it's a type of technogloy/program to electorincally identify human frontal faces in digital images, such as photos, camera or suvillance. 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.
My '10 LOC program'
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 actul run output.
Here is the origional image
Below is the result of execution of this tiny utility
Please be advised the red rectangle around faces.
Souce Code
Prerequite
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 (http://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. 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 above 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
Souce Code Dissection
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.
Expansion of computer vision knowledge
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.
Classifier
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.
Conceps of objects recognize
Generally speaking辛藻，, 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.
Cascade
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)匠襟，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
In summary
We've skimmed source codes and related knowledge. This is just one of bunch of use cases of this framework, hope this can bring some insights. ，such as hack of CAPTCHA, newly open sourced project form Yahoo, NSFW, Not Suitable for Work (NSFW)惭等，to detect images with pornagraphy, etc.
Finally阅爽，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
Any comments/suggestions, feel free to contact me
Contact me：
- phray.zhang@gmail.com (email糙及，whatsapp, linkedin)
- helloworld_2000 (wechat)
- weibo: cloudsdocker
- github
- jianshu
- wechat：vibex
