Download Create your Own haar cascade and more Study Guides, Projects, Research Digital Image Processing in PDF only on Docsity!
Creating a Cascade of Haar-Like Classifiers: Step by Step
Mahdi Rezaei Department of Computer Science, the University of Auckland [email protected] www.MahdiRezaei.com
Abstract: The tutorial provides a detailed discussion on what you need to create a cascade of classifiers based on Haar-like features, which is the most common technique in computer-vision for face and eye detection. This tutorial is designed as part of course 775- Advanced multimedia imaging; however, it is easy to understand for any CS student who is interested in the topic.
All the required tools and positive/negative image dataset are provided here: https://www.cs.auckland.ac.nz/~m.rezaei/Tutorials/Haar-Training.zip
After creating your classifier, you may continue with the tutorial āFace and Eye Detection Using OpenCVā, and the available source code: www.cs.auckland.ac.nz/~m.rezaei/Downloads.html as well as our published academic papers and articles listed in [1-8].
Keywords: Face Detection, Eye Detection, Haar Features, Haar-Wavelet, Image Processing, Computer Vision, Classification, Weak Classifiers, Markup Tool, Object marker, Haar-Training, XML file.
Training Steps to Create a Haar-like Classifier:
ļ· Collection of positive and negative training images ļ· Marking positive images using objectmarker.exe or ImageClipper tools ļ· Creating a .vec (vector) file based on positive marked images using createsamples.exe ļ· Training the classifier using haartraining.exe ļ· Running the classifier using cvHaarDetectObjects()
STEP 1: Collecting Image Database All the students will receive 200 positive and 200 negative sample images for training. You may like to add more positive and negative images by recording some sequences in HAKA1 or adding more public images from Internet resources.
The positive images are those images that contain the object (e.g. face or eye), and negatives are those ones which do not contain the object.
Having more number of positive and negative (back ground) images will normally cause a more accurate classifier.
STEP 2: Arranging Negative Images Put your background images in folder ā¦\training\negative and run the batch file create_list.bat
dir /b *.jpg >bg.txt
Running this batch file, you will get a text file each line looks as below:
image1200.jpg image1201.jpg image1202.jpg ...
Later, we need this negative data file for training the classifier.
STEP 3: Crop & Mark Positive Images In this step you need to create a data file (vector file) that contains the names of positive images as well as the location of the objects in each image. You can create this file via two utilities: Objectmarker or Image Clipper. The first one is simpler and faster, and the second one is a bit more versatile but more time consuming to work. We continue with Objectmaker which is straight forward; however, you may try Image Clipper later.
In folder ..\training\positive\rawdata put you positive images
In folder ..\training\positive there is a file objectmaker.exe that we need it for marking the objects in positive images. Note that in order to correctly run objectmaker.exe two files cv.dll and highgui.dll should also exist in the current directory.
Before running the objectmaker.exe make sure you are relax and you have enough time to carefully mark and crop tens or hundreds of images!
How to mark objects? Running the file objectmaker.exe you will see two windows like below: one shows the loaded image, and the other one shows the image name.
a- Click at the top left corner of the object area (e.g. face) and hold the mouse left-key down. b- While keeping the left-key down, drag the mouse to the bottom right corner of the object. Now you could be able to see a rectangle that surrounds the object (see below). If you are
works! Make a backup numbered file (e.g. info1.txt, info2.txt anytime you escape) and finally merge all your backups into a final info.txt
Within the info.txt there would be some information like below:
rawdata\image1200.bmp 1 34 12 74 24 rawdata\image1201.bmp 3 35 25 70 39 40 95 80 92 120 40 45 36 rawdata\image1202.bmp 2 10 24 90 90 45 68 99 82
The first number in each line defines the number of existing objects in the given image. For example, in second line, the number 3 means that you already selected three objects (e.g. face) within image1201.bmp. The next four numbers (shown in green) defines the location of first object in the image (top left vertex: x= 35 , y= 24 , width= 70 and height= 39 ). The red numbers identifies the data for the second object; blues ones are for the third object, and so forth.
: Errors like this line: rawdata/d19.bmp 3 83 119 185 183 can lead to a serious hidden issues while training your cascade. Before going further, make sure every line in info.txt is correct. The error in above example is the number 3 while it should be 1. This kind of error may happen when you merge info.txt files?
The next step is packing the object images into a vector-file.
STEP 4: Creating a vector of positive images
In folder ..\training\ there is a batch file named samples_creation.bat
The content of the bath file is:
createsamples.exe -info positive/info.txt -vec vector/facevector.vec -num 200 -w 24 -h 24
Main Parameters:
-info positive/info.txt Path for positive info file -vec vector/facevector.vec Path for the output vector file -num 200 Number of positive files to be packed in a vector file -w 24 Width of objects -h 24 Height of objects
The batch file loads info.txt and packs the object images into a vector file with the name of e.g. facevector.vec
After running the batch file, you will have the file facevector.vec in the folder..\training\vector
Note: To run creatsample.exe you also needs the files cv097.dll, cxcore097.dll, highgui097.dll, and libguide40.dll in the folder ..\training.
STEP 5: Haar-Training In folder ..\training , you can modify the haartraining.bat :
haartraining.exe -data cascades -vec vector/vector.vec -bg negative/bg.txt -npos 200 -nneg 200 -nstages 15 -mem 1024 -mode ALL -w 24 -h 24 ānonsym
-data cascades Path and for storing the cascade of classifiers -vec data/vector.vec Path which points the location of vector file -bg negative/bg.txt Path which points to background file -npos 200 Number of positive samples ā¤ no. positive bmp files -nneg 200 Number of negative samples (patches) ā„ npos -nstages 15 Number of intended stages for training -mem 1024 Quantity of memory assigned in MB -mode ALL Look literatures for more info about this parameter -w 24 -h 24 Sample size -nonsym Use this if your subject is not horizontally symmetrical
The size of āW and āH in harrtraining.bat should be same as what you defined on sample-creation.bat
Harrtraining.exe collects a new set of negative samples for each stage, and ānneg sets the limit for the size of the set. It uses the previous stagesā information to determine which of the "candidate samples" are misclassified. Training ends when the ratio of misclassified samples to candidate samples is lower than Ü“ÜØ ąÆ¦ąÆ§ąÆąÆąÆ ąÆ”ąÆ¢.^ So:
Regardless of the number of stages (nstages) that you define in haartraining.bat, the program may terminate early if we reach above condition. Although this is normally a good sign of accuracy in our training process, however this also may happen when the number of positive images is not enough (e.g. less than 500).
Note: To run haartaining.exe you also needs the files cv097.dll, cxcore097.dll, and highgui097.dll in the folder ..\training.
While running the haartraning.bat you will see some information similar to screen below.
- and, the computational time for training has increased
STEP 6: Creating the XML File After finishing Haar-training step, in folder ../training/cascades/ you should have catalogues named from ā0ā upto āN-1ā in which N is the number of stages you already defined in haartraining.bat.
In each of those catalogues there should be AdaBoostCARTHaarClassifier.txt file. Copy all the folders 0..N-1 into the folder ../cascade2xml/data/
Now we should combine all created stages (classifiers) into a single XML file which will be our final file a ācascade of Haar-like classifiersā.
Run the batch file convert.bat at ../cascade2xml/ Which is:
haarconv.exe data myfacedetector.xml 24 24
myfacedetecor.xml is the output file name and 24 24 are W and H respectively.
Now you have your own XML file. Copy it in MyCascade folder, point to this classifier from your project source code, and run your face (eye) detection program.
More information: You can find more advanced information via our published academic papers with various applications for face detection [1], [3], open or closed eye state detection [4], [6], vehicle detections and driver assistance systems [5], and even facial feature improvements in artistic painting and rendering [7],[8].
References:
[1] Mahdi Rezaei, Reinhard Klette, ā3D cascade of classifiers for open and closed eye detection in driver distraction monitoringā, In Proc. Computer Analysis of Images and Patterns , pp. 171-179,
[2] Mahdi Rezaei, Reinhard Klette, āSimultaneous Analysis of Driver Behaviour and Road Condition for Driver Distraction Detectionā, International Journal of Image and Data Fusion, 2 , 217-236, 2011.
[3] Mahdi Rezaei, Hossein Ziaei Nafchi, Sandino Morales, āGlobal Haar-like Features: A New Extension of Classic Haar Features for Efficient Face Detection in Noisy Imagesā, 6th Pacific-Rim Symposium on Image and Video Technology , PSIVT 2013.
[4] Mahdi Rezaei, Reinhard Klette, āNovel Adaptive Eye Detection and Tracking for Challenging Lighting Conditionsā, Asian Conference on Computer Vision Workshops (ACCV) , pp. 427-440,
[5] Mahdi Rezaei, Mutsuhiro Terauchi, Vehicle Detection Based on Multi-feature Clues and Dempster-Shafer Fusion Theory, 6th Pacific-Rim Symposium on Image and Video Technology , PSIVT 2013.
[6] Mahdi Rezaei, Reinhard Klette, āAdaptive Haar-like Classifier for Eye Status Detection Under Non-ideal Lighting Conditionsā, Proceedings of the 27th Conference on Image and Vision Computing New Zealand (IVCNZ), pp. 521- 526, 2012.
[7] Mahdi Rezaei, āArtistic Rendering of Human Portraits Paying Attention to Facial Featuresā, Arts and Technology, 101 , pp. 90-99, 2012.
[8] Mahdi Rezaei, Juan Lin, Reinhard Klette, āHybrid Filter Blending to Maintain Facial Expressions in Rendered Human Portraitsā, International Journal of Arts and Technology, 2014.
