Implemented Viola Jones Face Detection algorithmn.
About Viola Jones:
​
The Viola–Jones object detection framework was first proposed in 2001 by Paul Viola and Michael Jones.
The algorithm has four stages:
-
Haar Feature Selection
-
Creating an Integral Image
-
Adaboost Training
-
Cascading Classifiers
My Implementation of Viola Jones Detector:
5 Haar types features were taken:
Where haar==1: 5 are the following
All these Harr features are calculated with the help of integral images. The integral images are calculated using a function which calculates the integral values for the points taken with the changing dimensions of the harr feature. The haar dimensions changes as per the command line haarX = dimX:dimX:window-pixelX and haarY = dimY:dimY:window-pixelY inside the window size pixelX = 2:window-dimX and pixelY = 2:window-dimY.
​
Image weights are normalized by dividing 1/(facesize+nonfacesize) where face size is the number of face images and nonface is the number of non face images.
Image weights are updated in the adaboost function on the basis of the false positives and false negatives.
Adaboost assigns a capture value 1 to all correct captures on the basis of the harrVal that it generates. Also classifier weights are updated as per the error and new image weights are passed on to the main script.
Weakclassifiers obtained : 221
And strong classifiers are obtained by combining these weak classifiers to give one strong classifier.
For training purpose 507 face images were taken and 1870 non face images from the same data set was chosen.
These images were then resized to 24X24 size in order to get haar features on just that window. haarVector1 stores all the values for haar one in all the face images. And haarVector2 stores all the values for haar one in all non face images. The loops take care that the total error is stays less than 50% for all the time.
Results:
False negative Capture Rate : 42/07=0.082
False positive Capture Rate: 542/1860=0.291
Accuracy 57% (approx.)
.png)
