FaceDetection

Haar Classifiers - Adaboost for efficient face detection

Open CV offers two types of pre-trained classifiers which are trained on multiple positive and negative sample of images. We will use these classifiers and build algorithm for the sample of dataset – FDDB.

Two pre-trained face detection classifiers by open cv :

1. Haar classifier
2. LBP – Local Binary Patterns

Haar classifier:

Haar classifier learns haar features mentioned below which are windows just like convolutional kernels to learn image features. These windows move across image learning the image features and each window when placed on image returns a single value for the whole window by subtracting the sum of values of black portion from white portion. Larger images can be scaled down to learn the features. As we learn non-essential features which will not addon to classification accuracy so we use machine learning method Adaboost to build strong classifier from the sequence of weak classifiers. Adaboost discards group features which does not add up to classification accuracy.

LBP: Local Binary Patterns

For each pixel is compared to its neighbor pixels like 3*3 window with center being the pixel of interest and if the value of neighbor pixel is greater than pixel of interest then set to 1 else 0 then summed up to get the value of LBP. Uses histogram of these blocks to create a feature vector which contains features of interest.