Target Detection Images in Design of a

Target Detection Images in

Design of a classifier of the form

NN leads to a Complex Classifier

SVM leads to a Complex Classifier

Fast (and accurate) classifier

Input Blocks

+1 -1 Hyperplane

Non-Faces Rejected non-faces

Rejected points

Define a distance between a point and a PDF by

In the case of face detection in images we have We Should Maximize (GEP)

Maximize the distance between all the pairs of [face, non-face] The same Expression Minimize the distance between all the pairs of [face, face]

If the two PDF’s are assumed Gaussians, their KL distance is given by And we get a similar expression

The MRC algorithm idea is strongly dependent on these assumptions, and it leads to Fast & Accurate Classifier.

All faces detected with no false alarms

Assume that Minimize variances and Gaussians Maximize mean difference *FLD - Fisher Linear Discriminant

Maximize Minimize

In the MRC we got the expression for the distance If P(X)=P(Y)=0. 5 we maximize The distance of the Y points to the X-distribution The distance of the X points to the Y-distribution

Instead of maximizing the sum Minimize the inverse of the two expressions (the inverse represent the proximity)

Compute Minimize f(θ) & find thresholds Remove Sub-set END

No more Kernels Project onto the next Kernel Face Is value in Non Face

Input block Face/ Non. Face

Input block Face/ Non. Face

Searching targets in a given scale, for a 1000 by 1000 pixels image, the classifier is applied 1 e 6 times (even if no scale is involved!!)

While allowing outliers for better generalization behavior