1. Information Based Maximum Discrimination Learning

Visual pattern detection is a problem of significant importance and difficulty. Visual detection of targets is the first step in most automatic vision systems. Although it seems an easy task for the human eyes, machine detection of visual patterns is difficult. Most pattern recognition techniques have been successfully used for visual object detection, however the resulting detection strategies are extremely computational expensive.

We have developed a learning technique particularly attractive because of its fast detection strategy which has allowed us to implement real-time systems for face detection and tracking of multiple people.

The visual learning technique described here maximizes the information-theoretic divergence between two probability distributions. The Kullback divergence is a non-negative measure of the difference between two probability distributions that equals zero only when they are identical. It measures the discrimination power between the classes described by these distributions.

Object detection is carried out using maximum likelihood classifications between two classes: the class of objects in question and the class of all other objects (background, garbage, etc.). Using discrete observation spaces and probability models, the computation the log-likelihood ratio of observation vectors of fixed size is carried out with one operation per pixel. Scale-invariant object detection is performed via multi-scale search.