Gabor Feature Selection Based on Information Gain

Abstract

In the field of machine vision object detection has become a popular area over the past several years. It is applied on a large scale in scientific research such as bioinformatics, machine learning and computer vision or in everyday life, like traffic supervision, access control, identification and authentication systems and also in industry, robotics etc. Every application has its own particularities and works only in some well-defined conditions. The main difficulty of general object detection comes from the extreme diversity in which all objects appear. They have a large variety of appearance, aspect, form, dimension, color, position, rotation angle, illumination, shadow or occlusion. In this paper we use numerous Gabor filters for feature extraction, specially tuned for global face and local eye detection. Because the high dimensionality of the data the obtained features are hardly manageable. We propose to apply, in the training and test phases, feature selection. Feature selection is an important step in almost every data mining problem. The selection of the most representative feature-descriptors is done by measuring the pairwise entropy of the filter responses. The final classification result is given by the most informative filter responses obtained from information gain of a weak classifiers computed from the corresponding filter responses on the training set. Besides, this paper compares other learning methods used in our previous works with the currently proposed approach, comparing the role of measuring the information gain and the mutual information between the selected filters.