Dimensionality reduction in face detection: A genetic programming approach

The high number of features in many machine vision applications has a major impact on the performance of machine learning algorithms. Feature selection (FS) is an avenue to dimensionality reduction. Evolutionary search techniques have been very promising in finding solutions in the exponentially growing search space of FS problems. This paper proposes a genetic programming (GP) approach to FS where the building blocks are subsets of features and set operators. We use bit-mask representation for subsets and a set of set operators as primitive functions. The GP search, then combines these subsets and set operations to find an optimal subset of features. The task we study is a highly imbalanced face detection problem. A modified version of the Nai¿ve Bayes classification model is used as the fitness function. Our results show that the proposed algorithm can achieve a significant reduction in dimensionality and processing time. Using the GP-selected features, the performance of certain classifiers can also be improved.