Skip to Main Content
We present in this paper a new regression method adapted to problems dealing with a huge set of potential features like in pattern recognition. This method combines a boosted forward feature selection algorithm and a generalized regression neural network. The feature selection uses a new criterion, the fuzzy functional criterion, to evaluate the relevance of each feature. It is well suited to measure to what extent a random variable y can be viewed as a function of another random variable x. We explain how this measure is more appropriate than the classical mutual information. At each step, features are evaluated using weights on examples computed from the error produced by the neural network at the previous step. This boosting strategy helps our system to focus on hard examples during the feature selection process. The application is head pose estimation, a challenging problem in pattern recognition. Test are carried out on the commonly used Pointing 04 database and compared with state-of-the-art results.
Date of Conference: 14-19 June 2009