Face detection in complex backgrounds based on the integration of neural classifiers through similarity analysis of higher order representation patterns

Abstract

A new algorithm is herein developed for combining the classification decisions of different feedforward neural network models with applications to face detection in complex backgrounds. Instead of the usual approach for applying voting schemes or linear combinations schemes on the decisions of their output layer neurons, the proposed methodology integrates higher order representation patterns extracted by their upper hidden layer units. More specifically, different cases of each such model, corresponding to the same classification problem, derived from the same training process but with different training techniques and parameters, are involved in terms of their higher order features, through similarity analysis, in order to find out stable higher order representation features. This procedure is repeated for the various feature extraction methodologies involved in the original classification problem of face detection. Then, all such higher order features are integrated through a second stage Combiner neural network classifier having as inputs suitable similarity features of them. The proposed improved hierarchical modular neural system for pattern recognition presents increased classification performance in the face detection task in complex backgrounds. The two different feature extraction methods (FEM) involved in this problem deal with a specific eigenanalysis technique applied to both image itself and, secondly, to the image in the frequency domain. Such an eigenalysis aims at identifying principal characteristics in the image and the transform domain of uniquely identified classes relevant to face detection. The experimental study illustrates that such an approach, integrating higher order features through similarity analysis of a committee of the same classifier instances (cases) and a second stage neural classifier, outperforms other combination methods, like voting combination schemes as well as single neural network classifiers having as inputs all FEMs derived features. In addition, it outperforms hierarchical combination methods non performing integration of cases through similarity analysis.