Evolutionary Feature Preselection for Viola-Jones Classifier Training

Object detection and recognition still pose a major challenge for computer vision systems. The Viola-Jones framework is a popular means of building classifiers that can perform these tasks effectively and efficiently. Viola-Jones draws upon a set of simple, Haar-like image features at all scales and positions. As this set grows rapidly with image size, it can become costly to evaluate and also encourages overfitting of the classifier. Our paper presents a unique application of artificial evolution to pre-select features for use by the Viola-Jones training method. A population of classifiers is evolved with varying feature sets and tested on detecting human faces in images. Several search strategies are compared and a greedy approach is found to most reliably lead to classifiers with improved accuracy. Training time is particularly reduced by the use of smaller feature sets. However, the computational cost for evolving the sets is high and arguably only worthwhile if the set is to be reused to train further cascades, such as in a dynamically updated system.