Creating Classifiers using Artificial Neural Networks and the ADABOOST Principle

The problem of constructing various types of object detectors in images is still an urgent task, despite the relatively strong set of methods described in the literature. One of the methods that have become standard for the construction of efficient and fast classification, is a Viola-Jones cascade, which is still fundamental to search for objects in the image in real time and which implementation has been included in the open-source computer vision library OpenCV. For the experiments in this study we used the database of images CMU Face Database. In practice, when we use of the algorithms in computer vision the computational complexity becomes a significant factor. Preferably, one should use threshold decision rules or Haarfeatures as classifiers, which gives small the computational complexity. In this paper, the approach to the construction of classifiers of comparable performance for the problem of detecting faces. For the construction of the detector were studied approach involves separating detection process into two stages: construction the descriptor of image, and classification stage. For the phase, which responsible for the classification, were considered two possibilities: a two-layer neural network, i.e. using multilayer perceptron as a “strong” classifier, and a cascade of several such networks of different size. For the phase, which responsible for forming the descriptor, we also have investigated two possibilities. First one - fixed Haar-basis, which gives us a feature-vector of the descriptor of input image. This basis was constructed using the ADABOOST principle. The second possibility, investigated in this paper, was the construction of the basis of fewer required Haarfeatures, every of which more accurately reflects the object characteristics, which was obtained by using Karhunen-Loeve transform. In order to get Haar-features from eigenvectors, they have been quantized. As a result, the classifier built with efficiency which comparable to the Haar cascade.

ADABOOST, neural networks, ADABOOST, Haar-features, object detection, image processing