A SYSTEM FOR LOCALIZATION OF HUMAN FACES IN IMAGES USING NEURAL NETWORKS

Marián Beszédeš - Miloš Oravec

   In this paper, a neural network-based face detection system inspired by the work of H. A. Rowley [3] is presented. Faces are detected in an unprocessed input image. The system processes and normalizes small windows extracted from the input image. Digital image processing techniques, such as normalization of size, position and rotation, improvement of light conditions and contrast are used here. A neural network is applied in two parts of the system. In the first one it detects rotation of the input window and afterwards it decides whether the window contains a face or not. In both cases a multilayer perceptron is used. The choice of the best topology and training method is discussed here, too. A face detection neural network uses the method of distributing the decision among multiple subnetworks. A special bootstrap algorithm is used to train the network. The result of the face detection system is in the form of a set containing locations of human faces.

Keywords: biometrics, face recognition, neural network, multilayer perceptron, bootstrap algorithm, digital image processing, histogram equalization

[full-paper]