Track: Artificial Intelligence

Abstract

Over the last decade, great interest has been attributed to resolve faces detection and tracking problems that being very important in development of Human-Computer Interaction systems and audiovisual intelligent services; including face recognition, video conferencing, e-learning, security and telemonitoring. This particular interest given to faces is due on one hand to their wealth on information and their structure fixed relatively with little deformability and on the other hand to importance of faces in communication.

This task simple for humans is still very difficult and complex for machine where face can have different position and color representation.

Through this work we try to evaluate the color impact on performance (detection rate and processing time) of faces detection system. For this we tested several approaches (MLP, PCA, MLP-PCA ... etc.) on color and grayscale images in environments with varying complexity.

In color images our approach is based on the skin searching using a parametric model. Once skin regions (faces candidates) detected we apply the classification methods mentioned above in face / no face decision. In the case of gray scale images the process of face searching is throwing the entire image and as face can occur in different scales in the same image, searching is therefore based on multi-resolution processing (pyramidal image processing) where we scan all image to search faces candidates, using Yang and Huang faces model, which will be classified as in color images.

Preliminary results show that although faces detection rate in color and gray scale images are very close but the performance of these systems is largely based on false detection rate and processing times that are significantly lower in color images compared to grayscale image.