Face detection by color and multilayer feedforward neural network

Abstract

In this paper, we introduce a novel approach for automatic detection of human faces embedded in dissimilar lighting. The proposed system consists of two primary parts. The first part is to convert input RGB images to a binary image directly using segmentation. Because absolute values of r, g, and b are totally different with various skin colors in altered lighting conditions and relative value between r, g, and b are almost similar with different skin colors in changed brightness circumstances, we use relative value between r, g, and b in segmentation process to binarize RGB images directly instead of color images to gray level images, then binary ones. For this reason, our system is very robust for different lighting conditions. The second part of proposed system is to search potential face regions and perform task of face detection. In second part, each face candidate is obtained from isosceles-triangle criterion that is based on rules of the combination of two eyes and one mouth, and then to be normalized to a standard size (60*60 pixels). Next, each of these normalized potential face regions are fed to neural networks function to obtain location of face region. The proposed face detection system can detect multiple faces embedded in dissimilar lighting conditions. Moreover, it can conquer different size, varying pose and expression. Experimental results demonstrate that an approximately 97% success rate is achieved and relative false estimation rate is very low.