Multiview Face Detection Using Six Segmented Rectangular Filters And Skin Tone Information

Face detection is a difficult task in image analysis which has increasing applications. The existing methods for face detection can be divided into image based methods and features based methods. In this research, an intermediate system that combined both methods was developed. Firstly, the images were processed using skin tone information in order to extract face candidates; then a learning algorithm was used to classify these face candidates.The procedure started by preventing the problem of lighting condition, which may change the skin color, by compensating the light in image. Through removing non-skin pixels, the skin tone was extracted, so a set of face candidates were obtained. These face candidates were classified as face or non face via the adaptive boosting algorithm (Adaboost).The family of simple classifiers contained six segmented rectangular wavelets which were reminiscent of the Haar basis. Their simplicity and an image representation called integral image allowed a very quick computing of these Haar-like features. With these new feature sets the training time became significantly shorter: five times faster than using the previous feature sets. The detection of faces in input images was preceded using a scanning window at different scales which permitted detecting faces of every size without re-sampling the original image.The experimental results demonstrated the effectiveness of the method in detecting profile and rotated faces under a wide range of variant light conditions. The method detected 96% of positive faces while 6% was declared as false positive.