Comparative study on the influence of Mahalanobis distance and skin color range for face detection using AdaBoost

This paper reports a comparative study on using Mahalanobis distance and skin color range to segment the skin region. It also studies the effect of using skin region segmentation for face detection. Range of 1σ from the skin color mean in CbCr channel is used to segment the skin region, instead of using Mahalanobis distance. Ranges more than 1σ should be avoided, as they will give precision and recall worse than that of 1σ. Using skin color region to segment the face candidate does not necessarily reduce the detection time, and it is not always true that the detection rate of the detector will be improved. In most of the tests, the precision does not fall much when we increase the range. This means that the performance of the detector is `buffered' by the AdaBoost-based detector.