Traditional Adaboost face detection algorithm uses the Haar-like features to train face classifier, and has a low error rate in detecting the face region. However, as in the complex background, it’s prone to false positives for the background region similar to the face gray distribution, causing the traditional Adaboost face detection algorithm with a high false detection rate. As one of the important characteristics of human face, skin color has a better clustering within the YCgCr color space, so the non-face background region can be quickly filtered out by the color model.

An Adaboost face detection algorithm based on the skin color model of YCgCr color space is proposed by IOE’s researcher XU Zhiyong, who combines the advantages of Adaboost algorithm and skin color detection algorithm. In this research, to obtain the human face classifier, ORL, FERET face database, and more than 2,000 face samples each in the network front and side are chosen for training. Then, the photos to be detected are transformed into YCgCr color space to go through skin color segmentation, filter out the non-color background area, and use the human face classifier obtained during the training to detect the candidate shin color area, see whether containing human face. When the classifier slides over the sub-window of the photos to be detected, there is no interference from the background area anymore. Experiments show that on condition of ensuring the detect time is not increased significantly; this approach can obviously improve the detection accuracy and false detection rate than the traditional Adaboost face detection methods.

The results of this paper will be published SPIE Proceedings.