Face detection and facial feature points localization in complex background
Face detection and recognition technology is a key application technology of computer vision field which has great potential value in economic, security, social security, military, and so on. In recent years, many studies carried out and achieved great success at home and abroad. Since the 1990s, face detection achievements in this stage are mainly to improve detection accuracy and various perspectives face detection (Sung & Poggio 1998, Yang & Roth 1999). After 10 years of development, the face detection accuracy has been greatly improved, the detection rate can reach more than 90% (Rowley & Baluja 1998). However, face detection to move towards practical application, the detection speed is a key problem needed to resolve. Therefore, the detection accuracy is improved at the same time and researchers pay more and more attention to the detection speed (Sanderson & Palial 2003, Feraud & Olivier 2001). The adaBoost based method proposed by Viola gains great popularity due to better detection results in 2001, but the error detection rate is higher (Viola & Jones 2001). In order to achieve higher detection rates and lower false positive rates, in this paper, we proposed an improved cascade adaBoost algorithm model training face classifier. We also present a facial feature localization algorithm based on image gray statistic, achieving rapid positioning of facial features, and verify whether to face each other based on facial feature localization, effectively eliminate the phenomenon of false detection under
complex and multi-face detection background. Using this new weight update rules and improved adaBoost algorithm is more accurate classification results, can effectively remove the background interference. Comparative experiments on images from the FERET and the self-made image databases demonstrate that our method is more robust than traditional methods to scale variation, illumination changes, partial occlusion, and complex facial expressions. The tests also show that the proposed method improves the detection rate and achieves high precision of localization.