Face Detection Research Based On The Adaboost Algorithm

Abstract:Based on the background of static image study, this paper fuses the characteristics of human skin and algorithm of Adaboost cascade classification to design a face detection system. Because of nonrigid structure of human faces, present research situation and influence of image acquisition from external environments, all these make the development of face detection technology be in immature state. As an important development direction in pattern recognition and computer vision, face detection has important application values in the fields of image and video retrieval, information security, face recognition and detection, and video monitoring.This paper mainly discusses the principle and process of Adaboost algorithm which is applied in face detection. The whole process of detection mainly includes extraction of image rectangular characteristics, calculation of rectangular characteristic value, reference mechanism of integral map, training of weak classifier, ascending of strong classifier, and training of cascade classifier information. That just taking advantage of Adaboost algorithm to detect the human face in the pictures will cost high rate of residual, in additionally, adaptive adjustment of sample weights in Adaboost algorithm can lead to the phenomenon that non-face samples learn excessive training weights. Against question one, the fusion of face skin color feature will reduce the classifier detection area, narrow the key testing space of color area in human skin. Extraction of color area in human skin also involves the knowledge in image processing, because the distribution of human skin gray value in space is dense, so we can use color space to establish skin color model, then use threshold method binarize skin pixels and background pixels, finally get a binary image. After filtering the edge sharpening noise on the preprocessed binary image, we can get a image in which target region and background region have been separated. Against the question two, the solution is to set a threshold value on non-face samples, when samples’ weights are less than the threshold, it will increase the samples’ weights, if samples’ weights are out of the range of this threshold, it will decrease the weights correspondingly, which can effectively solve the over learning situation of error classified samples. In the field of face detection research, the biggest problems are the accuracy and speed of detection which doesn’t meet people’s expectations. In the future, face detection research will focus primarily on following situations:face detection system no longer use one certain kind of detection method, the direction of its development is definitely fusing two or more face detection methods, and improving precision and speed through fusion with each other. Finally, the tested object will be more complex, which is no longer a static or single face detection, but under a complicated background, or in the dynamic video.