Research And Realization Of Face Detection Based On Adaboost Algorithm

Face detection, with the convenient of acquisition data and the wide rangeof application,is a kind of non-contact detection of biological characteristics,becoming a concern topic in the field of data mining and pattern recognition. Inthe pursuit of efficiency of modern society, image and video have stored moreand more real-time information.Finding an efficient way to search and detect theface from a mass of data has become an urgent problem,which need to besolved. Adaboost algorithm is a popular machine learning algorithm in recentyears, and has been applied to face detection by many researchers. This paperwill be optimized from the process of training to testing,and applied to detect theface from image,video and the real-time camera, basing on the adaboostalgorithm.The research work is shown as follows:(1) The current research situation of domestic and international and theexisting problems of face detection technology will be summarized.(2) The Adaboost face detection algorithm will be deeply researched, withthe analysis of the basic idea and the realization of the process. The principle ofHaar feature, integral map and classifier are introduced in turn.(3) The improved Adaboost algorithm can improve the detectionperformance effectively.In order to avoid the problem of over training in thelearning process of the Adaboost algorithm,a method to enrich the trainingsamples is proposed，which add new samples obtained after image processing oforiginal samples to sample set. According to the problem ofcharacteristics involves a large amount of calculation in training process, amethod of cutting training samples is raised,with the priori knowledge that the position of face in the middle of training samples. For the problem that theimbalance of sample weight distribution will be occoured when there existssome samples hard to classify,a new method of weight distribution is putforward,combining positive and negative error ratio and the classified error rate.The improved Adaboost algorithm can not only effectively avoidthe imbalance of the weight distribution but also improve thedetection performance.(4) To reduce the number of false detection target of Adaboost algorithm inthe detection process, the method of false alarm reduction were putforward,which has two kinds. Front false alarm reduction refers to using edgeenergy detection to test the window before the classifier. Rear false alarmreduction refers to using false alarm mechanism,consisting of skin colordetection and edge board,to test the window after the adaboost algorithm.Experiments show that the proposed method can effectively reduce the numberof false alarm.(5) The core idea of PCA algorithm is analyzed. The real-time facerecognition system is realized by PCA algorithm and the improved Adaboostalgorithm.