Research On AdaBoost Algorithm In Face Detection

Face detection is the first and key step in face recognition. It is widely appliedin security system, Video Surveillance, Human-computer Interface and so on.AdaBoost algorithm becomes one of the most important algorithms because of itshigh detection rate and fast detection speed. It has important practical value tofurther improve the detection rate and reduce the testing error rate of the AdaBoostalgorithm. The problem of the fitness function optimization and sample spacepartition are further discussed in this paper, in which proposed algorithms simulatein the matlab. Experimental results indicate that the improved algorithm has asuccessful application in terms of the performance of the AdaBoost algorithm.The primary contributions of this paper are listed as follows:(1)A new fitness function is proposed which uses the absolute differencebetween the threshold and feature to measure the extent of misclassification andcombine with the relative entropy principle as the fitness function. The new methodselects the best weak classifiers according to the extent of the misclassification.Therefore, the proposed method can efficiently avoid the issue that the traditionaltarget function of the AdaBoost algorithm based on PSO can not adapt to theproblem of weak classifiers selection when they have the same minimum error rate.The experimental results on MIT face database and on image database from internetindicate that the proposed method can achieve both better performance and lessgeneralization error.(2) A real AdaBoost face detection algorithm based on MCV partition is given.The traditional finite division can not reflect the distribution of positive and negativesamples. Focusing on this problem, this paper measured the similarity of the samplethrough calculating the class variance of every finite division in the process of finitedivision and selected the best finite division corresponding to the minimum sum ofclass variance. Simulations show that the algorithm is of better detection rate, lessfalse detection rate and faster convergence.