Face Detection Algorithm Based On Improved Viola-Jones Framework

Face detection as a preparation work of face recognition has been the researchfocus in the field of pattern recognition, especially after the appearance of AdaBoostalgorithm and the establishment of the framework of Viola-Jones, which makes theefficiency and accuracy of face detection on a higher level.This paper starts from the mathematical proof of the on-line allocation algorithmand the AdaBoost algorithm to improve the traditional framework of Viola-Jones.This paper proposes a skin segmentation algorithm based on a combination of borderand region to conduct a pretreatment of skin segmentation of the detected images.This approach can reduce the false detection rate of the face detector based on theframework of Viola-Jones, particularly reducing the false detection rate of thecomplexity background image of no one face. Due to narrow the detection area, theefficiency of the detection is highly improved. Then we optimize the weak classifiersneeded by the AdaBoost algorithm in the framework, respectively training out theoptimal characteristics of the a rectangular feature template by a single layerperceptron, and then we can get the corresponding optimal threshold value by thetheory of conditional probability, then we also can get weak classifiers have bothoptimal characteristics and optimal threshold value. Such process makes the error ofthe strong classifier trained by AdaBoost algorithm reduced, and substantially reducethe number of features. All of that increases the detection accuracy and efficiency ofthe face detector based on the framework of Viola-Jones. Finally, we use a feature settrained by AdaBoost algorithm as the input of the RBF neural network which istrained through the proposed new fuzzy clustering algorithm. It makes the improvedRBF neural network better reflect the coherent connection of the facial feature data.Because the input of this RBF neural network is the feature set trained by AdaBoostalgorithm, it can strengthen the learning of the counter-examples of the extractedfeature to decrease the classification error of the classifier and the false detection rateof the face detector based on the framework of Viola-Jones.At last, through Matlab simulation, we realize the face detector based on theimproved framework of Viola-Jones. In the simulation, we compared the renderingsof the face detector based on thetraditional framework of Viola-Jones and providemulti-angel face detection、multi-expression face detection、face detection with a covering and back ground face detection with no one face. Through the experiments,we draw the conclusion that the proposed improved face detector based on theframework of Viola-Jones is superior than the traditional one. It has a significantpractical value.