Continuous Adaboost Face Detection Algorithm Based On

Against the application of Real Adaboost algorithm with Haar features in face detection, this paper analyzes the selection methods of Haar features as well as the criteria of cascade classifier, and proved the relationship between threshold in Real Adaboost algorithm and training error. For the shortage exist in the Real Adaboost algorithm, Real Adaboost algorithm smoothing factor is constant so that the smoothing effect to all Haar features are the same, which makes Real Adaboost algorithm is too prone to overfitting, we propose a Dynamic Smooth Adaboost algorithm, Referred to as DS-Adaboost algorithm.This paper describes the Haar feature selection method, and calculate the number of Haar features which is effective face detection for different scales template.Second, given conclusions relevant to the Real Adaboost algorithm, and gives a rigorous proof. Theorem1proved the relationship between threshold in Real Adaboost algorithm and training error, Theorem2gives the weight update strategy, and Theorem3consecutive the relationship between factor smoothing factor and normalization in Adaboost algorithm.Finally the value of the smoothing factor in Real Adaboost algorithm which is proposed by Schapire is constant, so that the general smoothing factor which is the number of training examples so that the smoothing effect to all Haar features are the same, which makes Real Adaboost algorithm is too prone to overfitting.we propose a DS-Adaboost algorithm, the algorithm is characterized by a smoothing factor which is the dynamic characteristics, and different smoothing factor corresponding to different values. Experiments show that, DS-Adaboost algorithm is largely suppressed overfitting and makes the detection rate has been improved to some extent.