Improved RealAdaboost Algorithm And Application In Face Detection

Face detection technology is the premise and basis for automatic face recognition,the goal of which is to search all candidate sub-windows in an input image,and determine whether there is a face,if there is,lable the position of the face in the image.At present,face detection technology has been widely applied in face recognition,video monitoring,web conferencing and many other areas.Face detection based on Adaboost algorithm is a mainstream direction currently,which has high detection rate and lower false detection rate with faster detection speed.As an improved algorithm of Adaboost,RealAdaboost algorithm expand the output of the weak classifier from discrete values to the real value space,which has a better classification performance than Adaboost algorithm.In RealAdaboost algorithm,through dividing the feature space of samples into several small areas,and statistics of the positive and negative samples probability distribution in small areas,the distribution of face in the feature space can be simulated.The feature space of samples is divided equally in traditional RealAdaboost algorithm,which can not reflect the boundary of the positive and negative samples clearly.The classification performance of the weak classifier still has room for improvement.Based on the analysis above,the way dividing the feature space of samples in traditional RealAdaboost is improved in this paper,which take minimizing the normalization factor Z as the optimization goal,and the thresholds of which divide the feature space of samples as decision variables,coding the thresholds and bringing genetic algorithm to build optimization model.At last,the optimal thresholds which divide the feature space of samples excellently are found out.Towards the statistical calculation of the sum of the weight of positive and negative samples in small areas,a weight integral list of samples is dsigned,which reduces a large number of calculations in the process of training,and optimizes the structure of the algorithm.Aimming at the output of the weak classifier with small areas which is divided non-equally,a new lookup table is constructed,which can avoid the seeking of the small areas of which the input features exists,and give the corresponding output directly.In the framework of waterfall cascade detector,a face detector is trained by the improved algorithm.In order to speed up the training,the scale of Haar features is restrained.On the one hand,the features which have good classification performance are reserved.on the other hand,the features with poor performance are trimmed.Then a multi-scale bootstrap of negative samples is designed,which not only solve the redundancy problem of negative samples for the training of each level of strong classifiers successfully,but also avoid missing representative negative samples.So each strong classifier can classify variety of samples correctly,and a further improvement of the performance of the face detector is achieved.Finally,the effectiveness of the improved scheme is verified by experiment.The weak classifier trained by the improved algorithm can achieve a smaller classification error rate,and the convergence speed of improved algorithm can be faster too.Compare to the traditional algorithm,the cascade detector trained by the improved algorithm can get a better detection result.