A Study On Performance Enhancement Of Facial Image Based Age Estimation Algorithm

The task of facial age estimation is to automatically predict the age values or the age ranges by utilizing the face images.As an important research branch of computer vision,facial age estimation has been widely applied in security monitoring,smart business,entertainment,cosmetology and so on.For the obtained face images,the sample distribution is sometimes unbalanced for different age ranges.Moreover,the appearances of the same individuals may have a larger variation due to the age variation,illumination variation,pose variation,and other negative factors.Thus,how to effectively improve the facial age estimation accuracy is still a relative difficult problem.Therefore,we have carried out the following two works.(1)An effective facial age estimation algorithm is proposed based on extreme learning machine(ELM).First,a group of weak estimators are constructed based on ELM.In order to increase the diversity,the feature spaces are randomly selected and the initial weighting matrices are varied for different weak estimators.Then,a weighting model is devised to derive the weighting coefficients of the weaker estimators.Finally,the estimation results of the various weaker estimators are integrated to obtain the final estimated age.The proposed method can effectively reduce the negative effects caused by the imbalance of sample distribution.(2)An accurate facial age estimation algorithm is proposed based on sample clipping and label distribution support vector regression(LDSVR).First,a sample clipping model based on the estimated mean square error,which consists of five steps:division,classification,clipping,mixing and iteration.A new training sample space can be obtained by clipping the error and repeated training samples.Then,the age distribution probabilities are estimated by the LDSVR method with the new training samples.Finally,the final estimation values are obtained by select the age with the maximum probability.The estimation accuracy of LDSVR can be effectively improved by the strategy of sample clipping.In our proposed works,the performance of facial age estimation can be enhanced by decreasing the affection caused by the unbalanced sample distribution,the large appearance variations,and other negative factors.The experimental results demonstrated the effectiveness and feasibility of the proposed methods on several widely used face image databases.