Research On Partial Least Squares Dimension Reduction Based Facial Age Estimation

Facial age estimation is a computer vision technique for automatically determining face age.It is based on the age image features extracted from the human face images,and processed and analysized with models and techniques relevant to pattern recognition.As an important application of pattern recognition,facial age estimation is a hotspot in the field of computer vision and human-computer interaction.But human aging is a very complex and independent process;it is not only related with health status,lifestyle,nature of work,life stresses and other factors,but also affected by some complicated factors,such as genetic genes.It has the characteristics of diversity and uncertainty.In general,different people will have different appearance in the same age,which imposes great difficulties and challenges to facial age estimation in the academic research.Usually,the feature dimension of face image is relatively high,it is necessary to take effective way to reduce the dimension.Principal Component Analysis(PCA)is the common way to reduce dimension.However,as an unsupervised method,PCA does not consider age information and the extracted principal components' interpretation abilities are insufficient.The dimension reduction method based on supervised partial least squares(PLS)is an effective way to deal with this problem.Partial least squares is not only an efficient data dimension reduction model,but also a regression analysis model;it can efficiently interpret the low dimensional space with better performance.Specifically,Action Appearance Model(AAM)is firstly adopted to extract human face features.Then the face features are processed by PLS based dimension reduction.Finally,Support Vector Regression(SVR)is used to estimate human ages.The combination of Partial Least Square dimension reduction and Support Vector Regression is a major innovation point in this paper.Compared with the method of directly estimate age with PLS regression,the method proposed in this paper enjoys obvious advantages.We conducted a large number of comparative experiments in FG-NET face image database,including Rank algorithm,Support Vector Regression Algorithm(SVR),Neural Network Regression(NNR),algorithm combining Principal Component Analysis(PCA)and Support Vector Regression algorithm(PCA+SVR),and Partial Least Squares Regression(PLSR)algorithm.The results show that the method proposed in this paper has better performance with smaller mean absolute error.