Age Estimation Method Based On Facial Image

Image based age estimation is a new challenging research topic in computervision field. It has considerable potential applications in human-computer interaction(HCI), multimedia communication and some other areas. As aging process presentsdiverse patterns in different people, people may age differently, which is determinedby gene, living condition or some unpredicted factors. Finding an appropriate agefeatures as well as an appropriate age model has always been open problems in ageestimation. Based on the existing research work, this dissertation carries out a seriesresearch work on the age estimation area.The major works of this dissertation are listed as follow:(1) Image pre-processing. Mainly focus on face detection and image enhancement.The Adaboost learning algorithm has been used with the Harr feature for the facedetection. The histogram equalization algorithm has been used to improve the imagequality after the face area has been detected.(2) For the age feature extraction problem, a global age feature, the activeappearance model (AAM) has been discussed including model building and modelfitting. The AAM is extracted as the age feature for the age estimation.(3) As the global age feature (AAM) missing details for the face wrinkles， Onelocal texture descriptor, the local binary patterns (LBP) has been discussed, includingthe extended LBP descriptor. The multi-scale LBP (MLBP) descriptor is introduced tothe AAM, which is the MLBP based AAM method. In this way, the texture model hasbeen improved to give more texture detail for the face area.(4) The support vector machine (SVM) theory has been studied for the ageestimation, while both support vector classification (SVC) and support vectorregression (SVR) are included. In this dissertation, a global classifier is trained for agerange classification, besides several local regressors are also trained to get the accurateage. In this way, the age estimation’s accuracy has been improved. Experimental results are given to demonstrate the MLBP-AAM feature and SVCbased SVR method performing a lower mean-absolute error (MAE) and high accuracyof estimation compare to other methods.