Age Estimation Using Multi-instance And Multi-label

In recent years, with the rapid development of intelligent machine vision computing,intelligent image processing fields have emerged a large number of new algorithms forsolving practical problems provides a good foundation, which makes face recognition,intelligent monitoring and friendly people applications in areas such as machine interactionhas attracted more and more attention. In the face of the many challenges of the current issueof machine vision and pattern recognition, computer estimates by age face age estimation isone of the challenges that the current age estimates based on face research work is stillrelatively weak, mainly includes two aspects:the first one characteristic facial imagerepresentation, although the face image feature extraction methods are many, but thecharacteristics of suitable age estimation methods are not fully represented;second ageestimation model of learning, in some recent papers are generally age estimation modelusing Support Vector Machines (SVM) and Support Vector Regression (SVR) and othermethods, but the problem of age estimation particularity makes the mere use of these twomethods do not necessarily yield good results.Thus, the study can be carried around these two problems. First, estimate the agelearning model, the use of multi-label classification theory to the age of the model. Due tothe age of the label change is gradual, and the age label to determine the existence of fuzzy,so consider the training sample face image assignment multiple approaching the age label, soafter every training set to get more than the estimated model to predict the age label, and Agecan calculate the probability of the label corresponding to the predicted, then the probabilityof solving these tags can linearly to obtain more accurate predictions of age label;Second,improved facial feature representation. Use a combination of multi-feature multi-instancelearning framework, the local feature BIF, LBP and AAM combined to characterize theglobal features of face features a combination of age, each of which corresponds to an ageof the sample is that the sample frame by means of a plurality of single multi-sampleframework for multi-label to learn the various features of the model are obtained by themulti-label multiple labels corresponding to predict the age and characteristics of the predicted probability of each label to do weighted solve this theoretical analysis to achievehigher estimation accuracy and has good scalability.In addition to these two aspects of improvement, age estimation details still have someimprovements, such as the multi-layered labels and lsLDA(sensitive label LinearDiscriminant Analysis) method to improve the accuracy of age estimation. At the same timein order to prove the validity of the proposed method, the need for a large number ofexperiments on FG-NET ages database.