Facial Age Estimation Method Based On Convolutional Neural Network

With the rapid development of computer vision,face detection and face recognition technology has become a hot topic in the field of computer vision.Facial feature is one of the most important biological features of human beings,which can convey large amounts of information,such as gender,age,mood,race,etc..As an important biological feature,age information has been widely used in the fields of human-computer interaction,security monitoring and so on.The age estimation technique based on face image has become an important issue in the field of computer vision and pattern recognition.In recent years,convolution neural networks have made excellent progress in many tasks and a large number of labeled training set data is one of success factors of its network.However,it is difficult to collect sufficient and accurate labeled training set data in facial age estimation.It is necessary to study how to make full use of limited labeled training data on visual recognition tasks.Image data augmentation is a way to expand the training data set.The existing image data augmentation methods are only a few,including image flipping,image cropping,image rotation and so on.In this dissertation,a new unsupervised image data augmentation method based on the saliency map is proposed.In addition,there are many prevalent network structures used in image recognition tasks,such as ZF network,VGG network.Considering the unbalanced distribution of the age labels of facial image database,this paper studies the transfer learning ability by finetuning the prevalent network structures on the facial images whose labels remain relatively rarely.It further improves the accuracy of facial age estimation.The contribution of this paper are as follows.(1)This dissertation proposes a new unsupervised image data augmentation method based on saliency map.By setting the background as random noise or others,the number of images will be expanded.What is more,this image data augmentation method can be applied not only to the facial age estimation task,but also to other visual tasks.(2)It improves the ability of facial age estimation by finetuning the prevalent network structures.And this dissertation achieves the state of the art facial age estimation results in two facial image age estimation database.