Facial Aging Based On Face Images

Age-related face image analysis based on machine learning has become one of the important research parts in the field of computer vision.Particularly,facial aging simulation based face images has been given increasingly attention these years.However,how to produce a credible and natural face image is still facing many challenges.This thesis surveys the state-of-art about the research of face aging,explores face aging simulation based on machine learning,and puts forward two new algorithms for face aging.(1)Face aging based on 2D-DCT and joint High-Low frequency dictionary learningGiven face images from both low-age group and high-age group,low frequency part from low-age group and the corresponding high frequency part from high-age group are extracted and formed into train set,and used for joint high-low frequency dictionary learning in the stage of face aging modeling.Given an input face image for aging simulation,the low frequency part is firstly separated based on 2D-DCT.The high frequency part corresponding to the high-age group can be further reconstructed based on sparse representation and the dictionary of high frequency.Finally the simulated aged face image is synthesized by combining the reconstructed high frequency and the low frequency part from low-age input image.Experiments on FG-NET dataset show that the aging effect of the method is more reliable.(2)Face aging based on Conditional Adversarial Autoencoder and low-quality face image enhancementFirstly,global face age simulation is conducted based on the Conditional Adversarial Autoencoder(CAAE).Secondly,block-based image super-resolution via dictionary learning and sparse representation is used to globally enhance texture details in order to get more realistic and high quality for the simulated results.Finally,age-related facial components including eyes,nose,mouth and forehead are extracted from the train samples in each age group.By means of similar facial component searching via mean SSIM index and Possion editing,the visual quality and the aging effects for the final simulated aged face image can be further improved locally.Experiments on UTKFace dataset demonstrate the effectiveness of the proposed algorithm.