Research On Face Generation Of Multi-attribute Based On GAN And Its Application In Assistant Recognition

Since Generative Adversarial Networks(GAN)was proposed and developed quickly in these two years,it had been applied to many research fields and solved existing problems,such as face disocclusion,image transformation,generation from texts to images and so on.From then on,GAN shows booming development trend in the field of image application.Facial attributes analysis is an active research topic in the pattern recognition for many years.Some researchers also turn the center of research into applying GAN on facial attributes transformation and face aging.However,in these works,compact ages are divided into groups and emphasize on generation of each group.Little works focus on face generation of fine-grained ages and application value of them on optimizing models of classifier in detail.Compared with traditional data augmentation and generative models,GAN does not rely on the form of Markov chain but directly generate samples through back propagation,which makes the efficiency and quality of generation higher and categories more diverse than ever before.Currently many public face databases are greatly restricted by limited samples.Based on great generative ability of GAN,this work proposes a novel FM-GAN for face generation of multi-attribute including fine-grained synthesis of different ages,such as a 35-year-old white man,and concentrate on the effectiveness of generated samples on data augmentation and improving performance of age estimation models.At the stage of experiments,we will evaluate our method on different public facial databases.Finally,on the basis of previous experiments,an active optimization architecture based on self-learning is further put forward.The main work of the paper is as follows:(1)Attempt different architectures and propose a novel FM-GAN for face generation of multi-attribute including fine-grained synthesis of different ages,gender and ethnicity.Synthetic images perform great visual fidelity and abundant diversity,and representations of gender,ethnicity and age are perfectly disentangled from latent vector.(2)A quantitative measure is proposed to weight the quality of generated images and effectiveness on improving models' performance.Experiments demonstrate that samples generated from FM-GAN are helpful to improve performance of pre-trained age estimation models on multiple public facial databases to the state-of-the-art..(3)In the procedure of optimizing model,there existing problems that generated samples for optimization are chosen completely by handwork without accordance,which wastes lots of time and effort,and takes up a lot of space.Then an active optimization architecture combining GAN and self-learning is further put forward.Model optimization is changed to online form.This method allows for GAN model directly involved in the optimization process and pre-trained models choosing generated samples that satisfy the requirements by themselves.In this paper,both of online and off-line optimization methods achieve the target of models' performance improvement and show state-of-the-art performance on multiple public facial databases.Compared with the off-line one,online form is more reasonable and effective.