Research On Key Technology And Applications Of Generative Adversarial Nets

Generative Adversarial Nets(GANs)proposed by Goodfellow in 2014 introduces a creative new method to generate data.The quality of samples generated via GANs is improved a lot compared with the traditional methods due to the adversarial training process.Moreover,the framework of GANs is flexible.The generator or discriminator can be any differentiable function,which lays a foundation for further applications of GANs.Data generation via GANs has made great progress.But there are still several chal-lenging problems that have not been totally solved.Such issues are conditional data gen-eration,vanishing gradient during discrete data training,mode collapse,etc.These disad-vantages have seriously limited its further applications,such as conditional facial attributes data generation,discrete musical data generation,and so on.Researchers are working hard on these tasks and rapid progress has been made.However,there is still space to improve GANs,especially in building more robust models in specific applications with a higher quality of results and avoiding its original defects based on some essential techniques.This dissertation has reviewed current research works of GANs,then analyzed significant problems and challenges about GANs,introduced related methods to solve or relieve the issues,performed a series of experiments to estimate the methods.The innovative works are as follows,(1)Iterative GANs is introduced to solve the problem of conditional facial data gen-eration based on GANs.The model tries to regress the latent variants and vectors with semantic information related to the input sample.Based on this,the generator can re-build the image or modify the attributes of the image by controlling the latent variants and conditional vectors.Sample rebuilding,attributes transformation and conditional data generation thus can be implemented in a unified GANs framework.The experiments over Celeb A facial data set have shown that the proposed method achieves outstanding results of these tasks.(2)Stack-Chord-GANs is introduced to generate musical data from coarse to fine to solve the discrete musical data generation problem.The prior musical knowledge is taken into consideration.The chord information is introduced into adversarial training,and two different encoding patterns of chords(dense coding based on prior probability and R-chord2vector based on root key attention)are proposed,which avoid vanishing gradient of discrete music data and pay attention to music-theoretical knowledge.The maximize variance regularization constraints based on the sliding mean value method is proposed to avoid the mode collapse issue.Experimental results show that our method generates more reasonable and pleasant music data in objective and subjective estimations.(3)GGS-GANs is introduced to explore efficient ways to apply GANs to specific ap-plication areas.The benefits of self-supervised learning in GANs is also estimated.The ability of the generator and discriminator is enhanced by applying the global and group strategy.Moreover,self-supervised learning is applied to classify the group information and make contributions to improve the efficiency of the whole system.Theoretical anal-yses and experimental results have shown that the proposed method achieves promising results among current SOTA(state-of-the-art)works with the help of the global and group strategy,adversarial training process,and self-supervised learning.