Research On Generative Performance And Mode Collapse In Conditional Generative Adversarial Networks

In the field of vision,Conditional Generative Adversarial Networks(cGANs)have achieved great success in many application scenarios,aroused widespread attention in the academic community,and become an important frontier topic in deep learning research at home and abroad.However,there is still a lot of room for improvement in the authenticity and diversity of the samples.Hence,this paper takes ACGAN,a representative method in cGANs,as the starting point to discuss the existing problems and improvement strategies of cGANs.This paper first studies the authenticity of ACGAN,and finds that two major problems restrict the authenticity of the generated samples,namely the problem of non-salient features and the problem of difficult fitting of data distribution.The former means that ACGAN does not pay attention to non-salient features,and the latter means that the prior distribution of the generator is difficult to fit the distribution of the data.Aiming at the problem of non-salient features and the difficulty of fitting data distribution,this paper proposes a normalized Soft Max improvement strategy and a conditional prior distribution sampling strategy,respectively,so that ACGAN can focus on non-salient features and set different prior distributions for different conditions.Make the prior distribution closer to the high-dimensional manifold of the data distribution,thereby reducing the fitting difficulty of the generator and improving the authenticity of the generated samples.This paper conducts experimental verification on CIFAR10,CIFAR100 and Tiny-Image Net datasets,and uses FID and Inception Score evaluation indicators to compare the quality of generated samples before and after improvement.The experimental results show that the improved model of the baseline method can obtain better performance.On the basis of the first work,this paper studies the diversity and mode collapse problems of ACGAN,and finds that the entropy drop effect of ACGAN auxiliary classifier seriously affects the diversity of generated samples.In response to this problem,this paper proposes a multi-center improvement strategy,that is,setting multiple cluster centers for the same condition to accommodate multiple patterns of samples under the same condition.The experimental results show that adding the improvements in this paper to the baseline model can effectively improve the diversity,and it can be compatible with other methods to jointly improve the model.