Research On Improved Algorithm Of Generative Adversarial Networks Based On Zero-Centered Gradient Penalty

Generative Adversarial Networks,as a new type of deep generative model,have super modeling capabilities that can fit arbitrarily complex data distributions,and especially in image generation tasks,which show excellent performance.Therefore,it quickly occupied the main position in the field of deep learning once it was proposed.However,while the generative adversarial network model has many advantages such as clear images,simple and efficient,it also has some disadvantages,and mode collapse is one of the main challenges.The mode collapse means that the data samples generated by the generator have a high degree of similarity,that is,the samples are distributed centrally on the partial modes of the target distribution,and the diversity is poor.Mode collapse is also an unavoidable phenomenon in most generative adversarial network models at present.Based on the above problems,the main research content of this paper is to alleviate the mode collapse phenomenon in the generative adversarial network model,that is,to enhance the diversity of the data samples generated by the generator.The specific work can be summarized as the following two parts.First of all,in the first part of the work,two improved models are proposed,namely generative adversarial networks model with zero-centered gradient penalty is based on the determinantal point process(DPPGAN-0GP)and generative adversarial networks model with zero-centered gradient penalty is based on the distance of the discriminator-score(DGAN-0GP).Among them,the DPPGAN-0GP model introduces the method of determinantal point process on the generative adversarial networks model with zero-centered gradient penalty.This model not only effectively avoids the gradient explosion problem caused by the discriminator's strong discriminatory ability,but also uses the features extracted by the discriminator to construct the determinantal point process kernel,which establishes a connection with the real data distribution information,thus encouraging the generator to be able to generate samples with similar diversity to real data samples.The DGAN-0GP model is improved from two aspects.On the one hand,by minimizing the distance between the discriminator-scores as the objective function of the generator,guiding the generated distribution to align with the target distribution.On the other hand,random noise is mixed through reparameterization techniques,which enhances the modeling ability of prior noise and is more conducive to fitting the target data distribution.Then,in the second part of the work,a packed generative adversarial network model based on zero-centered gradient penalty(PacGAN-0GP)is proposed.Although it is easier to detect mode collapse in the PacGAN,there is also a problem in it that the discriminator is too powerful,resulting in the imbalance between the generator and the discriminator.Therefore,the zero-centered gradient penalty term is introduced on the PacGAN model,namely PacGAN-0GP model.This model guarantees that the discriminator has certain generalization while maintaining the discriminative ability.In addition,the PacGAN-0GP model is further combined with the determinantal point process method in the first part of the work,and a more complete model is proposed.Finally,the effectiveness of the proposed model in alleviating mode collapse and improving image quality is verified through experiments on different datasets.