Generative Adversarial Network Based On Ensemble Learning And Applications In Image Generation

Generative adversarial network(GAN)is a hot research topic in the field of deep learning and has been successful in many aspects.Its performance is determined by both the generator and the discriminator,and to improve the quality of the generated image by the generator,the discriminator's discriminative ability also needs to be improved,and vice versa.There is a big gap between unsupervised learning based GAN and supervised learning in image generation.Supervised learning requires a large amount of labeled data for training,and it is usually expensive to obtain labeled data.Unsupervised learning uses unlabeled class data to train the network,which is prone to poor sample diversity,low quality,and long model training time.Therefore,it is a meaningful research work to improve the performance of GAN and well solve the problems of GAN in unsupervised scenarios.One is to modify the objective function of the GAN to obtain better local optimal solutions and thus increase the diversity of the generated samples;the other is to modify the network structure of the GAN to "force" the generation of more samples with more patterns.In this paper,the second approach is adopted,and the main work of this paper includes the following two aspects:(1)A multi-discriminator generative adversarial network method is proposed,which integrates the discriminative networks in GAN with the idea of selective integration learning,mainly to avoid the discriminative error caused by the poor discriminative performance of single discriminative network,which affects the generative network learning.To improve the performance of the discriminator,the discriminative networks with different network structures are employed.Meanwhile,a majority voting strategy with dynamic adjustment of the voting weight of the base discriminative network is introduced,which effectively prevents the discriminative networks from being too strong and thus inhibiting the learning of the generative networks.(2)A multi-generator generative adversarial network method is proposed,which mainly improves the performance of generator and loss function.Firstly,we adopt the idea of integrated learning to integrate the generative networks,because each generative network has the same goal to "cheat" the discriminate network,so the information can be "exchanged" among the generative networks in the model,and this information exchange in the specified layer can make the integrated generation system to learn the data features as soon as possible and shorten the training time of the network,and then fuse the feature maps of multiple generation networks as the final image input to the discriminator network,which greatly increases the detail information in the generated samples.In addition,since the use of simple convolutional superposition is not conducive to the image generation of GAN,the residual network is incorporated in the construction of the generative network.Finally,Wasserstein distance is used as the loss function,which greatly improves the gradient disappearance problem due to the discontinuity of JS.In order to verify the feasibility and effectiveness of the proposed two methods,experiments are conducted on multiple datasets and compared with various representative methods.By comparing the quality and diversity of the generated images and the time required for model training,it is demonstrated that both proposed methods outperform the compared methods.By comparing the quality and diversity of the generated images and the time required for model training,it is demonstrated that both proposed methods outperform the compared methods.