多生成器生成对抗网络

Multi-generator generative adversarial networks

生成对抗网络(Generative adversarial networks,GAN)广泛应用于各种领域,尤其在图像生成方面.该模型由生成网络与判别网络2部分组成,在无监督的训练方式下,2个网络相互竞争相互提高.然而,GAN在训练时经常出现模式崩溃问题,进而导致模型收敛较慢,生成样本多样性较差.为解决这一问题,在深度卷积神经网络的基础上提出了一种多生成器生成对抗网络模型.该模型包含多个生成网络,每个生成网络均使用残差网络进行搭建,同时在生成网络间引入协作机制,以加快模型获取信息并减少参数量,最后将各生成网络的特征图进行融合得到最终图像输入到判别网络中.GAN在训练过程中还会出现梯度消失、训练不稳定问题.为避免出现这些问题,将Wasserstein距离和梯度惩罚引入模型的损失函数.通过在多个数据集上与多种相关方法进行实验比较,结果表明提出的模型在缓解模式崩溃问题、加快模型收敛速度以及减少参数量上均明显优于其他几种方法.

Generative adversarial networks(GAN)are widely used in various fields,especially in image generation.The model consists of two parts,a generative network and a discriminative network,and the two networks compete with each other to improve each other in an unsupervised training method.However,GAN often suffers from pattern collapse during training,which leads to slow convergence of the model and poor diversity of generated samples.To solve this problem,a multi-generator generative adversarial network model based on deep convolutional neural networks is proposed.The model consists of multiple generative networks,each of which is built using residual networks,and a collaboration mechanism is introduced among the generative networks to speed up the information acquisition and reduce the number of parameters,and finally the feature maps of each generative network are fused to obtain the final image and input to the discriminative network.GAN also has the problems of gradient disappearance and training instability.To address the two problems,Wasserstein distance and gradient penalty are introduced into the loss function of the mode.Through experimental comparison with several related methods on multiple datasets,the results show that the proposed model significantly outperforms several-other methods in alleviating the pattern collapse problem,speeding up model convergence,and reducing the number of parameters.