结合频谱规范化与自注意力机制的DCGAN研究

针对基于深度卷积对抗式生成网络的图像生成方法存在训练过程稳定性亟待提高、图像生成质量效果欠佳等问题,提出一种将频谱规范化、自注意力机制与深度卷积对抗式生成网络结合的图像生成方法.在网络结构中,将频谱规范化的权重标准技术引入判别器,使判别器的参数矩阵满足Lipschitz约束,提高网络模型训练过程的稳定性;将自注意力机制...

RESEARCH OF DCGAN COMBINED WITH SPECTRAL NORMALIZATION AND SELF-ATTENTION MECHANISM

The image generation method based on the deep convolutional generation adversarial networks has problems such as the stability of the training process and the poor image generation quality.An image generation method combining spectral normalization,attention mechanism and deep convolutional adversarial generation network is proposed.In the network structure,the weighted standard technology of spectral normalization was introduced into the discriminator,so that the parameter matrix of the discriminator satisfied the Lipschitz constraint,and the stability of the training process of the network model was improved.The self-attention mechanism was introduced into the generator to obtain a better quality image.The experimental results show that the proposed method can effectively improve the convergence speed,training stability and image generation effect of the model on the CelebA and Cartooon datasets compared with the current generation model.