关键词: 深度学习, 生成对抗网络（GAN）, 生成模型, 对抗学习, 注意力机制

Abstract:

Generative adversarial network (GAN) has become a research hotspot in generation model, since it      can generate realistic images. To cope with the problem that the GAN cannot effectively capture the dependency between local and global features of the image and the dependency between different classes, this paper proposes a new generation model, named twins attention mechanism based generative adversarial network (TAGAN). Driven by twins attention mechanism, the real natural image is modeled by simulating the dependencies between local and global features and the dependencies between categories to create realistic fake images. TAGAN has feature attention and channel attention. The feature attention learns the correlation between similar features by selectively aggregating features. The channel attention learns the internal dependencies of each channel by integrating the relevant features of each channel dimension. The experiments implemented on the MNIST, CIFAR10 and CelebA64 datasets demonstrate that the proposed model is effective.

Key words: deep learning, generative adversarial network (GAN), generative model, against learning, attention mechanism