使用孪生注意力机制的生成对抗网络的研究

Research on Generative Adversarial Networks Using Twins Attention Mechanism

生成对抗网络(GAN)能够生成逼真的图像,已成为生成模型中的一个研究热点。针对生成对抗网络无法有效提取图像局部与全局特征间依赖关系以及各类别间的依赖关系,提出一种用于生成对抗网络的孪生注意力模型(TAGAN)。以孪生注意力机制为驱动,通过模拟局部与全局特征间的依赖关系以及各类别间依赖关系,对真实自然图像建模,创建逼真的非真实图像。孪生注意力机制包含特征注意力模型和通道注意力模型,特征注意力模型通过有选择地聚合特征,学习相似特征间的关联性,通道注意力模型通过整合各通道维度的相关特征,学习各通道的内部依赖关系。在MNIST、CIFAR10和CelebA64数据集上验证了所提出模型的有效性。

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.