基于GAN的文本生成图像算法研究

Research on GAN-based Text Generation Image Algorithm

针对文本生成图像的问题,提出了一种基于生成对抗网络(GAN,Generative Adversarial Networks)的文本生成图像算法。为了得到更多的图像细节,该算法利用带有注意力机制的标题匹配模型从先验知识中选择和充实兼容的候选标题。然后,使用自注意机制来提取候选标题的特征,利用多标题注意力生成对抗网络从这些特征生成图像。最后,使用公共数据集Caltech-UCSD Birds来验证本文算法的性能;与其它相关方法相比,本文所提的方法能够生成更高质量的图像。

Aiming at the problem of text-generated images,this paper proposes a text-generated image algorithm based on Generative Adversarial Networks(GAN,Generative Adversarial Networks).To get more image details,the algorithm uses the attention-based title matching model to select and enrich compatible candidate titles from prior knowledge.Then,the algorithm uses a self-attention mechanism to extract the features of candidate titles,and uses a multi-title attention generation adversarial network to generate images from these features.Finally,the Caltech-UCSD Birds dataset(CUB)is used to verify the performance of the algorithm;the experimental results show that the algorithm generates high-quality images based on the text.