Abstract:

Based on the Stacked Generative Adversarial Networks (StackGAN), a novel method is presented to solve the problem of insufficient diversity caused by non-uniformity of generated samples, which constructs the stacked text-to-image generation antagonistic network model by combining local-global mutual information maximization. In the method, the global vector is first decoupled from the generated model to obtain different scale feature maps. And then, the correlation between global features and text descriptions is enhanced by maximizing mutual information between feature maps and global vectors. Finally, in order to make the text-to-image mapping more relevant, we extract the feature map as a local position feature vector, and enhance the correlation between it and text description by maximizing the average mutual information between the local position feature vector and the global vector. Numerical results show that the proposed method can improve effectively the diversity of generated samples on the CUB dataset. Moreover, it is possible to generate samples with a higher semantic accuracy and the method is more realistic for subjective evaluation.

Key words: image generation, information, generative adversarial networks, local position feature vector