Abstract:

In the implementation of the traditional text image confrontation model, the convolution network in the discriminator is used to extract image features, but the convolution network can not consider the spatial relationship between the underlying objects, resulting in poor quality of the generated image, and the capsule network is an effective method. In this paper, the traditional text conditional generation adversarial model is improved based on the capsule network method. The convolution network in the discriminator is replaced by the capsule network to enhance its robustness to image size. The experimental results on data set show that the new model can effectively improve the quality of generation. The FID value of flower image generation is reduced by 14.49%, and the FID value of bird image generation is reduced by 7.18%. The inception scores of the images generated on Oxford-102 and CUB data sets are 22.60% and 26.28% higher, which shows that the image features generated by the improved model are more rich and meaningful.

Key words: generating images, capsule network, generation adversarial network, convolutional network, robustness

摘要：

在传统文本-图像对抗模型的实现中，判别器中的卷积网络用于提取图像特征，但是卷积网络无法考虑到底层对象之间的空间关系，导致生成图像的质量较差，而胶囊网络是一种有效的解决方法。基于胶囊网络的方法对传统的文本条件式生成对抗网络模型进行了改进，将判别器中卷积网络换为胶囊网络，增强其对图像的鲁棒性。在Oxford-102和CUB数据集上的实验结果表明新模型可以有效提高生成质量，生成花卉图像的FID的数值降低了14.49%，生成鸟类的图像的FID的数值降低了9.64%。在Oxford-102和CUB两个数据集上生成图像的Inception Score分别提高了22.60%和26.28%，说明改进后模型生成的图片特征更丰富、更有意义。

关键词: 生成图像, 胶囊网络, 生成对抗网络, 卷积网络, 鲁棒性