Abstract:

In order to solve the problem that the feature information of the generated image against the network is insufficient, so that the generated effect characteristic is not obvious, and the key feature information of the image is blurred, this paper proposes an image generating method for a conditional self-attention generative adversarial network. The network combines the advantages of the self-attention generative adversarial network, and adds additional conditional features to the generator and the discriminator. It is explicitly indicated that the model can generate corresponding iconic category information. The specific dimensions of the data are related to the semantic features. In this way, the generation model is extracted, so that the feature representations of the images of a particular type are more closely matched to the original data distribution. Experimental results show that the FID values of the proposed method on the CelebA and MNIST data sets are increased by 1.26 and 2.47, respectively, compared with the self-attention generative confrontation network. It is verified that compared with other supervised class generation models, the proposed method can effectively improve the image quality and diversity, and can make the network converge faster.

Key words: generative adversarial network, condition feature, self-attention, image generation