SA-SinGAN: self-attention for single-image generation adversarial networks

Single-image training is a research hotspot task of generating adversarial networks, especially in tasks such as image editing and image coordination. However, the existing network has a series of problems such as a long training time, poor image quality, and an unstable training model. Based on the research hot issues, we propose a single-image generation adversarial network of the self-attention mechanism and discuss the changes of the model when the self-attention mechanism is placed in different positions of the generator. We introduced the spectral normalization in the generator and discriminator networks to stabilize the training process and compared the influence of the learning rate on the network. We used artificial vision and model evaluation methods to test the performance of the model on three representative datasets and compared with the current more advanced models. Experiments show that our proposed model has better performance than single-sample generative adversarial networks, reducing Single Image Fréchet Inception Distance (SIFID) from 4.80 to 2.057 on the challenging Generation datasets, reducing SIFID from 0.06 to 0.02 on the Places datasets, and reducing SIFID from 0.23 to 0.04 on the LSUN datasets. The training time of our model is one-ninth of the single-sample generation adversarial network, which can obtain the overall structure of the single training sample, which has great research significance.