MCA-GAN: Text-to-Image Generation Adversarial Network Based on Multi-Channel Attention
Jingcong Sun, Bin Zhang · 2019
It is an extremely challenging task to text-to-image generation based on generative adversarial network(GAN). Although some researchers have made remarkable progress in the field of image generation in recent years, it is still a problem to be solved in the aspect of texture detail in image generation. An attention generation algorithm based on multi-channel is presented in this paper. The network gradually refine by three branches to produce high-resolution images that conform to the text description. By the iterative training of GAN, the multi-channel feature attention mechanism is used to guide the pixel value deviation and loss in the process of image generation. Through a lot of experiments and people's visual perception of the generated images, the algorithm based on multi-channel attention mechanism proposed in this paper has been significantly improved in the processing of details.