Abstract: Recent challenging task of synthesizing images from text descriptions based on the generative adversarial network(GAN) has shown encouraging results.These methods can produce images with general shapes and colors,but often produce global images with unnatural local details and distortions.This is due to the inefficiency of the convolutional neural network in capturing high-level semantic information for pixel-level image synthesis and the fact that the generator-discriminator in a rough state generates flawed results for lack of detail,which then serves as input to the final result.We propose a generative adversarial network based on feature fusion,which introduces multi-scale feature fusion by embedding residual block feature pyramid structure,generates the final fine image directly by adaptive fusion of these features,and produces a 256px×256px realistic image with only one discriminator.The proposed method is verified on the flower data set Oxford-102 and Caltech bird database CUB,and the quality of generated images is evaluated by using Inception Score and FID.The results show that the quality of the generated images produced by the proposed method is better than images produced by some classical methods.

Key words: Discriminator, Feature fusion, Generative adversarial network, Residual block feature pyramid