Image Generation Algorithm of Text Description Based on StyleGAN

The training of the text-to-image task relies on a large number of image-text pairs, and the problems of insufficient sample size of the dataset and inprominent image subject content will greatly affect the overall performance of the model. In addition, most of the existing methods are stacked multi-stage generation structure, and the quality of generated images is difficult to meet the actual application requirements. Aiming at the above problems, this paper enhances the dataset and designs a single-stage text generation image model based on StyleGAN. Based on the analysis of global semantic features and semantic diversity expressions, this paper carries out adaptive perturbation text data enhancement, and designs an image cropping method that fuses boundary information, and adopts random cropping and rotation to complete image data enhancement. In the generator, the conditional style construction network fuses global semantic information and hidden space vectors to obtain conditional style control vectors that control different visual features of the image. The generation network adopts a progressive generation structure, accepts the control of conditional style in the way of AdaIN, and adds random noise to each layer at the same time, which improves the authenticity of the image and enriches the details. At the same time, the semantic inner product is added to the discriminator, and the loss function of cyclic consistency loss and visual semantic matching loss are introduced to further improve the semantic consistency of graphics and texts. The IS, FID and R-precision values of this model on the CUB-200-2011 dataset reached 4.96, 14.11 and 75.20%, respectively.