Research On The Generation Of Independent Poster Based On Generative Adversarial Network(GAN)

Image generation technology is widely used in daily life and also a very important research direction in the field of computer vision.The generation of text to images is an important branch in this direction.The usual image generation tasks can only complete the image synthesis of a single category,and cannot meet the increasingly complex and delicate production requirements.It is especially important to generate images with rich content and detailed content according to the text description of the given conditions.Especially for poster design staff,the generation of independent posters based on the technical model of text-based image generation can not only greatly reduce the boring simplicity of poster staff.The poster design work frees them from a lot of repetitive mechanical labor,and can improve their design efficiency,allowing designers to engage in more creative work.To generate graphics that conform to the semantic expression of text content through text,first of all,it is necessary to solve how the computer can understand the semantic information contained in the text,and secondly,convert the semantic information of the text understood by the computer into the corresponding graphic output.In this paper,through the study of text generation image method,a text generation image generation method based on improved confrontation generation network with Chinese text as input is put forward and verified by corresponding experiments.The research on generating images by synthesizing existing text found that current text input is based on English characters as text input.In order to facilitate domestic design staff,the paper proposes text input using Chinese as a text image.Use the Word2 vec model to generate word vectors,and input the word vectors into the Seq2 Seq model for training to generate text vectors.In order to further improve the quality and diversity of the generated images,this paper proposes a text generation image model U-Net?GAN based on the improved U-Net network structure.The U-Net?GAN text generation image model first uses Chinese text as input and uses the Word2 vec model to generate word vectors.Input the word vector into the Seq2 Seq model for training to generate the text vector.The U-Net?GAN text generation image model is implemented in two stages.The first stage generates an initial blurred image with lower resolution from the text content;the second stage uses the image generated in the first stage as input and uses the improved U-Net network structure maintains the details of the first stage image,replaces the downsampling process in the confrontation network with the Dense Net network structure,and finally outputs a high-resolution image that expresses text information and has a higher resolution.The experimental results show that the text-generated image model U-Net?GAN has an increase of 0.2% and 0.3% on the Oxford Flower and California Bird Datasets respectively from Stack GAN in the Inception Score,indicating that U-Net?GAN can effectively improve the generated image quality and diversity.