Research On Text-to-Image Generation Technology Based On Attentional Generative Adversarial Networks

With the rise of artificial intelligence technology,the use of deep learning in the field of computer vision has achieved remarkable development and progress.The proposal of generative adversarial networks provides a new idea for the study of text-to-image generation task.This task requires the input of a text description and the output of realistic images that conform to the semantic information of the text.Although this work has achieved encouraging results,most existing methods focus on the text descriptions,and ignore some important fine-grained image information.Such an approach will not only hinder the generation of higher-quality images,but also affect the diversity of generated images.Therefore,this article has launched the following work to further improve the quality of the generated imagesFirst,paying attention to the improvement of the generative model.A brand new image attention mechanism is proposed and applied to the Attentional Generative Adversarial Networks,so that the generative network can spontaneously select important sub-regions in the image and draw more fine details for them.At the same time,when evaluating the matching of fine-grained text and images,the Deep Attentional Multimodal Similarity Model can pay more attention to the matching of important sub regions in the image,thereby providing support for the optimization of the generative network.After verification,this method can not only improve the image quality to a certain extent,but also increase the training speed by about 28.57%Secondly,in order to further optimize the generative network,a new Image Representativeness-Diversity Reward Model is proposed.Implementing this model in the generative network can optimize the more representative areas of the image and the content needs to be improved in the image,so as to generate images with richer details.A large number of experimental results have proved the effectiveness and advanced nature of the method Through testing on two datasets,this method makes the Attentional Generative Adversarial Networks improve the Inception Score of the image by 4.28%and 4.39%,respectively.And the generated image has richer details and diversified contentFinally,paying attention to the improvement of the discriminant model.By giving different levels of attention to the image sub-blocks input to the discriminator,the discriminant network can examine the image to be evaluated more strictly,thereby guiding the generative model to generate high-quality images that meet higher standards.Experimental results show that the scheme can optimize the image generation effect to a certain extent.