Research On Text-generated Image Algorithm Based On SANet’s RATGAN Model

Text generation image task has become a very hot direction in the field of image generation.The task of text generated images is mainly to generate natural and realistic and diverse images through the given simple text description,and the main difficulty of this task is how to effectively integrate text information into high-quality images.As the most popular text-generated image model,generative adversarial network has become the basic research method of text-generated image.At present,there have been many related studies in the field of text-generated images based on generative adversarial networks,but there are still many challenges,such as insufficient clarity of the generated images,poor diversity of images,unstable network training,and poor authenticity and diversity of generated image details.Therefore,how to effectively optimize the generative adversarial network so that it can generate higher quality images that conform to the given text has become an urgent problem to be solved.Aiming at the problems of difficult selection learning rate,overfitting and high risk in text-image synthesis task Chinese this information and generated images,this paper proposes and constructs the AdamW-SA-RATTAN model,which integrates the AdamW optimizer and SA spatial attention to improve and optimize the RATGAN network.(1)Aiming at the problems of difficult selection learning rate,overfitting and high risk in text-image synthesis task Chinese this information and generated images,this paper proposes and constructs the AdamW-SA-RATTAN model,which integrates the AdamW optimizer and SA spatial attention to improve and optimize the RATGAN network.(2)In view of the fact that the convergence in the Adam optimization algorithm in the RATGAN model cannot be effectively guaranteed,and there is a problem of non-convergence or too slow convergence,this paper proposes to replace the Adam optimization algorithm in the RATGAN model with the AdamW optimization algorithm,and continuously update the weights by adding the weight attenuation factor w when the parameters are updated,and the hyperparameters of the AdamW optimizer are optimized,so that the training and test sets are reduced and better generalization performance is produced.This makes the model have better accuracy.(3)Aiming at the problem of image diversity and clarity in text-generated images,this paper proposes a RATGAN model based on SANet fusion optimization,which adds SA spatial attention model to the image generation module in the image generation network,so that the text description is aware of the matching image region,supervises the generator to synthesize the relevant diversity image content,and the image generated by the fusion optimization model is compared with the original RATGAN model.A large number of experiments on the CUB dataset and the Oxford-102 dataset show that compared with the original RATGAN model,the IS index is improved by 0.05,the FID index is reduced by 2.56,the overall visual effect is more detailed and realistic,and the generated pictures are more diverse.