Text generated images based on dynamic memory attention

This article proposes a Dynamic Memory Attention Generative Adversarial Network (DMAttn-GAN) to address the issues of insufficient text image semantic alignment and weak ability to generate complex scene details in text to image generation tasks. This model is based on the AttnGAN framework and enhances the generation quality through the following core features: (1) dynamic memory attention mechanism, introducing a multi-stage dynamic attention module, dynamically adjusting word weights by combining text features and local image features, solving the problem of equal importance of words in traditional methods, and significantly enhancing the ability to generate key details. (2) Feature alignment enhanced discriminator, by introducing highly expressive features extracted by Inception-v3 network, achieves precise alignment of generated image features, thereby enhancing the discriminator’s feature discrimination ability; In addition, a multi-stage independent discrimination mechanism is designed to further enhance the training stability and convergence efficiency of the generative adversarial network. Our experiment validated the performance of the model on the CUB-200-2011 single scene dataset and the MS-COCO complex scene dataset. DMAttn-GAN outperforms the AttnGAN baseline model in metrics such as Inception Score, FID, and R-Precision. This study provides an effective solution for high-resolution and semantically consistent text to image generation tasks, with potential applications in virtual scene construction, art design, and other fields.