Text-to-Image Generation Method Based on Attention and Dual-Residual Module

The goal of text-based image generation is to create premium and authentic images conditioned on specific textual descriptions. The main challenges in this assignment include improving image quality and achieving a deep integration between the image and textual information. To address these issues, this paper presents a novel Generative Adversarial Network (ADR-GAN) model, which is optimized through the incorporation of attention mechanisms and dual residual modules. Specifically, ADR-GAN employs two innovative word-level attention modules-channel attention and spatial attention modules - to model word information from different perspectives, thereby significantly enhancing semantic consistency. To accelerate convergence and enhance image detail vividness, a dual residual module is developed, which supports deeper network architectures while preserving more original visual features. Furthermore, a word-level discriminator is introduced to furnish the generator with detailed feedback on each individual word. Compared to previous methods, ADR-GAN excels in generating realistic and text-aligned images, achieving a commendable balance between quality and semantic consistency, and demonstrates superior performance in benchmark tests.