MSRR-T2I: Multi-Scale Reverse Residual for Text-to-Image

Abstract Text-generated image is a very challenging task in the artificial intelligence and machine learning. To address the problems of low quality and no-diversity in images generated by conventional text-generated approaches, a multi-scale reverse residual text-generated image network (MSRR-T2I) is proposed in this paper. Firstly, the multi-scale residuals inverse structure is embedded into the fine image generation module to resolve the issue of loss of feature information when original residuals are reduced. Then, the information content of the feature map is extended, and the features of different resolutions are captured by multi-scale convolution. Finally, channel transformation technology is introduced to achieve cross-group information fusion and enhance image diversity. The test results of common CUB and COCO data sets indicate that compared with the original model ATTNGAN, the IS of MSRR-T2I increases by 8.7% and 8.4%, and the FID increases by 24.8% and 17.0%, respectively. Therefore, the proposed method is confirmed to be effective.