Text-to-image approach combining information blending structure and attention

The semantic consistency, loss of detail features, and other problems are caused by the traditional text generation image method's focus on extracting the information transmitted by the previous layer while ignoring the information of the layer's features that are lost during the subsequent propagation process. Text-to-images combining information blending structure and attention (CAGAN) with information fusion capabilities is suggested in order to address the aforementioned issues. The word and sentence levels of the text's characteristics are encoded using the pre-trained BiLSTM model. The network structure includes a feature fusion module that receives upper-level transmission information while maintaining feature information at the same level. The resulting image's semantic coherence is enhanced, and the image's features are refined, thanks to the application of the affine transformation to map the visual features based on natural language descriptions. According to the experimental findings, the CAGAN's FID score in the CUB dataset is 15.93, and its IS index is 4.84±0.04 in the CUB dataset. It has been greatly improved when compared to the current standard AttGAN and DMGAN models, demonstrating the usefulness of the suggested network.