Text-to-Image algorithm Based on Fusion Mechanism
Jinzhi Deng, Yan Wei, Jiangtao Liu · 2022
A multi-step text-generated image algorithm based on feature fusion is proposed to solve the problems of image blurring and missing image details in text-generated images. To enhance the ability of single-channel text features to guide multi-channel image features, the migration of text features to image feature is implemented using a feature fusion module that enhances text features while refining the details of the generated images. The fusion module is executed alternately with the upsampling operation in the generator to enhance the frequency of text feature usage. Placing the generator and discriminator is three pairs achieves the goal of generating fuzzy images from text to clear large images from fuzzy images.The experimental data of the model on the CUB dataset show an improvement in the Inception Score as well as the Frechet Inception Distance score, and the comparative analysis of the generated images shows that the images generated by this method have rich detail texture and sharpness.