Usage of Generative Adversarial Network to Improve Text to Image Synthesis

Nowadays generating high-quality image from text description is one of most challenging problem in computer vision. Generating images from text description has a wide range of applications, like computer-aided design. In this chapter, a new model called Knowledge Transfer Generative Adversarial Networks (KT GAN) is proposed for an exact synthesis of given image. To enhance the quality of generated image, an Alternate Attention Transfer Mechanism (AATM) and Semantic Distillation Mechanism (SDM) are introduced to help the generator better to traverse cross-domain area that exists between text and image. There are two main goals for using AATM: first, to make images more visually appealing by highlighting relevant words and second, to make images more visually appealing by highlighting essential sub-regions of an image. Use an image encoder that has been taught in image-image for training the text encoder for image-image. Improved text characteristics and photos of higher quality is possible with this method. The suggested KT GAN outperforms the traditional approach significantly and provides convincing results over a wide range of assessment measures after precise testing on two public datasets.