A Deep Learning Based Cross Model Text to Image Generation using DC-GAN
Gayathiri Kasi, S. Abirami, R. Dhana Lakshmi · 2023
In recent times, Generative Adversarial Networks have successfully synthesized images through text descriptions. In the domain of image processing, deep convolutional generative adversarial networks (DCGANs) have recently demonstrated promising results. In this technique, two major approaches are used. The first approach in this method utilizes a text transformer using XLNet to extract semantically significant text descriptions using a stronger text encoder. The better text encoder will be allowing the generator to gather more insightful text information with the goal of generating realistic and text-matching visuals. The second method uses the Deep Convolutional Block to efficiently merge text and picture information. It is known as the Deep Convolutional Text to Image Generative Adversarial Network (DCGAN). We experiment with the DCGAN with OXFORD 102 flower dataset to generate the relevant visuals according to the text descriptions. The quality of the images gained from DCGAN has improved with the OXFORD 102 Flower dataset of inception score of 4.580 in order to show the DCGAN network how deeply varies from other networks. The intensive experiments also provide a PSNR value with the Deep Convolutional Layer of 3.983 and an SSIM value of 0.508 and the stronger text encoder improves text-to-image synthesis performance. The proposed method is successfully boosted the inception score by 9.39% by depicting XLNet as text encoder.