High-Resolution Realistic Image Synthesis From Text Description Using Iteratively Generative Adversarial Network

Generative Adversarial Network(GAN)is the combination of two deep neural networks namely Generator(G)and a Discriminator(D),which are trained in a competitive way by pitting one against the other such as "G" generates new data,while "D" authenticate the data(i.e.,the generated data is real or fake).GAN is capable of synthesize or produce reasonable and realistic images,sketches,and videos from a text description,due to the power of the deep network and competitive training manner of G&D.Additionally GAN have shown a great capability in image,sketch,and video synthesis or generation applications.In this work,we proposed a novel iteratively Generative Adversarial Network(iGAN)that allows synthesizing high-resolution realistic images from their text descriptions.Synthe-sizing or generating high-resolution realistic images from their text descriptions in one stage mostly the blurry artifacts and mode collapse problems can occur.To mitigate these problems,we divide our model into three stages instead of one stage.In 1st stage GAN synthesized a low-resolution 64 x 64 pixels basic shape image from the text description with less mode collapse and blurry artifacts problems.In 2nd stage,GAN takes the result of 1st stage and text description again and synthesizing a better resolution 128 x 128 pixels more clear image with very less mode collapse and blurry artifacts problems.In the last stage,GAN takes the result of 2nd stage and text description as well and synthesized a high-resolution 256 × 256 well shape and clear image with almost no mode collapse and blurry artifacts problems.Our proposed iGAN shows a very good result on CUB birds and Oxford-102 flowers datasets.Moreover,iGAN shows a good inception score and human rank as compare to the other state-of-the-art methods.Additionally,we are applying our iGAN on MS COCO dataset to solve more complex problems of text-to-image synthesis.