Utilizing Binary Segmentation to Enhance Object of Interest Details in Images Synthesized Using Generative Adversarial Networks

Image synthesis, once considered a distant fantasy, has emerged as a rapidly evolving field, driven by advancements in generative models and computational power. Many researchers in academia and corporations have tackled this problem and achieved impressive results. Among the widely adopted models in this domain is the Generative Adversarial Network (GAN). In this work, we integrate Binary Segmentation with GANs to enhance the quality of the object of interest in generated images. The binary segmentation model guides the generative model to focus on the region of interest within the image rather than the background producing more detailed representations of the target object. Our approach was trained and validated using labeled samples from the well-known CUB dataset. For testing, we generated images corresponding to text descriptions from the dataset. Evaluation of these images demonstrated the effectiveness of our model, achieving a comparable Inception Score of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$5.243\pm 0.090$</tex> with significantly fewer training epochs (200 epochs).