Deep Learning for Image Upscaling: Exploring the Potential of ESRGAN

One useful tool in the field of image super-resolution learning techniques is the Enhanced Super-Resolution Generative Adversarial Network (ESRGAN). This model is solely intended to convert low-resolution images into its stunning high counterparts, relying heavily on the power of GAN-based architecture.The paper showcases the successful implementation and training of ESRGAN equipped with a specialized architecture. The model boasts 21 Residual in Residual Dense Blocks (RRDB) and harnesses 36 growth channels, endowing it with an extraordinary capacity for image resolution enhancement. This training procedure supported by Set14, while validation is performed on Set5. The resultant model is rigorously assessed using a bespoke dataset comprising a diverse array of images sourced from OST300, Urban100, BSD100, and real-life scenarios.