Publication

Abstract : Convolutional neural networks have achieved remarkable progress in enhancing the resolution of single images with high accuracy and speed. However, the primary challenge is to recover finer texture details when super-resolving the images. Optimization-based super-resolution methods have aimed to minimize MSRE, which often sacrifices perceptual quality and high-frequency characteristics in favor of high PSNR. This paper presents a comparative analysis of two prominent SISR techniques, SRGAN and ESRGAN, focusing on their performance metrics and visual quality. SRGAN employs a GAN architecture, which uses a perceptual loss function to generate photorealistic 4x upscaled images of natural images with improved visual appeal. However, SRGAN might introduce artifacts like patterns around high-frequency details, impacting image quality. To further enhance visual quality, ESRGAN, an enhanced variant of SRGAN, is employed, which introduces improvements in its architecture and loss function to overcome the limitations of SRGAN. ESRGAN is renowned for its ability to produce higher-quality visual output than SRGAN, with more realistic and natural textures. The quality of images generated by both SRGAN and ESRGAN is assessed using quantitative metrics, including Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM).