SRGAN

SRGAN, short for Super-Resolution Generative Adversarial Network, is a deep learning model developed for single-image super-resolution. Introduced by Ledig and colleagues in 2017, SRGAN aims to convert a low-resolution image into a higher-resolution version that preserves structure while exhibiting realistic texture details, addressing limitations of traditional pixel-wise loss methods.

At its core, SRGAN comprises two networks trained in opposition: a generator that upsamples a low-resolution input to a high-resolution image, and a discriminator that attempts to distinguish real high-resolution images from generated ones. The generator is built from residual blocks and employs a upsampling mechanism based on sub-pixel convolution (PixelShuffle) to increase resolution efficiently. The discriminator acts as a binary classifier, guiding the generator toward more natural textures.

The training objective combines a perceptual content loss with an adversarial loss. The content loss measures differences in high-level feature representations extracted from a pre-trained convolutional network such as VGG, while the adversarial loss encourages the generator to produce photo-realistic textures that fool the discriminator. This combination improves perceptual quality, particularly for 4x upscaling tasks, compared with traditional mean-squared-error-based methods.

SRGAN has influenced subsequent work in perceptual super-resolution and popularized the use of GANs for this task. It highlighted the trade-off between perceptual realism and traditional fidelity metrics like PSNR, prompting further research into perceptual losses, robust upsampling, and higher-fidelity texture synthesis.