Perceptual Single Image Super Resolution Using Residual Generative Adversarial Networks

Single-image super-resolution (SISR) seeks to reconstruct a high-resolution (HR) image from a single low-resolution (LR) observation, a task that becomes especially challenging at large upscaling factors where textures and fine structures are easily lost.Recent deep CNN methods largely optimize pixel-wise losses such as mean squared error (MSE), which correlates with peak signal-to-noise ratio (PSNR) but often suppresses high-frequency details and yields overly smooth images.Motivated by this limitation, we propose a residual generative adversarial network (ResGAN-SR) that combines deep residual learning with adversarial and perceptual losses for 4× SISR.The generator adopts a ResNet-style architecture with stacked residual blocks and sub-pixel up-sampling, while the discriminator learns to distinguish super-resolved images from real HR counterparts.A composite perceptual loss, formed from VGG-based content features and an adversarial term, guides the generator toward solutions that are both structurally faithful and visually realistic.Experiments on the Div2K dataset demonstrate that ResGAN-SR improves perceptual quality in terms of structural similarity (SSIM), mean opinion score (MOS), and visual sharpness, while remaining competitive in PSNR and MSE compared with a purely residual CNN baseline.These results indicate that explicitly modelling perceptual cues in a GAN-based residual framework is an effective strategy for photorealistic image upscaling in applications such as surveillance, medical imaging, and remote sensing.