Enhanced Image Super Resolution Using ResNet Generative Adversarial Networks

Significant advancements in SISR have been achieved through the use of deeper CNNs, enhancing both speed and accuracy. However, a crucial challenge persists in restoring finer texturing details at higher up-scaling factors. Recent research efforts have focused on lowering Mean Square error of reconstruction to achieve high PSNR. However, these methods frequently fail to capture the high-frequency details necessary for preserving fidelity at higher resolutions. This paper introduces ResNet GAN, a GAN customized with residual learning for enhanced super resolution. Specifically, it excels in generating realistic images at a 4x upscaling factor. Notably, proposed perceptual loss function, encompassing both adversarial and content losses. A trained discriminator is employed to differentiate super-resolved and actual photos based on the computed adversarial loss. In contrast to traditional pixel space resemblance, the content loss relies on perceptual similarity. The results demonstrate that ResNet GAN with the proposed perceptual loss function outperforms Deep Residual Learning on Div2k. The framework exhibits superior metrics such as PSNR, SSIM, MOS, and MSE. By prioritizing perceptual details over pixel space on highly down-sampled images, the proposed approach successfully recovers photorealistic features, addressing previous methods limitations. This advancement holds promising implications for applications requiring high-resolution image reconstruction.