Analysis of Super-Resolution Techniques: From Traditional Interpolation to FSRCNN

Abstract

We use single image super-resolution (SISR) techniques to reconstruct high-resolution images from low-resolution images using deep learning algorithms to display frames in high resolution displays like laptops and television screens. Modern upsampling methods like NVIDIA's Deep Learning Super Sampling (DLSS) show the power of deep learning-based upscaling in games where computing frames are very expensive as the resolution increases. In this study, we implemented and evaluated an upscaling model called Fast Super-Resolution Convolutional Neural Network (FSRCNN) as a lightweight and efficient alternative for real-time image upscaling for images. We trained a model on the DIV2K dataset and it achieved a Peak Signal-to-Noise Ratio (PSNR) of 31.96 dB with high structural similarity. The results show that FSRCNN improves image quality over traditional upscaling methods. The model's lightweight architecture and computational requirements shows its potential for deployment in super-resolution systems.