Deep Learning-Based Single-Image Super-Resolution: A Comprehensive Review

High-fidelity information, such as 4K quality videos and photographs, is increasing high-speed internet access becomes more widespread less expensive. Even though camera sensors’ performance constantly improving, artificially enhanced photos created by intelligent image processing algorithms have significantly improved fidelity in recent years. Single super-resolution a class of that produces high-resolution from given low-resolution image. Since the advent deep learning decade ago, this field has made significant strides. This paper presents comprehensive review assisted single domain including generative adversarial network (GAN) models discusses prominent architectures, used, their merits demerits. The reason behind covering GAN it been known to perform better than conventional methods resources time. For real-world applications with noise other issues can cause low-fidelity super resolution (SR) images, we examine another solution based on model. model-based technique popularly blind resilient. We examined various techniques varying scaling factors (i.e., 2x, 3x, 4x) measure PSNR SSIM metrics for different datasets. across datasets covered experimental section shows an average 14-17 % decrease score move up scale 2x 4x. observed all every model mentioned paper. results also show outperforms complex models. are preferred when upscale factor high, while residual dense recommended smaller upscaling factors. super-resolution, finally, concluded challenges future directions.