Deep Learning-based Image Super-resolution Algorithms – A Survey

Abstract

Image super-resolution (SR) is a technique of enhancing image by increasing spatial resolution of the image. By performing image SR, the pixel intensity in an image is increased. Based number of input images, the SR technique is categorized into two types as single-image SR (SISR) and multi-image SR (MISR). This work analyses different characteristics of SISR on different image datasets like medical and real-world images by using different quality factors such as peak signal to noise ratio (PSNR), structural similarity index (SSIM), and perceptual index (PI). Also, it examines the complexity of SISR schemes based on their computation time. Based on the detailed study it is identified that the deep-leaning-based SR using component learning is a best method in terms of quantitative, qualitative and computation time of generating SR image. Further, it is suggested that to improve the quality a greater number of convolutional network layers can be used in SR algorithms.