Different Techniques of Image SR Using Deep Learning: A Review

Image superresolution (SR) is a task to enhance low-resolution (LR) images to high resolution (HR) and is broadly used in applications, such as surveillance, medical diagnosis, and so on. With increasing number of imaging application, SR is better and more respected in efficiency and practical application. Therefore, this article aims to provide a comprehensive study of image SR using a deep convolution neural network (CNN). A brief discussion is presented about different network designs, such as linear networks, residual networks, recursive networks, and attention networks, for the image SR and also compared their performances and complexity. Then, the performance of upscaling techniques and loss functions is also highlighted in this article to observe reconstruction performance. A brief analytical study on standard datasets is also presented, which helps in getting information for less complex network design. Regardless of the advancement as of late, this article identified some shortcomings in existing models and gave future research direction to solve open issues related to image SR.

Automated summary

This article provides a comprehensive study of image superresolution (SR) using a deep convolution neural network (CNN). It explores different network designs and compares their performances and complexity. It also highlights the importance of upscaling techniques and loss functions in reconstruction performance and presents an analysis of standard datasets to inform network design. The article identifies shortcomings in existing models and suggests future research directions for image SR.