Image super resolution using deep learning

Abstract

Image Super-Resolution is the process of enhancing Low-Resolution (LR) images to High Resolution (HR) images, which are widely used in applications such as surveillance, medical diagnosis, and so on. With an increasing number of imaging applications, super-resolution is becoming more efficient and practical. As a result, the goal of this work is to provide a comprehensive study of image super-resolution using a deep convolution neural network. A brief discussion of frame works and different network designs, such as Linear networks, Residual networks, Recursive networks, Gan, and Attention networks, for image super-resolution using deep learning is presented, as well as a comparison of their performances and complexity. This work also highlights the use of various upscaling techniques and loss functions, as well as parametric evaluation, which assists new researchers in designing a new framework. A brief analytical study on standard datasets such as Set5, Set14, BSD100, and Urban100 is also presented, which aids in the gathering of data for less complex network design. Regardless of recent progress, this work identified some flaws in existing models and proposed future research directions to address open issues related to image super resolution. In this work, various deep learning architectures are designed for image multi-scaling using a progressively dilated convolution interleaved correlation filter, which improves image quality at the reconstruction end and is very useful in many applications such as medical images, generalized images, facial images, which are also important in computer vision tasks