Super-Resolution Image Restoration Using Convolutional Neural Network

Abstract

The main goal of the super resolution method is to create a higher resolution image from a lower resolution image. High-resolution images provide a high pixel density, hence more detail in the original image. The need for high resolution is widespread in computer vision techniques, pattern recognition applications, or general image analysis. However, high-resolution images are not always available. This is due to the fact that the conversion processes and processing methods require ultra-powerful processes, and the equipment for obtaining high-resolution images is expensive. These problems can be overcome by using image processing algorithms that are relatively inexpensive, which has led to the concept of super-resolution. This has the advantage that it can cost less and existing low-resolution imaging systems are readily available. High resolution is essential in medical imaging for diagnosis. Many applications require zooming into a specific image area, where high resolution becomes essential, such as surveillance, forensics, and satellite imaging. The method is presented in this paper, using a convolutional neural network to reproduce super-resolution images, directly performs the conversion from a low-resolution image to an image similar to the original. To speed up the output time, the proposed method performs most computational operations in low-resolution space, while reducing the sampling does not lead to information loss. The main task of the neural network is to reconstruct the distorted image and find the ideal reconstruction function, according to which, in fact, a neural network of a simple structure creates high-quality images with better performance, such as resolution, signal-to-noise ratio, with less time spent on image restoration. During the experiment, we determined an algorithm by which the proposed neural network can reconstruct any image with different types of distortion. The super-resolution method is implemented using the python 3.6 programming language and the tensorflow and tensorlayer software modules for convolutional neural networks. Graphical data of signal-to-noise ratio, structural similarity, and loss plots are obtained using the tensorboardX module.