IMAGE SUPER-RESOLUTION USING SUPER-RESOLUTION CONVOLUTIONAL NEURAL NETWORKS

Abstract

The primary goal of Super-Resolution (SR) is to create a higher resolution image by enhancing lower resolution images. This is crucial because high-resolution images contain more pixels, providing finer details of the original scene. The demand for high resolution is widespread in computer vision applications, as it improves pattern recognition and image analysis performance. In fields like medical imaging, high resolution is essential for accurate diagnosis. Moreover, various applications such as surveillance, forensics, and satellite imaging require high-resolution capabilities for zooming into specific areas of interest. However, obtaining high-resolution images can be challenging and costly due to limitations in sensor and optics manufacturing technology. To address these issues, Super-Resolution Convolutional Neural Networks (SRCNN) have emerged as a cost-effective solution. SRCNN enables the transformation of low-resolution images to high-resolution ones, allowing the utilization of existing low-resolution imaging systems. As implied by its name, SRCNN is a deep convolutional neural network that learns to map low-resolution images to high-resolution ones in an end-to-end manner. This approach significantly improves the quality of low-resolution images. Unlike traditional methods that handle different components separately, SRCNN optimizes all layers together. The performance of the network is assessed using image quality metrics such as peak signal-to-noise ratio (PSNR) and mean squared error (MSE). Additionally, the Open Source Computer Vision Library (OpenCV) is utilized for pre and post-processing of the images