No Reference Image Quality Assessment using Deep Learning Techniques

Abstract

Evaluating the quality of images is essential in the development of image processing newlinesystems, especially when considering potential visual degradation. This assessment newlineenables the developers to fine-tune designs, aiming for optimal quality while keeping newlinesystem costs to a minimum. These measures play a pivotal role in various applications, newlineincluding bench-marking and optimizing diverse image processing systems and newlinealgorithms, monitoring and adjusting image quality, and advancing perceptual image newlinecompression and restoration technologies. newlineHistorically, researchers have linked image quality to image fidelity, assessing how newlineclosely a distorted image matches a presumed reference image of perfect quality. This newlineproximity is commonly gauged through the modeling of the human visual system or the newlineapplication of diverse mathematical criteria for assessing signal similarity. newlineObjective measures for image quality have been developed to quantitatively anticipate newlinethe perceived image quality. A significant aspect of objective image quality newlineassessment is No-Reference Image Quality Assessment (NR-IQA) that aims to predict newlineperceived visual quality solely from a distorted image, eliminating the need for a reference (distortion-free) image. No-reference image quality measures are particularly valuable in scenarios where obtaining a reference image is either costly or simply not feasible. The intrinsic complexity and limited understanding of human visual perception present substantial challenges in developing no-reference image quality measures. newlineDespite being an area largely unexplored and not yet mature, the field of NR-IQA is newlineevolving rapidly, offering ample opportunities for creative thinking. newlineThere has been a notable increase in the introduction of novel NR-IQA measures newlinein recent years. The focus of this dissertation lies in the exploration of automated algorithms designed for the assessment of digital image quality. This thesis centers on the advancement of No Reference image quality measures utilizing Deep Convolutional Neural N