Perceptual Quality Assessment of Digital Images in the Absence of Reference Information

Abstract

Digital multimedia content experiences a variety of degradations during the process of acquisition, processing, and transmission. Automatic approaches are required to evaluate the performance of any of these processes regarding perceptual image quality assessment. Several human observers can judge an image based on its perceptual quality, and the results may then be averaged to obtain an image quality score, making it the most accurate way of assessing image quality. Because this approach requires the engagement of human resources, an automatic method based on machine learning modelling of the human visual system is necessary. Machine learning techniques may be trained in a supervised manner by providing a set of images along with their subjective quality score as ground truth. This subjective score is sometimes expressed as the Mean Opinion Score, which is the average of numerous human judgments (MOS). In this study, we have trained a deep Convolutional Neural Network (CNN) using labelled images. The study focuses on images that are naturally distorted during acquisition, processing, or transmission rather than on synthetically distorted images with discrete distortion levels. Using naturally distorted images will result in a more general-purpose image quality evaluation model can be obtained. Furthermore, collecting a significant number of labelled images for image quality is an expensive task, and no dataset with a sufficient number of images is available, thus several ways for training the CNN will be investigated. The trained model was evaluated by determining the correlation between the ground truth (MOS) and predicted quality scores, resulting in PLCC and SROCC values of 0.7729 and 0.7489, respectively of 0.7729 & 0.7489, respectively. However, when tested on the BIQ2021 dataset, the PLCC and SROCC values were 0.7845 and 0.7588, respectively.