Candidate: Shahrukh Athar
Title: Quality Assessment of Images undergoing Multiple Distortion Stages
Date: November 19, 2018
Time: 3:00 PM
Place: EIT 3142
Supervisor(s): Wang, Zhou
Abstract:
Contemporary image and video quality assessment algorithms suffer from a number of limitations when evaluating the quality of content that has been distorted by multiple distortions. These include the reliance of full-reference and reduced-reference quality assessment algorithms on pristine reference content which is mostly not available in the real world, inadequate performance of no-reference quality assessment algorithms, and the inability of all three categories to take into consideration the availability of content at various points in the content delivery chain. The purpose of this work is to address these challenges by proposing a new paradigm called degraded-reference image quality assessment and making the following three contributions. First, we conduct an extensive review of the field of image quality assessment, to ascertain whether it is possible to generate an acceptable synthetic ground truth for large datasets where subjective testing is not an option. Second, we build three datasets that are composed of images that have undergone up to two stages of distortions, the largest of which is composed of around 3.45 million images. These datasets contain practical distortion combinations and images are annotated with ground truth that has been synthesized through the reciprocal rank fusion method. Third, we design degraded-reference quality models that use data, such as full-reference and no-reference quality scores from various points of the content delivery chain or natural scene statistics of images, to perform the task of quality assessment of images undergoing multiple distortion stages.