No-reference image quality assessment using spatial and transform features and SVR computation
| Autor: | Chen, Kai-Wen, 陳凱文 |
|---|---|
| Rok vydání: | 2016 |
| Druh dokumentu: | 學位論文 ; thesis |
| Popis: | 104 Image quality assessment has become an important issue as more and more consumers are using 3C products. While in some cases, they are no reference images, hence it is difficult to predict image qualities using conventional full-reference (FR) or reduced-reference (RR) algorithms. Therefore Non-reference Image quality assessment (NRIQA) has drawn much of attentions these years, which contained three main trends: distortion-specific, natural scene statistic (NSS) based, and training-based algorithms. However, the first type of methods is highly predictable when the distortion types are pre-known, and the second and third type of methods work well as they were combined, hence, an NSS features combined with training based algorithm is proposed in this paper that can accurately predict image quality. In this study, support vector regression (SVR) is adapted to train extracted NSS features, and to pool image qualities using trained model. The primary advantage of combing NSS features with SVR is that NSS features can reveal the properties of natural images, so the distortions of unnatural images can be better measured. Moreover, SVR is common and robust machine learning technique, which is easily tuned and trained. Overall, the proposed approach was tested on some well-known NIQA databases, e.g., LIVE II, TID2008; Reaching high accuracy compared with the state-of-the-arts. |
| Databáze: | Networked Digital Library of Theses & Dissertations |
| Externí odkaz: |
|