Internal Generative Mechanism Driven Blind Quality Index for Deblocked Images
Bo Hu, Leida Li, Jiansheng Qian · 2018
Image deblocking has been widely studied. However, the relevant quality evaluation of deblocked images remains an open problem. The deblocked images are usually contaminated by multiple distortions, typically blocking artifacts and blur. Although various quality metrics have been reported, they are not designed specially for deblocked images, so they cannot accurately predict the quality of deblocked images. To fill this gap, we propose a new quality metric for deblocked images. With the guidance of the internal generative mechanism (IG- M) theory, a deblocked image is first decomposed into two portions, i.e., the predicted and disorderly portions. Then the distortions in the predicted portion are evaluated. Specifically, the distortion-specific features are extracted to evaluate blocking artifacts and blur in the spatial domain, separately. The joint effect of blocking artifacts and blur is evaluated by extracting energy-based features in the Curvelet domain. Finally, all features are combined to train a random forest model for quality prediction of deblocked images. Experimental results conducted on a newly released DeBlocked Image Database (DBID) demonstrate that the proposed metric outperforms the existing relevant quality metrics.