Vol. 49, Issue 4, pp. 527-543 (2019)

Abstract

High dynamic range imaging systems can offer a more complete representation of scene, aiming to capture all brightness information of a visible range of scene, even in extreme lighting conditions. This paper proposes a no-reference quality metric for high dynamic range image (HDRI), in which a combination of tensor decomposition and curvature analysis is used to construct an efficient feature set that is sensitive to degradation levels of patches in HDRIs. Tensor decomposition maintains the majority of color information of an HDRI, and the geometric structure information of the HDRI is then extracted by a curvature analysis. A quality-related label feature matrix is subsequently defined and obtained by using a feature set and a sparse dictionary with quality-related labels. Then, the multi-dimensional local feature set of the HDRI is determined from the quality-related label feature matrix. Finally, the local feature set and other global feature set are pooled to predict the quality of the HDRI. The prediction performance of the proposed metric is verified by three public test databases, and the experimental results indicate that both its Pearson linear correlation coefficient and Spearman rank-order correlation coefficient are better than those of other no-reference metrics. The proposed metric produces statistically better assessment results, implying a higher consistency with human visual perception.