No reference quality evaluation for screen content images considering texture feature based on sparse representation.

Highlights • Sparse representation produces the better prediction accuracy for SCIs. • A new blind scheme considering the relative gradient direction is designed. • HOG features as second-order derivatives represents the texture information well. Graphical abstract Abstract In this paper, an accurate blind metric evaluating screen content images (SCIs) considering texture feature via sparse representation is proposed. The existing theory on human vision tells us that human visual system (HVS) is sensitive to texture information of images. In addition, gradient direction has not been adequately explored as a predictive information source of SCIs. This method is first based on three kinds of gradient maps of SCIs—i.e., gradient magnitude map, relative gradient direction map, relative gradient magnitude map. Since the existing literatures about human neuroscience have revealed that image texture can be captured by high-order derivatives, texture information of SCIs can be represented by local histogram of oriented gradient (HOG) features extracted from the aforementioned gradient maps. Finally, it is composed of two important procedures: representing HOG features by means of sparse coding, then weighting subjective experimental quality scores via the sparse coding coefficients to predict objective perceptive quality scores of SCIs. Comparison experiments on the public SCI database demonstrate that the designed metric has the better superiority over existing the-state-of-art algorithms and delivers consistency and accuracy in relation with subjective evaluation. [ABSTRACT FROM AUTHOR]