Blind image quality assessment with the histogram sequences of high-order local derivative patterns

Automatic assessment of the perceptual quality of digital image is an important and challenging issue in computer vision. Although human visual system (HVS) is sensitive to degradations on spatial structures, most of the existing methods do not take into account the spatial distribution of local structures. This paper reports a novel approach coined high-order local derivative pattern (LDP) based metric (HOLDPM). In particular, HOLDPM extracts local image structures with LDPs in multi-directions to yield an accurate assessment of image quality. HOLDPM is extensively evaluated on three large-scale public databases. Experimental results demonstrate that HOLDPM is able to achieve high assessment accuracy. Besides, objective assessment result of the HOLDPM is consistent with the subjective assessment result of the HVS. Specifically, the experimental results also indicate that HOLDPM outperforms most of the state-of-the-art methods in distortion specific tests. Additionally, HOLDPM shows competitive overall performance when measured with the weighted average of Spearman rank-order correlation coefficient (SROCC) and the weighted average of Pearson linear correlation coefficient (PLCC) over the test databases.