Quality assessment of images with multiple distortions based on phase congruency and gradient magnitude

Highlights：

• Given that PC is contrast invariant while the contrast does affect perceptual image quality of the HVS, image gradient magnitude (GM) is employed as complementary features for PC to represent the contrast information and used as a weighting factor for LBP histogram creation.

• The support vector regression model is trained to map the gradient-weighted LBP histograms in PC domain at multi-scale to quality scores.

• Combining gradient magnitude (GM) and LBP in phase congruency (PC) domain at multiple scales to design the NR IQA method for multiply distorted images is the first work in NR IQA field. The proposed method achieves high consistency with subjective perception and outperforms other state-of-the-art full-reference (FF) and NR IQA methods.

摘要

•Image distortions not only destroy the intensity of low-level image features, but also alter their distribution. Based on these observations, the proposed method employs phase congruency (PC) to represent low-level image features, and extracts local binary pattern (LBP) in PC domain at multiple scales to represent the distribution of low-level features.•Given that PC is contrast invariant while the contrast does affect perceptual image quality of the HVS, image gradient magnitude (GM) is employed as complementary features for PC to represent the contrast information and used as a weighting factor for LBP histogram creation.•The support vector regression model is trained to map the gradient-weighted LBP histograms in PC domain at multi-scale to quality scores.•Combining gradient magnitude (GM) and LBP in phase congruency (PC) domain at multiple scales to design the NR IQA method for multiply distorted images is the first work in NR IQA field. The proposed method achieves high consistency with subjective perception and outperforms other state-of-the-art full-reference (FF) and NR IQA methods.