Hybrid NSS features for no-reference image quality assessment

A novel general-purpose no-reference image quality assessment (NR-IQA) model utilising hybrid natural scene statistics (HNSS) is proposed. Distinguished from existing NR-IQA approaches, the new model combines the statistics of locally mean subtracted and contrast normalised coefficients in the spatial domain and the statistics of image patch coefficients in a codebook space, which is constructed by codebooks extracted from pristine images using K-Means. The authors demonstrate that the coefficients in the codebook space keep the NSS characteristics as same as these in the spatial domain. After extracting the statistical features, a two-stage framework of distortion classification followed by quality assessment is applied. Experimental results show that the authors' predicted quality score well matches human perceptual image quality. The proposed model outperforms state-of-the-art general-purpose NR-IQA approaches when it is tested on the LIVE database.