LG-IQA: Integration of local and global features for no-reference image quality assessment

Abstract

The visual physiology research shows that the human visual system (HVS) first rapidly and unconsciously produces a global perception, and then gradually focuses on specific local areas for the perception of image quality. Therefore, in order to better simulate the perception of image quality by the HVS, we need to consider both local and global information when evaluating image quality. Based on above analysis, this paper proposes a new method by integrating local and global features for no-reference image quality assessment (IQA), which is named local and global IQA (LG-IQA). Specifically, a test image is transformed by the dual-tree complex wavelet transform (DTCWT) to get the wavelet coefficients, and then the local binary pattern (LBP) operator is applied on the reconstructed image of high-frequency wavelet sub-bands. The obtained results are used as local features. After the DTCWT, we estimate parameters using an asymmetric generalized Gaussian distribution (AGGD) model by fitting the wavelet coefficients, and take the results as global features. Finally, the extracted local and global features are combined together, and the relationship between image features and image quality scores will be learned by an AdaBoosting back-propagation (BP) neural network. It is worth mentioning that this is the first attempt to integrate the LBP and DTCWT together in the IQA research based on neural network. Experimental results demonstrate that the predicted scores of our proposed LG-IQA correlate well with the subjective quality scores.