Reduced-Reference Image Quality Assessment based on saliency region extraction

In this paper, a novel saliency theory based RR-IQA metric is introduced. As the human visual system is sensitive to the salient region, evaluating the image quality based on the salient region could increase the accuracy of the algorithm. In order to extract the salient regions, we use blob decomposition (BD) tool as a texture component descriptor. A new method for blob decomposition is proposed, which extracts blobs not only in different scales but also in different orientations. Different blob components consist of location of blobs, blob shape and color attributes are used to describe texture of the image accordance to the human visual system conception. A region Covariance matrix is calculated from extracted blob components which can be easily interpreted in terms of its eigenvalues. Therefore, the reference image is described as a squared covariance matrix and a good data reduction is achieved. The same process is used for describing the received image in the destination. Finally, the image quality is estimated by using the eigenvalues of two covariance matrices. The performance of the proposed metric is evaluated on different databases. Experimental results indicate that the proposed method performs in accordance with the human visual perception and uses few reference data (maximum 90 values).