Qpro: An improved no-reference image content metric using locally adapted SVD
Xiang Zhu, Peyman Milanfar · 2013
We present an improvement to our earlier Q metric, 1 a no-reference image content measure based on singular value decomposition (SVD) of local image gradient matrix. The new extension, Qpro, is capable of better measuring the amount of latent image content in the presence of both blur and random noise. As desired, its value drops monotonically when image becomes either blurry or noisy. Compared with our earlier metric Q which was computed using only anisotropic patches, Qpro implements SVD in transformed coordinates which are adapted to local estimated structure. In this way, it can measure a much wider variety of local image content, leading to significantly improved performance. Experiments demonstrate that this metric correlates with subjective quality evaluations even better than some full-reference quality metrics. It also outperforms other metrics in optimizing tuning parameters for image denoising filters.