Assessment of full color image quality with singular value decomposition
Aleksandr Shnayderman, Ahmet M. Eskicioğlu · Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE · 2005
In subjective evaluation of distorted images, human observers usually consider the type of distortion, the amount of distortion, and the distribution of error. We recently proposed an image quality measure, M-SVD, for gray-scale images that can be used as a graphical tool to predict the distortion based on these three factors, and also as a numerical tool to assess the overall visual quality of the distorted image. It performs better than two state-of-the-art metrics, Q and MSSIM, especially when we compute the correlation with mean opinion score across different types of noise. The test image was degraded using six types of noise (JPEG, JPEG 2000, Gaussian blur, Gaussian noise, sharpening and DC-shifting), each with five different levels of intensity. In this paper, we extend M-SVD to full color images using a color model which decouples the color and gray-scale information in an image. Our experiments show that using only the luminance component, the measure outperforms Q and MSSIM. When we also use the two chrominance layers, the performance of M-SVD becomes slightly higher whereas the performance of Q and MSSIM is degraded. This indicates that the color components may also contribute to the performance of the proposed measure.