What is Wrong and Right with MSE?

L. Guo and Y. Meng (USA)


Quality assessment, JPEG, color image coding


Mean squared error (MSE), or its equivalence, peak signal-to-noise ratio (PSNR), has been widely used as the quality metric in image compression for long time. Meanwhile, numerous evidences testified the loose correlation between MSE (or PSNR) and the subjective quality assessment result of the same distorted image. If MSE is such an unreliable quality metric, why do encoders based on it work well in most cases? But if it works well in image compression, why can’t people count on it in objective quality assessment? These questions are investigated in this paper under the framework of JPEG baseline compression on color images. We concluded that MSE generally works well in comparing two images compressed by the same encoder, since they normally share the same distortion structure. When comparing compressed images from different encoders or artificially distorted images, MSE will perform very poor, since there are significant discrepancies between the essence of MSE and contrast sensitivities of human visual system.

Important Links:

Go Back