Perceptual Distorsion Metrics for JPEG Images EE392 Project

JPEG is a standard compression algorithm used to reduce memory requirement for the storage of digital images. The JPEG standard allows to specify the desired quality of the encoded image by varying a quality factor (QFactor) between 0 (lowest quality) and 100 (perfect image-no compression). It is of interest to be able to assess the quality of a JPEG image. A common and very effective way for that is to use the mean square error (MSE) metric to compute the distortion caused by encoding. This metric has proven very powerful and yields very good results. Several questions can however be asked regarding the MSE metric:-why use a squaring criteria-can it be improved by taking into account properties of the human visual system (HVS)?-are there better or simpler metrics for JPEG images? The first possible modification to the MSE metric is to consider a more general one based, called Minkowski summation (or Holder norm): ([2],[3],[7]) For p=2, one simply obtains the MSE metric. These attempts showed that a value of 2 (hence the MSE) is often a very good choice. It should be noted that a high value for p (like 5) tends to take into account only the largest errors and discard the smaller ones. Values beyond 10 seem not to be used at all. The second possible modification to the MSE metric is to incorporate some known characteristics of the visual system ([7]). Numerous attempt have been made to include additional factors like:-visual masking effects, both spatial ([1], [2]) and temporal-the spectral response of the human system ([2],[3],[4],[6]) Minkowski 1 N-y n [ ] x n [ ] – p n 1 = N ∑       1 p / =