The Perceptual Image Distortion model doesn't appear to be particularly useful. From what I gather, they developed a model which closely matches empirical data gathered by other researchers, but did not actually conduct any human perceptual studies on images to determine the real-world usefulness of this model. Nor is there really any rating scale which could be applied to describe an image's perceived quality, unless you just want to throw around their calculated numbers. This model doesn't seem very promising, and they do state in their conclusion that future work needs to be done to correlate their model with some perceived quality.

The JNDmetrix, however, conducted real-person studies with trained image analysts and other people to compare how well JND values correlate to perceived quality. And also how badly root-mean-error fairs in the same tests. In short, the JNDmetrix VQM works extremely well at providing easy to understand ratings of an image's quality. They've won an Emmy for their work and the JNDmetrix is in use by the broadcast industry. Unfortunately, their description of how the JND values are computed is described without hard numbers of formulas. So I'll either need to contact them to see if they are willing to reveal their algorithms, or implement my own interpretation of what is discussed in their paper. Or purchase their JNDmetrix products. I'm not very optimistic about them giving me the algorithms, since they sell JNDmetrix through their products. At least there was no mention of a patent on this technology.

The third VQM is the one documented in NTIA Report 02-392. This paper is extremely detailed and provides all of the information you need to implement and use their model, including suggested parameter values, calculation descriptions, and equations specific to the type of video sequence involved. Their model also applies specifically to video sequences, and not individual images. All the equations appear to be linear, which is a good thing. It is specifically stated that this information is available for anyone to use anywhere for whatever. Like with the JNDmetrix, this model also has a specific number which can be used to compare the quality of two video sequences. Right now, this VQM looks like the best choice.

None of these models specifically tie any sort of network characteristics to the quality calculation, or deal with what type of artifacts were detected in the reconstructed frames. So, I'm guessing we'll have to go with actual calculations on sample video streams in a simulated network.

Post a comment