Full-reference metrics

About

A full-reference video quality metric measures the impact of visual distortions on perceived quality by representing them in a perceptual domain (a domain in which signals have a representation and a magnitude related to visual perception) and comparing the measured video with its reference video. The reference video is the video which represent the same content as the measured video but which doesn't contain the distortions which must be taken into account.

Typically, to benchmark a video encoder, the reference video is the one located at the input of the encoder and the measured video is the one located at the encoder's output.

Our full-reference video quality metrics are the result of many years of research on human vision modeling and on the application of models to perceived video quality measurement.

And because different codecs produce different distortions, we have developed three different full-reference video quality metrics:

  • a full-reference metric for HEVC (H.265) video: measurement of blur perception, measurement of picture flatness.
  • a full-reference metric for MPEG-4/AVC (H.264) video: measurement of blockiness visibility, measurement of blur perception, measurement of contrast between macroblocs (because of the deblocking filter).
  • a full-reference metric for MPEG-2 video: measurement of blockiness visibility, measurement of blur perception.

And apart from these features, our full-reference metrics also extract visual features that are similar to the ones used by the Human Visual System (HVS).

All the features extracted from the tested video are then compared to the features extracted from the reference video. This comparison produces a quality score expressed on a DMOS (Differential Mean Opinion Score) scale.

A DMOS represents the perceived quality loss of the tested video with respect to the quality of the reference video. That's why the reference video that are used should have very good quality or even an excellent quality.

Related products (evaluations are free):