Digital Signal Processing Reference
In-Depth Information
A
B
C
P
P = ½
×
C + ½
×
B
P = C - A + B
P = ¾
×
C - ½
×
A + ¾
×
B
Figure 11.2. Three examples of video predictors.
The usefulness of these predictors becomes apparent when used
in conjunction with differential encoding.
11.5 Differential Encoding
Suppose we take the actual pixel value and subtract the pre-
dicted pixel value. This is differential encoding, and this differ-
ence is used to represent the pixel value for that location. For
example, to store a video frame we would perform the differential
encoding process, and store the results.
To restore the original video data for display, we would simply
need to compute the predicted pixel value (using the previous
pixel data) and then add to the stored differential encoded value.
This is the original pixel data. (The three pixels on the upper left
corner of the frame are stored in their original representation, and
are used to create the initial predicted pixel data during the
restoration process).
All of this sounds unnecessarily complicated, so what is the
purpose here? The differential encoder outputs are most likely to
be much smaller values than the original pixel data, due to the
correlation between nearby pixels. Since statistically the differ-
entially encoded data is just representing the errors of the
predictor, this signal is likely to have the great majority of values
concentrated around zero, or to have a very compact histogram.
In contrast, the original pixel values are likely to span the entire
color value space, and therefore are a high entropy data source
with an equal or uniform probability distribution.
Differential encoding sends less information than if the video
frame was simply sent pixel by pixel across the whole frame. The
entropy of the data stream has been reduced through use of
differential encoding, as the correlation between adjacent pixels
has been largely eliminated. Without differential encoding, we
have no information as to the type of video frames that will be
processed, the initial pixel value outcomes are assumed to be
