Game Development Reference
In-Depth Information
The differential signal is also called residual signal and at the receiver side the original
signal can be reconstructed by adding the residual and the prediction. Compared to
the original signal, the residual signal has lower correlation. Prediction coding is an
efficient tool to reduce the spatial, temporal, and set redundancy existing within or
among the video signal(s), and many prediction coding tools have been proposed.
The earliest using of prediction coding is pixel-based DPCM (differential pulse
coding modulation) coding, where the difference between neighboring two pixels
are quantized and transmitted (Cutler
1950
). For video prediction, Harrison (
1952
)
proposed the first representative intra-prediction method, which takes the linear com-
bination of reconstructed pixels as the prediction of the current pixel. The modified
algorithm, named LOCO-I algorithm (Weinberger et al.
2000
), has been adopted in
JPEC-LS image compression standard. Afterwards, the AC/DC intra prediction in
transform domain (Grgi´cetal.
1997
) and the directional intra prediction in spatial
domain (Bjontegaard
1998
) have been proposed, and the latter becomes the prevalent
prediction method in video coding field. Many popular video coding standards adopt
the directional intra-prediction, e.g., AVC/H.264, HEVC/H.265 and AVS.
Later, the unit of prediction is extended from pixel to picture. In Seyler (
1962
),
a picture-based difference coding method was proposed, where only the difference
between two pictures is transmitted, and the data redundancy was reduced signifi-
cantly. Rocca first proposed block-basedmotion estimation inRocca (
1969
), inwhich
an arbitrary-shaped block-based motion-compensated technique was proposed. The
basic idea of Rocca's method is to model the scene as a set of constant-brightness
zones denoted with arbitrary-shaped block. These zones would move from one
frame to the next tracked by motion vectors, and difference values were transmitted
for picture reconstruction. Besides these methods, motion-compensating prediction
was further improved by employing the long-term statistical dependencies in coded
video sequence instead of only the immediately preceding frame used for prediction.
Wiegand et al. (
1997
) proposed a long-term memory scheme that used up to 50
previously decoded frames to determine the best motion vector. In addition, Puri
et al. (
1990
) first proposed the B picture concept, which interpolates any skipped
frame taking into account the movement between the two “end” frames, i.e., the
forward and backward frames. It can achieve higher compression ratio by more
effectively exploiting the correlation between reference pictures and current B pic-
ture, especially for copping with occlusion, uncovering problem caused by zooming,
nonlinear motion, and so on. The B picture is further generalized by linearly combin-
ing motion-compensated signals regardless of the reference picture selection, which
is referred to as multihypothesis motion-compensated prediction (Flierl and Girod
2003
).
Besides reducing the redundancies within image and video data, the compression
performance is further improved by reducing the set redundancies among similar
images and videos. Karadimitriou et al. first proposed the set redundancy concept
and proposed a series set of similar image compression methods, e.g., Min-Max
differential (MMD) method (Karadimitriou and Tyler
1997
) and centoid method
(Karadimitriou and Tyler
1998
). The centroid method generates one central image
by averaging the pixel values in the same position among all the images, then the
