Hot Research Topics in Video Coding and Systems - Advanced Video Coding Systems - page 213

Fig. 12.4 Foveal image at a fixation point and the original image
with distance from the fixation point as shown in Fig. 12.4 . The main idea of applying
attention model in video coding is to allocate more bits into the area near the fixation
point. The head and shoulder model is first applied in the video conference appli-
cations, and the face region is classified as the ROI region (Yang et al. 2005 ). The
foreground area always needsmore bits to code, as it attractsmore attention compared
to the background area (Chai and Ngan 1997 ). With the advance of the computer
vision techniques, the ROI region can be automatically detected (Itti 2004 ), and
thus high accuracy rate control can be achieved. In Chen et al. ( 2006 ), object-based
rate allocation scheme was proposed in the framework of MPEG-4 coding standard.
Moreover, the eye moving effect is also considered in the attention-based video cod-
ing. InNguyen andHwang ( 2002 ), the authors proposed a novel rate control algorithm
based on the smooth pursuit eye movement effect (SPEM) (Robinson 1965 ).
Though quantization parameter optimization can efficiently improve the percep-
tual quality in video coding, two issues should still be considered. First, the blocking
artifacts caused by different quantization parameter between coding blocks would
be more significant. Second, the prediction framework in video coding makes use of
the previous frame to code the current frame, and thus some research work has been
conducted to allocate more bits into the area that are more frequently referenced in
the following frames (Lou and Sun 2011 ). However, the quantization optimization
may not be an optimal way in the sense of prediction, and how to combine the visual
cues with the prediction structure to fully optimize the codec is still an open issue.
12.2.2.3 Perceptual Rate Distortion Optimization
In conventional video coding, distortion models such as MSE and SAD are used
as the distortion measure. However, they are widely criticized for not correlating
the perceptual quality. In perceptual video coding, there are usually two ways to
achieve the optimal rate perceptual distortion performance. One way is to change the
distortion measure, such as using advanced metric that correlates better compared
to MSE to achieve better rate distortion performance. The other way is to adaptively
adjust the Lagrangemultiplier to allocatemore bits to perceptually important regions.