Game Development Reference
In-Depth Information
suitable for evaluation anymore, and the perceptual visual quality assessment has
been a hot topic in recent years in the perceptual coding area (Mannos and Sakrison
1974
; Moorthy and Bovik
2011
).
In the years of past decade, machine learning techniques havemade great progress.
Inspired by the recent advancement on semi-supervised learning techniques, some
learning-based coding methods have been proposed (Li et al.
2007
; Sun and Wu
2009
; Bryt and Elad
2008
; Bai et al.
2009
). Compared with the preceding coding
methods, learning-based coding algorithm is more adaptive to the image content,
and more intelligent, which is referred to as intelligent coding by Harashima and
Kishino (
1991
). In that image coding systems, the intelligent coding is classified
as the highest level coding methods which can understand the semantics, intention,
motives, and other factors in an image. However, the difference is that the semantic
coding is classified as an interlayer coding method between knowledge-based coding
method and perceptual coding method. The intelligent coding is the kind of adaptive
coding algorithms which are the composite methods of signal processing, computer
vision, machine learning, and modeling of human physiological and psychological
system, etc. More important, the intelligent coding is expanded from kinds of simple
images, e.g., facial images, to more general image content intelligently. Three recent
typical intelligent methods are screen content model-based coding, cloud model-
based coding, and background model-based surveillance video coding.
8.3 Background Picture Model-Based Video Coding
8.3.1 General Background Picture Modeling
Background modeling is widely used in object detecting and tracking, and is a hot
research topic in designing better background picture model. In general, these back-
ground picture models can be classified as parametric and nonparametric methods.
Parametric methods typically do not rely on large memory to buffer all the training
data, but updates the background model frame by frame. For parametric background
modeling methods, the most commonly used assumption is that the underlying dis-
tribution of the intensity value of a pixel is Gaussian or mixture of Gaussian. In
Stauffer and Grimson (
2000
), Chris Stauffer dealt with the motion segmentation
problem based on an adaptive background subtraction method by modeling pixels as
a mixture of Gaussian and used an online approximation to update the model. Several
improvements on Gaussian mixture modeling have been made in Li et al. (
2002
). In
Toyama et al. (
1999
), a three-level Wallflower scheme was presented, which tried to
solvemany problems that exist in backgroundmaintenance, such as light switch, fore-
ground aperture, etc. In (Haritaoglu et al.
2000
), three values, maximum value (M),
minimum value (N) and the largest interframe absolute difference (D), were stored
for each pixel to model background. Besides, running average method (Haque et al.
2008a
) and codebook method (Wren et al.
1997
) are also parametric models.
