Digital Signal Processing Reference
In-Depth Information
entire segments are matched instead of single pixels, the initial match values are
more robust to image noise and intensity bias [
11
].
Videophone applications
, which achieve high perceptual quality by coding the
areas of interest with better quality. Fewer bits can be allocated for encoding
the background by using the higher quantization level. The motivation of this
application is that the foreground region is the most important for the viewer [
12
].
•
Digital entertainment
, such as video matting and video tooning, which employs
the segmented objects to generate fantastic effect, or puts them into a virtual
scene or game.
There are other possible applications, such as medical diagnosis, tele-education,
industrial inspection, environmental monitoring, or the association of metadata with
the segmented objects, etc.
Segmentation has become an efficient way to bridge the primary image data and
semantic content in image/video processing. In order to satisfy the future content-
based multimedia application, more and more researchers seek for efficient ways to
segment arbitrary object from multimedia data over the past decade. There are many
methods that addressed the segmentation problem, which can be categorized with
respect to various criteria:
(1)
Data-based mode
: Based on the original data types, segmentation can be clas-
sified into image (e.g., nature, medical, or remote sensing images, etc.), video,
audio, and text segmentations, which can be applied to different scenarios. For
example, we can use the text segmentation to extract captions displayed in
movies, or partition a document into interesting parts. In this chapter, we only
concentrate on the image and video segmentation.
(2)
Interaction-based mode
: Two main categories can be classified, namely super-
vised and unsupervised modes. Supervised methods require user intervention
for segmentation, which allow users to easily indicate the foreground across
space and time. These methods can provide the better performance than au-
tomatic ways because the prior knowledge of the object can be obtained by
selecting training data on the images. Unsupervised methods mean that there is
no contextual knowledge assumption regarding to the object being segmented.
Object segmentation is performed in fully automatic manner, which has become
the key technique in a large number of real-time application areas such as video
monitoring and surveillance.
(3)
Feature-based mode
: Feature extraction plays important role in image/video
segmentation. According to the selection of the feature space, segmentation
can be divided into color, texture, intensity, shape, or motion based segmenta-
tion method. These features are usually applied to evaluate the region property.
For example, for color segmentation, the grouping decision relies on the color
distance between neighboring pixels. For the motion segmentation, the main
problem is to find independently moving objects in a video in terms of the
motion cue.
(4)
Inference-based mode
: Segmentation can be formulated as two message passing
modes, namely bottom-up and top-down segmentations. The first performs seg-
mentation on the basis of low-level visual features (e.g., color, texture, intensity,
•
