Digital Signal Processing Reference
In-Depth Information
Fig. 1.4
An example of mean shift segmentation. (
a
) Original image
flower
.(
b
) Segmentation
result with different parameters
(3) For each group, assign a label.
(4) Eliminate those regions with less pixels.
Figure
1.4
shows an example of mean shift image segmentation. The original
image is shown in Fig.
1.4
a. Figure
1.3
b shows the segmentation results with differ-
ent kernel bandwidths. The spatial and range bandwidths are set to
h
s
=
4
,
h
r
=
10
and
h
s
=
15, respectively. We can see that with the increase of kernel band-
widths, more pixels are grouped together, which results in large regions partition.
This method is also implemented in an unsupervised manner.
10
,
h
r
=
1.2.3
Motion-Based Segmentation
In general, image segmentation algorithms mentioned above can be regarded as
spatial-based video segmentation. One can simply perform video segmentation
frame by frame using spatial segmentation methods. However, this will result in low
efficiency of video segmentation because high correlation between adjacent frames
in the temporal axis is neglected.
Temporal segmentation is usually based on change detection followed by mo-
tion analysis. The change detection masks can be defined as the absolute difference
between two consecutive frames, which are the most common forms of motion
information incorporated into the segmentation process. This algorithm employs in-
tensity changes produced by the motion of moving object to identify the position and
boundary of objects in time and space. Unlike the spatial segmentation approaches,
higher efficiency can be achieved because of small number of operations for the
segmented moving region instead of the whole image for every frame [
3
].
