Image/Video Segmentation: Current Status, Trends, and Challenges - Video Segmentation and Its Applications - page 13

Fig. 1.5 An example of superpixel segmentation. ( a ) Original image flower .( b ) Superpixel map
with the number of 200. ( c ) Segmentation result based on the superpixel map

In addition, there are many approaches that are proposed recently to partition an
image into meaningful regions. For example, an unsupervised segmentation algo-
rithm for color image was proposed in [ 38 ], which utilizes the gradient information
in CIE L*a*b* color space. The initial regions are first created by grouping those
nonedge pixels. The regions with similar color and texture are consequently merged
to obtain a final segmentation map. Using semantics information, an iterative re-
gion growing method was proposed in [ 39 ], which is characterized by edge penalty
functions within Markov random field context model and region growing technique.
This method allows various region features to be incorporated in the segmentation
process. In order to allow scene understanding, a region-based model that combines
appearance and scene geometry was proposed to partition a scene into semanti-
cally meaningful regions [ 40 ]. This model is defined in terms of a unified energy
function over scene appearance and structure. In addition, an interesting segmenta-
tion technique was proposed to partition multivariate mixed data from a lossy data
coding/compression viewpoint [ 41 ]. This work aims to search the optimal segmen-
tation that minimizes the overall coding length of the segmented data based on the
concepts in lossy data compression and rate-distortion theory. From above analysis,
we can see that the goal of these works is to achieve 'good' image segmentation in
terms of defined decision strategy.

1.3.2 Towards Machine Learning Based Segmentation

Another emerging trend in image/video segmentation is the learning based segmen-
tation, which seeks good segmentation for understanding images and their semantic
contents. These methods learn the optimal clustering algorithms from unsegmented,
cluttered images using a probabilistic model incorporating both shape model and
bottom-up cues (e.g., color, texture, or edge).
Generative models, as an important probabilistic graphical model, are usually
applied to represent the process by which images of objects are created. Unlike
the grid graph, in this model, each node represents a random variable, and the
links express probabilistic relationships between these variables. A typical prob-
abilistic graphical model for a collection of exchangeable discrete data is Latent