Image Processing Reference
In-Depth Information
ND-CA method uses CA to compute the Ford-Bellman shortest path for segmenta-
tion on GPU (Graphic Processing Unit). Although the Growcut method and Seeded
ND-CA method seem dissimilar, it states that their results are identical and proves
the equivalence between them.
One extension of CA is the GrowCut segmentation framework over video se-
quences [26]. This was done by introducing mean-shift filtering as a pre-processing
step, as well as providing improvements and addressing problematic areas in the
original formulation, providing a good increase in accuracy. Results showed that the
proposed scheme achieves better results than the original formulation, and good seg-
mentation results were achieved using a considerably less complicated framework.
However, there are many areas available for future research. Specifically, GrowCut
can be transferred to a parallel framework, such as OpenMP, to significantly de-
crease the run time. In addition, the mean-shift preprocessing step can done on a
temporal basis, or other colour clustering methods could be examined. Also, differ-
ent cost functions and colour spaces may provide better accuracy. The contribution
of these factors will inevitably increase accuracy, while achieving real-time results.
Another extension is interactive matting algorithm for natural images [19]. By
employing a cellular automaton to iteratively optimize an energy function, the matte
is gradually extracted. One attractive feature of the proposed approach is that it
does not require a well-defined trimap, but only a few user-specified foreground
and background strokes. What ars more, additional user input can be added to refine
the matte not only after the matting process is over but also in the middle of the
matting process. Experimental results show that the proposed approach is capable
of producing high quality mattes for many complex natural images. Due to the in-
trinsic parallelism of cellular automata, the proposed approach has the potential of
being accelerated by using parallel computing technique. The parallel implementa-
tion will be researched in the future. And the extension of the proposed approach to
video matting is also the future work.
In medical applications, no single method is suitable for all kinds of segmenta-
tion, and further, there is no general algorithm to segment special tissues from all
kinds of medical images. In segmentation, the region of our concern is labeled as
foreground and the rest is labeled as background. Mostly, manual segmentation is
time-consuming and needs a method that segments an image automatically or with
few interactions of the user to provide accurate results.
This chapter is drawn to the CA method for several reasons. First, the CA method
supports n-dimensions and m-label categories where the number of labels does not
increase computational time or complexity. This ability makes it suitable for simul-
taneously segmenting multiple tissues. Second, for an effective interactive segmen-
tation, speed and usability are crucial. Since CA is an iterative method where each
cell independently follows a set of rules, this naturally lends itself to an efficient
parallel implementation. Further, the iterative nature of CA method enables the user
to visualize the results during the segmentation process.
Search WWH ::
Custom Search