Image Processing Reference
In-Depth Information
5.3
Edge Detection in Intensity Images
For edge detection of intensity images, some approaches first convert the intensity
images into binary ones, and then evolve two-state cellular automata using specific
state transition rules to determine edge pixels, while the others directly update pixel
states based on the relationship of the central pixel with its neighbourhood, mostly
a 1-ring von Neumann or Moore neighbourhood.
Popovici and Popovici [26] proposed an edge detection approach based on the
state differences between the central pixel and the pixels in its von Neumann neigh-
bourhood. If all the absolute state differences are less than a threshold
, then the
state of the central pixel becomes 0, otherwise it remains unaltered. The rule can be
formulated as:
ε
0,if
|
|≤ ε , ∀
v i
v c
i
N c
v c =
v c ,otherwise.
where v c and v c are the current and the updated states of the central cell c , N c is the
von Neumann neighbourhood of the cell c ,and v i is the current state value of the
cell i in N c .
Gorsevski et al. [11] used Popovici and Popovici's approach to detect the grain
boundaries in deformed rocks, but they did not cite [26].
Wongthanavasu and Sadananda [45] proposed a conditional rule to update the
cellular state as:
v c
,if v c
v max
v min
v c =
v min ,otherwise.
where v max and v min are the maximum and minimum states, respectively, in the von
Neumann neighbourhood of the central cell c . The above rule can be described the-
oretically by a state transition table. However, it is hard to construct and use such a
table in practice, because there are 256 5 entries in the table for a 256 greyscale image
with the von Neumann neighbourhood. Wongthanavasu and Sadananda provided a
partial transition table for illustration. Our experiments show that only a single it-
eration of this state transition will produce reasonable results. Multiple iterations
actually degrade the results.
Wongthanavasu and Lursinsap [44] also extended the above-described condi-
tional rule into 3D image edge detection with the only difference that the neigh-
bourhood is now 3D von Neumann. Their experiments showed that the CA approach
exhibited better performance than the Sobel and the Laplacian detection algorithms
on average.
Kumar and Sahoo's method [15] also directly utilizes the intensities of the central
pixel and its neighbourhood to detect edges, but the algorithm description is unclear,
and it is hard to figure out how the algorithm is actually implemented.
Diwakar et al. [7] presented an approach that first convert an intensity image to
a binary image through thresholding, and then use rules similar to Conway's Game
v max
 
Search WWH ::




Custom Search