Information Technology Reference
In-Depth Information
12.3.2 Canny Edge Detector
The objective of this experiment is to determine the area of ice and water in the
images. With that purpose an algorithm of edge detection was used. Considering
the discrete description of image, several techniques can be used for edge
enhancement and detection in digital images. Here, the chosen technique was the
Canny method.
John Canny, in 1986, in his paper ''the computational Approach to edge
detection'' described a method for edge detection purpose [
25
]. In this paper,
Canny treats the problem of edge detection establishing suitable criteria, so that
one can achieve better results than other known methods.
The first criterion is the called low error rate, which considers that ''edges
occurrences can't be discarded and false edges can't be found.'' The second cri-
terion refers to ''the distance between the center of the edge and the points in its
neighborhood that should be minimized.'' The last one was implemented because
the first two were not sufficient to eliminate the possibility of multiple results for
only one edge. Therefore, the third criterion considers that the algorithm should
find only one result for a single edge.
The application of the Canny method is done by steps. The first step is the use
of a Gaussian filter. The objective of this filter is to remove undesirable noises and
enhance important features. This filter is accomplished by the convolution of the
image, using a mask that is different depending on the specific filter.
The convolution operation consists in the application of a bidimensional filter
taking into account the image grayscale intensity domain. Each pixel in the ori-
ginal image x(n
1
,n
2
) is transformed in a new one y(n
1
,n
2
) (Eq.
12.14
).
Þ
¼h
ð
n
1
;
n
2
Þ
x
ð
n
1
;
n
2
Þ
¼
X
1
X
1
yn
1
;
n
2
ð
x
ð
k
1
;
k
2
Þ
h
ð
n
1
k
1
;
n
2
k
2
Þ
k
1
¼
1
k
2
¼
1
ð
12
:
14
Þ
The term h represents the convolution mask of the used filter.
In order to implement the filtering of images through convolution, it is necessary
to slide a mask (kernel) across the image (Fig.
12.13
). The new pixel value y(n
1
,n
2
)
is the weighted sum of the input pixels x(n
1
,n
2
) within the mask where the weights
are the values of the filter assigned to every pixel of the window itself [
21
].
The choice of convolution mask is done according to the filter and the reso-
lution desired. If the mask is large, the effect of the algorithm will be less observed.
On the other hand, when the mask is small we will spend more computational
processing.
For the Gaussian filter, its mask is obtained by a discreet Gaussian fit. The
Gaussian curve has the characteristic of augment values that are close to the
average value and to reduce the intensity values that are far from it. The Gaussian
Search WWH ::
Custom Search