Improving the Canny Edge Detector Using Automatic Programming

Improving the Canny Edge Detector Using Automatic
Programming: Improving Non-Max Suppression
Lars Vidar Magnusson
Roland Olsson
Østfold University College
Østfold University College
[email protected]
[email protected]
ABSTRACT
In this paper, we employ automatic programming, a relatively unknown evolutionary computation strategy, to improve the non-max suppression step in the popular Canny
edge detector. The new version of the algorithm has been
tested on a dataset widely used to benchmark edge detection
algorithms. The performance has increased by 1.9%, and a
pairwise student-t comparison with the original algorithm
gives a p-value of 6.45 × 10−9 . We show that the changes to
the algorithm have made it better at detecting weak edges,
without increasing the computational complexity or changing the overall design.
Previous attempts have been made to improve the filter
stage of the Canny algorithm using evolutionary computation, but, to our knowledge, this is the first time it has been
used to improve the non-max suppression algorithm.
The fact that we have found a heuristic improvement
to the algorithm with significantly better performance on
a dedicated test set of natural images suggests that our
method should be used as a standard part of image analysis platforms, and that our methodology could be used to
improve the performance of image analysis algorithms in
general.
1.
INTRODUCTION
The Canny edge detector [3] is a popular algorithm that
can be found in most image analysis platforms. The algorithm is based on the gradient image, which is calculated
by convolving the image with a filter designed to approximate either the first or second order derivatives. This is a
common approach in traditional edge detectors, so the success of the Canny algorithm can be attributed to the other
stages of the algorithm; non-max suppression and hysteresis
thresholding.
It is a well known fact that traditional gradient based
edge detectors like the Canny algorithm can suffer from poor
performance in textured image regions. The best algorithms
today handle this issue by incorporating more complex inforPermission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for profit or commercial advantage and that copies bear this notice and the full citation
on the first page. Copyrights for components of this work owned by others than the
author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or
republish, to post on servers or to redistribute to lists, requires prior specific permission
and/or a fee. Request permissions from [email protected]
GECCO ’16, July 20 - 24, 2016, Denver, CO, USA
c 2016 Copyright held by the owner/author(s). Publication rights licensed to ACM.
ISBN 978-1-4503-4206-3/16/07. . . $15.00
DOI: http://dx.doi.org/10.1145/2908812.2908926
mation and processing schemas. There is also an increased
tendency to utilize some form of machine learning strategy
to help distinguish between different textures [1, 17].
There have been several attempts at improving the Canny
algorithm. Wang and Fan [16] introduced an adaptive filter to replace the standard filter in the algorithm. They
also introduced an additional noise reduction technique, and
they have replaced non-max suppression with morphological
thinning. They demonstrate the performance by qualitative
analysis of a single image and a simple evaluation function.
Ding and Goshtasby [4] proposed an improvement to the
non-max suppression stage in the Canny algorithm based
on an observation that the original approach will incorrectly
suppress corner edges. They introduce an extra processing
stage to identify what they refer to as minor edges, which
are joined together with the edges produced by the original
Canny algorithm. There have also been attempts to create
an adaptive threshold mechanism for the Canny algorithm
[7].
Evolutionary computation (EC) strategies such as genetic
algorithms (GA) and genetic programming (GP) have been
used to evolve improved filters [6] and to infer logic filters for
sliding windows [14, 18], but, as far as we know, it has never
been used to evolve improved solutions for the remaining
stages in the Canny algorithm.
We want to investigate the possibility of making the nonmax suppression stage in the Canny edge detector better
at suppressing edges in textured areas, and to increase the
ability of the algorithm to identify contours around objects.
We aim to do this without adding additional information or
processing—to make the original algorithm better by changing the internal calculations. We have employed automatic
programming (AP), a relatively unknown evolutionary computation strategy related to genetic programming. AP has
the ability both to invent brand new solutions and to improve existing algorithms by evolving suitable heuristics, and
it has been used with success on similar problems in the past
[2, 10, 8].
2.
AUTOMATIC PROGRAMMING
Automatic programming is like Genetic Programming (GP)
an evolutionary computation strategy, but, unlike GP, automatic programming systems use a systematic search process
rather than a random one. A major benefit of this approach
is the ability to manage complex code like loops and recursive patterns in a better way. In this paper we have employed Automatic Design of Algorithms Through Evolution
(ADATE) [12, 13] which has been continuously developed
over the last two decades.
The ADATE system infers and evaluates new programs
based on problem specifications written in a combination of
Standard ML [11] and ADATE ML, where ADATE ML is
a minimum subset of SML used for evolution to keep the
syntactic complexity to a minimum.
2.1
The Evolutionary Search Strategy
ADATE maintains a hierarchical structure called a kingdom that contains the individuals during evolution. The
structure is inspired by Linnean taxonomy [9], and it has
three levels; families, genera and species. A family is a collection of genera that has comparable syntactic complexities.
A genus is a set of one or more potential parent programs. A
species is a group of programs created from the same parent
program in one of the genera.
At each step in the evolution the individual in the kingdom with the lowest cost limit and syntactic complexity is
selected. The cost limit associated with each program controls the number of programs to create from a parent program, and it is doubled every time a program is expanded.
Based on the cost limit and syntactic complexity of the selected program a number of new programs are synthesized
by transforming the program using compound transformations. Each new program is evaluated to determine whether
it should be inserted into the kingdom or not. The search
process then proceeds to select the next program to expand,
and the process continues in this manner until the user terminates.
A program is inserted into the kingdom if it is smaller
and better than any of the other individuals already in the
kingdom. When a program is inserted, all programs that
are bigger without being better are removed. The result is
a tree structure with genealogical chains of gradually bigger
and better programs.
2.2
Program Transformations
Programs are transformed using so called compound transformations, which are sequences of atomic transformations.
There are four atomic transformation types; replacement
(R), abstraction (ABSTR), case-distribution (CASE-DIST),
and embedding (EMB).
All of the atomic transformation types maintain the semantics of the parent program except for Rs. This makes
them essential to the evolution of new individuals, but also
makes them the most computationally challenging. The system infers replacement transformations by searching through
type correct replacements of increasing complexity. All replacements are tested to identify replacements that do not
make the program worse. Replacements that fulfill this requirement are marked as replacements preserving equality
(REQs), and they play an important role in the exploration
of plateaus in a fitness landscape.
New auxiliary functions are created by the ABSTR transformation. A source expression is extracted from the program and inserted into a new function with an appropriate
domain. A function call to the new function, along with
the corresponding arguments, is inserted in the location of
the source expression. The two remaining transformation
types, CASE-DIST and EMB, are responsible for changing
the scope of function and variables, and for changing the
domain of an auxiliary function, respectively.
Compound transformations group together atomic trans-
Figure 1: A test image similar to the one used by
Ding and Goshtasby [4] and the result of running
the Matlab implementation of the Canny algorithm
on the image.
formations according to a set of pre-calculated forms designed to create sequences of transformations that are aimed
towards a common target. To illustrate this concept consider the example of a R following an ABSTR. The ABSTR
creates a new function, and the R will be applied to the
right hand side of the new function. With this setup, only
the first transformation in a compound transformation can
chosen freely; the remaining transformations are restricted
to the ones allowed by the pre-calculated forms.
3.
THE CANNY EDGE DETECTOR
The Canny edge detector [3] was introduced three decades
ago, and it is still the standard detector in most image analysis platforms. In essence, it was designed to disregard discontinuities caused by noise and to give a single response to
each edge. In this respect, it is an effective algorithm, both
in terms of the quality of the response and the processing
time needed.
There are three major steps in the algorithm. The first
step is responsible for producing the gradient image, which
is calculated by convolving the input image with one or more
filters both to remove noise and to find the differences. The
gradient image provides the input to non-max suppression—
the target of our investigation. The final step is called hysteresis thresholding. It uses the output of non-max suppression to identify the strong and weak edges, corresponding
to pixels greater than a high and low threshold respectively.
The final edge image contains all the strong edges in addition
to all 8-connected weak edges.
3.1
Non-Max Suppression
Non-max suppression was designed to reduce multiple responses to a single edge [3], and it has played an essential
role in the success of the Canny edge detector. So much so
that the step has become a standard post-processing step
for edge detectors in general.
The idea is to suppress gradient magnitudes that are less
than either of the magnitudes along the gradient angle. There
are several ways this could be done, but due to both the
quality and the popularity of the implementation we have
decided to use the implementation in Matlab as a reference.
We tested the implementation to see if it has problems
with corner edges as observed by Ding and Goshtasby [4],
and the results can be seen in Figure 1. While it is clear that
a few pixels have been erroneously suppressed, the problem
is not as significant as illustrated in [4].
fun f ( d1 , d2 , m, m1, m2, m3, m4 ) =
let
fun l e r p ( ( x , y , t ) =
x ∗ ( 1.0 − t )+y∗ t
in
case abs ( d1 /d2 ) of t =>
case l e r p ( m1, m2, t ) of tm1 =>
case l e r p ( m3, m4, t ) of tm2 =>
case m < tm1 of
f a l s e => (
case m < tm2 of
f a l s e => m
| t r u e => 0 . 0 )
| t r u e => 0 . 0
end
Figure 3: The base individual written in ADATE
ML.
Figure 2: The eight possible sectors for the gradient
angle. Two and two sectors are grouped together to
form 4 possible configurations.
The Matlab implementation generates two gradient images, one for each axis. These are used both to find the
final magnitude and to determine the angle of a gradient.
The angle belong in one of the eight sectors that can be
seen in Figure 2. The angles in sector 1 are equivalent to
the ones in sector 5, the angles in sector 2 with the ones
in sector 6, and so on. For the purpose of this discussion
we define the x-axis to point right, and the y-axis to point
up. If the y gradient dy is positive and the x gradient dx
larger than dy , or if dy is negative and dx is less than dy ,
the gradient angle is in sector 1 or 5 respectively. In this
case, the neighbor magnitudes are calculated by means of
linear interpolating between neighboring magnitudes with
the interpolating parameter t set to dy /dx . The first neighbor magnitude is found by interpolating between the right
and upper-right neighbors, and the second neighbor by interpolating between the left and the lower-left. The tests to
determine if a gradient magnitude falls into the remaining
sectors and the corresponding linear interpolations between
the neighbor magnitudes are conducted in a similar manner.
4.
EVOLUTIONARY PREREQUISITES
The evolutionary prerequisites presented in this section
define the world in which new programs will be evolved by
the ADATE system. The final problem specification, along
with all the relevant resources, can be found on our web site.
4.1
The Image Dataset
There are several available datasets that could be used to
train object contour detection, but the BSDS500 dataset [1]
stand out when it comes to the quality of the ground truth
annotations. It has also been widely used as a benchmark
for both edge detectors and image segmentation algorithms.
The BSDS500 is a dataset of 500 natural image with
ground truth annotations by multiple human subjects, and
the images feature a number of different motives and scenarios. There are 200 training images, 100 validation images
and 200 test images, both in grayscale and color. We have
selected to use only the grayscale images in our experiments,
since this will help keep the runtime low and speed up evolution. This can be changed in future experiments to allow
the ADATE system access to more information to utilize in
the evolved programs.
ADATE evaluates the performance of the programs on
the training set at a higher frequency than on the validation
set. Previous experience with using the dataset has shown
good overfitting characteristics, which has allowed us to use
a smaller training set than intended by the dataset creators.
We used 50 of the training images for training. The remaining 150 training images were used along with the 100
validation images for validation.
4.2
The Fitness Function
The fitness function used for the experiments is inspired
by the benchmark provided in the BSDS500 [1]. They use
the popular F-measure [15] to evaluate the performance of
edge detectors by accumulating the precision and recall counts
over all the images, and calculates the F-measure based on
the total counts. While this is a reasonable approach, we
decided instead to use the average F-measure. This makes
each image equally important to the final score, regardless
of the number of edge pixels in the image.
The dataset contains on average 5 ground truth annotations for each image. Determining whether or not a proposed edge pixel matches an edge pixel in one of the ground
truth annotations is reduced to an assignment problem and
solved using the CSA algorithm [5]. However, due to the
size of the graphs, evaluating an edge map against a single
ground truth is significantly slower than finding the edges.
We therefor decided to minimize the time needed to evaluate each image by using only the best ground truth during
evolution. We define the best ground truth to be the ground
truth with the highest F-measure when evaluated against
the other ground truths in the set. This approach was used
during evolution to speed up the process, but the final evaluation of a program was conducted using the entire ground
truth set.
4.3
The Base Individual
The base individual for the original non-max algorithm is
listed in Figure 3. The program is invoked once per pixel in
the gradient image. The arguments passed to the program
depend on the gradient angle. The are four possible con-
figurations, corresponding to the eight possible directions
that can be seen in Figure 2. Determining which of the four
configurations that apply to the current position in the gradient image is left outside the evolution environment. This
ensures that the overall structure of the algorithm maintains
the same.
To illustrate the alternative argument configurations for
a particular position we will give two examples, representing two different variants. In the first configuration, where
the gradient angle is in either sector 1 or 5, d1 and d2 will
contain the gradient for the y and x direction respectively.
The parameter m will contain the magnitude of the current
pixel, and the remaining parameters m1, m2, m3 and m4
will contain the magnitude of the right, the upper-right, the
left and the lower-left neighbor respectively. In the second
configuration, where the gradient angle is in either sector 2
or 6, the d1 and d2 parameters have switched contents so
that d1 contains the gradient in the x direction and vice
versa. The m argument is still the magnitude of the current pixel, but m1, m2, m3 and m4 have changed to the
magnitude of the upper, the upper-right, the bottom and
the bottom-left neighbor. This pattern is repeated for the
two remaining configurations, i.e., d1 will contain the lower
of the two axis gradients, m1 and m3 will contain the magnitudes of the axis-aligned neighbors and m2 and m4 will
contain the magnitude of the diagonal neighbors.
4.4
Functions and Constructors
We decided to restrict the functions and constructors available during evolution to include only the functions and constructors already present in the base individual. The only
exception is the inclusion of the hyperbolic tangent function,
which has proved to be useful in several experiments in the
past.
4.5
The Constants
The return characteristic of the Canny algorithm is dependent on three constants. The first is the standard deviation to use for the noise reduction filter in the first stage of
the algorithm, and the remaining two are the high and low
thresholds used during hysteresis thresholding. Note that
none of these are used directly in the base individual, or in
the non-max algorithm, but they still affect the overall performance. We evaluated different values for these constant
using both the training and validation images along with a
simple grid search algorithm to find the configuration used
during evolution.
5.
RESULTS
In this section the ADATE-improved non-max suppression algorithm is described and analyzed, and we present the
results of running the Canny algorithm with the improved
non-max suppression algorithm on a popular test set.
5.1
ADATE-Improved Non-Max Suppression
The improved individual can be seen in Figure 4. At first
glance it is quite similar to the original program listed in
Figure 3, but there are several important differences.
The ADATE-improved algorithm is slightly smaller than
the original. The first reason for this is that the calls to
lerp have been moved into the case-expressions which are
dependent on their return. This is just a syntactic rewrite
and has no effect on the performance of the algorithm. The
fun f ( d1 , d2 , m, m1, m2, m3, m4 ) =
let
fun l e r p ( x , y ) = x ∗ ( 1.0 −m3 ) + y∗m3
in
case m < l e r p ( m1, m2 ) of
f a l s e => (
case m < l e r p ( m3, m4 ) of
f a l s e =>
abs ( m/ tanh ( m/ d2 ) )
| t r u e => 0 . 0 )
| t r u e => 0 . 0
end
Figure 4: The improved program.
second reason however is that the calculations of t—the interpolation parameter in the original algorithm—has been
removed. The calculation performed by lerp no longer uses
t to interpolate between the values, but instead uses the contents of the m3 parameter. This has a profound effect on
the behavior of the algorithm, as we will discuss below. The
final change is the return value when the current magnitude
m is bigger than both neighbors. The original behavior is to
return m, but this has been changed to m/(tanh (m/d2)),
which has introduced a dependency on d2—the largest of
the two axis gradients.
In order to make the difference between the two implementations more clear we have plotted the differences of the
most significant values, when running on three separate images with OD constants, in Figure 6. The histograms show
the distributions of m3 in red and t in blue. Based on the
histograms we can see that the m3 parameter is close to
normally distributed, while t is considerably more uniform.
It is clear that on average m3 is significantly smaller than t.
This has the effect that the linear interpolation will prioritize the axis-aligned neighbors over the diagonal, which in
turn will cause fewer pixels to be suppressed—particularly
gradients with angles close to the diagonal.
The scatter plots illustrate the difference between m and
m/(tanh (m/d2)). There is considerable correlation between
the two, but the latter is slightly larger than the former. The
consequence is that the magnitudes that pass the ADATEimproved non-max suppression will on average be slightly
larger than in the original algorithm. The variation between the two is caused by the denominator tanh (m/d2).
The value of this expression will increase as m/d2 increases,
which depends on the other axis gradient d1. The result is
that a non-suppressed value will be largest when the gradient angle is axis aligned and reduced when approaching the
diagonal.
This is an extremely interesting heuristics. It will suppress fewer pixels with steep angles due to the change in
interpolation, and at the same time reduce the resulting
magnitude of these pixels due to the change in the return of
non-suppressed magnitudes.
5.2
Performance
The performance of the algorithms have been measured
on the test set in BSDS500 [1]. We have copied their evaluation function in order to for the results to be comparable
to previous results on the dataset. We test each of the algorithms using two constant configurations; one optimized for
the entire dataset (OD), and one optimized for each image
Figure 5: A collection of images and corresponding edge maps where the ADATE-improved algorithm is
better than the original algorithm. The edge maps produced with the ADATE-improved algorithm are in
the middle.
10
×10 4
5
×10 4
5
9
4.5
8
4
4
7
3.5
3.5
4.5
6
3
3
5
2.5
2.5
4
2
2
3
1.5
1.5
2
1
1
1
0.5
0.5
0
0
0
0.2
0.4
0.6
0.8
1
0
0
0.2
0.4
0.6
0.8
1
1.2
1.2
1.2
1
1
1
0.8
0.8
0.8
0.6
0.6
0.6
0.4
0.4
0.4
0.2
0.2
0.2
0
0
0
0.2
0.4
0.6
0.8
1
×10 4
0
0.2
0.4
0.6
0.8
1
0
0.2
0.4
0.6
0.8
1
0
0
0.2
0.4
0.6
0.8
1
Figure 6: Histograms and scatter plots of the differences between the two algorithms on three images.
The histograms contains the distribution of m3 in red and t in blue. The scatter plots compares m to
m/(tanh (m/d2)). The three images used are the same as the three first images in Figure 5.
(OI). The results for the two configurations can be seen in
Table 1. The F-measure with OD constants is listed under
ODF, and the F-measure with OI constants is listed under
OIF. We have included the corresponding results of the best
known algorithm for the dataset, the SCG algorithm [17].
50
Table 1: The Evaluation Results
ODF OIF
SCG
0.71
0.73
ADATE-Improved 0.618 0.657
Canny
0.606 0.652
45
40
35
30
25
20
15
10
5
0
−0.05
0
0.05
0.1
0.15
0.2
Figure 7: An histogram plot of the difference in Fmeasure between the two algorithms on each of the
images.
The ADATE-improved algorithm has been improved by
1.1 percentage points or 1.9% with OD constants, and by
0.5 percentage points or 0.8% with OI constants. The improvement is reduced to about half using OI constant, which
can be interpreted as being less dependent on the constants
used, but it can also be interpreted as being optimized for
the conditions under which the program was evolved. More
testing is required to reach a definite conclusion. The performance gap up to the SCG algorithm is to be expected when
considering that the SCG algorithm uses considerably more
information and processing, but the ADATE- improved algorithm has closed about one fifth of the gap with OD constants.
We have performed two statistical tests to ensure that
the improvement is statistically significant. Both algorithms
were tested using their respective OD threshold setup. The
differences in F-measure between the two algorithms can be
seen in Figure 7. The distribution is close to a normal distribution, so for the first test we used a paired student-t test.
1
0.9
0.8
Precision
0.7
0.6
0.5
0.4
0.3
0.2
0
0.1
0.2
0.3
0.4
0.5 0.6
Recall
0.7
0.8
0.9
1
Figure 8: The receiver operating characteristics of
the two algorithms. The ADATE-improved algorithm is in solid red.
This gave a p-value of 6.45 × 10−9 . We also performed a
Wilcoxon signed-rank test, in case our assumption of normally distributed differences is wrong. This gave a p-value
of 1.649 × 10−9 .
The receiver operating characteristic curves for the original Canny and the ADATE-improved algorithm can be seen
in Figure 8. The ADATE-improved algorithm is consistently
better in the entire range. The precision characteristics is
significantly better in parts of the curve. The recall characteristics however is only slightly better. Based on this we can
draw the conclusion that the improved algorithm is slightly
better at finding edges, but that it makes noticeably fewer
mistakes.
5.3
Visual Analysis
We have included a collection of images where the ADATEimproved algorithm outperform the original algorithm in
Figure 5. We also have included the corresponding edge
maps for both the algorithms for visual analysis. We used
the OD constants for both algorithms to produce the edge
maps.
The improvements made to the algorithm are easily noticeable in the first image in the collection. The improved
algorithm has correctly identified a large part of the duck,
while the original algorithm has only found the most prominent edges.
The improvements in the edge maps of the remaining images are far less visible, but significant nevertheless. The
improvements in the second image have been in the details
of the ice and landscape. In the third image the improved
algorithm has identified more of the glass and some of the
details in the label of the bottle. The edge on the floor in
front of the men in the fourth image has been partly identified, and so has more of the penguin in the last image.
6.
CONCLUSIONS
We have successfully used automatic programming to improve the non-max suppression stage in the Canny edge detector. We have evaluated the improved algorithm on a popular benchmark for edge detectors, and the performance has
increased with 1.9%. Based on a paired student-t test and a
Wilcoxon signed-rank test we can say that the improvement
is statistically significant.
The improved algorithm contains novel and interesting
heuristics that have made the overall algorithm better at
identifying weak edges. This is an impressive feat considering the size of the changes—the overall design of the algorithm has been maintained. Two minor changes have caused
the algorithm to suppress fewer gradients with angles close
to the diagonal, and to boost the gradients that pass suppression by an amount inversely proportional to how close
the gradient angle is to the diagonal.
The fact that we have used automatic programming to
infer interesting heuristics that enhance the performance of
the non-max algorithm on a set natural images is further
evidence that automatic programming is ideally suited for
image analysis problems.
We are planning to use the same methodology to improve
hysteresis thresholding, the last step in the Canny edge detector, and we want to investigate the possibility of inferring
a new algorithm for automatically determining the thresholds. There is also a possibility of trying to expand on the
algorithm by including additional information to help out
with textured regions.
7.
REFERENCES
[1] P. Arbelaez, M. Maire, C. Fowlkes, and J. Malik.
Contour detection and hierarchical image
segmentation. IEEE Transactions on Pattern Analysis
and Machine Intelligence, 33:898–916, 2011.
[2] H. Berg, R. Olsson, T. Lindblad, and J. Chilo.
Automatic design of pulse coupled neurons for image
segmentation. Neurocomputing, 71(10-12):1980–1993,
2008. Neurocomputing for Vision Research; Advances
in Blind Signal Processing.
[3] J. Canny. A computational approach to edge
detection. Pattern Analysis and Machine Intelligence,
IEEE Transactions on, (6):679–698, 1986.
[4] L. Ding and A. Goshtasby. On the canny edge
detector. Pattern Recognition, 34(3):721–725, 2001.
[5] A. V. Goldberg and R. Kennedy. An efficient cost
scaling algorithm for the assignment problem.
Mathematical Programming, 71(2):153–177, 1995.
[6] C. Harris and B. Buxton. Evolving edge detectors
with genetic programming. In Proceedings of the 1st
annual conference on genetic programming, pages
309–314. MIT Press, 1996.
[7] Y.-K. Huo, G. Wei, Y.-D. Zhang, and L. nan Wu. An
adaptive threshold for the canny operator of edge
detection. In Image Analysis and Signal Processing
(IASP), 2010 International Conference on, pages
371–374, April 2010.
[8] K. Larsen, L. V. Magnusson, and R. Olsson. Edge
pixel classification using automatic programming.
Norsk Informatikkonferanse (NIK), 2014.
[9] C. Linnaeus. Systema naturae per regna tria naturae
secundum classes, ordines, genera, species,...,
volume 1. impensis Georg Emanuel Beer, 1788.
[10] L. V. Magnusson and R. Olsson. Improving
graph-based image segmentation using automatic
programming. In Applications of Evolutionary
Computation, pages 464–475. Springer, 2014.
[11] R. Milner, M. Tofte, R. Harper, and D. MacQueen.
The Definition of Standard ML - Revised. The MIT
Press, 1997.
[12] R. Olsson. Inductive functional programming using
incremental program tranformation. Artificial
Intelligence, 74:55–81, 1995.
[13] R. Olsson. Population management for automatic
design of algorithms through evolution. In
Evolutionary Computation Proceedings, 1998. IEEE
World Congress on Computational Intelligence., The
1998 IEEE International Conference on, pages
592–597. IEEE, 1998.
[14] R. Poli. Genetic programming for feature detection
and image segmentation. In Evolutionary Computing,
pages 110–125. Springer, 1996.
[15] C. J. V. Rijsbergen. Information Retrieval.
Butterworth-Heinemann, Newton, MA, USA, 2nd
edition, 1979.
[16] B. Wang and S. Fan. An improved canny edge
detection algorithm. In 2009 second international
workshop on computer science and engineering, pages
497–500. IEEE, 2009.
[17] R. Xiaofeng and L. Bo. Discriminatively trained
sparse code gradients for contour detection. In
Advances in neural information processing systems,
pages 584–592, 2012.
[18] Y. Zhang and P. I. Rockett. Evolving optimal feature
extraction using multi-objective genetic programming:
a methodology and preliminary study on edge
detection. In Proceedings of the 7th annual conference
on Genetic and evolutionary computation, pages
795–802. ACM, 2005.