AD ALTA
JOURNAL OF INTERDISCIPLINARY RESEARCH
We can argue that with a different setting of the Hough
transform algorithm, the best setting of the Canny edge detector
changes as well, but the experiments verified also, that even
when the minimum number of intersections in Hough space for
line detection is arbitrary, the range of intensity values is not
always sufficient for the best result of line detection. The
reduction of this minimum number of intersections in an image,
where the edge detection was done inefficiently, often leads to
poor identification of lines (Mukhopadhyay, Chaudhuri, 2015).
Table 1: Information about the test images and the results of the
experiments
Image
Maximum Mean
Median
3:1
2:1
Airplane 1034.2
61.3
28.0
836–838 756–762
Banana
487.2
25.1
11.7
408–487 380–487
Basket
659.5
100.2
81.0
558–562 448–538
Beehive 481.7
42.5
31.6
276–352 226–352
Briefcase 807.2
53.4
25.6
581–599 401–455
Brush
567.3
37.5
16.0
482–490 434–446
Coffee
Maker
856.7
42.2
15.2
348–394 288–294
Feather
821.4
80.0
58.1
448–478 374–386
Golf-cart 792.9
62.9
40.5
528–550 370–384
Grater
786.5
39.4
15.6
324–358 272–278
Mailbox 782.3
84.4
61.1
522–526 404–426
Pillow
584.2
44.3
27.8
300–534 260–442
Pitcher
623.8
44.4
14.1
428–520 300–354
Grocery
Cart
910.8
81.8
55.3
547–593 448–554
Stairs
775.5
79.6
48.8
622–634 459–463
Stapler
575.0
23.4
11.4
288–370 193–205
Tire
943.6
61.1
26.0
631–735 421–543
Traffic
Cone
802.9
59.5
41.2
350–368 320–368
Trashcan 862.1
67.9
38.9
592–604 468–484
Video
Camera
782.7
43.5
19.2
500–502 500–518
4 Discussion
As we can see, the best threshold values often go beyond the
range of intensity values of the original image. So if we want to
apply statistical functions to determine the threshold values, we
should work with all the values of the gradient image.
The experiments also revealed that in some cases, the 2:1 ratio
allowed a more accurate line detection and in most other cases,
where the two tested ratios yielded comparable results, the 2:1
ratio had a greater interval, thus a higher probability of correctly
determined hysteresis thresholds. Based on the table I would
suggest for line detection to use the 2:1 ratio between the upper
and lower threshold for hysteresis of the Canny edge detector.
Then, the upper threshold to be set around 55 % of the maximum
value of gradient image. With this setting, namely 52 - 56 %, the
thresholds scored the best results in seven experiments and
provided sufficiently good results in many others (as we can see
in figure 1).
Similarly, we could also use the mean, where best results
provided the upper threshold equal to approximately seven times
the mean value of the gradient image, more precisely 691 - 696
%. I do not recommend, however, to use the median, because
good results were quite scattered.
5 Conclusion
Based on the results of this paper, I believe that it is a bad
practice to determine the Canny edge detector hysteresis
threshold values solely on the basis of the intensity values of the
input image. The values of the gradient image after performing
the convolution with the Sobel operator and then consequentially
calculating the gradient magnitude, provide much better results
when used with detection lines using the Hough transform.
There is no reason to believe that these values would not provide
better results in other cases.
With the repetition of experiments on a larger set of test images,
we could find the thresholds that give required results for many
images, a sort of a starting point of setting for the Canny edge
detector, which can be further modified according to the specific
line detection problem.
Resolving this problem would simplify the line detection not
only in OpenCV by the Canny edge detector, but also it would
be possible to apply the results to other edge detectors based on
the hysteresis, more of which appear in literature (Heath et al.,
1997). It is likely that the results could also be used with a line
detection algorithm other than the Hough transform, for example
with the PClines method (Dubská et al., 2011).
Literature:
1. Canny, J.: A computational approach to edge detection. In:
IEEE Transactions on pattern analysis and machine intelligence.
VI. issue. IEEE, 1986. p. 679-698. ISSN 0162-8828.
2. Dubská, M., Herout, A., Havel, J.: PClines—line detection
using parallel coordinates. In: Computer Vision and Pattern
Recognition (CVPR), 2011 IEEE Conference on. IEEE, 2011. p.
1489-1494. ISBN 978-1-4577-0394-2.
3. Fang, M., Yue, G., Yu, Q.: The study on an application of otsu
method in canny operator. In: International Symposium on
Information Processing (ISIP). 2009. p. 109-112.
4. Heath. M., et al.: A robust visual method for assessing the
relative performance of edge-detection algorithms. In: IEEE
Transactions on Pattern Analysis and Machine Intelligence. XII.
issue. IEEE, 1997. p. 1338-1359. ISSN 0162-8828.
5. Medina-Carnicer, R., et al.: Solving the process of hysteresis
without determining the optimal thresholds. In: Pattern
Recognition. IV. issue. 2010. p. 1224-1232. ISSN 0031-3203.
6. Mukhopadhyay, P., Chaudhuri, B.: A survey of Hough
Transform. In: Pattern Recognition. III. issue. 2015. p. 993-
1010. ISSN 0031-3203.
Primary Paper Section: I
Secondary Paper Section: IN
Figure 1: Results of the experiments with 2:1 ratio between the upper and lower threshold. Test images (x-axis) are described by intervals of
the best upper thresholds to detect lines in range of possible gradient values (y-axis). Results (green line) are upper thresholds set to 55 % of
the maximum value of gradient image.
- 258 -