Danny King (kxrs26): Computer Vision: Description of Results
0 Danny King (kxrs26): Computer Vision: Description of Results
Page 1 of 5
04/02/2012
Danny King (kxrs26): Computer Vision: Description of Results Throughout this document and will refer to the low and high thresholds respectively for the double thresholding stage of the Canny edge detection algorithm. will represent the Gaussian filter kernel radius and will represent the standard deviation of the Gaussian function. We will describe six results using various parameter values of the implementation. Two of these results are wellbalanced, two are overly noisy and two are understated. Finally, we will demonstrate how the Gaussian sigma parameter affects the algorithm using one of the well-balanced results. For extra interest, we have elaborated on the Canny edge detection technique and have used a probabilistic averaging approach in order to improve upon the results from the Canny algorithm which will be briefly discussed at the end. Features of my implementation All filters and algorithms programmed from scratch including the Gaussian filter; no third party image processing functions were used. Support for images of every size (including non-square images), all major image formats, all major colour models. Gaussian kernels of any size are supported; algorithms adapt dynamically to deal with different kernel sizes.
Well-balanced results
Most major edges are accounted for and there is minimal noise in the final images. There are very few false edges detected. There are some disconnected edges. The strong edges have all been identified although some of the weaker edges are missing. Overall this is a good result using mid-range values for parameters.
Danny King (undergraduate of Computer Science at Durham University) [email protected] www.dannyking.eu
0 Danny King (kxrs26): Computer Vision: Description of Results
Page 2 of 5
04/02/2012
Again, most major edges are accounted and the edges are well-connected. More details have been retained than in the previous example. There is very little noise in the images. There are very few false edges detected. There are some disconnected edges and some double edges in the pepper image. This is another good result, although there are some duplicate edges and a little noise.
Overly noisy results
The very low threshold values caused the edge detector to count almost any pixel as an edge. There are double, triple and quadruple edges and much detail has remained. These parameter settings did not produce good edge detection results.
Danny King (undergraduate of Computer Science at Durham University) [email protected] www.dannyking.eu
0 Danny King (kxrs26): Computer Vision: Description of Results
Page 3 of 5
04/02/2012
These results are significantly better than the previous example for the peppers although there is still too much noise due to the very low sigma value. It is significantly worse for the coins, almost every edge is triplicated or more.
Understated results
The high ‘low threshold’, combined with the relatively high sigma value has caused most of the detail to be lost from the images. Many important edges have not been detected. One coin and several peppers have been missed entirely.
Danny King (undergraduate of Computer Science at Durham University) [email protected] www.dannyking.eu
0 Danny King (kxrs26): Computer Vision: Description of Results
Page 4 of 5
04/02/2012
This is a better example than the previous example however there are still some major edges missing from the peppers image. The coin outer edges have been detected well but no details from the coin faces have been identified. The high threshold has caused the loss of detail, again combined with a fairly high sigma value. Although the low threshold is very low, the hugeness of the high threshold has still caused detail loss.
Gaussian parameters 0.2
0.3
0.4
0.5
0.9
2.0
These images demonstrate the impact that changing the sigma value of the Gaussian function has on the end result. The higher the value of sigma, the less detail in the final image.
Danny King (undergraduate of Computer Science at Durham University) [email protected] www.dannyking.eu
0 Danny King (kxrs26): Computer Vision: Description of Results
Page 5 of 5
04/02/2012
Extra interest: probabalistic averaging of multiple Canny results I decided to take the process one set further and expand upon the Canny method. I devised an algorithm that created a number of Canny images, each using random parameter values. Once these images had been generated, they were combined to create the final images below. The images below were built using the following algorithm: for each pixel in each randomly generated Canny image, if it has a maximum intensity (equal to 255) then paint that pixel slightly brighter in the output image. This causes pixels in the output image to be brighter the more often they were white in the randomly generated Canny images. This has the effect of causing those pixels most likely to be on an edge to be brightest in the output image. The first six images below are the results of applying this algorithm 50, 100, 500, 1000, 2000 and 2500 times respectively (so that, for example, the last image was the combination of 2500 canny images).
50
100
500
1000
2000
2500
The image to the left is a thresholded post-processing of the 2500 iteration image above; all pixels above the threshold were made white and all below were made black, which has had the effect of creating thick, curved lines for the most predominant edges.
Danny King (undergraduate of Computer Science at Durham University) [email protected] www.dannyking.eu
Page 1 of 5
04/02/2012
Danny King (kxrs26): Computer Vision: Description of Results Throughout this document and will refer to the low and high thresholds respectively for the double thresholding stage of the Canny edge detection algorithm. will represent the Gaussian filter kernel radius and will represent the standard deviation of the Gaussian function. We will describe six results using various parameter values of the implementation. Two of these results are wellbalanced, two are overly noisy and two are understated. Finally, we will demonstrate how the Gaussian sigma parameter affects the algorithm using one of the well-balanced results. For extra interest, we have elaborated on the Canny edge detection technique and have used a probabilistic averaging approach in order to improve upon the results from the Canny algorithm which will be briefly discussed at the end. Features of my implementation All filters and algorithms programmed from scratch including the Gaussian filter; no third party image processing functions were used. Support for images of every size (including non-square images), all major image formats, all major colour models. Gaussian kernels of any size are supported; algorithms adapt dynamically to deal with different kernel sizes.
Well-balanced results
Most major edges are accounted for and there is minimal noise in the final images. There are very few false edges detected. There are some disconnected edges. The strong edges have all been identified although some of the weaker edges are missing. Overall this is a good result using mid-range values for parameters.
Danny King (undergraduate of Computer Science at Durham University) [email protected] www.dannyking.eu
0 Danny King (kxrs26): Computer Vision: Description of Results
Page 2 of 5
04/02/2012
Again, most major edges are accounted and the edges are well-connected. More details have been retained than in the previous example. There is very little noise in the images. There are very few false edges detected. There are some disconnected edges and some double edges in the pepper image. This is another good result, although there are some duplicate edges and a little noise.
Overly noisy results
The very low threshold values caused the edge detector to count almost any pixel as an edge. There are double, triple and quadruple edges and much detail has remained. These parameter settings did not produce good edge detection results.
Danny King (undergraduate of Computer Science at Durham University) [email protected] www.dannyking.eu
0 Danny King (kxrs26): Computer Vision: Description of Results
Page 3 of 5
04/02/2012
These results are significantly better than the previous example for the peppers although there is still too much noise due to the very low sigma value. It is significantly worse for the coins, almost every edge is triplicated or more.
Understated results
The high ‘low threshold’, combined with the relatively high sigma value has caused most of the detail to be lost from the images. Many important edges have not been detected. One coin and several peppers have been missed entirely.
Danny King (undergraduate of Computer Science at Durham University) [email protected] www.dannyking.eu
0 Danny King (kxrs26): Computer Vision: Description of Results
Page 4 of 5
04/02/2012
This is a better example than the previous example however there are still some major edges missing from the peppers image. The coin outer edges have been detected well but no details from the coin faces have been identified. The high threshold has caused the loss of detail, again combined with a fairly high sigma value. Although the low threshold is very low, the hugeness of the high threshold has still caused detail loss.
Gaussian parameters 0.2
0.3
0.4
0.5
0.9
2.0
These images demonstrate the impact that changing the sigma value of the Gaussian function has on the end result. The higher the value of sigma, the less detail in the final image.
Danny King (undergraduate of Computer Science at Durham University) [email protected] www.dannyking.eu
0 Danny King (kxrs26): Computer Vision: Description of Results
Page 5 of 5
04/02/2012
Extra interest: probabalistic averaging of multiple Canny results I decided to take the process one set further and expand upon the Canny method. I devised an algorithm that created a number of Canny images, each using random parameter values. Once these images had been generated, they were combined to create the final images below. The images below were built using the following algorithm: for each pixel in each randomly generated Canny image, if it has a maximum intensity (equal to 255) then paint that pixel slightly brighter in the output image. This causes pixels in the output image to be brighter the more often they were white in the randomly generated Canny images. This has the effect of causing those pixels most likely to be on an edge to be brightest in the output image. The first six images below are the results of applying this algorithm 50, 100, 500, 1000, 2000 and 2500 times respectively (so that, for example, the last image was the combination of 2500 canny images).
50
100
500
1000
2000
2500
The image to the left is a thresholded post-processing of the 2500 iteration image above; all pixels above the threshold were made white and all below were made black, which has had the effect of creating thick, curved lines for the most predominant edges.
Danny King (undergraduate of Computer Science at Durham University) [email protected] www.dannyking.eu
