## SUSAN - A New Approach to Low Level Image Processing (1995)

Venue: | International Journal of Computer Vision |

Citations: | 208 - 3 self |

### BibTeX

@ARTICLE{Smith95susan-,

author = {S. M. Smith and J. M. Brady},

title = {SUSAN - A New Approach to Low Level Image Processing},

journal = {International Journal of Computer Vision},

year = {1995},

volume = {23},

pages = {45--78}

}

### Years of Citing Articles

### OpenURL

### Abstract

This paper describes a new approach to low level image processing; in particular, edge and corner detection and structure preserving noise reduction.

### Citations

3098 | A computational approach to edge detection
- Canny
- 1986
(Show Context)
Citation Context ... Quality of Feature Detectors In this section the desired qualities of feature detectors are explained. The criteria eventually given have a slightly different emphasis from those previously used. In =-=[9]-=- three criteria for edge detection are given. These have been used in similar form in a good deal of vision research. They are: 1. Good detection. There should be a minimum number of false negatives a... |

1721 | A combined corner and edge detector
- Harris, Stephens
- 1988
(Show Context)
Citation Context ...e, this is not sufficient to prevent false responses on strong diagonal edges. The corners are found at different smoothing levels allowing an estimation of the corner positions at zero smoothing. In =-=[26]-=- Harris and Stephens described what has become known as the Plessey feature point detector. This is built on similar ideas to the Moravec interest operator, but the measurement of local autocorrelatio... |

881 |
Exploratory Data Analysis
- Tukey
- 1977
(Show Context)
Citation Context ... are sorted into ascending order (according to their value) and this list is processed to give an estimate of the underlying image brightness. The simplest order statistic operator is the median (see =-=[69]-=-), where the central value in the ordered list is used for the new value of the brightness. The median filter is much better at preserving straight edge structure than Gaussian smoothing, but if the e... |

849 | E.: “Theory of edge detection
- Marr, Hildreth
- 1980
(Show Context)
Citation Context ...ks to approximate the first derivative of the image brightness function, thus enhancing edges; e.g., see [50] and [67]. These filters give very little control over smoothing and edge localization. In =-=[34]-=- Marr and Hildreth proposed the use of zero crossings of the Laplacian of a Gaussian (LoG). Contours produced using the LoG filter have the property, convenient for some purposes, of being closed. How... |

737 | Visual Reconstruction
- Blake, Zisserman
- 1987
(Show Context)
Citation Context ...ing the signal to noise ratio. (In [9] Canny makes a similar point, when discussing the symmetric and antisymmetric components of a filter.) Other work taking this approach includes [48] and [53]. In =-=[6]-=- Blake and Zisserman used a weak membrane model in the framework of statistical regularization. The elements of the cost functional (which is to be globally minimized) are the smoothness of the estima... |

539 |
Scale-Space Filtering
- Witkin
- 1983
(Show Context)
Citation Context ...iour of the SUSAN edge detector, scale-space graphs showing edge localization against mask size (e.g., plotting a single horizontal line from the edge image against mask size, in the manner of Witkin =-=[75]-=-) give vertical lines. (Most edge detectors do not give scale-invariant edge positions, thus producing curves in scale-space graphs.) This is obviously a desirable feature, as it means that accuracy d... |

386 |
Scale space and edge detection using anisotropic diffusion
- Perona, Malik
(Show Context)
Citation Context ... between these two quantities and also on the local brightness variance. This is similar to the contrast-dependent thresholded median filter, and again, is not very good at structure preservation. In =-=[47]-=- Perona and Malik use local image gradient to control anisotropic diffusion; smoothing is prevented from crossing edges. This is, in a sense, an opposite approach to the K-nearest neighbour, sigma and... |

244 |
Machine perception of three-dimensional solids
- Roberts
- 1965
(Show Context)
Citation Context ...scheme, comparing the output from the first stage of the SUSAN edge detector with the four filters tried in [71], namely, Sobel, Prewitt, Roberts and the "three-point energy model" 3 (see [6=-=7], [50], [51]-=- and [72]). A test image identical to the image described was created (i.e. a vertical step edge with added Gaussian noise) giving a signal to noise ratio of 14.79. The performance of the initial stag... |

235 |
Using canny's criteria to derive a recursively implemented optimal edge detector
- Deriche
- 1987
(Show Context)
Citation Context ...uristic extensions to the Canny algorithm to enable the joining of open contour ends with nearby contours. This however produces some false edge extensions. 7 work well on edges of high curvature. In =-=[14]-=- and [56] similar analytical approaches to that of Canny are taken, resulting in efficient algorithms which have exact recursive implementations. These algorithms give results which are very similar t... |

208 |
E.: A fast operator for detection and precise location of distinct points, corners and centres of circular features
- Förstner, Gülch
- 1987
(Show Context)
Citation Context ...ct position of the junction. In [43] Noble relates the Plessey operator to the four "greylevel corner detectors" mentioned above, and shows that it can be viewed as measuring image curvature=-= also. In [19] Forstner -=-and Gulch describe a method which appears to use exactly the same measure of "cornerness " as the Plessey operator (note the earlier statement that this work predates the Plessey paper). How... |

191 | The curvature primal sketch
- Asada, Brady
- 1986
(Show Context)
Citation Context ...s. A few methods of using binary edge maps to find corners have been suggested. The edges are detected and then edge curvature is calculated in order to find corner locations. See, for example, [20], =-=[3] and [36].-=- Clearly this sort of approach has severe problems with junctions. In [38] and [39], Moravec developed the idea of using "points of interest". These are defined as occurring when there are l... |

186 |
Towards automatic visual obstacle avoidance
- Moravec
- 1977
(Show Context)
Citation Context ... edges are detected and then edge curvature is calculated in order to find corner locations. See, for example, [20], [3] and [36]. Clearly this sort of approach has severe problems with junctions. In =-=[38] and [39],-=- Moravec developed the idea of using "points of interest". These are defined as occurring when there are large intensity variations in every direction. This definition is realized by computi... |

144 |
Parallel and Deterministic Algorithms from MFRs: Surface Reconstruction
- Geiger, Girosi
- 1991
(Show Context)
Citation Context ... i.e., the discontinuities in the membrane. This method does not pick up fine complicated detail very well, and is computationally expensive. This approach has been developed further, for example, in =-=[21], [44] and-=- [68]. Other methods of using "global" support for edge detection include [45] and [23]. The edge detector described in this paper uses a completely new definition of edges, and in doing so,... |

143 |
Finding edges and lines in images
- Canny
- 1983
(Show Context)
Citation Context ...he quality of the results is adversely affected. However a fast algorithm performing well in the other criteria is more desirable than a slow one. Canny has formulated his criteria mathematically (in =-=[8] and [9]). He uses t-=-he criteria functionals to derive "optimized" edge filters for each image type. Criteria 1 and 2 are clearly the most important ones in terms of "optimizing" the filter. Later anal... |

139 |
Digital step edges from zero crossing of second directional derivatives
- Haralick
- 1984
(Show Context)
Citation Context ...lar analytical approaches to that of Canny are taken, resulting in efficient algorithms which have exact recursive implementations. These algorithms give results which are very similar to Canny's. In =-=[22]-=-, Haralick proposes the use of zero crossings of the second directional derivative of the image brightness function. This is theoretically the same as using maxima in the first directional derivatives... |

123 |
Gray-level corner detection
- Kitchen, Rosenfeld
- 1982
(Show Context)
Citation Context ...In [4] Beaudet enhanced high curvature edges (i.e., looked for saddle points in the image brightness surface) by calculating image Gaussian curvature (the product of the two principle curvatures). In =-=[29]-=- Kitchen and Rosenfeld used a local quadratic surface fit to find corners. The parameters of the surface were used to find the gradient magnitude and the rate of change of gradient direction; the prod... |

99 |
Rotationally invariant image operators
- Beaudet
- 1978
(Show Context)
Citation Context ...finding the autocorrelation is square, and the operator is sensitive to strong edges, as the response is given by the minimum in the autocorrelation measurements, and not by the variation in them. In =-=[4]-=- Beaudet enhanced high curvature edges (i.e., looked for saddle points in the image brightness surface) by calculating image Gaussian curvature (the product of the two principle curvatures). In [29] K... |

95 |
Adaptive smoothing: A general tool for early vision
- Saint-Marc, Chen, et al.
- 1991
(Show Context)
Citation Context ...nse, an opposite approach to the K-nearest neighbour, sigma and SUSAN filters, which work out which neighbours to include in smoothing; anisotropic diffusion works out which neighbours to exclude. In =-=[54] Saint-Mar-=-c et al. "refine and generalize" the approach of Perona and Malik, using many iterations of a 3 by 3 mask which weights a mean of the neighbours' values according to the edge gradient found ... |

83 |
On Detecting Edges
- Nalwa, Binford
- 1986
(Show Context)
Citation Context ...oblem with the mathematical model approach is that deviations from the surface model cannot (by definition) be accommodated by varying the model's parameters. Work taking this approach includes [28], =-=[42]-=- and [59]. In [43] Noble uses mathematical morphology to find image structure. Several different morphological operations are described; these are used to enhance edges and find two dimensional featur... |

80 | Detecting and localizing edges composed of steps, peaks and roofs,” ICCV
- Perona, Malik
- 1990
(Show Context)
Citation Context ...expense of optimizing the signal to noise ratio. (In [9] Canny makes a similar point, when discussing the symmetric and antisymmetric components of a filter.) Other work taking this approach includes =-=[48]-=- and [53]. In [6] Blake and Zisserman used a weak membrane model in the framework of statistical regularization. The elements of the cost functional (which is to be globally minimized) are the smoothn... |

77 |
Volumetric model and 3d trajectory of a moving car derived from monocular tv frame sequences of a street scene
- Dreschler, Nagel
- 1982
(Show Context)
Citation Context ...e were used to find the gradient magnitude and the rate of change of gradient direction; the product of these quantities was used to define "cornerness", and local maxima were reported as co=-=rners. In [16] Dres-=-chler and Nagel defined corners to be points lying between extrema of Gaussian curvature. Later, (for example, see [41]) Nagel used the definition (of corner position) of "the point of maximum pl... |

76 |
Object enhancement and extraction
- Prewitt
- 1994
(Show Context)
Citation Context ...he FM scheme, comparing the output from the first stage of the SUSAN edge detector with the four filters tried in [71], namely, Sobel, Prewitt, Roberts and the "three-point energy model" 3 (=-=see [67], [50]-=-, [51] and [72]). A test image identical to the image described was created (i.e. a vertical step edge with added Gaussian noise) giving a signal to noise ratio of 14.79. The performance of the initia... |

72 |
Digital image smoothing and the sigma filter
- Lee
- 1983
(Show Context)
Citation Context ...lar filters, the first two of which are typical of more complicated order statistic operators, are the K-nearest neighbour operator (see [13]), the ff-trimmed mean (see [5]) and the sigma filter (see =-=[30]-=-). The K-nearest neighbour operator takes the mean of the K nearest neighbours from within the ordered list. The value which K is usually set to (5 or 6 when a 3 by 3 mask is used) means that corners ... |

62 |
The weighted median filter
- Brownrigg
- 1984
(Show Context)
Citation Context ...salt and pepper" noise) where the noisy pixels contain no information about their original values. There have been several variations on the median filter, for example the weighted median filter =-=(see [7]) selectiv-=-ely gives the neighbouring pixels multiple entries to the ordered list, usually with the centre pixels 17 Sometimes even some structure in the imaged world is treated as unwanted "noise". 18... |

62 | Biased Anisotropic Diffusion - a Unified Regularization and Diffusion Approach to Edge Detection - Nordstrom - 1990 |

54 |
An operator which locates edges in digitized pictures
- Hueckel
- 1971
(Show Context)
Citation Context ...The problem with the mathematical model approach is that deviations from the surface model cannot (by definition) be accommodated by varying the model's parameters. Work taking this approach includes =-=[28]-=-, [42] and [59]. In [43] Noble uses mathematical morphology to find image structure. Several different morphological operations are described; these are used to enhance edges and find two dimensional ... |

52 |
Edge preserving smoothing
- Nagao, Matsuyama
- 1979
(Show Context)
Citation Context ...his approach, tested in [24], were not very encouraging, showing a decrease in the signal to noise ratio with each of four different types of noise. 38 The selected-neighbourhood averaging (e.g., see =-=[40]) method c-=-hooses a local neighbourhood to smooth over based on a variety of criteria. Each pixel can be part of several different "windows"; the one chosen to smooth over is the most homogeneous windo... |

49 |
a new approach to low level image processing
- Susan
- 1995
(Show Context)
Citation Context ...oach to Low Level Image Processing Technical Report TR95SMS1c (Shorter versions have now been published in IJCV [66] and ICPR96 [64], and the relevant patent is [60]. To reference this research, cite =-=[66]-=-) 1995 S.M. Smith Oxford Centre for Functional Magnetic Resonance Imaging of the Brain (FMRIB), Department of Clinical Neurology, Oxford University, Oxford, UK (Previously in Computer Vision and Image... |

48 |
Visual Mapping by a Robot Rover
- Moravec
- 1979
(Show Context)
Citation Context ...e detected and then edge curvature is calculated in order to find corner locations. See, for example, [20], [3] and [36]. Clearly this sort of approach has severe problems with junctions. In [38] and =-=[39], Moravec -=-developed the idea of using "points of interest". These are defined as occurring when there are large intensity variations in every direction. This definition is realized by computing an un-... |

48 |
An optimal linear operator for step edge detection
- Shen, Castan
- 1992
(Show Context)
Citation Context ...xtensions to the Canny algorithm to enable the joining of open contour ends with nearby contours. This however produces some false edge extensions. 7 work well on edges of high curvature. In [14] and =-=[56]-=- similar analytical approaches to that of Canny are taken, resulting in efficient algorithms which have exact recursive implementations. These algorithms give results which are very similar to Canny's... |

46 |
Corner detection and curve representation using cubic b-splines
- Medioni, Yasumoto
- 1987
(Show Context)
Citation Context ... methods of using binary edge maps to find corners have been suggested. The edges are detected and then edge curvature is calculated in order to find corner locations. See, for example, [20], [3] and =-=[36]. Clearly -=-this sort of approach has severe problems with junctions. In [38] and [39], Moravec developed the idea of using "points of interest". These are defined as occurring when there are large inte... |

43 | 3D positional integration from image sequences
- Harris, Pike
- 1987
(Show Context)
Citation Context ...ner finders tested. For a more quantitative test of stability, the output from 10 consecutive frames was used as the first stage of the DROID three dimensional vision system developed by Plessey (see =-=[25]-=- and [10]). This program tracks corners through time in order to reconstruct a three dimensional description of the world. The results obtained when the Plessey corner finder was used were compared wi... |

40 |
Accurate corner detection: An analytical study
- Deriche, Giraudon
- 1990
(Show Context)
Citation Context ...measurement, that is, the product of the gradient magnitude and the edge contour curvature. They are all sensitive to noise, and have relatively low reliability, as shown in various tests in [24]. In =-=[15] Deriche a-=-nd Giraudon use Beaudet's "cornerness" measure in conjunction with edge detection to obtain accurate corner localization. Corners are found at two different scales, and lines are drawn betwe... |

39 |
Corner detection using the facet model
- Zuniga, Haralick
- 1983
(Show Context)
Citation Context ...ying between extrema of Gaussian curvature. Later, (for example, see [41]) Nagel used the definition (of corner position) of "the point of maximum planar curvature for the line of steepest slope&=-=quot;. In [76]-=- Zuniga and Haralick found corners at significant changes in curvature along edges. In [43] Noble shows that these four approaches are all making the same underlying measurement, that is, the product ... |

36 |
Alpha-trimmed means and their relationship to median filters
- Bednar, Watt
- 1984
(Show Context)
Citation Context ...l brightness values. Three similar filters, the first two of which are typical of more complicated order statistic operators, are the K-nearest neighbour operator (see [13]), the ff-trimmed mean (see =-=[5]-=-) and the sigma filter (see [30]). The K-nearest neighbour operator takes the mean of the K nearest neighbours from within the ordered list. The value which K is usually set to (5 or 6 when a 3 by 3 m... |

36 |
A Two-Stage Algorithm for DiscontinuityPreserving Surface Reconstruction
- Sinha, Schunck
- 1992
(Show Context)
Citation Context ...h the mathematical model approach is that deviations from the surface model cannot (by definition) be accommodated by varying the model's parameters. Work taking this approach includes [28], [42] and =-=[59]. In -=-[43] Noble uses mathematical morphology to find image structure. Several different morphological operations are described; these are used to enhance edges and find two dimensional features. The "... |

33 |
Descriptions of Image Surfaces
- Noble
- 1989
(Show Context)
Citation Context ...hematical model approach is that deviations from the surface model cannot (by definition) be accommodated by varying the model's parameters. Work taking this approach includes [28], [42] and [59]. In =-=[43] Nobl-=-e uses mathematical morphology to find image structure. Several different morphological operations are described; these are used to enhance edges and find two dimensional features. The "erodedila... |

30 | Scale{space ltering - Witkin - 1983 |

26 |
Early Jump-Out Corner Detectors
- Cooper, Venkatesh, et al.
- 1993
(Show Context)
Citation Context ...e, and feature location within selected windows. This results in improved localization accuracy with some corner types, but the algorithm is, as one would expect, slower than the Plessey detector. In =-=[12]-=- Cooper et. al. use image patch similarity tests in two edge based corner finding algorithms. In the first method the similarity function (an un-normalized local autocorrelation function) is found at ... |

25 |
A Cost Minimization Approach to Edge Detection Using Simulated Annealing
- Tan, Gelfand
- 1992
(Show Context)
Citation Context ...ontinuities in the membrane. This method does not pick up fine complicated detail very well, and is computationally expensive. This approach has been developed further, for example, in [21], [44] and =-=[68]. Other me-=-thods of using "global" support for edge detection include [45] and [23]. The edge detector described in this paper uses a completely new definition of edges, and in doing so, solves many of... |

24 |
Noise cleaning by iterated local averaging
- Davis, Rosenfeld
- 1978
(Show Context)
Citation Context ...een the maximum and minimum local brightness values. Three similar filters, the first two of which are typical of more complicated order statistic operators, are the K-nearest neighbour operator (see =-=[13]-=-), the ff-trimmed mean (see [5]) and the sigma filter (see [30]). The K-nearest neighbour operator takes the mean of the K nearest neighbours from within the ordered list. The value which K is usually... |

22 |
Modelling and identification of characteristic intensity variations
- Rohr
- 1992
(Show Context)
Citation Context ...used at eight orientations.) Then for each edge point the difference in orientation between the neighbouring edge points is used to find corners. Results are presented for very simple images only. In =-=[52]-=- Rohr uses strict models to detect and localize corners and junctions. The assumptions are made that regions making up junctions have constant underlying values and that junctions are well isolated in... |

22 |
Gradient inverse weighted smoothing scheme and the evaluation of its performance
- Wang, Vagnucci, et al.
- 1981
(Show Context)
Citation Context ...ract to find a stable best estimate of the underlying signal. The results given suggest that this method does not in general work as well as the first two. The gradient inverse weighted operator (see =-=[73]-=-) forms a weighted mean of the local pixels, with the weights depending on the difference between the central pixel's value and the value of the local pixels; J(x; y) = P I(x+i;y+j) max f 1 2 ;jI(x+i;... |

19 |
Box-filtering techniques
- McDonnell
- 1981
(Show Context)
Citation Context ...ructure. The method is based on image region determination using a non-rigid (in fact, spatially undetermined) region model. The simplest noise reduction method is the sliding mean or box filter (see =-=[35]-=-). Each pixel value is replaced by the mean of its local neighbours. The Gaussian filter is similar to the box filter, except that the values of the neighbouring pixels are given different weighting, ... |

18 | Corner detection - Mehrotra, Nichani - 1990 |

14 |
level corner detection a generalization and a robust real time implementation
- Singh, Shneier, et al.
- 1990
(Show Context)
Citation Context ...urring, and the relative positions of the edges forming the junction. Thus feature detection from a large scale viewpoint is performed. Results are only presented for very simple low noise images. In =-=[58]-=- Singh and Shneier attempt to overcome the problem of the trade-off between reliable detection and accurate localization by fusing various corner finding methods. Initially a template based corner det... |

13 |
Iterative enhancement of noisy images
- Lev, Zucker, et al.
- 1977
(Show Context)
Citation Context ...oach takes each symmetric pair of neighbouring pixels and uses the closer value to the centre pixel from each pair to form a mean output value. Again this is not well suited to preserving corners. In =-=[31]-=- three methods of noise reduction using different local image measurements to avoid destroying image structure are presented, all based on the assumption that image regions bounded by edges have const... |

13 |
An energy feature detection scheme
- Venkatesh, Owens
- 1989
(Show Context)
Citation Context ...omparing the output from the first stage of the SUSAN edge detector with the four filters tried in [71], namely, Sobel, Prewitt, Roberts and the "three-point energy model" 3 (see [67], [50],=-= [51] and [72]-=-). A test image identical to the image described was created (i.e. a vertical step edge with added Gaussian noise) giving a signal to noise ratio of 14.79. The performance of the initial stage of the ... |

12 |
Quantitative evaluation of some edge-preserving noise-smoothing techniques
- Chin, Yeh
- 1983
(Show Context)
Citation Context ... amount, only one image was generated, and used to test each of the filtering methods. The Algorithms Tested 20 All brightness values mentioned are within an image brightness scale of 0 to 255. 41 In =-=[11]-=- Chin and Yeh perform testing of the mean, K-nearest neighbour method, one of Lev's approaches (see [31]), the gradient inverse weighted method, selective neighbourhood weighting, Haralick's facet app... |

12 |
der Heydt, “Detection of general edges and keypoints
- Rosenthaler, Heitger, et al.
- 1992
(Show Context)
Citation Context ...f optimizing the signal to noise ratio. (In [9] Canny makes a similar point, when discussing the symmetric and antisymmetric components of a filter.) Other work taking this approach includes [48] and =-=[53]-=-. In [6] Blake and Zisserman used a weak membrane model in the framework of statistical regularization. The elements of the cost functional (which is to be globally minimized) are the smoothness of th... |