## Deformable Kernels for Early Vision (1991)

### Cached

### Download Links

- [authors.library.caltech.edu]
- [vision.caltech.edu]
- [dspace.mit.edu]
- DBLP

### Other Repositories/Bibliography

Venue: | IEEE Transactions on Pattern Analysis and Machine Intelligence |

Citations: | 132 - 9 self |

### BibTeX

@ARTICLE{Perona91deformablekernels,

author = {Pietro Perona},

title = {Deformable Kernels for Early Vision},

journal = {IEEE Transactions on Pattern Analysis and Machine Intelligence},

year = {1991},

volume = {17},

pages = {488--499}

}

### Years of Citing Articles

### OpenURL

### Abstract

Early vision algorithms often have a first stage of linear-filtering that `extracts' from the image information at multiple scales of resolution and multiple orientations. A common difficulty in the design and implementation of such schemes is that one feels compelled to discretize coarsely the space of scales and orientations in order to reduce computation and storage costs. This discretization produces anisotropies due to a loss of traslation-, rotation-, scaling-invariance that makes early vision algorithms less precise and more difficult to design. This need not be so: one can compute and store efficiently the response of families of linear filters defined on a continuum of orientations and scales. A technique is presented that allows (1) to compute the best approximation of a given family using linear combinations of a small number of `basis' functions; (2) to describe all finite-dimensional families, i.e. the families of filters for which a finite dimensional representation is p...

### Citations

3436 | A Computational Approach To Edge Detection
- Canny
- 1986
(Show Context)
Citation Context ...orientations, phases, elongation, scale. This approach is attractive because it is simple to describe, implement and analyze. It has been proposed and demonstrated for a variety of early vision tasks =-=[28, 27, 6, 3, 7, 18, 44, 34, 32, 35, 12, 31, 5, 45, 25, 26, 15, 39, 2]-=-. Various `general' computational justifications have been proposed for basing visual processing on the output of a rich set of linar filters: (a) Koenderink has argued that a structure of this type i... |

1622 | and Complex Analysis - Real - 1974 |

942 |
Receptive fields, binocular interaction and functional architecture in the cat’s visual cortex
- Hubel, Wiesel
- 1962
(Show Context)
Citation Context ...k special interest loci such as edges and junctions. The idea that biological visual systems might analyze images along dimensions such as orientation and scale dates back to work by Hubel and Wiesel =-=[21, 20]-=- in the 1960's. In the computational vision literature the idea of analyzing images along multiple orientations appears at the beginning of the seventies with the Binford-Horn linefinder [19, 4] and l... |

903 | The design and use of steerable filters - Freeman, Adelson - 1991 |

684 | Spatiotemporal energy models for the perception of motion
- Adelson, Bergen
- 1985
(Show Context)
Citation Context ...orientations, phases, elongation, scale. This approach is attractive because it is simple to describe, implement and analyze. It has been proposed and demonstrated for a variety of early vision tasks =-=[28, 27, 6, 3, 7, 18, 44, 34, 32, 35, 12, 31, 5, 45, 25, 26, 15, 39, 2]-=-. Various `general' computational justifications have been proposed for basing visual processing on the output of a rich set of linar filters: (a) Koenderink has argued that a structure of this type i... |

466 | AND HEEGER,D.J. Shiftable multiscale transforms
- SIMONCELLI, FREEMAN, et al.
- 1992
(Show Context)
Citation Context ... the interpolating funtions b(`), (a.2) one would like to build pyramid implementations of the decomposition for speeding up the filtering stage (work on this issue has been done by Simoncelli et al. =-=[11]-=-). 2. Another interesting question mentioned in the introduction is the synthesis of the discrete decomposition directly from the specification of an early vision task, rather than passing through the... |

425 | Multichannel texture analysis using localized spatial filters - Bovik, Clark, et al. - 1990 |

391 | Complete discrete 2D Gabor transforms by neural networks for image analysis and compression - Daugman - 1988 |

338 | Preattentive texture discrimination with early vision mechanisms
- Malik, Perona
- 1990
(Show Context)
Citation Context ...orientations, phases, elongation, scale. This approach is attractive because it is simple to describe, implement and analyze. It has been proposed and demonstrated for a variety of early vision tasks =-=[28, 27, 6, 3, 7, 18, 44, 34, 32, 35, 12, 31, 5, 45, 25, 26, 15, 39, 2]-=-. Various `general' computational justifications have been proposed for basing visual processing on the output of a rich set of linar filters: (a) Koenderink has argued that a structure of this type i... |

291 |
Foundations of Modern Analysis
- Dieudonné
- 1960
(Show Context)
Citation Context ... well defined and is a compact operator. Such a kernel is commonly called a Hilbert-Schmidt kernel. A second result tells us that if a linear operator is compact, then it has a discrete spectrum (see =-=[9]-=- pag. 323): Theorem 3 Let L be a compact operator on (complex) normed spaces, then the spectrum S of L is at most denumerable. A third result says that if L is continuous and operates on Hilbert space... |

261 |
Representation of local geometry in the visual system
- Koenderink, Doorn
- 1987
(Show Context)
Citation Context ...proposed for basing visual processing on the output of a rich set of linar filters: (a) Koenderink has argued that a structure of this type is an adequate substrate for local geometrical computations =-=[29]-=- on the image brightness, (b) Adelson and Bergen [2] have derived it from the `first principle' that the visual system computes derivatives of the image along the dimensions of wavelength, parallax, p... |

229 | Two-dimensional spectral analysis of cortical receptive field profiles - Daugman - 1980 |

228 |
Receptive fields of single neurons in the cat’s striate cortex
- Hubel, Wiesel
- 1959
(Show Context)
Citation Context ...k special interest loci such as edges and junctions. The idea that biological visual systems might analyze images along dimensions such as orientation and scale dates back to work by Hubel and Wiesel =-=[21, 20]-=- in the 1960's. In the computational vision literature the idea of analyzing images along multiple orientations appears at the beginning of the seventies with the Binford-Horn linefinder [19, 4] and l... |

208 | Trace inference, curvature consistency, and curve detection
- Parent, Zucker
- 1989
(Show Context)
Citation Context |

165 | Feature detection in human vision: a phase dependent energy model - Morrone, Burr - 1988 |

134 | Weierstrass and approximation theory - PINKUS - 2000 |

126 |
Feature detection from local energy
- Morrone, Owens
- 1986
(Show Context)
Citation Context |

122 | A computational framework for determining stereo correspondence from a set of linear spatial filters - Jones, Malik - 1992 |

120 | Texture discrimination by Gabor functions
- TURNER
- 1986
(Show Context)
Citation Context |

118 |
Gabor filters as texture discriminator
- Fogel, Sagi
- 1989
(Show Context)
Citation Context |

87 |
Inferring Surfaces from Images
- Binford
- 1981
(Show Context)
Citation Context ...esel [21, 20] in the 1960's. In the computational vision literature the idea of analyzing images along multiple orientations appears at the beginning of the seventies with the Binford-Horn linefinder =-=[19, 4]-=- and later work by Granlund [16]. A computational framework that may be used to performs this proto-visual analysis is the convolution of the image with kernels of various shapes, orientations, phases... |

84 | Detecting and localizing edges composed of steps, peaks and roofs
- Perona, Malik
(Show Context)
Citation Context ...5]. However, one must be aware of the fact that for most kernels F of interest a finite decomposition of F ` as in Eq. (1) cannot be found. For example the elongated kernels used in edge detection by =-=[38, 39]-=- (see Fig. 1 top right) do not have a finite decomposition as in Eq. (1). One needs an approximation technique that, given an F ` , allows one to generate a function G [n] ` which is sufficiently simi... |

74 |
Texture analysis using two-dimensional quadrature
- Knuttson, Granlund
(Show Context)
Citation Context |

62 |
An operator which locates edges in digitized pictures
- Hueckel
- 1971
(Show Context)
Citation Context ...m the specification of an early vision task, rather than passing through the synthesis of a 2D (nD) kernel which then is deformed somewhat arbitrarily. Work in this direction has been done by Hueckel =-=[22, 23]-=-, Hummel [24], and Haralick [17] who approached the problem of feature (step edge, line in [23]) detection and localization as one of projecting image neighbourhoods on small-dimension linear subspace... |

58 |
Optical flow from spatiotemporal filters
- Heeger
- 1987
(Show Context)
Citation Context |

56 | Receptive field families - Koenderink, Doorn - 1990 |

51 | The Gaussian derivative theory of spatial vision: Analysis of cortical cell receptive field line-weighting profiles”. General Motors - Young - 1986 |

36 |
search of a general picture processing operator
- In
- 1978
(Show Context)
Citation Context ... computational vision literature the idea of analyzing images along multiple orientations appears at the beginning of the seventies with the Binford-Horn linefinder [19, 4] and later work by Granlund =-=[16]-=-. A computational framework that may be used to performs this proto-visual analysis is the convolution of the image with kernels of various shapes, orientations, phases, elongation, scale. This approa... |

31 |
A local visual operator which recognizes edges and line
- Hueckel
- 1973
(Show Context)
Citation Context ...m the specification of an early vision task, rather than passing through the synthesis of a 2D (nD) kernel which then is deformed somewhat arbitrarily. Work in this direction has been done by Hueckel =-=[22, 23]-=-, Hummel [24], and Haralick [17] who approached the problem of feature (step edge, line in [23]) detection and localization as one of projecting image neighbourhoods on small-dimension linear subspace... |

27 |
Feature detection using basis functions
- Hummel
- 1979
(Show Context)
Citation Context ...on of an early vision task, rather than passing through the synthesis of a 2D (nD) kernel which then is deformed somewhat arbitrarily. Work in this direction has been done by Hueckel [22, 23], Hummel =-=[24]-=-, and Haralick [17] who approached the problem of feature (step edge, line in [23]) detection and localization as one of projecting image neighbourhoods on small-dimension linear subspaces, and derivi... |

23 | The design of multistage separable planar filters - Treitel, Shanks - 1971 |

22 |
Computational models of visual processing
- Adelson, Bergen
- 1991
(Show Context)
Citation Context |

21 |
Steerable filters for early vision, image analysis and wavelet decomposition
- Freeman, Adelson
- 1990
(Show Context)
Citation Context |

19 |
The design and use of steerable filters for image analysis
- Freeman, Adelson
- 1991
(Show Context)
Citation Context ...suited for edge detection if one wants to recover edge-junctions (see in Fig. 1 the comparison with a detector that uses narrow orientationselective filters). Freeman and Adelson have recently argued =-=[15, 14] that it w-=-ould be desirable to construct orientation-selective kernels that can be exactly rotated by interpolation (they call this property "steerability" and the term will be used in this paper) and... |

16 |
The Binford-Horn line-finder
- Horn
- 1971
(Show Context)
Citation Context ...esel [21, 20] in the 1960's. In the computational vision literature the idea of analyzing images along multiple orientations appears at the beginning of the seventies with the Binford-Horn linefinder =-=[19, 4]-=- and later work by Granlund [16]. A computational framework that may be used to performs this proto-visual analysis is the convolution of the image with kernels of various shapes, orientations, phases... |

15 | Scalespace, its natural operators and differential invariants - Romeny, Rorack, et al. - 1991 |

14 |
Computing visual correspondence
- Kass
- 1983
(Show Context)
Citation Context |

10 |
The laplacian algorithm as a compact image code
- Burt, Adelson
- 1983
(Show Context)
Citation Context |

10 |
Compact image representation from multiscale edges
- Zhong, Mallat
- 1990
(Show Context)
Citation Context |

9 |
Finite representation of deformable functions
- Perona
(Show Context)
Citation Context ...ns T ` belonging to a compact group. This situation arises in the case of n-dimensional rotations and is useful for edge detection in tomographic data and spatiotemporal filtering. It is discussed in =-=[36, 37]-=-. 4 Steerable approximations: practical issues and experimental results In this section the formalism described so far is applied to the problem of generating steerable and scalable approximations to ... |

6 |
Steerable filters for image analysis
- Freeman, Adelson
- 1990
(Show Context)
Citation Context ...depend on the particular group of transformations (the rotations of the plane in this case) that is used to generate F ` from F . 3. The `if ' part of statement 3 is equivalent to the main theorem of =-=[13]-=- -- Freeman and Adelson show that all functions that one can write as a linear combination of functions like the a i 's (polar-separable with sinusoidal ` component) are steerable. The `only if ' part... |

4 |
Digital step edges from zero crossing of second directional derivatives
- Haralik
(Show Context)
Citation Context ...on task, rather than passing through the synthesis of a 2D (nD) kernel which then is deformed somewhat arbitrarily. Work in this direction has been done by Hueckel [22, 23], Hummel [24], and Haralick =-=[17]-=- who approached the problem of feature (step edge, line in [23]) detection and localization as one of projecting image neighbourhoods on small-dimension linear subspaces, and deriving the relevant par... |

4 |
Computational stereopsis{beyond zero-crossings
- Jones, Malik
- 1990
(Show Context)
Citation Context |

3 |
Sequential convolution techniques for image filtering
- Abramatic
- 1982
(Show Context)
Citation Context ... as discussed at the end of Sec. 2. One must remember that beyond X-Y-separation there are a number of techniques for speeding up 2D FIR filtering, for example small generating kernel (SGK) filtering =-=[1]-=-, that could further speed up the convolutions necessary to implement deformable filtering. 5 Rotation and scale A number of filter-based early vision and signal processing algorithms analyze the imag... |

3 |
Using orientation and spatial frequency disparities to recover 3d surface shape { a computational model
- Jones, Malik
- 1991
(Show Context)
Citation Context |

3 |
A local visual operator which recognizes edges and
- Hueckel
- 1973
(Show Context)
Citation Context ...m the specification of an early vision task, rather than passing through the synthesis of a 2D (nD) kernel which then is deformed somewhat arbitrarily. Work in this direction has been done by Hueckel =-=[22, 23]-=-, Hummel [24], and Haralick [17] who approached the problem of feature (step edge, line in [23]) detection and localization as one of projecting image neighbourhoods on small-dimension linear subspace... |

2 |
Lectures on analysis, volume I
- Choquet
- 1969
(Show Context)
Citation Context ...the operator L : A \Gamma! B defined by F as (La(\Delta))(`) = hF (\Delta; `); a(\Delta)i A . A first theorem says that if the kernel F has bounded norm then the associated operator L is compact (see =-=[8]-=- pag. 316): Theorem 2 Let X and T be locally compact Hausdorff spaces and F 2 L 2 (X \Theta T ). Then L is well defined and is a compact operator. Such a kernel is commonly called a Hilbert-Schmidt ke... |

2 |
and Z Gigus, “A model for curvilinear segregation
- Malik
- 1991
(Show Context)
Citation Context ...d phases unspecified) and Jones and Malik [25, 26] (6 scales, 2-6 orientations, 2 phases). Work on curved line grouping by Parent and Zucker [35] (1 scale, 8 orientations, 1phase) and Malik and Gigus =-=[30]-=- (9 curvatures, 1 scale, 18 orientations, 2 phases). Work on brightness edge detection by Binford and Horn [19, 4] (24 orientations), Canny [7] (1-2 scales, 1-6 orientations, 1 phase), Morrone,Owens a... |

2 |
An intelligent image processing display terminal
- Pratt
- 1979
(Show Context)
Citation Context ...f this is true the 2D convolution can be implemented cheaply as a sequence of two 1D convolutions. Even when the kernel F is not X-Y-separable it may be the sum of a small number of separable kernels =-=[41]-=-: F (x; y) = P i f x i (x)f y i (y). One may notice the analogy of this decomposition with the one expressed in Eq. (1): the function F (x; y) may be thought of as a kernel defining an operator from a... |

2 |
bands depend on spatial phase. Nature
- Mach
- 1986
(Show Context)
Citation Context ...tion technique for calculating the value of the filtered image at any point in a continuum; in the case of phases a pair of filters in quadrature can be used for calculating the response at any phase =-=[3, 33]-=-. Rotation, scalings and other deformations are less well understood. An example of 'rotating' families of kernels that have a finite representation is well known: the first derivative along an arbitr... |