## Distortion Invariant Object Recognition in the Dynamic Link Architecture (1993)

### Download From

IEEE### Download Links

- [ftp.neuroinformatik.ruhr-uni-bochum.de]
- [www.vision.caltech.edu]
- DBLP

### Other Repositories/Bibliography

Venue: | IEEE Trans. Computers |

Citations: | 491 - 55 self |

### BibTeX

@ARTICLE{Lades93distortioninvariant,

author = {Martin Lades and Jan C. Vorbrüggen and Joachim Buhmann and Jörg Lange and Christoph V. D. Malsburg and Rolf P. Würtz and Wolfgang Konen},

title = {Distortion Invariant Object Recognition in the Dynamic Link Architecture},

journal = {IEEE Trans. Computers},

year = {1993},

volume = {42},

pages = {300--311}

}

### Years of Citing Articles

### OpenURL

### Abstract

We present an object recognition system based on the Dynamic Link Architecture, which is an extension to classical Artificial Neural Networks. The Dynamic Link Architecture exploits correlations in the fine-scale temporal structure of cellular signals in order to group neurons dynamically into higher-order entities. These entities represent a very rich structure and can code for high level objects. In order to demonstrate the capabilities of the Dynamic Link Architecture we implemented a program that can recognize human faces and other objects from video images. Memorized objects are represented by sparse graphs, whose vertices are labeled by a multi-resolution description in terms of a local power spectrum, and whose edges are labeled by geometrical distance vectors. Object recognition can be formulated as elastic graph matching, which is performed here by stochastic optimization of a matching cost function. Our implementation on a transputer network successfully achieves recognition ...

### Citations

1313 |
Self-Organization and Associative Memory
- Kohonen
- 1984
(Show Context)
Citation Context ...ich our system may be compared are methods using correlation or template matching. In the simplest version, correlation is performed by rigidly comparing stored patterns to the image, either directly =-=[27]-=- or by rst comparing a small set of canonical patterns (eigen-patterns characteristic for an object class) and then treating their set of amplitudes as a global feature vector for the recognition prop... |

607 | Relations between the statistics of natural images and the response properties of cortical cells
- Field
- 1987
(Show Context)
Citation Context ...exponential removes the dc-component of ~ k . Equation (3) does not normalize the energy picked up by akernel in the convolution. Consequently, this energy will be proportional to j ~ kj 2 . D. Field =-=[19]-=- noted that the power spectrum of \natural images" decreases like 1=j ~ kj 2 . The energy in the resulting components of our image transform should therefore be roughly independent ofj ~ kj, an assump... |

541 |
Application of the karhunen-loeve procedure for the characterization of human faces
- Kirby, Sirovich
- 1990
(Show Context)
Citation Context ...t comparing a small set of canonical patterns (eigen-patterns characteristic for an object class) and then treating their set of amplitudes as a global feature vector for the recognition proper [28], =-=[29]-=-. These methods have the weakness that they cannot deal even with position invariance. Position (and size) invariance can be achieved by pre-processing the image by taking the magnitude of the Fourier... |

436 |
Low-dimensional procedure for the characterization of human faces
- Sirovich, Kirby
- 1987
(Show Context)
Citation Context ... by rst comparing a small set of canonical patterns (eigen-patterns characteristic for an object class) and then treating their set of amplitudes as a global feature vector for the recognition proper =-=[28]-=-, [29]. These methods have the weakness that they cannot deal even with position invariance. Position (and size) invariance can be achieved by pre-processing the image by taking the magnitude of the F... |

328 |
Oscillatory responses in cat visual cortex exhibit inter–columnar synchronization which re�ects global stimulus properties
- Gray, Konig, et al.
- 1989
(Show Context)
Citation Context ...ivity patterns there is a number of useful structures. A large network can spontaneously decompose into smaller blocks|segments|, a process that is important for scene segmentation [6]{[9], (see also =-=[10]-=-). Another type of network pattern has the form of two-dimensional graphs with short-range connections, just as are needed for the representation of images. Finally, network self-organization can acti... |

262 | An evaluation of the two-dimensional Gabor filter model of simple receptive fields in the cat striate cortex - Jones, Palmer - 1987 |

236 |
Computers and Intractability
- Garey, Johnson
- 1982
(Show Context)
Citation Context ...ons Many researchers feel that the computational complexity of attributed graph matching is prohibitive. Their argument is mainly based on the NP-completeness of the general subgraph matching problem =-=[22]-=-. However, this worst case argument isnot applicable (and convincing) for two reasons: i) pattern recognition problems like the one addressed here are solved satisfactorily in almost all situations if... |

169 |
An Analogue Approach to the Traveling Salesman Problem using an Elastic Net Method", Nature 326
- Durbin, Willshaw
- 1987
(Show Context)
Citation Context ...akes the general subgraph matching problem intractable. Neural network systems are especially designed for an e cient search of average case approximations due to their stochastic and analogue nature =-=[23]-=-. Furthermore, ii) the combinatorial explosion of possible matching solutions is dramatically reduced by the constraint that nodes are distinguishable by their attributes and that we are mainly dealin... |

125 |
A neural cocktail-party processor
- Malsburg, Schneider
- 1986
(Show Context)
Citation Context ...rns." Among connectivity patterns there is a number of useful structures. A large network can spontaneously decompose into smaller blocks|segments|, a process that is important for scene segmentation =-=[6]-=-{[9], (see also [10]). Another type of network pattern has the form of two-dimensional graphs with short-range connections, just as are needed for the representation of images. Finally, network self-o... |

66 |
der Malsburg. The correlation theory of brain function
- von
- 1981
(Show Context)
Citation Context ...bject recognition, neural network, wavelet. I. Introduction THIS paper describes an object recognition system based on a new neural information processing concept, the Dynamic Link Architecture (DLA) =-=[1]-=-{[3]. The DLA, rst proposed in 1981, attempts to solve certain conceptual problems of conventional Arti cial Neural Networks. One of the most prominent among these is the expression of syntactical rel... |

53 |
Sensory segmentation with coupled neural oscillators
- Malsburg, Buhmann
- 1992
(Show Context)
Citation Context ..." Among connectivity patterns there is a number of useful structures. A large network can spontaneously decompose into smaller blocks|segments|, a process that is important for scene segmentation [6]{=-=[9]-=-, (see also [10]). Another type of network pattern has the form of two-dimensional graphs with short-range connections, just as are needed for the representation of images. Finally, network self-organ... |

49 |
Position, rotation, and scale invariant optical correlation
- Casasent, Psaltis
- 1976
(Show Context)
Citation Context ...ey cannot deal even with position invariance. Position (and size) invariance can be achieved by pre-processing the image by taking the magnitude of the Fourier-transform (resp., the Mellin transform) =-=[30]-=-{[32]. Systems based on the magnitude of Fourier components, however, have the drawback of being extremely sensitive to image distortions, as produced by perspective transformations, for example. An a... |

45 |
Statistical Mechanics as the Underlying Theory of Elastic and Neural Optimisations
- Simic
- 1990
(Show Context)
Citation Context ...been proposed which represent each potential link between image and model by a neuron [14]. Topology constraints can then be implemented in a soft fashion by connections between those neurons. Simic (=-=[33]-=-, [34]) has analytically compared that class of graph matching methods to another one that is called elastic matching and that includes our type of approach. He found that graph matching systems that ... |

44 |
Elastic matching of line drawings
- Burr
- 1981
(Show Context)
Citation Context ...ng has often been discussed and advocated in the context of object recognition. Our pattern recognition system is based on a special form of attributed graph matching which resembles elastic matching =-=[11]-=-, [12]: an attributed graph in the model domain, encoding an object, is locally distorted to cope with deformations and changes in perspective. In the mathematical literature, the relation to be estab... |

44 | Modeling perceptual grouping and figure-ground segregation by means of active reentrant connections - Sporns, Tononi, et al. - 1991 |

38 |
Isomorphism testing for graphs of bounded genus
- Miller
- 1980
(Show Context)
Citation Context ... matching solutions is dramatically reduced by the constraint that nodes are distinguishable by their attributes and that we are mainly dealing with planar or other low-dimensional graphs (see, e.g., =-=[24]-=- for a discussion of e cient algorithms). The hierarchical representation of our image data suggests that the average running time of our matching algorithm scales linearly with the number of resoluti... |

34 |
der Malsburg. How patterned neural connections can be set up by self-organization
- Willshaw, von
- 1976
(Show Context)
Citation Context ...3], [13], [14]. The third process, paring down many-to-many connections between topological connectivity patterns to one-to-one connections, has been studied extensively, both in computer simulations =-=[15]-=-, [16] and analytically [17]. Essential for this process are events with simultaneous activation of a block of neighboring cells inI and a block of neighboring cells inM, in corresponding positions. A... |

22 |
Evidence for edge and bar detectors in human vision
- Burr, Morrone, et al.
- 1989
(Show Context)
Citation Context ...ximation to the sensitivity pro les of neurons found in visual cortex of higher vertebrates (see, e.g., [20]). There is evidence that those cells tend to come in pairs with even and odd symmetry (see =-=[21]-=- and references from [19]), similar to the real and imaginary parts of (2). In the choice of the actual values of relative bandwidth and the sampling density of ~ k, which are closely related, we dive... |

20 |
An evaluation of the twodimensional Gabor ®lter model of simple receptive ®elds in cat striate cortex
- Jones, Palmer
- 1987
(Show Context)
Citation Context ...er of oscillations under the envelope function. The Gabor-based wavelets seem to be a good approximation to the sensitivity pro les of neurons found in visual cortex of higher vertebrates (see, e.g., =-=[20]-=-). There is evidence that those cells tend to come in pairs with even and odd symmetry (see [21] and references from [19]), similar to the real and imaginary parts of (2). In the choice of the actual ... |

18 | Pattern recognition and associative memory as dynamical processes in a synergetic system. i. translational invariance, selective attention, and decomposition ofscenes - Fuchs, Haken - 1988 |

17 |
der Malsburg, “Pattern recognition by labeled graph matching
- von
- 1988
(Show Context)
Citation Context ...been exploited and proved to be e ective inanumber of simulations [6]{[9]. The second of these processes, retrieval of a graph from an associative memory for graphs, has been demonstrated before [3], =-=[13]-=-, [14]. The third process, paring down many-to-many connections between topological connectivity patterns to one-to-one connections, has been studied extensively, both in computer simulations [15], [1... |

17 |
der Malsburg, "Pattern segmentation in associative memory
- Wang, Buhmann, et al.
- 1990
(Show Context)
Citation Context ...tural extensions. For instance, visual segmentation can be achieved with the help of mechanisms that induce temporal signal correlations within segments and anticorrelations between segments [7]{[9], =-=[35]-=-, and composite objects and scenes can be represented in a hierarchically structured system where the nodes of graphs are themselves smaller graphs to some depth of recursion. Once developed to its fu... |

15 |
der Malsburg. Distortion invariant object recognition by matching hierarchically labeled graphs
- Buhmann, Lange, et al.
- 1989
(Show Context)
Citation Context ...phasis on a neural style of formulation. Our speci c implementation in more conventional algorithmic fashion is described in Section III, which can be read independently. For a preliminary report see =-=[4]-=-. A. The Representation Domains Although object representation and object recognition eventually will have to be implemented in a multi-level structure, we will restrict ourselves here for reasons of ... |

14 |
der Malsburg. Size and distortion invariant object recognition by hierarchical graph matching
- Buhmann, Lades, et al.
- 1990
(Show Context)
Citation Context ...age domain and depends only on the correspondence of feature types inI andM. More generally, further invariances like scale or rotational invariance can be implemented (for an algorithmic version see =-=[18]-=-) by extending the meaning of \feature correspondence" in the construction of the T IM -connections: if, e.g., scale changes are to be accommodated, feature correspondence has to allow for connections... |

9 |
Recognition of topological features of graphs and images in neural networks
- Kree, Zippelius
- 1988
(Show Context)
Citation Context ...xploited and proved to be e ective inanumber of simulations [6]{[9]. The second of these processes, retrieval of a graph from an associative memory for graphs, has been demonstrated before [3], [13], =-=[14]-=-. The third process, paring down many-to-many connections between topological connectivity patterns to one-to-one connections, has been studied extensively, both in computer simulations [15], [16] and... |

8 | der Malsburg: "Distortion invariant object recognition by matching hierarchically labeled graphs - Buhmann, Lange, et al. - 1989 |

7 |
Issues of representation in neural networks
- Bienenstock, Doursat
- 1991
(Show Context)
Citation Context ...ral correlationsC I x ;y = h I x I y it. These correlations are induced by the excitatory connections withinI. The key feature introduced with such correlations is a higher-order neural coding scheme =-=[5]-=- as opposed to rst-order coding schemes (mean ring rates) used in conventional neural networks. Four types of bindings are relevant for the task of object recognition and object representation: i) bin... |

7 | der Malsburg, "The correlation theory of brain function". Max-PlanckInstitute for Biophysical Chemistry (Reprinted in Models of neural networks - von - 1994 |

7 | Statistical mechanics as the underlying theory of "elastic" and "neural - Simic - 1990 |

6 |
der Malsburg. Development of retinotopic projections – an analytical treatment
- Häussler, von
- 1983
(Show Context)
Citation Context ...ocess, paring down many-to-many connections between topological connectivity patterns to one-to-one connections, has been studied extensively, both in computer simulations [15], [16] and analytically =-=[17]-=-. Essential for this process are events with simultaneous activation of a block of neighboring cells inI and a block of neighboring cells inM, in corresponding positions. All these cells have potentia... |

5 | der Malsburg, "How patterned neural connections can be set up by selforganization - Willshaw, von - 1976 |

4 |
der Malsburg, “A marker induction mechanism for the establishment of ordered neural mappings: Its application to the retinotectal problem
- Willshaw, von
- 1979
(Show Context)
Citation Context ...3], [14]. The third process, paring down many-to-many connections between topological connectivity patterns to one-to-one connections, has been studied extensively, both in computer simulations [15], =-=[16]-=- and analytically [17]. Essential for this process are events with simultaneous activation of a block of neighboring cells inI and a block of neighboring cells inM, in corresponding positions. All the... |

3 |
Anwendung der Korrelationstheorie der Hirnfunktion auf das akustische Figur-Hintergrund-Problem (Cocktailparty-Effekt
- Schneider
- 1986
(Show Context)
Citation Context ...elp of natural extensions. For instance, visual segmentation can be achieved with the help of mechanisms that induce temporal signal correlations within segments and anticorrelations between segments =-=[7]-=-{[9], [35], and composite objects and scenes can be represented in a hierarchically structured system where the nodes of graphs are themselves smaller graphs to some depth of recursion. Once developed... |

3 | Modeling perceptual grouping and gure-ground segregation by means of active reentrant connections - Sporns, Tononi, et al. - 1991 |

2 |
nets for graph matching and other quadratic assignment problems
- “Constrained
- 1991
(Show Context)
Citation Context ...roposed which represent each potential link between image and model by a neuron [14]. Topology constraints can then be implemented in a soft fashion by connections between those neurons. Simic ([33], =-=[34]-=-) has analytically compared that class of graph matching methods to another one that is called elastic matching and that includes our type of approach. He found that graph matching systems that enforc... |

2 | der Malsburg, "Pattern recognition by labeled graph matching - von - 1988 |

2 | der Malsburg, "Pattern segmentation in associative memory - Wang, Buhmann, et al. - 1990 |

1 | are nervous structures organized?," in Synergetics of the - How - 1983 |

1 |
structures with dynamical links," Berichte der Bunsengesellschaft fur
- \Nervous
- 1985
(Show Context)
Citation Context ...t recognition, neural network, wavelet. I. Introduction THIS paper describes an object recognition system based on a new neural information processing concept, the Dynamic Link Architecture (DLA) [1]{=-=[3]-=-. The DLA, rst proposed in 1981, attempts to solve certain conceptual problems of conventional Arti cial Neural Networks. One of the most prominent among these is the expression of syntactical relatio... |

1 |
dynamic model for image registration
- \A
- 1981
(Show Context)
Citation Context ... often been discussed and advocated in the context of object recognition. Our pattern recognition system is based on a special form of attributed graph matching which resembles elastic matching [11], =-=[12]-=-: an attributed graph in the model domain, encoding an object, is locally distorted to cope with deformations and changes in perspective. In the mathematical literature, the relation to be established... |

1 |
A Transputer-based neural object recognition system," in From Pixels to Features II---Parallelism
- Wurtz, Vorbruggen, et al.
- 1991
(Show Context)
Citation Context ...it is therefore easily implemented on parallel computers. We use a system based on the transputer, a microprocessor with integrated support for a message-passing, distributed memory MIMD architecture =-=[25]-=-. Our machine consists of 23 transputers, one of which hosts the development system, one is a combined frame grabber/graphics display, and the others are used as a processor farm. We programmed in occ... |

1 | der Malsburg, "A marker induction mechanism for the establishment of ordered neural mappings - Willshaw, von - 1979 |

1 | der Malsburg, "Development of retinotopic projections---An analytical treatment - Haussler, von - 1983 |

1 | der Malsburg, "Size and distortion invariant object recognition by hierarchical graph matching - Buhmann, Lades, et al. - 1990 |