Catalogue of Artificial Intelligence Techniques


Jump to: Top | Entry | References | Comments

View Maths as: Images | MathML

Dynamic Link Matching

Aliases: Face Recognition by Dynamic Link Matching

Keywords: dynamic, face, gabor, image, link, mapping, matching, network, neural, neuron, node, similarities

Categories: Pattern Recognition and Image Processing

Author(s): Roderick Hodgson

Dynamic Link Matching is a neural system used for the recognition of objects from realistic images, particularly for face recognition. It was developed by Laurenz Wiskott and Christopher von der Malsburg from the Institut für Neuroinformatik at the Ruhr Universität Bochum. It relies principally on rules of neural networks to find matches between an image domain and a model domain.

The advantages of this system is that it requires very little genetic or learned structure. As it uses rules of rapid synaptic plasticity and the pre-existing constraint of preservation of topography. It is inherently invariant to shift and robust against many other variations, particularly rotation in depth and deformation.

The objects to compare are composed of separate rectangular two-dimensional graphs with a network of simple feature detectors called neurons. Each neuron is a node in the graph and has a local description of the grey value distribution based on the Gabor transform. Topographical relationships are encoded by excitatory (over short range) and inhibitory (over long range) lateral connections between the nodes. The graphs are scaled and aligned manually, such that certain nodes of the graphs are consistent with certain features (such as nose and mouth, in the case of face recognition). The connectivity matrices between the nodes are determined using the similarities between the grey value distribution of the connected nodes.

The Dynamic Link Matching system serves as a process to restructure the connectivity matrices to find the correct mapping between the models and the image. This is done by network self-organisation, using the principle of the synaptic plasticity of the network, that is the change of strength of the connection between nodes (what is called the dynamic link), controlled by signal correlations between nodes.

The models will cooperate with the image depending on their similarity, the system then sequentially rules out the least active and least similar models, until the best fitting model remains.



Add Comment

No comments.