Oscillatory Dynamic Link Matching for Pattern Recognition Ramin Pichevar and Jean Rouat D´epartement GEI, Universit´e de Sherbrooke, J1K 2R1, QC, Canada ERMETIS, Universit´e du Qu´ebec, Chicoutimi, G7H 2B1, QC, Canada pichevar@hermes.usherb.ca, jean.rouat@ieee.org ABSTRACT 1 Introduction The ”dynamic link matching” (DLM) has been first proposed by Konen et al. [1] to solve the visual correspondence problem. The approach consists of two layers of neurons connected to each other through synaptic connections constrained to some normalization. The reference pattern is applied to one of the layers and the pattern to be recognized to the other. The dynamics of the neurons are chosen in such a way that ”blobs” are formed randomly in the layers. If the features of two blobs, each belonging to a different layer, are sufficiently similar, then weight strengthening between the blobs and activity similarity will be observed between sets of similar blobs. The size of the blobs remains fixed during all the simulation. In the original DLM network, the behavior is based on rate coding (averaged neuron activity over time is encoded). Here, we propose the Oscillatory Dynamic Link Matching algorithm (ODLM) that uses models of conventional spiking neurons and for which the coding is based on phase (place coding). We observe that the network is capable of performing motion analysis without optical flow computation and no additional signal processing should be made between layers unlike in [2] (translation, rotation, etc. between the patterns in the first and second layers can be seen as motion). More generally, our proposed network can solve the correspondence problem, and at the same time, performs the segmentation of the scene, which is in accordance with the Gestalt theory of perception. 2 The oscillatory dynamic link matcher The building blocks of this network are oscillatory neurons [3]. The dynamics of this kind of neurons is governed by a modified version of the Van der Pol relaxation oscillator (called the Wang-Terman oscillator). There is an active phase when the neuron spikes and a relaxation phase when the neuron is silent. A neighborhood of 4 is chosen in each layer for the connections (Fig. 1). A global controller is connected to all neurons in the first and second layers as in [4]. During the first processing stage, the two layers are not connected and image segmentation is performed in each layer. Then, during the pattern matching stage, each neuron in the first layer is connected to all neurons in the second layer and vice-versa (extra-layer connections). The intra and extra-layer connection weights w i,j,k,m (t) are defined as follows: w i,j,k,m (t)= 0.25 Card{N int (i,j ) ∪ N ext (i,j )}e λ|p(i,j;t)-p(k,m;t)| (1) where p(i,j ; t) is the external input to neuron i,j . Card{N int (i,j )} is a normalization factor and is equal to the cardinal number (number of elements) of the set N int (i,j ) that comprises the neighbor neurons connected to neuron i,j . It is equal to 4, 3 or 2 depending on the location of the neuron on the map (i.e. center, corner, etc.) and on the number of active connections. A connection is active when H (w i,j,k,m − 0.01) = 1, and is true both for intra-layer and extra-layer