Zhu, Junmei
Coauthors(s): Christoph von der Malsburg Computer Science Department USC
USC
Computer Science
HNB228 USC
selforg.usc.edu:8376/~junmeizh

Fast Dynamic Link Matching by Communicating Synapses

This work aims to provide a framework for finding continuous mappings between patterns. The importance of mapping can best be demonstrated in one of its subproblems, the correspondence problem in vision, including stereo correspondence, correspondence between consecutive images in time for motion estimation, and the correspondence between image and model for object recognition. As a start, we present a system for creating a mapping between two patterns for object recognition. The two patterns, identified as image and model, can have variations in size and orientation. Among various object recognition approaches, the feature based methods that do not consider spatial relationship suffer from rearrangement ambiguity, and global template matching can not deal with deformations efficiently. Dynamic link matching (DLM) has the advantage that it is intrinsicly invariant to translation and is robust against many other variations, such as deformation. However, DLM is too slow to be the mechanism for object recognition in adults, because mappings are created by running blobs in a sequential manner. The idea is to improve the speed of DLM, by letting synapses talk to synapses directly, and by storing and activating the mappings directly. Moreover, additional mechanisms are introduced for dealing with invariance in scale and orientation. The system is as following. 1. The input is two grey level patterns. Here we use 1D patterns because the result is easier to demonstrate. The goal is to create a continuous mapping so that the correspondence points in these two patterns can be connected. 2. The features we use are Gabor wavelet responses. In each position in the image and the model we compute the Gabor responses with different orientation and scale, forming a jet. 3. Initially the connection pattern is all-to-all. Each connection weight is computed as the similarity of the two jets. Because the similarity function depends on the relative scale and orientation between the image and the model, we have separate connection matrices for a number of discrete scales and orientations, covering the whole range of interest. 4. System dynamics. Connection growth is controlled by cooperation and competition. Cooperation for a synapse comes from synapses that connect neighboring cells in the image and neighboring cells in the model, and from the same scale and orientation only. The region of cooperativity is specific to the scale and orientation. Competition is between synapses diverging from one point in the image, and synapses converging to one point in the model. There is global competition between different scale and orientation. 5. Simulation result. One scale wins out rapidly, and a correct continuous mapping is formed. Even though the scale is sampled discretely, the system is very robust to the scale change, and can deal with continuous scale variation. In summary, we presented a very fast and scale invariant 1D mapping system. Ongoing work includes extending this toy system to two-dimensional patterns and to other variations such as rotation in depth.