Abstract: We present here a biologically motivated system for invariant and
robust recognition of objects from camera images. It originally arose
from a homework assignment for a course of neural network
self-organization at USC, and in a way it can be seen as a serious test of
NSL's maturity as a (neural) simulation tool. Formulated as a large
system of coupled non-linear differential equations comprising altogether
approximately 3 million variables, its development required extensive
series of experiments and continuous graphical monitoring of large sets of
variables. Not only did NSL support this process, requiring just minor
extensions, but it now makes our system directly accessible to students
and colleagues for close inspection and for further development.
Our model is based on the principles of temporal feature binding and
dynamic link matching. Objects are stored in the form of
two-dimensional aspects. These are competitively matched against
current images. During the matching process, complete matrices of
dynamic links between the image and all models are refined by a
process of rapid self-organization, the final state connecting only
corresponding points in image and object models. As data format for
representing images we use local sets (jets) of Gabor-based
wavelets. We have tested the performance of our system by having it
recognize human faces against data bases of more than one hundred
images. The system is invariant with respect to retinal position,
and it is robust with respect to head rotation, scale, facial
deformation, and illumination.