2D object recognition by adaptive feature extraction and dynamical link graph matching
Kenneth A. Flaton · University of Southern California Digital Library · 2015
We present a mid-level system for 2D object recognition, comprised of an adaptive feature extractor and a dynamical link graph matcher. The system is robust to variations caused by translation, scale, perspective, lighting, and partial occlusion. The feature extractor performs a Morlet wavelet decomposition that results in a feature vector for each point in the image. The feature vectors are quantized to a set of learned model vectors and labeled. Using a saliency measure derived during the decomposition, an object graph is formed and compared to graphs stored in the associative memory of the dynamical link graph matcher. The graph matcher that has formed the basis for our work comes from (von der Malsburg and Bienenstock, 1987). That system, not intended as a robust recognition system, was the first to use dynamical links in graph matching and demonstrated a rudimentary invariant recognition capability. We have improved on this graph matcher by providing the capacity for a richer (than binary) feature set and by increasing the processing speed through improved dynamics and a reduction in the number of nodes needed to represent an object. Our system uses a number of biologically plausible mechanisms (Gabor filters, temporal correlations, neural networks) and is in other ways guided by biological principles (local connections, saliency, internal focus of attention, self-organizing topological maps, orientation column tuning). The approach is that of an unstructured system, that is, a system in which no domain knowledge is initially embedded but which the system may learn; consequently, the system may be applied to a variety of domains with little or no alteration in architecture or dynamics. Modest extensions to the current system may include multi-modal recognition (fusion), expectation-driven perception, and 3D object recognition. (Copies available exclusively from Micrographics Department, Doheny Library, USC, Los Angeles, CA 90089-0182.)