A Biologically Plausible Architecture for Shape Recognition
No Thumbnail Available
Files
Date
2006-08-10
Authors
Journal Title
Series/Report No.
Journal ISSN
Volume Title
Publisher
Abstract
This thesis develops an algorithm for shape representation and matching. The algorithm is an object centered, boundary based method for shape recognition. Global features of the shape are utilized to define a frame of reference relative to which local shape features are characterized. The curvature of the boundary at a point is the local feature used. Curvature is computed by the Digital Straight Segments algorithm. Matching is done using the process of evidence accumulation similar in approach to the Generalized Hough Transform. The algorithm is tested for invariance to similarity transforms. Its robustness to noise and blurring is also tested. A multi-layer, feed-forward neural network architecture that implements the algorithm is proposed.
Description
Keywords
boundary based methods, neural networks, shape recognition
Citation
Degree
MS
Discipline
Electrical Engineering