Date of Graduation
This dissertation explores the Counter Propagation Network (CPN) to implement an image feature finder for automatic image matching systems. This research concludes that the CPN paradigm does not appear to be capable of extracting distinctive features for matching complex images such as fingerprints. There are three possible factors explored in the document that contribute to this inability to match. These are: (1) Image Enhancement - Most of fingerprint images have certain levels of localized blur or smear, locally low contrast, or other noise. Currently available image enhancement methods are shown to not suppress these noises effectively. (2) Image Invariance - Consistently finding the acceptable fingerprint center is difficult. The successful matching rate of the CPN-based fingerprint matching system is 50% or less if the centers between two rollings are off by 4 pixels or more. (3) CPN Feature Finder - It is shown that the CPN paradigm does not effectively extract closely resembled fingerprint features. It is shown that the feature distances among fingerprint images may not be large enough to distinguish them.
Lin, Qiang, "A neural network approach for implementing a high dimension reduction feature finder for automatic image matching." (1996). Graduate Theses, Dissertations, and Problem Reports. 9306.