Author: Professor George Nagy Rensselaer Polytechnic Institute, nagy@ecse.rpi.edu
Advisor: Nikolaos P. Papanikolopoulos (University of Minnesota, USA)
Date of Defense: March 1995
Abstract: This project is part of the long-term goal of optical character recognition (OCR) to automatically design a system that improves with use, and can be adapted to new symbols, fonts, or even foreign scripts. This ambitious goal calls for the automatic design not only of the classifier, but also of the features. Furthermore, in order to satisfy the requirement of adaptivity on short segments of text, we must design the entire system using only a few training samples of each class.
We address the design of the kernel of an adaptable OCR system, i.e., the feature extractor and the pattern classifier. We present a new paradigm in which n-tuple features are automatically designed in conjunction with a probabilistic decision tree to classify isolated printed characters. An n-tuple generating program that enforces more than marginal fit and misfit of the n-tuples on the design samples produces more robust features than earlier Monte Carlo methods. The resulting feature probabilities are accurately estimated using a compound Bayesian procedure in order to delay the fall-off in classification accuracy with tree size that has been the bane of decision trees designed on a small sample set.
On a ten-class confusion set of eight-point x-height characters scanned at 300 dpi, the method yields error rates under 1% with a training set of only 3 samples per class.
Author: Ioannis Pavlidis; Honeywell Technology Center, Honeywell
Inc., MN65-2500, 3660 Technology Dr., Minneapolis, MN 55418;
(pavlidis@htc.honeywell.com)
Advisor: Nikolaos P. Papanikolopoulos (University of Minnesota, USA)
Date of Defense: November 11, 1996
URL: http://www.cs.umn.edu/ pavlidis
Abstract: Significant progress has been achieved in the area of Optical Character Recognition (OCR) the last twenty years. Nevertheless, there are still many open research issues in the field. In particular, results are far from satisfactory in the areas of cursive handwriting and signature recognition. This dissertation addresses the open research problem of recognizing difficult handwritten patterns such as those mentioned above. Two methods have been developed to that effect. One is an off-line, parallel, physics-based method that performs gross classification of handwritten signatures. The innovation of the method is that it avoids elegantly the segmentation problem, which is especially pronounced in the off-line case, while it retains significant discriminating power. This method is well suited for a distributed environment. That is what makes it an ideal search space pruning tool in a pyramidal signature recognition system. The other method is an on-line, physics-based technique that performs accurate classification of handwritten words and simple shapes. The innovation of the approach is that it involves a segmentation method that does not try to locate letters but instead performs the significantly easier task of locating corners and some key low curvature points. This is part of the method's strategy to see the word as a generic on-line shape. Thanks to this strategy, this is the only method that can handle collectively cursive words, and hand-drawn line figures, both useful forms of communication in pen-based computing. Most importantly, the proposed system achieves high recognition rates without ever resorting to complex statistical models. This is made possible because the segmentation data are utilized within the framework of shape metamorphosis, a graphics technique innovatively transplanted to the field of handwriting recognition. Surprisingly, the on-line curve segmentation algorithm, has also been applied with success to an area as diverse as the area of deformable-model-based object tracking.
Melanie A. Sutton Dept. of Computer Science University of
West Florida (msutton@dcsuwf.dcsnod.uwf.edu)
Advisors: Kevin W. Bowyer (University of South Florida) and
Louise Stark (University of the Pacific)
Date of the defense: May 2, 1997
HTTP ADDRESS: http://marathon.csee.usf.edu/
Abstract: The GRUFF-I (Generic Recognition Using Form, Function and Interaction) system reasons about and generates plans for interaction with 3-D shapes for the purpose of generic object recognition. A researcher selects an object and places it in an observation area. An initial intensity and range image are acquired and provided as input to a three-stage recognition system. The first stage builds a 3-D model. The second stage considers the shape-suggested functionality of this model by applying concepts of physics and causation (e.g., to infer stability) to label the object's potential functionality. The third stage uses this labeling to instantiate a plan for interaction to confirm the object's functional use in a task by incorporating feedback from both visual and robotic sensors. Results of this work are presented for eighteen chair-like and cup-like objects. Major conclusions from this work include: (1) metrically accurate representations of the world can be built and used for higher level reasoning, (2) shape-based reasoning prior to interaction-based reasoning provides an efficient methodology for object recognition, in terms of the judicious use of system resources, and (3) interaction-based reasoning helps to confirm the functionality of a categorized object without explicitly determining the object's material composition.
{
Please submit by email ( sarkar@cs.csee.usf.edu) an abstract conforming to the guidelines given below.