Ph.D. Dissertation Proposal Defense "Recognition of Incomplete Objects using LG Graph Modeling" By Michael Christopher Robbeloth
Ph.D. Committee: Drs. Nikolaos Bourbakis (advisor), Soon Chung, Yong Pei, and Arnab Shaw (EE)
ABSTRACT:
The problem of recognizing incomplete objects, is a challenging problem for a variety of reasons. For starters, even complete, fully-separated input images may have textured backgrounds, non-differentiated foreground objects, or artifacts from the capture process (e.g., noise, dust, light, etc.) that interfere with the identification process and can lead to misclassification. Furthermore, if an object is introduced into the system with key areas missing (particularly its borders) or is inseparable due to obstruction or merging within the proximity of a larger object’s extent, the problem is made more difficult. Sometimes, even with a whole object present, it may be altered from its original form in such a manner that its deformations make accurate recognition difficult or impossible. Accurate object recognition, particularly of incomplete or occluded objects, is a problem that is of ongoing interest to researchers in the computer vision domain.
In this proposal we offer an approach for handling the recognition of incomplete objects. The approach deconstructs the obstructed image into clusters, which are then refined into segments or nodes that be matched in a probabilistic manner using a variety of methods such as string similarity methods of the chain codes describing the borders of the nodes. The deconstruction itself involves k-means segmentation followed by quantification of the borders of the image using the local-global graph approach. The resulting graph description is matched to a set of preprocessed model images stored using six different views to find the closest probable matching side. The model images were preprocessed in the same manner as any obstructed images submitted to the system for matching. The use of synthesis of deconstructed nodes into macro nodes might aid in the effort to produce a better match. In addition, it might be helpful to use secondary characteristics such as the relationship between angle measures among the nodes’ centroids or the properties of the lines making of the border of a node to produce matches.
BRIEF BIOSCKECH: Michael Robbeloth is an assistant professor of Computer Science at Mount Vernon Nazarene University (MVNU) in Mount Vernon, Ohio. In the past, Mr. Robbeloth has worked in a variety of positions as a software engineer, supervisor of embedded software engineering, and as a computer scientist at companies including PDi Communication Systems in Springboro, Ohio; Data Science Automation (as a contractor at WPAFB for AFRL/RBSD), and Eastman Kodak in Dayton, Ohio among other companies. Presently Mr. Robbeloth is working with Dr. Nikolaos Bourbakis, Distinguished Professor in Informatics & Technology and the Director of the Center of Assistive Research Technologies (CART) at Wright State, on methodologies for working with incomplete objects in the areas of computer vision and computational geometry. Mr. Robbeloth hopes to be able to continue his work in these areas with his own students in the future.
