From nobody@cs.Buffalo.EDU Sun Apr 12 22:14 EDT 1998 From: nobody@cs.Buffalo.EDU Date: Sun, 12 Apr 1998 22:13:53 -0400 (EDT) To: techreps@cs.Buffalo.EDU Subject: techrep: POST request Content-Type: text Content-Length: 3077 Comments: The file is on hadar and a gzip'ed PS file. ContactPerson: soh@eagle.seri.re.kr Remote host: emberiza.seri.re.kr Remote ident: unknown ### Begin Citation ### Do not delete this line ### %R 98-03 %U soh.ps %A Soh, Jung %T A Theory of Document Object Locator Combination %D March 25, 1998 %I Department of Computer Science, SUNY Buffalo %K document analysis, object location, address block location, combination, Korean mail %X Traditional approaches to document object location use a single locator that is expected to locate as many instances of the object class of interest as possible. However, if the class includes subclasses with diverse visual characteristics or is not characterized by easily computable visual features, it is difficult for the single locator to account for wide variation in object characteristics within the class. As a result, increasingly complex models of objects to be located are used. An alternative approach is to combine the decisions of multiple locators, each of which is suitable for certain image conditions. This approach utilizes a collection of simple locators that complement one another, rather than relying on one complex locator. An effective method for combining the location results is vital to the success of this approach. This thesis presents a theory of combining the results of multiple document object locators tuned to different object characteristics. The purpose of the combination is to achieve integrated location performance which is better than that of any individual locator. Location results are represented by objective confidence values, regardless of their types. This allows for combining locators at any output information level. Since different locators use different segmentation algorithms, multiple decisions on a single object are not readily available. Object equivalence and correspondence are used to determine what decisions to combine. The highest confidence, confidence summation, weighted confidence summation, probabilistic confidence estimation, and fuzzy integral methods are proposed for combining confidence values. A method for dynamic selection of locators based on training image set partitioning is proposed. Differences in locator performance levels are represented by locator relevance values, which are used in a combination method and dynamic selection. Measures for evaluating combination performance are proposed and compared. The theory is applied to the address block location problem. A multiple locator algorithm employing a machine-printed address block locator, a handwritten address block locator, and a clustering-based locator is created and tested on a database of 1,100 mail piece images, in two separate experiments. The experimental results show that combining location results using the theory provides location performance that excels any individual locator performance, and therefore is a viable approach to document object location problems.