| Project/Area Number |
13680470
|
|---|
| Research Category |
Grant-in-Aid for Scientific Research (C)
|
| Allocation Type | Single-year Grants |
|---|
| Section | 一般 |
|---|
| Research Field |
Intelligent informatics
|
|---|
| Research Institution | Okayama University of Science |
|---|
Principal Investigator |
LASHICIA George Okayama University of Science Dept. of Infor and Comp. Eng. Associate Professor, 工学部, 助教授 (10251745)
|
|---|
| Project Period (FY) |
2001 – 2002
|
| Project Status |
Completed (Fiscal Year 2002)
|
| Budget Amount *help |
¥1,400,000 (Direct Cost: ¥1,400,000)
Fiscal Year 2002: ¥700,000 (Direct Cost: ¥700,000)
Fiscal Year 2001: ¥700,000 (Direct Cost: ¥700,000)
|
|---|
| Keywords | crack detection / feature selection / concept learning / relevant features / relevant examples / noise filtering / 特徴量 / 有利な特徴量 / 有利なサンプル |
|---|
| Research Abstract |
The first part of the research project was dedicated to the selection of features that are relevant for cracks detection. Since the number of relevant features is large, the use of feature reduction methods is necessary in purpose to develop a practical real time processing system. The focus was moved to the selection of relevant irredundant features. An algorithm called PTD was developed for the selection of relevant, indundant features and it is available at http://lacom2.ice.ous.ac.jp/lash/rel.html. A research was also conducted on selection of relevantexamples. A new noise elimination method which is based on the filtering of the so called pattern frequency domain and which resembles frequency domain filtering in signal and image processing was proposed. Noise elimination is achieved by identifying examples that are non-typical in the determination of irredundant sets of relevant attributes. Empirical results show the effectiveness of the proposed example selection method on artificial and real databases. A novel approach to inductive learning based on a 'conflict estimation based learning' (CEL) algorithm was also proposed. CEL is a new learning strategy, and unlike conventional methods CEL does not construct explicit abstractions of the target concept. Instead, CEL classifies unknown examples by adding them to each class of the training examples and measuring how much noise is generated. Empirical results showed that CEL could generate improved classification accuracy over popular, conventional classifiers
|
