Optimized texture operators for the automated design of image analysis systems: Non-linear and oriented kernels vs. gray value co-occurrence matrices

Issue title: Hybridization of Intelligent Systems

Guest editors: M. Köppenx and R. Webery

Article type: Research Article

Authors: Peters, Stefanie^a | Koenig, Andreas^b

Affiliations: [a] Fraunhofer Institut Techno- und Wirtschaftsmathematik, Fraunhofer-Platz 1, D-67663 Kaiserslautern, Germany. Tel.: +49 631 316004574; E-mail: [email protected] | [b] TU Kaiserslautern, Lehrstuhl Integrierte Sensorsysteme, Erwin-Schroedinger Str. 12, D-67663 Kaiserslautern, Germany. Tel.: +49 631 2053403; Fax: +49 631 2053889; E-mail: [email protected] | [x] Kyushu Institute of Technology | [y] University of Chile

Abstract: The rapid development in image processing technology allows to tackle applications of increasing complexity. For efficient design of application specific systems, design automation techniques are required. This paper reports on activities for automated texture classification system design employing non-linear oriented kernels (NLOK) configured by evolutionary optimization techniques and swarm optimization (PSO). First and second order statistical features of the dominating kernels in automatically adapted regions of interests serve as features for the texture classification. Our approach was tested with benchmark and application data from leather inspection and was found to be viable and competitive in both cases. The optimized feature set was tested versus features computed from gray value co-occurrence matrices (COOC) with non-optimized parameters. The classification rates for NLOK were significantly higher than for COOC (> 75% vs. < 65%) for benchmark data and slightly higher for application data (> 90% vs. < 90%). Regarding the two optimization approaches, PSO outperformed genetic algorithms (GA) in optimization time and in accuracy for the benchmark data and in accuracy for application data. Additionally, the COOC parameters were adapted as well and yielded higher classification rates than without adaptation.

DOI: 10.3233/HIS-2007-4305

Journal: International Journal of Hybrid Intelligent Systems, vol. 4, no. 3, pp. 185-202, 2007