Method for solving nonlinearity in recognising tropical wood species

Abstract

Classifying tropical wood species pose a considerable economic challenge and failure to classify the wood species accurately can have significant effects on timber industries. Hence, an automatic tropical wood species recognition system was developed at Centre for Artificial Intelligence and Robotics (CAIRO), Universiti Teknologi Malaysia. The system classifies wood species based on texture analysis whereby wood surface images are captured and wood features are extracted from these images which will be used for classification. Previous research on tropical wood species recognition systems considered methods for wood species classification based on linear features. Since wood species are known to exhibit nonlinear features, a Kernel-Genetic Algorithm (Kernel-GA) is proposed in this thesis to perform nonlinear feature selection. This method combines the Kernel Discriminant Analysis (KDA) technique with Genetic Algorithm (GA) to generate nonlinear wood features and also reduce dimension of the wood database. The proposed system achieved classification accuracy of 98.69%, showing marked improvement to the work done previously. Besides, a fuzzy logic-based pre-classifier is also proposed in this thesis to mimic human interpretation on wood pores which have been proven to aid the data acquisition bottleneck and serve as a clustering mechanism for large database simplifying the classification. The fuzzy logic-based pre-classifier managed to reduce the processing time for training and testing by more than 75% and 26% respectively. Finally, the fuzzy pre-classifier is combined with the Kernal-GA algorithm to improve the performance of the tropical wood species recognition system. The experimental results show that the combination of fuzzy preclassifier and nonlinear feature selection improves the performance of the tropical wood species recognition system in terms of memory space, processing time and classification accuracy.