TEXTURE CLASSIFICATION ON WOOD IMAGES FOR SPECIES RECOGNITION By TOU

ii To my family and friends iii ABSTRACT Surface textures are the most salient characteristics of an object as it encode surface details. Texture classification is the process to classify the images into different classes of textures and has been widely used in various implementations based on the textural information of the subjects, such as face detection, defects detection and rock classification. The implementation of it in the identification of wood species is a recent research and has yet to be widely researched on. The primary aim of this work is to study the identification of various wood species. Three texture classification techniques were investigated: 1) grey level co-occurrence matrices (GLCM); 2) Gabor filters, and; 3) covariance matrix; on three different datasets, namely: 1) the 32 Brodatz textures, 2) the wood dataset from the Centre for Artificial Intelligence and Robotics (CAIRO), and 3) the wood dataset from the Forestry and Forest Products Research Institute (FFPRI). Later, a framework was proposed to deploy the wood species recognition system onto an embedded platform to provide mobility and compactness. Here, the Embedded Computer Vision (ECV) platform used, which includes an ARM processing board, a VGA webcam and a network card, were specifically designed for this work and experimental results were encouraging even though the computational capability and speed are limited due to its processing power in comparison to regular PC desktops. Three major work were conducted: 1) on the wood species datasets using GLCM and covariance matrix with the verification-based recognition; 2) on the 32 Brodatz textures using GLCM, Gabor filters, covariance matrix and a few combined algorithms to investigate their accuracy and speed; 3) to determine the time duration of processing raw GLCM on the ECV platform. Experimental results show that the covariance matrix using feature images generated by Gabor filters implemented with the verification-based recognition has the best accuracy of 98.33% on six wood species from the CAIRO dataset. Experimental results also show that the covariance matrix using feature images generated by Gabor filters provide the best accuracy of 91.86% while the raw GLCM has the shortest time duration of 3708 ms for an image of 64 × 64 pixels on the ECV platform with a slightly lower accuracy of 90.86% among all the experimented algorithms on the 32 Brodatz textures. These results have shown huge potential for implementing texture classification techniques on wood species recognition in real time and the …