CCIS Seminar: Texture Classification on Wood Images for Species Recognition

Dear all,

You are cordially invited to the 2nd CCIS Seminar on this coming Friday. The topic of the seminar is on texture analysis and classification.

Date: 27 Feb 2009 (Friday)

Time: 11.30am - 12.30pm 

Venue: PD206 

Speaker: Tou Jing Yi 

Title: Texture Classification on Wood Images for Species Recognition 

Abstract: 

Textures contain details and characteristics that can be physically seen on the surface of an object. Texture classification is a process to analyze and to classify which texture they belong to. It has been widely used in various implementations which use the texture information of the subjects, such as face detection, defects detection and rock classification. Identifying wood species using texture classification techniques is a very recent research and has yet to be widely researched on. The primary aim of this work is to analyze various texture classification techniques to create an algorithm to identify which species a wood belongs to. This thesis provides a framework on how wood species classification could be deployed onto an embedded platform to create an embedded system that provides mobility and compactness. Many embedded platforms are less powerful than regular PC desktops; therefore it limits the performance of applications that can be deployed onto the platform, including the computational capability and its speed. The algorithm is deployed into the Embedded Computer Vision (ECV) platform which is an embedded platform designed for deploying computer vision (CV) algorithms. The ECV platform includes an ARM processing board, a VGA webcam and it communicates with the user through the network. In this work, three different algorithms are implemented on 32 Brodatz textures and wood samples from the Centre for Artificial Intelligence and Robotics (CAIRO) dataset and Forestry and Forest Products Research Institute (FFPRI) dataset: 1) grey level co-occurrence matrices (GLCM); 2) Gabor filters and 3) covariance matrix. The experiments being conducted can be divided into four main phases: 1) GLCM features on a small sample of five wood species; 2) GLCM features on larger number of samples as well as species; 3) GLCM, Gabor filters and covariance matrix on 32 Brodatz textures and; 4) verification-based recognition algorithm on six wood species on the CAIRO dataset. Experimental results show that covariance matrix using Gabor filters to generate feature images provide the best accuracy on the 32 Brodatz textures which is 91.86% while the raw GLCM has the shortest time duration of 50 ms for a 64 × 64 image on the PC platform with an accuracy of 90.86% compared to 924 ms for the combined feature of normalized GLCM and Gabor filter with an accuracy of 91.06%. The best experimental result is 80% on six wood species from the CAIRO dataset. Due to the lack of wood samples, experimental results does not show what can be achieved in the real world with more species involved but has shown possibilities of implementing a high accuracy embedded wood species recognition system in the future.

*This is in conjuction with the work completion seminar for Mr Tou Jing Yi