3D Computation of Texture Features for Hyperspectral Image Cubes

• Main Authors: Chun-kai Chang, 張鈞凱
• Other Authors: 蔡富安
• Format: Others
• Language: zh-TW
• Published: 2007
• Online Access: http://ndltd.ncl.edu.tw/handle/40116266311872231161

碩士 === 國立中央大學 === 土木工程研究所 === 95 === In recent years, three-dimensional (3D) image formats have become more and more popular, providing the possibility of examining texture as volumetric characteristics. For example, hyperspectral images of remote sensing applications, magnetic resonance imaging and computerized tomography of medical imagery, and seismic data in geology and earthquake researches. Texture-based algorithms are important methods for feature extraction and image analysis. However, traditional texture analysis concentrates on 2D texture properties, few have truly explored the possibility of extending it to 3D forms for volumetric data analysis. This study extended traditional 2D Grey Level Co-occurrence Matrix (GLCM) to a 3D form (Grey Level Co-occurrence Matrix for Volumetric Data, GLCMVD) for extracting useful texture features in hyperspectral image cubes. For traditional 2D GLCM analysis, a primary issue was to determine the optimal window (kernel) sizes in the computational process. Previous studies demonstrated that the window size could account for 90% of the variability in the results of classification. During the evaluation, it usually requires a large window size in order to obtain meaningful description of the whole data set. However, for texture segmentation, a small window size is preferred in order to accurately locate the boundaries between different textured regions. Therefore, how to determine the most appropriate box size for GLCMVD computation has become a critical issue. In order to solve this problem, an extended semi-variance analysis was proposed to determine the optimal kernel size for GLCMVD. Experimental results of this study indicated that the proposed extended semi-variance analysis could successfully identify appropriate kernel sizes for the GLCMVD computation of different targets. In addition, the results also indicated that texture information derived directly from volumetric data performed better in discriminating individual image features than 2D texture derived form sliced data.