Textured Image Segmentation Based on Wavelet Decomposition

• Main Authors: Chiang, Kai Wei, 江凱偉
• Other Authors: Lo King Chang, Tsay Jaay Rong
• Format: Others
• Language: zh-TW
• Published: 1997
• Online Access: http://ndltd.ncl.edu.tw/handle/69334476585571614459

碩士 === 國立成功大學 === 測量工程學系 === 85 === Remote sensing images have some characteristics ，such as real timeand automatic data extraction，which can provide an important resourceof data collection and adapting for geographical information system.If image segmentation techniques are used to process remote sensing images,such as image segmentation ,we can integrate the segmentation result withpolygon data and corresponding attribute data from the database ofgeographical information system ,and overlay analysis is applied to helpthe recognition of land-cover, the classification result of remote sensingimages can be improved also.Regions in an image can be distinguished from the others based on the differences of their textures. For example, different land-cover categories(residential district, water body ,forest area)in remote sensing images havedifferent textures. We use a asymmetric Daubechies orthogonal wavelet functionfor wavelet decomposition to extract texture features from images by means ofmultiresolution analysis. This method provides some advantages, such assimple computational model, orientation sensitive and multiresolutionanalysis of texture features.After using wavelet decomposition to extract texture features from images ,ISODATA clustering algorithm is used to accomplish image segmentation ,whichcan be regarded as one of the multichannel segmentation algorithms .Finally, the accuracy of segmentation result is estimated by comparing with reference images.Thsee kinds of images are selected for experiment. The first experimentalimage comprises three kinds of artificial textures . The overall accuracyof the segmentation is 93%,and the κindex is 90%. The second experimentalimage comprises four kinds of ground surface textures . The overall accuracyof the segmentation is 94%,and the κindex is 92%. The third experimentalimage is the multispectral remote sensing image of experimental area . Theoverall accuracy of the segmentation is 82%,and the κindex is 72%.This research examines the feasibility of using wavelet decomposition toaccomplish textured image segmentation. We expect that the segmentationresult can be improved by integrating region-based segmentation techniquewhich is used in this research and edge-based segmentation techniques.