Image segmentation via Spatially Chunking and Resolving Label Inconsistency

• Main Authors: Ming-Jui Kuo, 郭明瑞
• Other Authors: Jung-Hua Wang
• Format: Others
• Language: zh-TW
• Published: 2005
• Online Access: http://ndltd.ncl.edu.tw/handle/33517690786033843668

碩士 === 國立臺灣海洋大學 === 電機工程學系 === 94 === This thesis proposes a technique called Resolving Label Inconsistency (RLI) useful for performing image segmentation. We show that by incorporating RLI and Support Vector Clustering (SVC), accurate segmentation results can be obtained for complex input images. Because each image pixel is taken as a data point, in conducting clustering task the data points naturally contain the spatial information of the input image. Consider that the most significant characteristics of SVC is its capability of extracting the boundary of data structure, and the pixels in digital image are arranged regularly as a 2-D lattice structure, i.e. no other pixels can exit between any two neighboring pixels. Thus, clustering the image data points can extract the object boundary. However, because SVC suffers from the large data size problem, clustering on the whole input image would be computationally impractical. In light of these, we employ the chunking concept to overcome the large data size problem. The idea is to apply a sliding window to the input image so that each time only a 3�e3 sub-image set undergoes the SVC process. Following that, clustering results of sub-images are combined and cross-validated through RLI, and the contour extraction algorithm extract the boundary information. Simulation results show that the RLI can overcome the large data size problem, even if the input image contains drastic variations.