| Summary: | 碩士 === 國立海洋大學 === 電機工程學系 === 89 === Over-segmentation is a serious problem in conventional watershed analysis owing to the topographic relief inherent in the input image. To this problem, currently existing watershed methods merge two regions in sequence. However, sequential merging would require heavy computation load. This thesis presents two novel approaches that incorporate the watershed analysis and fuzzy theory, namely the synchronous Fuzzy-based Feature Tuning (FFT) and Clustering Merging (CM), to perform image segmentation.
Both FFT and CM need not pre-specify the final number of regions. Each region Ri obtained from watershed analysis is first represented by the mean intensity (noted as mi) of gray pixels in Ri. FFT simultaneously adjust mi values of all regions by referencing their adjacent neighboring regions. Due to the use of synchronous strategy, FFT can achieve fast merging and provides great potentiality for a fully parallel hardware implementation. The iterative algorithm of FFT is terminated when the number of merged regions of two successive iterations is identical.
In the CM method, the region merging processing has been formulated as clustering with special constraint. Each small region is regarded as a virtual data point and all the small regions are clustered if they share great similarity. When two small regions are adjacent and are clustered into an identical cluster, we say that they are of the same object and can be merged. Finally, empirical results are provided to show that the proposed approaches outperform other methods in terms of computation efficiency and segmentation accuracy.
|
|---|
