Image Segmentation Based on Improved Watershed Techniques

• Main Authors: Yi-Wei Yu, 俞亦偉
• Other Authors: Jung-Hua Wang
• Format: Others
• Language: zh-TW
• Published: 2000
• Online Access: http://ndltd.ncl.edu.tw/handle/96273458160481042340

碩士 === 國立海洋大學 === 電機工程學系 === 88 === This thesis presents a watershed-based algorithm, namely Improved Watershed Algorithm (IWA), to solve the over-segmentation problem encountered in Watershed Analysis. The advantages of the proposed IWA are twofold. Firstly, a simple mean filter is applied to the gradient map of input image to reduce the number of primitive regions of Watershed Analysis. Secondly, the presented region-smoothing method can merge major portion of primitive regions in the first few iterations and it needs not pre-specify the number of regions, the final number of regions is autonomously determined by input nature. The local closed valleys and ridges, that is, local variance of gray level in input image, will result in many small regions. Applying the mean filter to the gradient map can smooth small valleys and ridges, and therefore the number of primitive regions of Watershed Analysis can be reduced effectively. Moreover, the discontinuity of boundary between two neighboring regions in the gradient map will incur merging the two regions erroneously. The mean filter can patch the discontinuity and decrease the possibility of false merging. In addition, the computation time of traditional region-merging algorithm [12] dominates the total execution time of the watershed-based algorithm. Thus, to save computation time, in IWA the presented region-smoothing method will update (at the current iteration) the mean gray level of each region in accordance to the neighborhood of each region, and similar regions will be grouped to the same region in the next iteration. Experimental results have verified that the IWA outperforms traditional watershed-based algorithm in computation time.