Combined Unsupervised Image Segmentation Algorithm

Watershed transform is a fast morphologic-based image segmentation algorithm which gained a lot interest recently. But pure watershed always produces an over-segmented result. This thesis contains research on three different ways for modifying the watershed algorithm, which represent three different stages of a complete algorithm. The first stage is to use multi-scale morphologic filter to smooth the gradient image, the second stage is to control the forming procedure of watershed lines, while the third stage is to merge regions with a hierarchical clustering algorithm.The second and third stages contain creative researches of this thesis. In the second stage, a modified order-invariant algorithm was proposed, which uses lake minimum and watershed fall to represent topographical information so as to deduce the number of RIDGE labels in Lin's algorithm[3]. Here order-invariant means that algorithm result is independent of pixels processing order, so the two literate implementations of watershed algorithm– immersion approach and toboggan approach– can get the same segmentation result. And in the third stage, a new heuristic region-based model which is derived from Bayesian method and Markov random field is used to perform hierarchical clustering. This model is easily computed and contains two different parts, one works as a factor that merges the adjacent regions, and the other keeps region independently. It is the interaction of these two parts that produce optimal segmentation result under Bayesian criteria.The proposed modified order-invariant algorithm in the second stage can deduce the number of Ridge labels by about 80% as well as the number of segmented regions by about 10%. And the hierarchy model in the third stage can merge regions in a similar way as human perception. Both of these creative researches can benefit complete multistage segmentation algorithm by eliminating the uncertainty and deducing the computation complexity.