Research Of Image Segmentation Algorithm Based On Watershed Algorithm And Spectral Clustering

Image segmentation is dividing an image into non-overlapping areas. It is theintermediate step of the image processing to image analysis and also the link of lower visualto higher visual in computer vision research. Spectral theory for image segmentation is a hottopic nowadays and had developed many spectral clustering algorithms for imagesegmentation, which normalized cut algorithm is the most widely used one. The criteria of thealgorithm is to find the global optimal solution, the shortcoming of the algorithm is highcomputational complexity and huge memory requirements. Watershed algorithm is anunsupervised image segmentation algorithm, the advantage of this algorithm is fast, thesegmented part is continuous and closed, it can get a better segmentation result even for theweak edge. However, the weak point is that the algorithm is over-segmented and sensitive tonoise.For characteristics of normalization algorithm and watershed algorithm, we design amorphological watershed combined with spectral clustering image segmentation algorithm toreduce the computation and interference of noise, in this process, the main contents of thispaper are as follows:（1）Introduced the watershed algorithm, including physical model of immersionwatershed algorithm and the steps of immersion watershed algorithm, through simulation thewatershed algorithm we found that the algorithm is over-segmentation and sensitivity to noise.Introduced the relevant basic theory of graph, given five image segmentation criteria basedgraph theory, then Introduce solution of normalized cut and2-way Ncut algorithm and K-wayNcut algorithm, and finally given the process of K-MEANS algorithm which could be used inK-way Ncut algorithm.（2）To improve the problem of classical watershed algorithm, we introduced thethreshold morphological watershed algorithm reduced the "areas" segmented by classicalwatershed algorithm, and assigned the "watershed" to its nearby "area" and called those"areas" regional pixels. Construct a weighted undirected graph to describe the relationshipbetween regional pixels. We set grayscale variance as grayscale Gaussian kernel and distancevariance as distance Gaussian kernel avoiding set them manually. （3）Improves initial point selection of traditional K-MEANS algorithm, avoiding selectthem randomly resulting into a situation of local optima. Using the new K-MEANS algorithmto k-way Ncut algorithm. The proposed algorithm does not use the pixels for clusteringalgorithm as it used to do, used regional pixel instead, thus decreasing the running time ofimage segmentation. And some experiment proved the algorithm is superior to traditionalNcut algorithm.