Intelligent Algorithm To Optimize The Normalized Cut Image Segmentation

Image segmentation based on graph theory becomes one of the hot topics of the image segmentation in recent years, and Normalized Cut is a standardized form. It regards every pixel in image as a node in graph, and regards the similarity between each pixel in image as the weight between each node in graph. And then an image is transformed into a weighted undirected graph. But minimizing the Normalized Cut criterion is an NP-hard problem. The approximate solution about this NP-hard problem is using spectral decomposition to similarity matrix, and this approximate optimal solution is the eigenvector, which can guide the graph's Normalized Cut partition, and finally obtain the image segmentation result.Spectral clustering has the disadvantage of the high computational complexity and the low precision. Base on Normalized Cut theory and relevant knowledge, main work of this paper is described as follows:(1) Put forward a double threshold image segmentation method based on Normalized Cut criterion. The traditional single threshold image segmentation method based on Normalized Cut only can get a satisfying segmentation result in single object grey image, and isn't suitable for segmenting multiple objects grey image. And the traditional weight computational formula doesn't consider the pixel point of its neighborhoods. In this paper, firstly we use B-type correlation to compute the similarity of pixels. This computational formula not only considers the pixel information, but also considers the pixel's neighborhoods information. Secondly, we deduce double threshold image segmentation from single threshold image segmentation based on Normalized Cut criterion. And finally we use particle swarm optimization calculating the two thresholds. This method can partition multiple objects image effective, quickly and high recurrence.(2) Put forward a grey image segmentation method using genetic algorithm to minimize Normalized Cut criterion. The traditional method minimized Normalized Cut criterion is spectral clustering algorithm, which can only obtain approximate solution. This paper preprocesses the image using fuzzy c-means clustering, which can cut down the size of similarity matrix and reduce the algorithm complexity. And then use genetic algorithm instead of spectral clustering algorithm minimizing the Normalized Cut criterion. After the genetic algorithm finishes its evolution, the optimal chromosome can direct the graph's partition instead of eigenvector, and get the segmentation result of grey image. This algorithm's optimization ability is better than the traditional spectral clustering algorithm, and can get the satisfying segmentation result of grey image.(3) Put forward a color image segmentation method using binary discrete particle swarm algorithm to minimize Normalized Cut criterion. Because color image has too many color information which leads to a large time-consuming. This paper firstly uses fuzzy c-means clustering to deal with the R channel, G channel and B channel of the color image separately. And then makes an intersection set operation to these three processing results, so gets the preprocess image which is made up of several biggest similar zones. Finally, we use binary discrete particle swarm algorithm to minimize the Normalized Cut criterion instead of spectral algorithm. After the optimization algorithm finishes its iteration, the optimal particle can direct the graph's partition instead of eigenvector, and obtain the segmentation result of color image. This algorithm's optimization ability is better than the traditional spectral clustering algorithm, and can get the satisfying segmentation result of color image fast and.precisely.