Complex Image Segmentation Based On Immune Clonal Selection Optimization And Spectral Clustering

In recent years, digital image processing has been a top object of study and research in the field of information science, computer science, biomedicine, military science, even social science. Image segmentation is very critical and essential to digital image processing, computer vision and pattern recognition. The quality of image segmentation determines the quality of the final result of image analysis and image understanding. Image segmentation problem can be modeled as different mathematical models, then using different optimization methods to solving it. Recently, immune clonal selection optimization is a new hotspot and new area of research in artificial intelligence researcher. Inspired by biological immune systems which contains a wealth of information processing mechanisms and functions, immune clonal selction optimization may provide novel solutions and approaches to the problem of image segmentation. At present, spectral clustering is widely used in pattern recognition. Compared with traditional clustering, it can group the non-convex and serious overlapping data sets. However, applying it to image segmentation are still several difficulties. To solving a number of problems currently prevalent in image segmentation, several new algorithms and strategies are proposed, and the main content of this dissertation is summarized as follows:1. When many clustering methods segmenting the image which has large size and contains lots of noise and outliers, the results is poor and the running time is terrible. According to the theory of immune clonal selection, an image segmentation algorithm based on hybrid immune clonal k-medoids clustering is proposed. There are three characteristics in the algorithm. First, using Turbo Pixels superpixels method can reduce the time and space complexity. Second, using immune clonal selection algorithm with a new designed mutation strategy and hybridizing it with an effective local search heuristic operator can get global optimum quickly. Third, the optimized target is k-medoids model which is more robust to noise and outliers. Experimental results on several artificial data sets, complex images show that our algorithm outperforms the k-means algorithm, the RARWGA algorithm, the GCA algorithm.2. There are several difficulties in using partitional clustering algorithm to deal withimage segmentation problem including choosing the correct number of clusters without any prior knowledge, measuring the image data sets which have complicated manifold structures, reducing the computation time. In order to solve the above problems, two automatic immune clonal clustering methods using manifold distance for image segmentation are proposed. Firstly, the two methods can automatically determine the number of clusters. Secondly, manifold distance is suitable to measure complicated manifold data sets. Finally, using the SLIC superpixels instead of pixels as the operating unit leads to less computation time. Experimental results on five typical artificial data sets and five complex natural images show that the two novel methods outperform the k-means algorithm and the GCUK algorithm, hence have some practicability and advancement.3. Color image segmentation may be view as image classification problem based on color feature. Therefore, there are two key technologies need to be resolved in color image segmentation, how to choose the best color space and how to choose the appropriate classifier. A novel color segmentation method is proposed. Multidimensional space is defined by using PCA technique to computing the most discriminating color components for a given color image among a set of conventional color spaces. Then, training samples for every region in the given color image is selected and these samples is trained by clonal selection algorithm to obtain clustering center of every region. Finally, output the segmentation result according to these clustering centers. Due to the nonlinear classification property of clonal selection algorithm and adaptive definition of a multidimensional space for a given image, the segmentation result can be obtained accurately and quickly. In experiments, different color images are used to test the performance of the suggested method. The result indicated that this method performs more robustness and adaptability.4. When using spectral clustering algorithm for SAR image segmentation, spectral clustering using the method is very popular due to its lower computational complexity, but there are mainly three disadvantages including segmentation result easily affected by random sampling, sensitive to the scaling parameter, selecting the weight of spatial information is cumbersome. In order to solve the above problems, a novel SAR image segmentation method based on spectral clustering ensemble using nonnegative matrix factorization is proposed. According to the theory of ensemble Nystrom&& learning, the random selection of the set of sampling points, scaling parameter, weight of spatial information are used to generated diverse components, then these multiple components are combined by using nonnegative matrix factorization technique which is superior to other combining methods to obtain the final segmentation result. By qualitative analysis and quantitative analysis, the experimental results indicated that our method is better.