SAR image segmentation based on Immune Greedy Spectral Clustering

In this paper we propose a novel spectral clustering algorithm called immune greedy spectral clustering algorithm, which introduces immune clone selection algorithm instead of greedy selection to choose a subset before greedy spectral embedding without any prior experiment knowledge. As we know, Nystro¿m algorithm is a random selecting method which depends so much on the selecting result so that the clustering result fluctuates obviously. Greedy spectral clustering algorithm acquires a subset called dictionary by greedy selection algorithm so that the result could be more stable and better. However, to choose an appropriate input tolerance, we need prior experiment knowledge about the relationship between tolerance and select number. Moreover, since the criterion used for greedy selection is the distance in feature space between a candidate example and its projection on the subspace spanned by selected examples, we need to compute every example point one by one to get the error of using the selected examples to approximate the candidate example and then decide whether to choose it. So the time expense increases inevitable. Considering all above, we present a new method called immune greedy spectral clustering algorithm. The experimental results show that immune greedy spectral clustering algorithm need no prior experiment knowledge and could save time compared with greedy spectral clustering while getting a better accuracy rate compared with Nystro¿m algorithm.