Color image segmentation using a cosine affinity based spectral clustering algorithm

There are two problems in the spectral clustering algorithm based on Gaussian kernel affinity matrix. First, the algorithm needs to accurately set scale factors. Second, it requires the eigendecomposition of the affinity matrix, the computational complexity of which is proportional to O(n³), with n representing the total number of vertices in a graph. Therefore, the algorithm may be unsuitable for large data sets. To solve the above two problems, a color image segmentation using a Cosine Affinity based Spectral Clustering Algorithm (CASCA) is designed in this paper. By incorporating the cosine function for the construction of the affinity matrix and Nystr̈om method for the solution of the integral eigenvalue problem, CASCA requires low computational complexity and is therefore very feasible for real-time image segmentation processing. The experimental results in several Berkeley images show that CASCA can significantly improve the efficiency of image segmentation and the segmentation effects of it outperforms those of several other image segmentation algorithms.