A novel spectral clustering method with superpixels for image segmentation

Similarity measure is critical to the performance of spectral clustering. The most commonly used similarity measure for spectral clustering is Gaussian kernel similarity measure. However, the selection of accurate scaling parameter in Gaussian kernel function is difficult. To reduce the sensitivity of scaling parameter, in this paper, a novel spectral clustering method with superpixels for image segmentation (SCS) is proposed. In particular, a novel kernel fuzzy similarity measure is presented, which uses membership distribution in partition matrix obtained by kernel fuzzy C-means clustering(KFCM). In addition, the superpixel is introduced into image segmentation to alleviate the computational burden of affinity matrix. The experimental results show that our approach is able to perform steadily under different parameters, and obtain good clustering results on various natural images. Moreover, the evaluation comparisons also indicate that our method can achieve comparable accuracy and significantly outperform most state-of-the-art algorithms.