摘要
为解决谱聚类算法应用于图像分割时,相似矩阵内存占用较大甚至满溢以及后续计算量大的问题,利用Nystrom方法随机获取一部分样本点,根据样本点和样本点、样本点和非样本点2种相似关系近似表征所有像素点的相似性,得到原图像的近似相似矩阵。在构建上述所需2种相似关系的相似矩阵时,距离度量采用余弦函数。结果表明,采用近邻传播聚类算法代替k-means算法对得到的低维向量子空间聚类,克服了聚类过程对初始值的敏感性,得到的分割结果较稳定,4幅真实图片也验证了研究算法的优越性。改进的谱聚类算法为图像分割的稳定性研究提供了依据。
To solve the problem of large usage of the similar matrix memory,even overflowing,and the large amount of calculation when spectral clustering algorithm is applied in image segmentation,a part of sample points are obtained by using the Nystrom method,according to the two similarity relations between sample points and sample points,sample points and non-sample points to get the similarity relation of all pixel points,then the approximate similarity matrix of the original image is got. In building the two block similar matrixs required in Nystrom,cosine function is used to overcome the absoluteness of Euclidean distance in the distance measurement.In this paper,the nearest neighbor propagation clustering algorithm is used to replace the k-means algorithm to overcome the sensitivity to the initial value in the clustering process,thus obtaining a more stable effect. The experimental results of the 4 real images show that the method is far superior,and the improved spectral clustering algorithm provides reference for the stable study of image segmentation.
出处
《河北工业科技》
CAS
2018年第1期55-60,共6页
Hebei Journal of Industrial Science and Technology
基金
国家自然科学基金(61601412
61571404
61471325)
山西省自然科学基金(2015021099)
