摘要
针对目前流行的显著性检测算法不能精确反映显著性信息的问题,提出一种基于超像素融合方法的显著性检测算法.首先对图像进行超像素分割,在保证高质量的图像目标边缘信息前提下,建立以超像素为节点的图模型;然后计算超像素邻接矩阵,将该图模型转化为最小生成树模型.通过OTSU算法自适应地确定最佳阈值,根据该阈值将最小生成树模型的部分节点进行融合,获得大超像素分割区域;最后利用大超像素的颜色和相互距离信息,获得高质量的显著性图.实验结果表明,相对于其他检测方法,该算法可以更有效地检测出图像中的显著目标,并能达到接近分割的效果.
According to the problem that some saliency detection algorithm can't reflect saliencyinformation exactly, a new saliency detection method based on superpixel-fusion was proposed.Firstly, superpixel segmentation operation was executed for the input image, then the graphmodel with superpixels as nodes was built. Secondly, by computing the superpixel adjacencymatrix, the graph model was transfered to minimum spanning tree model. Thirdly, the thresholdcan be fixed by using OTSU algorithm which was the standard to fuse part of nodes of the MSTto gain big superpixel region. At last, the saliency maps were computed via the color and spacedistance between big superpixels. Compared with other detection methods, experimental resultsshowed that this algorithm can detect the salient object more efficient and nearly reach thesegmentation effect.
出处
《北京理工大学学报》
EI
CAS
CSCD
北大核心
2015年第8期836-841,共6页
Transactions of Beijing Institute of Technology
关键词
超像素融合
图模型
最小生成树
显著性检测
superpixel-fusion
graph model
minimum spanning tree
saliency detection