利用超图图割的图像共分割算法

An Image Co-Segmentation Algorithm Using Hyper-Graphcut

针对基于传统图割的图像共分割算法存在计算复杂度高而导致分割效率低的问题,提出了一种基于超图的图像共分割算法,利用超图能更简洁有效地代表图像中特征关系的特性来提高图像共分割的效率。该算法首先对具有相似前景的2幅图像分别进行Mean-shift过分割,并将得到的过分割区域分块作为超图的节点;然后利用分块的颜色直方图计算所有分块间的相似性,并将相似的分块对应的相似节点集合和单幅图像中相邻节点集合作为超边并计算其权值,构造超图;最后利用基于谱分析的近似算法求解超图归一化分割问题,获得图像对的共分割结果。实验结果表明,所提算法相比于单幅图像的归一化分割算法以及基于传统图割的图像共分割算法具有更好的分割效果,同时分割时间缩短至少45%。

An image co-segmentation algorithm using hyper-graphcut is proposed to improve the efficiency of image co-segmentation algorithms based on traditional graph-cut for the high computational complexity in optimizing energy functions. The algorithm bases on the property that hyper-graphs represent the relationship among images better. Mean-shift over-segmentation is applied to the input image pair to obtain image patches and these patches are then expressed as the nodes of a hyper-graph. Then the similarities between patches are calculated through their color histogram features, and the collections of similar patches and their neighbor patches are set as hyper-edges and a hyper-graph is constructed. The spectral analysis algorithm is performed on the constructed hyper-graph to obtain the final co-segmentation result of the image pair. Experimental results and comparisons with the normalized-cut of single image and the co- segmentation algorithm based on graph-cut show that the proposed algorithm has a notable improvement on image co-segmentation efficiency with a calculation cost reduction at least 45 %.