融合特征的自适应超像素图割算法

目的针对图割（GrabCut）算法对于前景与背景颜色特征相差不大容易发生分割错误，SLIC（simple linear iterative clustering）预分割在对应情况下边缘不够准确以及时间复杂度较高等问题，提出一种融合特征的自适应超像素GrabCut算法。方法该算法首先将图像转化到Lab色彩空间，并对原图像提取Gabor纹理特征，综合得到融合特征；再利用融合特征改进SuC方法，使用改进方法对图像进行预分割，提取超像素区域，构建区域邻接图；然后保存每个超像素区域的融合特征，对两种特征分别进行高斯混合模型（Gaussian mixture model，简称GMM）建模，并利用相对熵自适应调整分割过程中混合特征的权重，优化Gibbs能量函数；最后执行迭代图割算法，得出分割结果。结论实验结果表明，本算法对颜色特征不佳的情况下有较好的分割效果，并通过改进的SLIC预分割提高了算法的执行效率，降低了迭代次数，前景物体边缘也得到较好的保护。

An adaptive super-pixels graph-cut algorithm by fusion features

Objective To solve some problems of segmentation and pre-segmentation, such as the GrabCut algorithm is sometime to get wrong answer when the foreground color characteristics and the background characteristics are almost the same. SLIC algorithm gets inaccurate edges and use high time complexity. Above all, we propose an adaptive super-pixels graph-cut algotithm by fusion features. Methods Firstly, we convert the image to Lab color space, and extract the original image＇s Gabor tex- ture features, fuse the texture features and color features, to obtain fusion features. Re-use integration fea-tures to improve SLIC （Simple Linear Iterative Clustering ） method. Using the improved method to pre-segment the original image, extract super-pixel area, and build area adjacent diagram. Then save the fusion characteristics of each super pixel region, respectively, build Gaussian mixture model of two features, and using the relative entropy to adapt the weight of fusion features in segmentation process, and optimize Gibbs energy function. Finally, perform iterative graph cut algorithm, and the segmentation re- sults are obtained. Conclusions Experimental results show that tation of images with poor color features, and by improving the prove the efficiency, reduce the number of iterations, and can ground objects our algorithm has good result of segmen- SLIC pre-segmentation algorithm to im- get a better edge protection of the fore-ground objects.