摘要
由于自然图像中包含丰富的颜色信息与多尺度的纹理信息,伴随多个同质目标区域的出现,依靠半监督手动交互标记的图像分割方法难以实现自动分割,因此提出一种多类无监督彩色纹理图像分割方法.首先,提取了颜色特征(向量结构)与多尺度纹理特征(矩阵集合),并对两者分别进行能量描述;对于2种具有不同特征结构的能量函数,通过多类融合策略计算两者的融合因子,并自适应地融合;再将融合后能量函数的全局最优化问题转化为其对应的多层Graph Cut图割模型,利用最大流?最小割理论计算得到全局近似最优解.为了自适应地控制分割过程,提出了自适应迭代分割收敛准则,并最终在自然图像及合成的彩色纹理图像上进行了质量评估与量化分析.实验结果表明,该方法具有较好的目标区域完整性与一致性,并具有较高的准确率.
For semi-supervised and interactive image segmentation methods, they are usually confronted with difficulty with the segmentation task for natural color-texture image automatically. The reason for that is the natural image contained abundant color information and multi-scale texture information, and meanwhile some homogeneous color-texture objects are appeared. An unsupervised color-texture image segmentation approach is proposed in this article. The features of color （vector） and multi-scale texture （set of matrix） are extracted firstly, and then the corresponded energy functions of color and texture are constructed; Meanwhile, a multi-class merging strategy is designed for computing the corresponded merge factor, so that we can integrate the two different kinds of energy functions adaptively. In the following, the optimal solution of merged energy function can be acquired through converting the corresponded energy function as multilayer graph, and then an approximate optimal solution can be calculated in max-flow/rain-cut algorithm. Lastly, an adaptive iteration convergence criterion is designed to control the convergence of segmentation process, and meanwhile some experiments are carried out on a large number of synthesis color-texture images and natural images. The experiments demonstrate the superiority of our proposed method with better region entirety, consistency, and high accuracy.
出处
《计算机辅助设计与图形学学报》
EI
CSCD
北大核心
2014年第5期812-825,共14页
Journal of Computer-Aided Design & Computer Graphics
基金
国家自然科学基金(61073094
61073093)
郑州市智能图像处理与识别重点实验室项目
郑州市科技局攻关项目(14A520054
20120410
20110993
20120411)
河南省科技厅项目(20130704)
关键词
结构张量
彩色纹理
多层图割模型
图像分割
structure tensor
color-texture
multi-layer graph model
image segmentation
