结合稀疏编码和空间约束的红外图像聚类分割研究

Infrared image segmentation based on clustering combined with sparse coding and spatial constraints

提出了结合稀疏编码和空间约束的红外图像聚类分割新算法,在稀疏编码的基础上融合聚类算法,扩展了传统的基于K-means聚类的图像分割方法.结合稀疏编码的聚类分割算法能有效融合图像的局部信息,便于利用像素之间的内在相关性,但是对于分割会出现过分割和像素难以归类的问题.为此,在字典的学习过程中,将原子的聚类算法引入其中,有助于缩减字典中原子所属类别的数目,防止出现过分割;考虑到像素及其邻域像素具有类别属性一致性的特点,引入了空间类别属性约束信息,并给出了一种交替优化算法.联合学习字典、稀疏系数、聚类中心和隶属度,将稀疏编码系数同原子对聚类中心的隶属程度相结合,构造像素归属度来判断像素所属的类别.实验结果表明,该方法能够有效提高红外图像重要区域的分割效果,具有较好的鲁棒性.

A new algorithm for infrared image segmentation is proposed based on clustering combined with sparse coding and spatial constraints. The clustering algorithm is fused on the basis of sparse coding. The traditional image segmentation method based on K-means clustering is extended. The clustering algorithm combined with sparse coding can fuse the local information of image. The inner relationships between pixels are used. However, the problem of over-segmentation and difficulty in pixels classification for segmentation arise. The clustering method is introduced for atoms into dictionary learning. The class number of atoms in dictionary is reduced in order to avoid over-segmentation. The spatial class property information is also introduced by considering the property of the pixel, and the pixels in the neighbor region should have class coherent constraints. An altemate optimization algorithm is proposed to learn the dictionary, sparse coefficients, cluster center and degrees of membership jointly. Then the classes of pixels are estimated by constructing pixel ownership degrees, combining the sparse coefficients and the degrees of membership with the atoms to cluster center. The experimental results show that the important area can be separated well, and the proposed method has good robustness.