Abstract: To improve the performance of spectral segmentation with single feature in image segmentation, this paper proposes a multiscale spectral segmentation algorithm. Firstly, this paper defines diffusion map based multiscale feature descriptor (DMFD) which can efficiently describe the geometry construction and intrinsic feature of an image. To utilize the fast algorithm in spectral graph wavelet (SGW) and compute the eigenvector of every pixel, this paper builds the connection between DMFD and SGW. Then, this paper proposes the DMFD based spectral segmentation method (DMFD-Ncut) and the pointwise self-adaptive optimal scale (PSOS) method to adaptively compute the affinity matrix in the optimal scale. At last, the experiments with Berkeley dataset show DMFD-Ncut and DMFD-Ncut with PSOS can both improve the performance of image segmentation.

Key words: spectral segmentation, diffusion map, spectral graph wavelet, self-adaptive optimal scale, image segmentation

摘要： 为提高使用单一特征信息谱分割的图像分割效果，提出了一种多尺度谱分割算法。基于扩散映射构建一种多尺度特征描述器（diffusion map based multiscale feature descriptor，DMFD），高效地描述图像数据的几何结构，反映出图像的本质特征。随后建立DMFD与谱图小波之间的联系，利用谱图小波的快速计算方法，计算输入图像每个像素点对应的特征向量。提出基于DMFD的谱分割算法（DMFD-Ncut）及逐点自适应尺度选择方法（pointwise self-adaptive optimal scale，PSOS），自适应地在最优尺度下计算相似矩阵。最后，在Berkeley数据库上的实验结果显示，DMFD-Ncut算法和使用PSOS的DMFD-Ncut算法都能够显著地提高图像分割结果的精度。

关键词: 谱分割, 扩散映射, 谱图小波, 自适应尺度选择, 图像分割