Pointwise Multi-resolution Feature Descriptor for Spectral Segmentation

Abstract

In the process of spectral segmentation, it is crucial to compute a reliable affinity matrix with different features of an image. In this paper, we present a method of constructing the affinity matrix based on multi-resolution features extracted from the original features. A pointwise multi-resolution feature descriptor (PMFD) is designed based on spectral graph wavelets, which characterize the topology of the image centered at different pixels. After choosing the scales of interest in our descriptor, a new affinity matrix is constructed based on the extracted features. For large-size affinity matrixes, it is difficult to compute the proposed PMFD for all pixels of an image. Therefore, an approximate algorithm is proposed to compute the PMFD. To demonstrate the effectiveness of our method, a series of experiments on the Berkeley image segmentation dataset are implemented using the PMFD-based spectral segmentation algorithm. A comparison with other image segmentation techniques demonstrates that our method offers significantly improved pointwise spectral segmentation performance.

from #Head and Neck by Sfakianakis via simeraentaxei on Inoreader http://bit.ly/2DAssmZ