Normalized Texture Motifs and Their Application to Statistical Object Modeling

S. D. Newsam* B. S. Manjunath**

* Center for Applied Scientific Computing Electrical and Computer Engineering
Lawrence Livermore National Laboratory University of California
Livermore, CA 94550 Santa Barbara, CA 93106
newsam1 [at] llnl.gov manj [at] ece.ucsb.edu

** ´╗┐Electrical and Computer Engineering
University of California
Santa Barbara, CA 93106
manj [at] ece.ucsb.edu

Abstract

A fundamental challenge in applying texture features to statistical object modeling is recognizing differently oriented spatial patterns. Rows of moored boats in remote sensed images of harbors should be consistently labeled regardless of the orientation of the harbors, or of the boats within the harbors. This is not straightforward to do, however, when using anisotropic texture features to characterize the spatial patterns. We here propose an elegant solution, termed normalized texture motifs, that uses a parametric statistical model to characterize the patterns regardless of their orientation. The models are learned in an unsupervised fashion from arbitrarily orientated training samples. The proposed approach is general enough to be used with a large category of orientation-selective texture features.
[PDF] [BibTex]
S. Newsam and B. S. Manjunath,
IEEE Computer Society Conference on Computer Vision and Pattern Recognition, Workshop on Perceptual Organization in Computer Vision, Washington, DC, Jun. 2004.
Node ID: 385 , DB ID: 184 , VRLID: 131 , Lab: VRL , Target: Proceedings
Subject: [Image Texture] « Look up more