The TSW features are the tree-structured wavelet decomposition of a
textured image. A fairly complex algorithm is needed to compare trees (not
Euclidean distance).
T. Chang and C.-C. J. Kuo,
"Texture Analysis and Classification with Tree-Structured Wavelet Transform",
IEEE Trans. Image Processing, vol. 2, no. 4, pp. 429-441, 1993.
The MR-SAR features are the parameters of a multiscale autoregressive
model, estimated using least-squares. For robustness, the parameters
are estimated over many subwindows of the image. The sample mean and
sample covariance of these estimates are used with the Mahalanobis distance.
@ARTICLE{Mao92,
AUTHOR = "J. Mao and A. K. Jain",
TITLE = "Texture Classification and Segmentation Using Multiresolution Simultaneous Autoregressive Models",
JOURNAL = "Patt. Rec.",
VOLUME = "25",
NUMBER = "2",
YEAR = "1992",
PAGES = "173--188"
}
The Tamura features are an attempt to model human notions of
"coarseness," "contrast," and "directionality."
The Euclidean distance is used to compare images.
@ARTICLE{Tamura78,
AUTHOR = "H. Tamura and S. Mori and T. Yamawaki",
TITLE = "Textural Features Corresponding to Visual Perception",
JOURNAL = tsmc,
VOLUME = "SMC-8",
NUMBER = "6",
YEAR = "1978",
PAGES = "460--473"
}
The Histogram features are pixel counts for each of 256 gray-levels.
For a color image, the three per-channel histograms are concatenated
to get one 768-bin histogram.
The Euclidean distance is used to compare images.
tpminka@media.mit.edu
1/26/99