Results by Timo Ojala

These results were submitted by Timo Ojala

		Machine Vision Group, Infotech Oulu and
		Department of Electrical Engineering
		University of Oulu, Finland

For details of the algorithms, the reader is directed to the references below.

It should be noted that although Timo used the MeasTex test suites, he did not use the testing framework.

REFERENCES

The texture classification method I used is described in detail in the following references:

[1] Ojala, T., Pietikainen, M., and Harwood, D., A comparative study of texture measures with classification based on feature distributions, Pattern Recognition, vol. 29, pp. 51-59, 1996.

[2] Ojala, T., Multichannel approach to texture description with feature distributions, Technical Report CAR-TR-846, Center for Automation Research, University of Maryland, December 1996. Submitted to ICIP 97.

[1] presents the basic methodology, and introduces the Local Binary Pattern (LBP) operator. In [2] the method is generalized to multichannel texture description.

EXPERIMENTAL RESULTS

I used confidence of 100% for the classification result given by the texture classification method. (winner-take-all scoring)

'Real World' Test Suites

Paramaters
	pseudo-metric = G2
	k = 9
	scale = 1 (i.e. 3x3 operator)
	bins = default 256 for LBP, 32 for other operators

The operators used with the multichannel-approach listed beside the result. I did not conduct an extensive search through all possible combinations of different operators, but just tried few combinations I thought could be useful. In any case, I believe that the listed results for multichannel are pretty much the maximum which can be obtained with the operators listed in reference [2].


Test Suite	LBP	Multi- channel	Multichannel Operators
visTex	0.9225	0.9349	LBP,DIFFX,DIFFY,DIFF4,SCOV,SRAC,VAR,S3L3,L3E3,S3E3
grass	0.8837	0.9751	LBP,DIFF2,DIFF4,SCOV,BVAR,WVAR,L3E3,L3S3,L3E3,E3S3
material	0.9713	0.9814	LBP,SCOV,L3L3,DIFF4
brodatz	0.9575	0.9930	LBP,DIFFX,DIFFY,DIFF2,DIFF4,E3E3,E3L3,E3S3,L3E3,L3L3,L3S3,S3E3,S3L3,S3S3,SAC,SCOV,SRAC,SVR,VAR,WVAR,BVAR

Artificial Test Suites

I restricted my analysis to single operator (LBP).

Parameters
	pseudo-metric = G2
	k = 9
	scale = 1 (i.e. 3x3 operator)


Test Suite	LBP
bomb	0.9688
bombRot	0.9459
lattice	0.7212
latticeRot	1
mortar	0.7387
mortarRot	0.9938
mortarRotS	0.9938