A comparison of the GLCM, GMRF, Gabor Energy, and Fractal Dimension methods over the ohanDubes test suite is presented. These results are based on the MeasTex implementations of these methods with, where possible, parameter choices similar to those used in Ohanian and Dubes (1992). Specifically, the following implementation choices have been used.
Program: gaborClass Wavelengths: 2, 4, 8 and 16 pixels Angles: 0, 45, 90, 135 degrees Mask Size: 17x17 pixels Gaussian Window: texton interpretation (sd = wavelength/2) Command Line: gaborClass -texton -lambda 2,4,8,16 -theta 0,45,90,135
Program: glcmClass Distances: 1 pixel Angles: 0, 45, 90, 135 degrees Re-quantization: 32 grey levels Rotation Invariance: average features over angles Features: Energy, Entropy, Inertia, Haralick's Correlation Command Line: glcmClass -q 32 -af -d 1 -impl WDR76 -theta 0,45,90,135
Program: markovClass Mask: standard 2nd order symmetric Command Line: markovClass -mask std2s
Program: fractalClass Command Line: fractalClass
The tables below give the winner-take-all and probability-based
scores for the implementations described above for the ohanDube test suite. Both
individual test problem scores and overall summary scores are shown.
These results differ significantly from those of Ohanian and Dubes (1992)[OD92]. OD92 found that fractal features gave the highest results with 85% correct for the fractal textures. We see from Tables 1 and 2 that both the Gabor and GMRF algorithms exceed this significantly with perfect results attained with GMRF features.
The results shown here for GMRF images are marginally worse than OD92 who obtained perfect classification using GLCM features, 98% correct for fractal features, 97% for GMRF features and 80% for Gabor Energy features. All our results are above 95% correct classification.
OD92 achieved an maximum 89% correct classification on the Leather textures with GLCM features. Equivalent maximum classification results were achieved here with GMRF Features, closely followed by Gabor Energy featurs. In general, the probability results shown here are slightly superior.
Painted surface texture results scored well for Gabor, GLCM and GMRF features (95% and above). OD92 obtained a maximum of 96% with GLCM and 93% with Fractal features.
The full 16 class problem caused some problems for OD92 with a best result of 90% correct using GLCM features. Fractal features gave the next best results with 85% correct. The results given here again show the superiority of GMRF features with 97% correct followed by GLCM features with 97% correct.
OD92 also gives results using features selected from the combined features of GLCM, Gabor Energy, Fractal Dimension and GMRF random fields. The results shown here generally exceed OD92's best results. The only exceptions are the results for the painted surface textures which are marginally worse and the leather textures which show approximately 5% lower classification.
It should be remembered that we have not implemented exactly the algorithms used by Ohanian and Dubes. The algorithms used here (distributed with MeasTex) were implemented with functionality common in the literature. Also, we have not gone to anywhere near as much trouble as Ohanian and Dubes (complicated feature selection) and present very similar (and often superiour) results.
These results indicate the care that must be taken when accepting comparison results. We use the same texture images and the same paradigms (albeit different implementations) and obtain significantly different results. In fact, if we were to rank these paradigms based on this test suite we would have (best to worse) GMRF, Gabor Energy, GLCM, and Fractal Dimension. This is in stark contrast to rankings based on OD92's conclusions of GLCM, Fractal, GMRF, and Gabor Energy.