On Morphological Color Texture Characterization Erchan Aptoula and

On Morphological Color Texture Characterization Erchan Aptoula and Sébastien Lefèvre Image Sciences, Computer Sciences and Remote Sensing Laboratory Louis Pasteur University Strasbourg, France {aptoula, lefevre}@lsiit. u-strasbg. fr October 12, 2007 ISMM

Contents > Morphological tools for texture analysis > Granulometry & covariance > A combination of SE size-direction-distance > Implementation on color images > Application results 1/15

Morphological texture description • Texture characteristics: regularity (periodicity), directionality, complexity, overall color and color purity • A rich variety of tools: granulometry, morphological covariance-variogram, orientation maps, etc • Main advantage of morphological approaches: their inherent capacity to exploit spatial pixel relations 2/15

Granulometry • Standard granulometry of an image f : • Extracts information on the granularity of its input • Has several extensions: attribute based, multivariate, spatial, etc 3/15

Morphological covariance • Morphological covariance of an image f : P 2, v: a pair of points separated by a vector v • Extracts information on the regularity, directionality and coarseness of its input 4/15

Covariance + granulometry = ? Granulometry • granularity Covariance • regularity • directionality • coarseness • They extract complementary information, how should they be combined, by concatenation, . . . or? 5/15

Covariance + granulometry = ? • Employ 3 structuring element variables: size, direction and distance. Size granularity Direction Distance regularity directionality 6/15

Covariance + “granulometry” Pλ, v: a pair of SEs of size λ, separated by a vector v • However: SE is not convex Strongly ordered texture pseudo granulometry Disordered texture 7/15

Extending to color images Requirements: • A suitable color space • A color ordering scheme (preferably total), to impose a lattice structure R G B 8/15

Color space choice • Color space choice : perceptual, polar, etc. . . • Polar color spaces : (+) intuitive components (-) manipulation of hue Luminance (-) multiple implementations Saturation 9/15

Color ordering • Luminance: contains the majority of variational information • Color: auxiliary component 1. For which levels of luminance does color become more important ? 2. How should the balance between luminance and “color” use be determined ? 10/15

Color ordering 1. For which levels of luminance does color become more important ? a b c 11/15

Color ordering 2. How should the balance between luminance and “color” use be determined ? • Image or vector specific configurations are better suited for intra-image applications • Here, an image database specific approach is used, by means of genetic optimization: 12/15

Application • Outex 13 texture database: 1360 images (128 x 128) of 68 colour textures • Four directions (0°, 45°, 90°, 135°), 15 different SE sizes (k: 1 to 30, 2 k+1) and 20 distances • Results in a feature of size 20 x 4 x 15, which was reduced to 20 x 4 x 2 by PCA • k. NN classifier (k=1) and the Euclidean distance 13/15

Classification accuracies Features Grayscale Color Optimized Color Granulometry 67. 53 68. 78 72. 03 Covariance 73. 82 76. 92 80. 46 Concatenated 77. 75 79. 93 83. 74 Combined 83. 53 85. 49 88. 13 14/15

Conclusion and perspectives • A way of combining the complementary information provided by granulometry and covariance • However, it leads to a pseudo granulometry • Genetic optimization aids in exploiting color • Shape variations, as well as the role of hue remain to be investigated 15/15

Thank you for your attention E. Aptoula and S. Lefèvre {aptoula, lefevre}@lsiit. u-strasbg. fr