Color and texture are of considerable importance for image analysis. Color as well as texture have been discussed in literature intensively. But most of the known texture models are based on grey-level images. This is an inadequate restriction in many real world applications of computer vision. We are working on color texture models for micro-textures in natural scenes.
Figure 1: Examples of textures.
First row: D17 (Brodatz), Food.0007 (VisTex, MIT), beeren (TU München).
Second row: fish002_col2b (Uni Koblenz), Flowers.0002 (VisTex, MIT), Water.0002 (VisTex, MIT).
The Color Covariance (CC) texture model is derived from the auto-covariance model for grey-level images. This texture model is defined for color images with three color planes, e.g. RGB-images. Interactive relations between different color planes (e.g. R-G) are calculated by a second order statistic.
The synthesis of color texture is examined in order to find an optimal set of color texture parameters describing natural color textures. If a good re-synthesis of color textures is possible then an adequate parameter set is found.
The texture synthesis in the CC model follows an evolutionary strategy. First a random color texture is generated with the same color histogram as the original texture. In several iterations all pixels of the image are exchanged in their position to make the synthesized texture more similar to the original. Step by step the distance between the CC parameters of the artificial texture and the original is reduced. The significance of the color texture parameters have been examined in order to find a minimal set of color texture parameters. The number of the CC parameters used for synthesis and the quantity of the color histograms used for initialization have been reduced in many experiments.
Figure 2: Examples of color texture synthesis.
First row: fish002_col2b (Uni Koblenz) original, random image, synthesis after one iteration, four iterations.
Second row: Water.0002 (VisTex, MIT) original, random image, synthesis after one iteration, four iterations.
Also classification is used in order to find a reduced parameter set of the CC texture model. Therefore a large set of images was built showing the bark of 408 different trees of six species. The color texture of these images are classified by a K-Nearest-Neighbour Classifier. In various classification experiments a new set of color texture parameters was proofed to work well. These parameters are related to the human perception of texture. For further reduction of the dimension of the feature space the Principal Component Analysis from statistics is applied among other methods.
Raimund Lakmann, Lutz Priese: A Reduced Covariance Color Texture Model for Micro-Textures
1997. Pattern Recognition Society of Finland. 947-953. Proceedings "10th Scandinavian Conference on Image Analysis (SCIA)",Lappeenranta, 9-11th June 1997.
Raimund Lakmann, Lutz Priese: Ein Farbkovarianzmodell zur Analyse und Synthese von Farbtexturen
1997. 9. 55-62. Proceedings "Mustererkennung 1997", 19. DAGM Symposium, Braunschweig, 15.-17. September 1997.
Raimund Lakmann: Farbtexturen - Analyse und Synthese
1997. 3. 49-66. Proceedings "Heidelberger Bildverabeitungsforum", Schwerpunktthema: Farb- und Mehrkanalbildverarbeitung,Heidelberg, 4. März 1997.
Raimund Lakmann, Lutz Priese: Klassifikation von Farbtexturen mit Farbkovarianzmerkmalen
1997. 9. 17-23. Proceedings "3. Workshop Farbbildverarbeitung", Erlangen, 25.-26. September 1997.
Raimund Lakmann, Lutz Priese: Farbtextursynthese mit Farbkovarianzmerkmalen
1996. 10. 21-24. Proceedings "2. Workshop Farbbildverarbeitung", Ilmenau, 10.-11. Oktober 1996. Zentrum für Bild- und Signalverarbeitung.