One of the main problems with previous work on textons is lacking variability in the learned image elements. We propose to define "texton" as a mini-template that consists of a number of bases at some geometric and photometric configurations. Figure 1 shows one example of star pattern. A "star" is represented explicitly by a generative model of several bases. In addition, these bases are not assumed to be independently distributed any more. A sophisticated probabilistic model which accounts for the spatial relation of the bases is built and learned.
Figure 1. The composition by images bases for a star pattern, which illustrates the texton idea that a star is represented explicitly by a generative model of several bases..
1. Star Pattern
Figure 2. The texton learned from star image, which consists of six bases.
Figure 3. The learning process of reconstructed image.
Figure 4. The base map and reconstructed image after learning, which is more regular than the base map in Figure 1.
Figure 5. How bases compose the image of a star.
2. Birds
Figure 6. The base maps and reconstructed images for the bird image, by bottom up only, and after the learning process.
Figure 7. The learning process of reconstructed image.
Figure 8. The learned three textons of the birds and five instances of each of them.
Figure 9. How bases compose the images of birds.
3. Cheetah dots pattern
Figure 10. The base maps and reconstructed images for the cheetah dot image, by bottom up only, and after the learning process.
Figure 11. The learning process of reconstructed image.
Figure 12. The learned two textons of the dots and several instances of each of them.
Figure 13. How bases compose the image of elongated dots.
4. Heart shape
Figure 14. Another example of heart shape image.
Figure 15. The learning process of reconstructed image.
Figure 16. How bases compose the image of the heart.
References:
-
Song-Chun Zhu, Cheng-en Guo, Yizhou Wang, and Zijian Xu, "What are Textons?" (.pdf 1.7M), International Journal of Computer Vision. (accepted).