Revisiting the texton theory (Julesz and Bergen, 1963) we have developed a computational approach to represent coloured textured images, that is faithful to the original definition of textons, defined as the attributes of image blobs. We use a refined procedure to extract blobs and compute their colour (three attributes) and shape attributes (size, length and orientation). Coloured texture images are represented by the histogram of these features following the modern framework of bag-of-words (BoW), thus density of features perfect match first order statistics of the texton theory.
Without any learning stage (usual step in BoW) we build a vocabulary by direct quantisation on the spaces of blob attributes. This quantization provides universal texture vocabularies whose visual words have a direct translation to linguistic terms.

We propose two different descriptors (JTD and STD) whose differences rely on how attributes are combined. JTD descriptor is early fusing the six blob attributes (full blob co-ocurrence) and STD descriptor concatenates shape and colour blob attributes.

Below we show an example of JTD descriptor, every bin in the histogram represent a specific shape and colour of the image.

Here we show the representation of a real image with JTD descriptor:

 

In STD descriptor the co-ocurrence of shape and colour attributes is removed and the descriptor is build concatenating shape and colour histograms, STD=[Ps, Pc].

 

STD and JTD have shown an efficient performance in representing coloured texture images in image retrieval and classification applications.

 

Code

A matlab implementation of these descriptors can be downloaded here. See the README file for further details.

In a short time a optimization of this code (in terms of speed) can be found here. In this case part of the code is implemented in C, therefore it needs to be compiled, in the README file there are the instructions and requeriments to obtain the executable code.

References

Susana Alvarez and Maria Vanrell. Texton theory revisited: a bag-of words approach to combine textons. To appear in Pattern Recognition, 2012.

Susana Alvarez, Anna Salvatella, Maria Vanrell and Xavier Otazu. Perceptual color texture codebooks for retrieving in hightly texture datasets. ICPR 2010. pp. 866-869.

© 2008 Colour in context Group | Computer Vision Center. All rights reserved | Contact webmaster |  Last updated: Monday 11 May 2009