Change search
ReferencesLink to record
Permanent link

Direct link
The Texture-Transform: An Operator for Texture Detection and Discrimination
KTH, School of Computer Science and Communication (CSC), Computer Vision and Active Perception, CVAP.
2009 (English)Doctoral thesis, monograph (Other academic)
Abstract [en]

In this thesis we present contributions related to texture detection and discrimination to be used for analyzing real world images. Many computer vision applications can benefit from a fast and low dimensional texture descriptor. Several texture descriptors have been introduced and used for texture image classification and texture segmentation on images with a single or a mixture of textures. For evaluation of these descriptors a number of texture image databases (e.g. CURet, Photex, KTH-TIPS2, ALOT) have been introduced containing images of different types of natural and virtual texture samples. Classification and segmentation experiments have often been performed on such databases. In real world images we have a variety of textured and non textured objects with different backgrounds. Many of the existing texture descriptors (e.g. filter banks, textons) due to their nature fire on brightness edges. Therefore they are not always applicable for texture detection and discrimination in such real world images, especially indoor images which in general contain non textured structures mixed with textured objects.

In the thesis we introduce a texture descriptor, the Texture-transform, with the following properties that are desirable for bottom-up processing in real-world applications: (i) It captures small-scale structure in terms of roughness or smoothness of the image patch. (ii) It provides a low dimensional output (usually just a single dimension) which is easy to store and perform calculations on. (iii) It generally does not fire on brightness edges. This is in contrast to for instance filters which tend to identify a strip around a brightness edge as a separate region. (iv) It has few parameters which need tuning. The most significant parameter that unavoidably appears is scale. It is here simply provided by the size of the local image patch. (v) It can be computed fast and used in real-time systems and easily be incorporated in multiple cue vision systems. Last but not least, it is extremely easy to implement, for example in just a few lines of Matlab.

The Texture-Transform is derived in a manner different from other descriptors reviewed in this thesis, but related to other frequency based methods. The key idea is to investigate the variability of a window of an image by considering the singular values or eigenvalues of matrices formed directly from grey values of local patches.

We show that these properties satisfy the requirements for many applications by extensive experiments in two main tests, one of detection and another of discrimination, as in [Kruizinga and Petkov, 1999]. We also demonstrate that the Texture-transform allows us to identify and segment out natural textures in images, without yielding too many spurious regions from brightness edges. In these experiments we perform comparisons with other descriptors of a similar low-dimensional type. Due to the nature of our descriptor it of course lacks invariance. Hence, it cannot by itself be used for classification, since the results do not carry over from one image to another. However, as a proof of concept we show experimentally that the detected textured regions can be used in a subsequent classification task. Invariance is not needed in all tasks of detection and discrimination, at least with regard to orientation and contrast, as we discuss and demonstrate in the thesis. As examples of real word applications, we show the function of the Texture-transform on detection of street plate names, visual attention, and vegetation segmentation. Moreover, we study the application of texture features to animal detection and also address learning the visual appearance of textured surfaces from very few training samples using a photometric stereo technique to artificially generate new samples.

Place, publisher, year, edition, pages
Stockholm: KTH , 2009. , xxi, 170 p.
Trita-CSC-A, ISSN 1653-5723 ; 2009:19
National Category
Computer Science
URN: urn:nbn:se:kth:diva-11384ISBN: 978-91-7415-485-6OAI: diva2:275069
Public defence
2009-11-25, Sal F3, Lindstedtsvägen 26, KTH, Stockholm, 10:15 (English)
QC 20100811Available from: 2009-11-03 Created: 2009-11-03 Last updated: 2010-08-11Bibliographically approved

Open Access in DiVA

fulltext(48771 kB)1134 downloads
File information
File name FULLTEXT01.pdfFile size 48771 kBChecksum SHA-512
Type fulltextMimetype application/pdf

Search in DiVA

By author/editor
Tavakoli Targhi, Alireza
By organisation
Computer Vision and Active Perception, CVAP
Computer Science

Search outside of DiVA

GoogleGoogle Scholar
Total: 1134 downloads
The number of downloads is the sum of all downloads of full texts. It may include eg previous versions that are now no longer available

Total: 328 hits
ReferencesLink to record
Permanent link

Direct link