Image Texture Modeling and Analysis


Texture analysis using modulation features and dominant components

Overview

Texture is ubiquitous in natural images and constitutes a powerful cue for a variety of image analysis and computer vision applications, like segmentation, shape from texture, and image retrieval. We pursue the construction of a concise texture analysis and segmentation system for generic natural images. Our research directions are in the areas of feature extraction, feature interpretation and texture segmentation.

1. Feature extraction: Building on the generic multicomponent AM-FM image modeling framework we research into more efficient algorithms for component demodulation and compact, descriptive representations of texture. We introduce regularized algorithms for demodulation based on energy separation using combinations of Gabor filtering and the nonlinear energy operators. Moreover, dominant modulation components are seeked through alternative channel selection energy measures.

2. Probabilistic analysis: The dominant AM-FM component channel selection is formulated through a probabilistic procedure by modeling observations in terms of sinusoids and introducing locality in likelihood expressions. This facilitates the interpretation of Gabor filtering in terms of model fitting, which is a formulation we also use to phrase edge detection in common terms with texture analysis. This lays the ground for a probabilistic discrimination between edges, textured and smooth areas, which is a practically important problem for image segmentation.

3. Image segmentation: We developed an unsupervised, variational segmentation scheme based on dominant component features that uses curve evolution implemented with level set methods. The method is an enhancement of the original Region Competition – Geodesic Active \Regions that combines heterogenous cues based on probabilistic feature interpretetion. This Weighted Curve Evolution method incorporates the posterior probabilities of the texture and edge classes in the evolution law.

Modeling

The focus here is on the development and efficient use of descriptive models for a broad class of texture images. Current research includes the topics of:

expansion and application of the multicomponent AM-FM model for images
multiband Gabor filtering and filterbank design
non-linear energy operators (EO) and energy tracking
demodulation of narrowband components via the Energy Separation Algorithm (ESA)
advanced and enhanced schemes (Complex, Regularized, GaborESA)
Dominant Components Analysis (DCA) with alternative selection criteria (Teager Energy)
low-dimensional feature vectors: dominant amplitude, frequencies and modulation energy


Figure 2. Filtering and modulation components (amplitude and cosine of phase signals) example

Probabilistic interpretation

Research has focused on the following aspects:

casting the DCA algorithm in a detection-theoretic framework
use of generative models to phrase the channel selection problem
probabilistic interpretation of Gabor filtering and Energy-based channel selection
building equivalent generative models for edges, thus casting edge detection in the same setting
allowing the discrimination between edges, texture and smooth image areas



Figure 3. Reconstruction results using the texture (middle) and the edge (right) generative model

See more examples from reconstruction results, i.e. synthesis through the two alternative models of texture and edges on the Berkeley Image Test set here.


Figure 4. Images and posterior probabilities for texture (middle) and edge (left) models

Segmentation

Combining the above two aspects of the problem, the texture and non-texture cues are fused in a curve evolution scheme for natural image segmentation.

DCA features are used as cues for variational unsupervised texture segmentation
formulation of a Weighted Curve Evolution as a method that allows to combine texture, edge and intensity cues in a locally adaptive manner


Figure 5. Segmentation with normal (left) and weighted (right) curve evolution

Based on these ideas, visually appealing results were obtained on natural images containing both textured and non-textured areas. Additionally, the method resulted in systematically better segmentation results on the Berkeley Segmentation Benchmark, consisting of 100 natural images.


Figure 6. Natural textured-image segmentation examples

For a more detailed overview with examples of the probabilistic interpretation and segmentation approach please refer to the presentation here.

People

Publications

- I. Kokkinos, G. Evangelopoulos and P. Maragos,
  Texture Analysis and Segmentation using Modulation Features, Generative Models and Weighted Curve Evolution,
  IEEE Transactions on Pattern Analysis & Machine Intelligence, vol. 31, no. 1, pp. 142-157, Jan. 2009.[pdf] [slides] [bib]
- G. Evangelopoulos, I. Kokkinos and P. Maragos,
  Advances in Variational Image Segmentation using AM-FM Models: Regularized Demodulation and Probabilistic Cue Integration,
  Proc. Int’ l Workshop on Variational and Level Set Methods (VLSM-05), Beijing, China, Oct. 2005, Springer LNCS, vol. 3275, pp. 121-136.[pdf] [bib]
- I. Kokkinos and P. Maragos,
  A Detection-Theoretic Approach to Texture and Edge Discrimination,
  Proc. 4th Int’ l Workshop on Texture Analysis and Synthesis, in conjunction with ICCV-05, Beijing, China.[pdf]
- I. Kokkinos, G. Evangelopoulos and P. Maragos,
  Advances in Texture Analysis: Energy Dominant Component and Multiple Hypothesis Testing,
  Proc. IEEE Int’ l Conf. on Image Processing (ICIP-04), Singapore, Oct. 2004, vol. 3, pp. 1509-1512.[pdf] [bib]
- I. Kokkinos, G. Evangelopoulos and P. Maragos,
  Modulation-Feature Based Textured Image Segmentation Using Curve Evolution,
  Proc. IEEE Int’ l Conf. on Image Processing (ICIP-04), Singapore, Oct. 2004, vol. 2, pp. 1204-1207.[pdf] [bib]

Software

A toolbox for image texture analysis with modulation features and generative models, implements the ideas of probabilistic feature interpretation and can be found here.