Texture classification using Gabor filters

https://doi.org/10.1016/S0167-8655(02)00056-9Get rights and content

Abstract

An unsupervised texture classification scheme is proposed in this paper. The texture features are based on the image local spectrum which is obtained by a bank of Gabor filters. The fuzzy clustering algorithm is used for unsupervised classification. In many applications, this algorithm depends on assumptions made about the number of subgroups present in the data. Therefore we discuss ideas behind cluster validity measures and propose a method for choosing the optimal number of clusters.

Introduction

In satellite image interpretation, classification is the operation by which an operator would like to detect different kinds of region like forest, urban zone, waterways, etc. As the scene is a set of points (pixels) with intensity in grey scale values in several bands, most methods use these grey scale values to determine the kind of terrain. However, a single ground cover usually occupies a number of pixels with some variability in their grey scale values. A more satisfactory interpretation of the scene should thus include textural aspects of regions.

In general, texture is characterized by invariance of certain local attributes that are periodically or quasi-periodically distributed over a region. There are many approaches for analyzing image texture. Haralick (1979) proposed a set of features (energy, entropy, maximum probability, correlation, etc.) based on grey level cooccurrence matrices. Some statistical techniques use Markov Random Field models to characterize textures (Cross and Jain, 1983). The spectral approach (Bovik et al., 1990; Randen and Husoy, 1999) to texture analysis is referred to as the multi-channel filtering approach. Textures are characterized by their responses to filters, each one being selective in spatial frequency and orientation.

In this paper, we present a spectral approach to extract texture features. The textured input image is decomposed into feature images using a bank of Gabor filters. These feature images are used to form feature vectors and each of them corresponds to one dimension of the feature space. Then, we present the Fuzzy c-means clustering algorithm used for unsupervised classification of the input pixels based on their associated feature vector. This method considers clusters as fuzzy sets, while membership function measures the possibility that each feature vector belongs to a cluster. At last, we present methods for evaluating how well different textures are separated in feature space, as well as measuring classification performance. In most applications, the number of classes is unknown. Here we propose a method for choosing the best number of classes and we apply it to a synthetic and a real texture representation problem.

Section snippets

Gabor filters and feature extraction

Gabor filters perform a local Fourier analysis and are essentially sine and cosine functions modulated by a Gaussian window. In the complex space these filters are defined asG(x,y,kx,ky)=exp−(x−X)2+(x−Y)22·ej(kxx+kyy),where x, y represent the spatial coordinates while kx, ky represent the frequency coordinates. X and Y are the spatial localizations of the Gaussian window. The filter's selectivity in spatial frequency and orientation is given byω=2πkx2+ky2andθ=arctankxky.

Since the signals to

Classification

Many algorithms have been developed for supervised and unsupervised classification. In supervised classification, training sets are needed whereas unsupervised classification classifies images automatically by finding clusters in the feature space. One of the unsupervised data clustering methods is the hard k-means clustering algorithm. It assigns each sample (feature vector) to one and only one of the clusters. This method assumes that boundaries between clusters are well defined. The model

Classification validity

Classification of data should be of high quality, i.e., all samples should have a large membership degree for at least one cluster. This problem is related to how many classes there are in the data. In fuzzy c-means algorithm, the number of clusters is required though in many applications this information is unknown. A method for measuring performance is needed to compare the goodness of different classification results. Gath and Geva (1989) have defined an `optimal partition' of the data into

Synthetic images

For experimental results, we present first a visual evaluation of Gabor filters in texture characterization and classification. To do this, we use two synthetic 256×256 textured images, one containing three Gaussian Markov Random Field (GMRF) textures (Fig. 5) and the other containing five textures from the Brodatz album (Brodatz, 1966) (Fig. 7). As the number of texture categories is known for these images, we also present the efficiency of the method proposed for classification validity

Conclusions

A fuzzy clustering approach to textured image classification has been presented. The texture features are extracted using a set of Gabor filters with different frequencies and orientations.

The fuzzy c-means algorithm has been successfully used for discriminating different types of textured image but the drawback is that one has to specify the number of clusters. We thus discussed the use of cluster validity parameters. A modification of compactness and separation validity function is proposed

Acknowledgements

The authors are grateful to colleagues: Dr. V. Lacroix for the helpful and encouraging discussions during the preparation of this paper and X. Neyt who develops the Gabor filter tools. They also would like to thank the International Institute for Aerospace Survey and Earth Sciences (ITC, The Netherlands) for providing the minefield aerial image in the frame of European Pilot Project entitled: Airborne Minefield Detection.

References (14)

  • I. Bloch

    On fuzzy distance and their use in image processing under imprecision

    Pattern Recognition

    (1999)
  • J.B. Bezdek

    Cluster validity with fuzzy sets

    J. Cybern.

    (1973)
  • J.C. Bezdek et al.

    Fuzzy Models for Pattern Recognition

    (1992)
  • A.C. Bovik et al.

    Multichannel texture analysis using localized spatial filters

    IEEE Trans. Pattern Anal. Machine Intell.

    (1990)
  • P. Brodatz

    Textures: A Photographic Album for Artists and Designers

    (1966)
  • G.R. Cross et al.

    Markov random field texture models

    IEEE Trans. Pattern Anal. Machine Intell.

    (1983)
  • I. Gath et al.

    Unsupervised optimal fuzzy clustering

    IEEE Trans. Pattern Anal. Machine Intell.

    (1989)
There are more references available in the full text version of this article.

Cited by (153)

  • Wavelet transform for cardiac image retrieval

    2021, Image Processing for Automated Diagnosis of Cardiac Diseases
  • Ependymoma and pilocytic astrocytoma: Differentiation using radiomics approach based on machine learning

    2020, Journal of Clinical Neuroscience
    Citation Excerpt :

    Transformation-based features are higher-order features encoding structural and frequency-based information of an image. As a windowed Fourier transform based on Gabor wavelet, Gabor transform can extract related features under different scales and directions in the frequency domain [27]. Through the convolution of two-dimensional images, Gabor transform shows good performance in capturing information of local space and frequency domain.

View all citing articles on Scopus
1

The first author's research is funded by the HUDEM project: HUDEM (HUmanitarian DEMining) is a technology exploration project on humanitarian demining launched by the Belgian Minister of Defense with funding provided by his Department, the Ministry of Foreign Affairs and the State Secretariat for Development Aid.

View full text