Elsevier

Pattern Recognition Letters

Volume 48, 15 October 2014, Pages 34-41
Pattern Recognition Letters

Rotation invariant co-occurrence features based on digital circles and discrete Fourier transform

https://doi.org/10.1016/j.patrec.2014.04.006Get rights and content

Highlights

  • We present a method to improve accuracy and robustness of co-occurrence features in rotation invariant image classification.

  • Co-occurrences are computed through digital circles.

  • Rotation invariance is obtained through discrete Fourier transform normalisation.

  • We tested or method on four different datasets.

  • Experiments show that our approach is more accurate and robust against rotation than the traditional one.

Abstract

Grey-level co-occurrence matrices (GLCM) have been on the scene for almost forty years and continue to be widely used today. In this paper we present a method to improve accuracy and robustness against rotation of GLCM features for image classification. In our approach co-occurrences are computed through digital circles as an alternative to the standard four directions. We use discrete Fourier transform normalisation to convert rotation dependent features into rotation invariant ones. We tested our method on four different datasets of natural and synthetic images. Experimental results show that our approach is more accurate and robust against rotation than the standard GLCM features.

Introduction

Grey-level co-occurrence matrices are among the most long-standing texture descriptors in use, their origin dating back to the pioneering work of Haralick et al. [23]. Though many other methods have been proposed since their appearance – see Ref. [44] for a comprehensive overview – GLCM continue to be very common and widely adopted still today. Bibliometric data reveal that the number of relevant scientific papers has even increased during the last years (see Table 1). GLCM features are particularly appealing for their conceptual simplicity, ease of implementation and the low number of features they produce. A recent comparative experiment on image classification under non-ideal conditions [27] showed that GLCM features tend to perform better when few classes (10 or less) are involved, a situation in which they can compete with newer and more powerful methods. Besides, co-occurrence features can be combined with other descriptors that convey complementary information through suitable fusion schemes [13]. Recent applications of GLCM span very diverse areas of image processing, including surface inspection [17], [7], environmental monitoring [3], [29], content-based image retrieval [39] and image reconstruction [4]. Among the numerous application areas, co-occurrence matrices seem to be particularly common in medical image analysis [24], [28], [6], [37], [20] and remote sensing [8], [42], [30], [26].

Co-occurrence matrices have been extended in various directions, leading to several variations such as generalised co-occurrence matrices [16], which consider the distribution of local maxima; integrative co-occurrence matrices [36], which operate on colour images and, more recently, pattern co-occurrence matrices [40], [21], which analyse the co-occurrence of local patterns. By contrast, the original formulation has not changed significantly since its appearance. This is not uncommon: when a method matures and new ones appear, scientific interest tends to switch from the former to the latter. Newer methods receive more attention, and the older becomes frozen, somewhat immutable, with few chances of improvement. Something of this type we believe has happened with co-occurrence matrices, at least for what it concerns rotation invariant features.

Motivated by the wide diffusion of the method – even in very critical areas like medical image analysis and computer-assisted diagnosis, we wished to investigate whether it was possible to improve robustness and accuracy of the method in rotation invariant classification tasks. This is a major concern, for in many applications images can occur in different and uncontrolled rotation angles. The common approach to obtaining rotationally-invariant features from co-occurrence matrices consists of averaging [23], [1], [11] or – equivalently – summing up [38, p. 215] the matrices corresponding to the same distance and different directions. We believe that this procedure reduces significantly and somewhat unnecessarily the discrimination capability of the resulting features. We therefore propose some improvements to compute more efficient rotationally-invariant features from GLCM. Our study considers the effects of two design factors that determine how GLCM features are computed. These are: (1) the spatial arrangement of pairs of pixels; and (2) the way to convert GLCM features into rotation invariant ones. In the remainder of the paper we first discuss such design factors (Section 2), then evaluate their effects through an image classification experiment (Section 3). We present and analyse the results in Section 4 and conclude with final considerations in Section 5.

Section snippets

Design factors

Grey-level co-occurrence matrices estimate the joint occurrence probability of grey levels at a given distance and direction. The method is intrinsically directional, hence sensitive to rotation. In order to achieve rotation invariant descriptors, we need to remove the dependence on direction and obtain features that depend on distance only. Such a goal can be obtained through the following steps: (1) for each pixel in the image, consider all pixels that are located approximately at a given

Experiments

We performed a set of image classification experiments to evaluate accuracy and robustness against rotation of the descriptors presented in the preceding section. Datasets and procedure used in the experiments are detailed in the following subsections.

Results and discussion

The overall results of the classification experiment are summarised in Table 3. For each descriptor and dataset the table reports the results in the form μ±σ, where μ and σ are, respectively, the mean and standard deviation of the classification accuracy (in %) over the different rotation angles. The figures in boldface highlight the highest accuracy values. Numeric subscripts indicate the distance (d) at which the treatments have been computed. The complete results for each rotation angle are

Conclusions

Co-occurrence matrices are among the most used image descriptors, with applications covering almost any area of image analysis. They are also one of the oldest – they first appeared on the scene forty years ago. In this study we investigated whether it was possible to improve robustness and accuracy of the method for rotation invariant image classification. The results of our study are in the affirmative: there is room for improvement.

As for the type of neighbourhood, digital circles proved

References (45)

  • F. Alam et al.

    Optimized calculations of Haralick texture features

    Eur. J. Sci. Res.

    (2011)
  • All-free-download, 2012, All-free-download,...
  • N. Azzabou et al.

    Image reconstruction using particle filters and multiple hypotheses testing

    IEEE Trans. Image Process.

    (2010)
  • M. Barrera, M. Andrade, H. Kim, Texture-based fuzzy system for rotation-invariant classification of aerial orthoimage...
  • A. Barwad et al.

    Artificial neural network in diagnosis of metastatic carcinoma in effusion cytology

    Cytometry Part B (Clinical Cytometry)

    (2012)
  • Y. Ben Salem et al.

    Automatic recognition of woven fabrics based on texture and using SVM

    Signal Image Video Process.

    (2010)
  • F. Bianconi et al.

    Robust colour texture features based on ranklets and discrete Fourier transform

    J. Electron. Imaging

    (2009)
  • V. Bino Sebastian et al.

    Gray level co-occurrence matrices: generalisation and some new features

    Int. J. Comput. Sci. Eng. Inform. Technol.

    (2012)
  • P. Brodatz

    Textures: A Photographic Album for Artists and Designers

    (1966)
  • D.A. Clausi et al.

    Design-based texture feature fusion using Gabor filters and co-occurrence probabilities

    IEEE Trans. Image Process.

    (2005)
  • W. Cochran et al.

    Experimental Designs

    (1957)
  • M. Crosier et al.

    Using basic image features for texture classification

    Int. J. Comput. Vision

    (2010)
  • Cited by (32)

    • Autoencoder-based anomaly detection for surface defect inspection

      2021, Advanced Engineering Informatics
      Citation Excerpt :

      Discriminant classifiers are then applied to identify defects. The statistical features or texture features based on local variance [3], local binary pattern [4] and co-occurrence matrices [5,6] are the commonly used spatial approach for surface defect detection. The spectral features based on Fourier transforms [7,8], Gabor transforms [9,10] and wavelet transforms [11,12] are the typical approach for defect detection in textured surfaces.

    • A novel local contrast fusion-based fuzzy model for color image multilevel thresholding using grasshopper optimization

      2019, Applied Soft Computing Journal
      Citation Excerpt :

      In addition, the computational complexity is better than those of many optimization techniques in the literature. These powerful features motivated our attempts to propose a multilevel thresholding algorithm inspired from the social behavior of grasshoppers in nature [41]. There are also other algorithms with different inspiration in the literature.

    • Fast and accurate computation of orthogonal moments for texture analysis

      2018, Pattern Recognition
      Citation Excerpt :

      We did not use the whole collection, since the original images included in the album are not rotated, but just a subset of 13 textures. This subset has been proposed by Bianconi [40], who acquired hardware-rotated images directly from the original book. The images have been acquired with angular steps of 10°, like in the previous case.

    • An optimal color image multilevel thresholding technique using grey-level co-occurrence matrix

      2017, Expert Systems with Applications
      Citation Excerpt :

      In Haralick's formulation, pixel's neighborhood is created using the central pixel and four peripheral pixels, which are spaced equally at angular intervals of 45°. For a given distance d, the relative coordinates of pixels in peripheral of central pixel are: (0, d), (-d, d), (-d, 0), (-d, -d) (Bianconi & Fernandez, 2014). Value of d is considered as unity for simplifying and reducing the computational complexity.

    View all citing articles on Scopus

    This paper has been recommended for acceptance by Edwin Hancock.

    View full text