Elsevier

Pattern Recognition Letters

Volume 78, 15 July 2016, Pages 22-27
Pattern Recognition Letters

Local fractal dimension and binary patterns in texture recognition

https://doi.org/10.1016/j.patrec.2016.03.025Get rights and content

Highlights

  • A new approach is proposed to extract features from grey-level texture images.

  • Based on computing binary patterns from the local fractal dimension.

  • The method was applied to the classification of two well-known benchmark databases.

  • The proposal outperformed other state-of-the-art and classical approaches.

  • The new descriptors are potentially useful for general purpose applications.

Abstract

The present work proposes a new texture image descriptor, combining the local binary patterns extracted from the grey-level image (classic approach) with those extracted from the local fractal dimension at each point of the image. In this way, these descriptors express two important measurements from the image, i.e., the variation among pixel intensities in each local neighbourhood and the local complexity (pixel arrangement) at each point. Such combination provides a rich and robust descriptor even for the most complex textures. The effectiveness of the proposed solution is evaluated in the classification of two well-known benchmark databases: UIUC and USPTex, showing that the combined features outperform all the other compared approaches in terms of correctness rates in the classification of grey-scale texture images.

Introduction

Fractal-based techniques [4], [15] and Local Binary Patterns (LBP) [18] have been applied to a large number of problems in different areas, particularly, those problems that can be addressed by the analysis of texture images [7], [8].

Since the initial works, the literature has presented a number of improvements for both techniques. Regarding LBP, several adaptations have been described both to enhance the image analysis as well as to best fit some particular applications. For instance, in [19], an enhanced rotation-invariant descriptor is proposed. In [10], the authors propose the completed LBP (CLBP) by separately considering the magnitudes and signs of the differences among neighbour pixels. Still, in [20] an extension to a ternary system to represent local patterns is proposed. Besides, adaptations of the method can be easily found for face recognition [1], dynamic textures [24], iris recognition [14], video surveillance [2], etc.

On the other hand, fractal-based analysis has also evolved in several ways since Mandelbrot [15]. The fractal dimension was extended to more complete descriptors in methods like multifractals [23], where the fractal dimension is computed over sets of pixels in the image satisfying certain regularity criteria. In multiscale fractal dimension [17], some geometrical features are extracted from the self-similarity curve of an object to compose the descriptors. In fractal descriptors [4], the entire self-similarity curve is used. In local fractal dimension approaches [13], a fractal measure is estimated over each pixel by taking into account the respective neighbourhood.

Although both fractal-based and LBP techniques present interesting results, the combination of them is still not explored in texture analysis. A particular motivation to verify such combination is the local approach efficiently utilized by LBP, but that could be leveraged by considering measures more advanced than the simple pixel intensity. In fact, the grey-level of a pixel is a simplistic measure of the image and cannot faithfully express important properties of the image, such as light changing, roughness, etc. On the other hand, fractal measures provide a realistic framework for describing natural structures and can richly represent these properties.

In this context, this work proposes to combine the LBP features computed over the grey-level values with those computed over the local fractal dimension of the image. The effectiveness of the proposed features is assessed on classifying two well-known texture data sets used as benchmark, that is, UIUC [12] and USPTex [3]. The success rate of our method is compared to other texture descriptors in the literature and our approach outperforms all of them.

Section snippets

Local binary patterns

Local binary patterns, introduced in [18], is one of the most successful approaches to texture analysis, providing great results in many applications. The method constructs a histogram of patterns of threshold in the neighbourhood of each pixel in the image. To obtain the code of a pixel, the intensity of this pixel is compared to the pixels in its 8-neighbourhood in clockwise direction. If the neighbourhood pixel intensity is greater than the central pixel, we assign the value 1, otherwise, we

Fractal theory

Fractals are geometrical structures characterized by two key properties: infinite self-similarity and infinite complexity. The first property implies that a fractal is composed of repeated copies of itself at reduced scales, while the second one means that the object has different structural details at any scale. The geometry of such unconventional objects, named fractal geometry, is an interesting alternative to the classic Euclidean geometry. As well as in the Euclidean reference we can

Proposed method

The effectiveness of LBP method is well known and discussed in the literature. Nevertheless, while the pixel intensities provide an essential and direct view of the texture, there are other perspectives of the same texture described in the literature and that could be better explored in the LBP method.

One of such alternative views from the image is given by the fractal dimension and more particularly the local fractal dimension in each pixel of the texture. The fractal dimension measures how

Experiments

The proposed features are assessed on the classification of two data sets of textures widely used in the literature as benchmark data, that is, UIUC [12] and USPTex [3].

UIUC is a set of grey-level textures extracted from photographies of natural scenarios (uncontrolled conditions) with high variance in scale and viewpoint. The images have a resolution of 512 × 512 and the set contains 25 classes with 40 texture images in each one.

USPTex is a database of colour textures obtained by photographing

Results and discussion

Table 1 shows the success rates achieved by each compared descriptor on the benchmark data. The proposed approach outperformed the original LBP method and all the other approaches. The addition of the fractal measurement enhanced the LBP performance as now the descriptors become less sensitive to unexpected variation in the pixel level, as those caused by noises. A single corrupted pixel in LBP classic descriptors provides a different decimal code. However, the local fractal dimension will

Conclusion

This work proposed a texture descriptor that applied the LBP encoding to the local fractal dimension at each point in the image and combined these features with the classic grey-level LBP.

We assessed the effectiveness of the proposed approach in the classification of well-known texture databases and compared the outcomes with other state-of-the-art texture descriptors. The results showed that our method outperformed all the other compared methods. They also pointed out that our approach

Acknowledgements

O.M. Bruno gratefully acknowledges the financial support of CNPq (National Council for Scientific and Technological Development, Brazil) (Grants #307797/2014-7 and #484312/2013-8) and FAPESP (The State of São Paulo Research Foundation) (Grant #11/01523-1). J.B. Florindo gratefully acknowledges the financial support of FAPESP Proc. 2012/19143-3.

References (24)

  • R.C. Gonzalez et al.

    Digital Image Processing

    (2002)
  • Z. Guo et al.

    A completed m odeling of local binary pattern operator for texture classification

    IEEE Trans. Image Process.

    (2010)
  • Cited by (17)

    • Texture characterization via deterministic walks’ direction histogram applied to a complex network-based image transformation

      2017, Pattern Recognition Letters
      Citation Excerpt :

      Works such as [5,18–20] use deterministic walks (sometimes referred to as crawlers [21]) to characterize images, employing mainly an analysis of the size of the walks. Recent efforts such as in [21–25] have explored analyzing textures using measures of fractal dimension as a robust discriminating feature, often combined with walks in an endeavor to go beyond the size of the walks and analyze their shapes. This work proposes a novel approach to texture classification by extracting directional information from deterministic walks applied on a proposed image transformation.

    View all citing articles on Scopus

    This paper has been recommended for acceptance by J. Laaksonen.

    View full text