Heterogeneous patterns enhancing static and dynamic texture classification

04/16/2013 ∙ by Núbia Rosa da Silva, et al. ∙ Universidade de São Paulo 0

Some mixtures, such as colloids like milk, blood, and gelatin, have homogeneous appearance when viewed with the naked eye, however, to observe them at the nanoscale is possible to understand the heterogeneity of its components. The same phenomenon can occur in pattern recognition in which it is possible to see heterogeneous patterns in texture images. However, current methods of texture analysis can not adequately describe such heterogeneous patterns. Common methods used by researchers analyse the image information in a global way, taking all its features in an integrated manner. Furthermore, multi-scale analysis verifies the patterns at different scales, but still preserving the homogeneous analysis. On the other hand various methods use textons to represent the texture, breaking texture down into its smallest unit. To tackle this problem, we propose a method to identify texture patterns not small as textons at distinct scales enhancing the separability among different types of texture. We find sub patterns of texture according to the scale and then group similar patterns for a more refined analysis. Tests were performed in four static texture databases and one dynamic one. Results show that our method provides better classification rate compared with conventional approaches both in static and in dynamic texture.

READ FULL TEXT VIEW PDF
POST COMMENT

Comments

There are no comments yet.

Authors

page 3

page 4

This week in AI

Get the week's most popular data science and artificial intelligence research sent straight to your inbox every Saturday.

I Introduction

Pattern recognition is the identification and interpretation of patterns in images, in order to extract relevant information on the image to identify and classify your content. Classification of patterns can be used in a variety of applications in different fields such as nanotechnology

Florindo et al. (2012, 2013); Annampedu and Wagh (2007); Oh et al. (2011), biology de Mesquita Sá Junior et al. (2013); Rossatto et al. (2011); Backes et al. (2011); Galas et al. (1985); Isasi et al. (2011), medicine Landeweerd et al. (1981) and computer science Este et al. (2009); Canals et al. (2010). Different approaches have been developed according to the application, however, most of them analyze the information in a global way, using all the features in an integrated manner.

One side in the pattern classification approach uses multi-scale analysis of patterns, that is, different scales of observation are used to perform the analysis and find similar patterns, because important structures in an image usually occur at different spatial scales Koenderink (1984) Witkin (1983). Methods based on textons Julesz (1981); Leung and Malik (1999) represent each pixel of a texture as the convolution of a multi-scale and multi-orientation filter bank producing a texton vocabulary. Thus, these methods of texture analysis characterize the image homogeneously on the scale to be analyzed. Both the overall analysis, such as multi-scale approach is appropriate for the vast majority of problems in pattern recognition. However, in some problems, due to the heterogeneous nature of the composition of the objects under consideration, it is needed one more step in the process of pattern recognition, the analysis of heterogeneous patterns.

The aim of this paper is to demonstrate how can apply heterogeneous analysis to improve results regarding the homogeneous analysis. To validate our proposal experiments were performed on four static e one dynamic texture databases using the same texture descriptor but with different approaches. In all tests, heterogeneous analysis proved to be better than the homogeneous analysis enhancing the rate classification.

This paper is organized as follows: Section II describes the heterogeneous pattern analysis. Section III shows the results and discussions and in Section IV the conclusions.

Ii Method

Heterogeneous pattern arises when the object under analysis presents combined patterns in its composition. Similar to the classic definition of heterogeneous compositions in chemical compounds Staley et al. (2004), which is the characteristic of presenting a different appearance or composition when analyzed in parts. The same analogy can be applied in recognizing patterns in images. In each image can be found heterogeneous patterns, i. e., it is possible to distinguish the different patterns in each of the texture images. Figure 1 shows an image from Brodatz, which analyzed by conventional methods, would be defined only a pattern for this image. Using the approach of heterogeneous patterns, two different types of texture patterns are identified in its formation.

This new view to analyze the various patterns requires a new approach for analysis of similarity between images. The first step is to segment the texture by defining regions where the image belong to a given pattern while defining the patterns in the image. Figures

1 and 1 illustrate the segmentation of the image of Figure 1 in two patterns.

Figure 1: Segmentation according to the heterogeneous patterns. (a) Original image. (b) and (c) Two patterns obtained from (a).

To obtain this result the image was divided into smaller sized windows of 8 8 pixels, for windows with texture sufficiently homogeneous, that is, where only one pattern is found. Furthermore, there was obtained a characteristic to represent each window. Windows with similar characteristics remain the same pattern, and those with distinct characteristics were separated into different patterns. Figures 1 and 1 show the windows stayed grouped using k-means algorithm according to their pattern with the remaining windows in each pattern. Textural features from Haralick Haralick et al. (1973) descriptors were used to characterize each window.

Haralick descriptors Haralick et al. (1973) are based on the spatial gray level dependence matrices, or co-occurrence matrix. Contrast, Correlation, Energy and Homogeneity were computed from resulting co-occurrence matrices to obtain a set of 32 descriptors for each window. Let be the gray levels and a matrix of relative frequencies of two neighboring resolution cell with intensity and separated by distance and direction . Extracted features are described as follows:

(1)
(2)
(3)
(4)

and

are, respectively, means and standard deviations of the sum of elements of each row and column of the co-occurrence matrix. Distances of 1 and 2 pixels with angles of

, , and were used.

After the windows were divided into patterns, we performed a survival analysis of windows, 25% of the windows with features farthest from the group average were discarded making the representation of pattern more consistent. Each pattern was characterized by a feature vector average that is the average of the characteristics of all windows belonging to the pattern. The next step is to find the best matching among the patterns to define the image similarity as Figures

2 and 2 exemplify. They show two examples where it has an optimal matching among the image patterns of the same class. When the best fit among all the patterns in each image is found, we have the degree of similarity, defining whether the images belong to the same class. Figure 2 shows an example where it is not possible to perform the fitting patterns and all possibilities of fitting are tested. In this case the images belong to different classes.

Figure 2: Matching patterns. (a) and (b) matching of two images from the same class. (c) Matching of patterns can not be made because the images belong to different classes.

However, this approach has a special feature when the same pattern can be found in different classes of texture (See Figure 3). This one influences the analysis of the similarity degree between classes, because different classes can obtain a high rate of similarity, since some patterns will have great matching.

Figure 3: Same pattern in different classes.

Iii Results and Discussions

To validate the method of heterogeneous patterns analysis, this technique was applied in four different texture databases, each with its peculiarities: Brodatz, USPTex, Vistex and Outex and one dynamic texture database: Dyntex. Figure 4 shows some samples of the texture databases and Figure 5 shows examples of dynamic texture database Dyntex. The classification method -Nearest Neighbor () with 10-fold cross-validation scheme was used in all experiments.

Brodatz

Brodatz (1966) contains 1110 natural textures of 200 200 pixels divided into 111 classes (Figure 4).

Vistex

vis (2009) contains 864 images of 128 128 pixels size with 54 texture classes (Figure 4).

USPTex

Backes et al. (2012) has 3984 natural texture images of 128 128 pixels size divided into 332 classes (Figure 4).

Outex

Ojala et al. (2002) has 1360 images of 128 128 pixels size divided into 68 classes (Figure 4).

Dyntex

Péteri et al. (2010) consists of 1230 videos with 250 frames with 400 300 pixels size divided into 123 classes of dynamic texture (Figure 5).

Figure 4: Samples of (a) Brodatz, (b) Vistex, (c) Usptex and (d) Outex.
Figure 5: Samples images from Dyntex, a dynamic texture database.

The method used as classical approach was Haralick descriptors with the same configurations of characterization of windows. This way we can compare the same descriptor but with different perspectives. One of the most important parameters is the window size that directly interferes in the results. Large windows remain with the general representation of the image, preserving the homogeneous analysis of patterns. However, becomes more homogeneous window, or small windows, the classification rate increases, as the representative pattern of the window increases. For this experiment two patterns were analyzed in each image. Table 1 shows the results obtained when using the conventional method in comparison with heterogeneous patterns, in all cases heterogeneous pattern method improved the classification rate proving to be better than the standard approach in which only one pattern is established for the entire image.

Database Brodatz   Vistex Usptex   Outex Dyntex
Classical approach 88.92 89.47 69.58 76.62 76.89
Heterogeneous pattern 93.06 93.63 78.49 76.99 98.29
Table 1: Accuracy rate by comparing the traditional method of analysis with the heterogeneous method.

Iv Conclusions

Usually, analysis in texture pattern recognition using the homogeneous approach taking all the image information in a global way, that is, the entire image is defined as one pattern. However, it is possible to find more than one texture pattern on the image that we called here, heterogeneous pattern. It is necessary to map the patterns in each image and then check the similarity of patterns among them. This method improved the accuracy rate of classification using the same method to characterize the image showing robustness evaluating heterogeneous patterns on images.

Acknowledgments

Núbia Rosa da Silva acknowledges support from FAPESP (2011/21467-9) and Odemir Martinez Bruno acknowledges the financial support of CNPq (308449/2010-0 and 473893/2010-0) and FAPESP (2011/01523-1).

References

  • Florindo et al. (2012) J. B. Florindo, M. S. Sikora, E. C. Pereira,  and O. M. Bruno, “Multiscale fractal descriptors applied to nanoscale images,” Journal of Superconductivity and Novel Magnetism, 1–6 (2012).
  • Florindo et al. (2013) J. B. Florindo, M. S. Sikora, E. C. Pereira,  and O. M. Bruno, “Characterization of nanostructured material images using fractal descriptors,” Physica A: Statistical Mechanics and its Applications, 392, 1694–1701 (2013), ISSN 0378-4371.
  • Annampedu and Wagh (2007)

    V. Annampedu and M. D. Wagh, “Reconfigurable approximate pattern matching architectures for nanotechnology,” Microelectronics Journal, 

    38, 430 – 438 (2007), ISSN 0026-2692.
  • Oh et al. (2011) E. H. Oh, H. S. Song,  and T. H. Park, “Recent advances in electronic and bioelectronic noses and their biomedical applications,” Enzyme and Microbial Technology, 48, 427 – 437 (2011), ISSN 0141-0229.
  • de Mesquita Sá Junior et al. (2013)

    J. J. de Mesquita Sá Junior, D. R. Rossatto, R. M. Kolb,  and O. M. Bruno, “A computer vision approach to quantify leaf anatomical plasticity: a case study on gochnatia polymorpha (less.) cabrera,” Ecological Informatics, 

    15, 34 – 43 (2013), ISSN 1574-9541.
  • Rossatto et al. (2011) D. Rossatto, D. Casanova, R. Kolb,  and O. M. Bruno, “Fractal analysis of leaf-texture properties as a tool for taxonomic and identification purposes: a case study with species from neotropical melastomataceae (miconieae tribe),” Plant Systematics and Evolution, 291, 103–116 (2011).
  • Backes et al. (2011) A. R. Backes, D. Casanova,  and O. M. Bruno, “IdentificaÁ„o de plantas por an·lise de textura foliar,” Learning and Nonlinear Models, 9, 84–90 (2011).
  • Galas et al. (1985) D. J. Galas, M. Eggert,  and M. S. Waterman, “Rigorous pattern-recognition methods for dna sequences: Analysis of promoter sequences from escherichia coli,” Journal of Molecular Biology, 186, 117 – 128 (1985), ISSN 0022-2836.
  • Isasi et al. (2011) A. G. Isasi, B. G. Zapirain,  and A. M. Zorrilla, “Melanomas non-invasive diagnosis application based on the abcd rule and pattern recognition image processing algorithms,” Computers in Biology and Medicine, 41, 742 – 755 (2011), ISSN 0010-4825.
  • Landeweerd et al. (1981) G. Landeweerd, E. Gelsema, M. Bins,  and M. Halie, “Interactive pattern recognition of blood cells in malignant lymphomas,” Pattern Recognition, 14, 239 – 244 (1981), ISSN 0031-3203.
  • Este et al. (2009)

    A. Este, F. Gringoli,  and L. Salgarelli, “Support vector machines for tcp traffic classification,” Comput. Netw., 

    53, 2476–2490 (2009), ISSN 1389-1286.
  • Canals et al. (2010) V. Canals, A. Morro,  and J. L. Rossella, “Stochastic-based pattern-recognition analysis,” Pattern Recognition Letters, 31, 2353 – 2356 (2010), ISSN 0167-8655.
  • Koenderink (1984) J. Koenderink, “The structure of images,” Biological Cybernetics, 50, 363–370 (1984), ISSN 0340-1200.
  • Witkin (1983) A. P. Witkin, “Scale-space filtering,” in IJCAI (1983) pp. 1019–1022.
  • Julesz (1981) B. Julesz, “Textons, the elements of texture perception, and their interactions,” Nature, 290, 91–97 (1981).
  • Leung and Malik (1999) T. Leung and J. Malik, “Recognizing surfaces using three-dimensional textons,” in Computer Vision, 1999. The Proceedings of the Seventh IEEE International Conference on, Vol. 2 (1999) pp. 1010–1017.
  • Staley et al. (2004) D. D. Staley, M. S. Matta,  and E. L. Waterman, Chemistry, edited by A. C. Wilbraham (Pearson Prentice Hall, 2004).
  • Haralick et al. (1973) R. M. Haralick, K. Shanmugam,  and I. Dinstein, “Textural features for image classification,” IEEE Transactions on Systems, Man and Cybernetics, 3, 610–621 (1973), ISSN 0018-9472.
  • Brodatz (1966) P. Brodatz, Textures, a photographic album for artists and designers (Dover Publications New York, 1966) ISBN 0486216691, p. 112.
  • vis (2009) “Vision texture database,”  (2009).
  • Backes et al. (2012) A. R. Backes, D. Casanova,  and O. M. Bruno, “Color texture analysis based on fractal descriptors,” Pattern Recognition, 45, 1984 – 1992 (2012), ISSN 0031-3203.
  • Ojala et al. (2002) T. Ojala, T. Mäenpää, M. Pietikäinen, J. Viertola, J. Kyllönen,  and S. Huovinen, “Outex - new framework for empirical evaluation of texture analysis algorithms,” in Proceedings of the 16 th International Conference on Pattern Recognition (ICPR 2002), ICPR 2002, Vol. 1 (IEEE Computer Society, 2002) p. 10701.
  • Péteri et al. (2010) R. Péteri, S. Fazekas,  and M. J. Huiskes, “DynTex : a Comprehensive Database of Dynamic Textures,” Pattern Recognition Letters, 31, 1627–1632 (2010), http://projects.cwi.nl/dyntex/.