Enhancing Volumetric Bouligand-Minkowski Fractal Descriptors by using Functional Data Analysis

by   João Batista Florindo, et al.

This work proposes and study the concept of Functional Data Analysis transform, applying it to the performance improving of volumetric Bouligand-Minkowski fractal descriptors. The proposed transform consists essentially in changing the descriptors originally defined in the space of the calculus of fractal dimension into the space of coefficients used in the functional data representation of these descriptors. The transformed decriptors are used here in texture classification problems. The enhancement provided by the FDA transform is measured by comparing the transformed to the original descriptors in terms of the correctness rate in the classification of well known datasets.


page 4

page 8

page 11

page 17

page 18


Fractal Descriptors Based on Fourier Spectrum Applied to Texture Analysis

This work proposes the development and study of a novel technique for th...

Deterministic Decoupling of Global Features and its Application to Data Analysis

We introduce a method for deterministic decoupling of global features an...

Fractal Descriptors in the Fourier Domain Applied to Color Texture Analysis

The present work proposes the development of a novel method to provide d...

Multiscale Fractal Descriptors Applied to Nanoscale Images

This work proposes the application of fractal descriptors to the analysi...

Functional dissipation microarrays for classification

In this article, we describe a new method of extracting information from...

Enhancing fractal descriptors on images by combining boundary and interior of Minkowski dilation

This work proposes to obtain novel fractal descriptors from gray-level t...

Binary Distance Transform to Improve Feature Extraction

To recognize textures many methods have been developed along the years. ...

1 Introduction

In recent years, the literature has presented a lot of applications of fractal theory to the solution of problems from distinct areas. As examples we may cite applications in Botany [1, 2, 3], Medicine [4, 5, 6] and Geology [7, 8, 9]. Particularly, in Physics, we may find applications of fractal theory in Optics [10, 11, 12], Materials Science [13, 14, 15] and Electromagnetism [16, 17, 18], among many other areas. Such large amount of works exploring tools from fractal theory is fully justified by an interesting observation already pointed out in [19]. This observation states that systems observed in the nature generally may be modelled by fractal measures rather than by classical formalisms.

Among the applications of fractal theory, most of them aim at using the fractal modeling in order to extract features from objects of interest according to the problem domain, like textures, contours, surfaces, etc. Such features are then provided as input data, for example, to methods for segmentation, classification and description of objects. A classical example of such fractal feature is the fractal dimension.

As in the most of cases the simple use of fractal dimension is still not sufficient to well represent the complexity of an object or scenario from the real world, the literature developed techniques for the extraction of a set of features based on the fractal dimension. Examples of such approaches are Multifractal theory [20, 21, 22], Multiscale Fractal Dimension (MFD) [23, 24] and Fractal Descriptors [25, 26, 27, 28].

Here, we are focused on fractal descriptors approach. Several authors, like in [27, 25, 26, 28], obtained interesting results in different applications of fractal descriptors technique to texture and shape analysis, mainly in the description of natural objects. Particularly, here we are focused on an approach developed in [26] which uses the volumetric Bouligand-Minkowski fractal dimension to generate a set of descriptors. Such descriptors obtained a high performance in an application to a task of plant leaves classification based on texture.

Nevertheless, an important drawback of fractal descriptors technique, particularly that based on Bouligand-Minkowski, is that the curve formed by the set of descriptors present a high correlation, that is, each descriptor is strongly dependent on each other. This correlation does their performance decrease drastically in problems of classification and segmentation with a high number of samples and classes. In such situations, volumetric Bouligand-Minkowski descriptors have severe limitations.

Aimimg at enhancing Bouligand-Minkowski descriptors, preserving the reliability of the results, this work proposes the development and use of Functional Data Analysis (FDA) transform concept. Functional Data Analysis is a powerful statistical tool developed in [29]. It represents an alternative to the traditional multivariate approach and deals with complex data as being a simple analytical function: the functional data. FDA approach presents certain advantages in this kind of application, like the easy handling of data in nonlinear domains (as the case in Bouligand-Minkowski descriptors) and the intuitive notion of functional operations, like derivatives and smoothing, employed in the definition of fractal descriptors.

Up our knowledge, Florindo et al. [28] is the first work to apply the FDA approach to fractal descriptors. In that work, functional data representation is used for reducing the dimensionality of the descriptors set in shape recognition problems. Here, we propose a different paradigm for FDA use, by defining the concept of FDA transform. The FDA transform is defined as the operation which changes the original data set (in this case, descriptors) space into the space of coefficients of functional data. The transform still presents two variants: the first uses the coefficient directly, the second performs a second algebraic transform, described in [30].

The relevance of the FDA transform is verified in experiments of classification of two well known datasets, that is, Brodatz [31] and OuTex [32]

. The results are compared in terms of classification correctness rate. It was considered two variants of the FDA transform and it was compared through three classifiers very well known in the literature: Linear Discriminant Analysis (LDA), K-Nearest Neighbors (KNN) and Bayesian

[33, 34, 35].

This work is divided into seven sections, including this introduction. The following explains the concepts of fractal theory, fractal dimension and fractal descriptors. The third introduces the Functional Data Analysis theory and definitions. The fourth shows the proposed method. The fifth describes the experiments. The sixth section shows the results and the last section concludes the work.

2 Fractal Analysis

The literature shows a lot of applications of fractal geometry involving the characterization of natural objects and scenarios. Examples of such applications may be found in [10, 13, 16, 2, 8, 11, 15, 18]. Most of these works use the fractal dimension as a metric for describe the object. This strategy is justified by the fact that fractal dimension measures the complexity of a structure. Physically, the complexity corresponds to the irregularity or to the spatial occupation. These properties are tightly related to constitution aspects which allow the identification of such objects.

An important drawback of using only fractal dimension is that it is a unique global value and is not capable of extract information about intricate details of a structure. With the aim of exploring fully the potential of fractal theory, the literature shows the development of techniques which provide not only a unique value but a set of values capable of describing in a richer way an object, based on the fractal theory. Among these techniques, we have the Multifractal [20, 21, 22], the Multiscale Fractal Dimension [23, 24] and the Fractal Descriptors [25, 26, 27, 28].

Multifractal theory replaces the fractal dimension analysis by the concept of fractal spectrum, capable of modeling objects which cannot be represented by a single fractal measure. Multifractal demonstrates to be an interesting tool to capture the different power-law scaling present in a system [20, 21, 22].

The literature still shows an alternative technique for the modeling of objects with fractal theory. This approach is the Multiscale Fractal Dimension (MFD) [23, 24]. In MFD approach, instead of simply calculate the fractal dimension from interest objects, a set of features is extracted from the derivative of the whole power-law curve used to provide the fractal dimension.

An extension of MFD are the fractal descriptors [25, 26, 27, 28]

. In this case, we extract features (descriptors) from an object through the calculus of the fractal dimension taking the object under different observation scales. These descriptors are used to compose a feature vector that could be mean as a “signature” to characterize the object. Particularly, fractal descriptors demonstrate to be an efficient tool for the discrimination of natural textures like that analyzed in the present work. The Figure

1 illustrates the discrimination power of fractal descriptors by showing two distinct textures whose fractal dimensions are identical but the curve of fractal descriptors is visually distinct.

Figure 1: Two textures with the same fractal dimension present fractal descriptors totally different. (a) Original textures (both with fractal dimension 2.618). (b) Fractal descriptors from the same textures.

The following sections describe in more details the aspects involved in fractal descriptors technique, starting from the fractal dimension definition.

2.1 Fractal Dimension

Fractal dimension is a real positive number constituting the main measure extracted from a fractal object. There is no absolute definition for the concept of fractal dimension. The most used and classical one is the Hausdorff-Besicovitch dimension.

Hausdorff-Besicovitch dimension is a concept derived from the measure theory and is defined over a set as


where is the dimensional Hausdorff-Besicovitch measure, defined by




In above equations, expresses for the diameter in , that is, .

In many situations, the calculus of Hausdorff-Besicovitch dimension is very complex and even impracticable. In such cases, we can calculate it by generalizing the concept of classical Euclidean dimension [19]. In this way, we obtain the following expression


where is the minimum number of objects with linear size needed to cover [36].

Most of different definitions of fractal dimension are based on a generalization of Equation 4, expressed through


where is a set measure depending on the specific fractal dimension method and is the scale parameter. As example of fractal dimensions defined from the previous expression we can cite the box-counting, the packing dimension, the Renyi dimension, etc [36].

Particularly, here we are focused on the Bouligand-Minkowski fractal dimension [36]. As the Hausdorff-Besicovitch dimension, the Bouligand-Minkowski dimension also is based on a topological measure, in this case, the Bouligand-Minkowski measure calculated through


where is the object (set) of interest, is a structuring element with radius and is the volume of the dilation between and the boudary of . The Bouligand-Minkowski dimension itself is given by


For an application to discrete objects represented in a digital image, the calculus is significantly simplified through the use of neighborhood techniques. In this way, the above expression becomes


in which is a disk with diameter (also called dilation radius), is the number of points pertaining to the dilation region and is the topological dimension of the space in which is immersed.

2.2 Multiscale Fractal Dimension

Although fractal dimension is an important measure, it is insufficient for a good representation of complex systems which present different fractal dimension depending on the observation scale taken into account. In order to provide a richer fractal-based information from an object, the literature shows the Multiscale Fractal Dimension (MFD) [23, 24].

MFD consists in the application of a multiscale transform to the fractal dimension. The multiscale transform of a signal is the function , where is directly associated with and is the scale variable. Essentially, the multiscale is performed through three approaches: scale-space, time-frequency and time-scale. In the following, we describe the approach used in MFD, e.g., scale-space. More details are found in [24].

Scale-space is a particular case of multiscale transform. It is based on the derivative of the signal followed by a convolution with a smoothing gaussian filter [37]:

where expresses the zero-crossings and represents the convolution of the original signal with the first derivative of the gaussian , that is:

In [23], the MFD is obtained from the Bouligand-Minkowski fractal dimension in the following manner:


where is the dilation area for each dilation radius . In MFD technique, some characteristics of MFD curve, like maximum, minimum and area below the curve graph, are extracted to compose a feature vector for the analyzed object.

2.3 Fractal Descriptors

Fractal descriptors [25, 26, 27, 28] are an extension of MFD concept where a feature vector is extracted from the fractal dimension calculated over a whole interval of scales.

Generally speaking, fractal descriptors are obtained from the function :


is a measure depending on the fractal dimension estimation method and

is the scale parameter.

The function must be used directly, as in [26], or may be summited to a particular transform. For instance, in [25], the descriptors are extracted from the Fourier derivative of :

where is equivalent to ,

is the Fourier transform of

and is Fourier derivative:

where is the imaginary number. In order to attenuate noises inherent to the derivative operation, one may still apply a convolution with a gaussian filter embedded in the Fourier derivative, as employed in [28]. Thus, the above expression becomes:

where is the derivative of the gaussian in the Fourier domain. The Figure 2 shows the aspect of descriptors curve of an object.

Figure 2: Fractal descriptors curve. (a) Object analyzed. (b) Curve . (c) Descriptors curve after Fourier derivative, followed by gaussian convolution.

In [27]

, the descriptors are obtained from the Fourier derivative, followed by a Principal Component Analysis (PCA) transform, aiming at reduce correlation among descriptors. In this way, a more reliable and consistent set of descriptors are provided to characterize plant leaf shapes analyzed in that work.

Here, we propose the application of Functional Data Analysis, described in the following, as a transform to , in order to generate more robust and precise fractal descriptors.

2.4 Volumetric Bouligand-Minkowski Fractal Descriptors

In this work we focus on a specific fractal descriptors approach developed in [26] called Volumetric Bouligand-Minkowski Fractal Descriptors (VBFD). The main idea is the calculus of Bouligand-Minkowski fractal dimension of a 3D surface taken under a range of observation scales. These descriptors are employed to describe texture images, that is, analysis of images based on spatial and color arrangement of pixels.

In the first step, we map the intensity image onto a threedimensional surface


such that


where is the maximum pixel intensity. This transform is illustrated in the Figure 3.

Figure 3: Texture mapped onto a surface. (a) Original texture. (b) Threedimensional surface.

In the following, each point of the surface is dilated by a sphere with variable radius , like illustrated in the Figure 4. Finally, we analyze the dilation volume , that is the number of points inside the structure composed by the dilation with each radius . also corresponds to the number of points with a distance at most from the object. Thus, the exact Euclidean Distance Transform (EDT) [38] becomes an efficient tool for this calculus.

Figure 4: Dilated surfaces with different radii. (a) Points from original surface. (b) Radius 2. (c) Radius 5. (d) Radius 10.

In 3D space, EDT is defined as the distance of each point in the space to a subset of it. In our case, this subset is the surface and the EDT for each point outside is given by


in which is the Euclidean distance.

Exact EDT is characterized by the fact that distances present discrete values




The dilation volume is provided by




such that




The Bouligand-Minkowski fractal dimension of the surface is simply given through


where is the slope of a straight line fit to the curve . The technique in [26] went beyond the simple fractal dimension calculus and uses all values as descriptors for texture: the VBFD descriptors. Notice that is directly related to the Bouligand-Minkowski dimension for maximum radius . Moreover, each radius corresponds to an observation scale, from further (greater radius) to closer (smaller radius).

In that work, they employed VBFD in the discrimination of plant leaves images, achieving excellent results. The Figure 5 shows the capability of MFD curve in discriminating two textures from different materials.

Figure 5: Discrimination power of fractal descriptors. (a) Textures from two different materials. (b) Fractal descriptors from each texture in a single graph.

Here, we propose the use of Functional Data Analysis [29] in order to enhance the performance of volumetric Bouligand-Minkowski descriptors.

3 Functional Data Analysis

Functional Data Analysis (FDA) [29, 39]

is a statistical tool alternative to multivariate analysis. While in multivariate statistic, we are interested in relations among observations of variables, in FDA, each observation of a set of variables is handled as a unique analytical function. Thus we extract measures from those functions, like derivatives, curvature, etc. In FDA terminology, each observation is called “data” and the function is called “functional data”.

FDA has found applications in Economy, Biology, Meteorology, etc. like synthesized in [39]. The functional representation has some noticeable advantages in practical applications. For example, we simplify the global analysis of an observation with missing values, irregular sampling domain or noise. Besides, the analytical representation allows the application of operations like derivatives of different orders, curvatures, integrals, etc. These operations turn possible, for example, a more accurate analysis of the variability level of the data among other important characteristics.

The main result of FDA theory attests that any observation of statistical variables with analytically smooth behavior may be perfectly represented by some analytical function. In practice, this function is not exactly known but we may obtain efficient approximations. The most common strategy for the obtainment of such approximations is the interpolation of the data through specific basis functions, like Fourier, wavelets, polynomials. Here we choose the use of B-splines functions, due to their flexibility in data representation. This process is also known as basis function development.

For the computational statistical analysis we must represent the analytical function numerically in some manner. Then, the coefficients of basis functions in the interpolation are used as the effective functional representation and statistical metrics, like mean and variance, are calculated from these coefficients. Here we also use a more complex representation in which the coefficients are previously summited to an algebraic transform. The Figure

6 illustrates FDA representation.

Figure 6: Functional data representation. From above, the curve of an original measure, the basis functions and the coefficients used to represent computationally the functional data.

3.1 Mathematical Foundations

For computational purposes, the data which must be analyzed consists in the observation of discrete pairs of variables , in which , are the pair values in the variable and is the number of observed pairs for each variable. For example, an observation may be the measure of temperature along a day and each pair may correspond to the hour in which the temperature is measured and the respective temperature value.

The fact is that in this hypothetical example, the temperature cannot be perfectly modelled by any continous function, capable of predicting exactly the temperature for any day and hour. In order to apply the FDA concept to such data, we need to use a very important result from FDA theory which affirms that a data capable of being analyzed through FDA may be represented by an analytical function if we take into consideration a tolerable approximation noise. The result demonstrates that we can find an analytical function such that


where measures the noise inherent to the acquisition process.

Although the function is not known explicitly, some classical function interpolation techniques have been applied to the observation pairs yielding an approximation of . Generally, this approximation method is based on the development of into functions basis. This technique consists in the projection of approximating function in a subspace with linearly independent basis functions . In this way, the approximating function is represented by . The values of are called the projection coefficients and are calculated as those which minimizes


Particularly, in this work, the functions basis choosen were the B-splines, which already demonstrated good results in the data here analyzed as shown in [28]. In the following, we describe briefly the B-splines concept.

3.2 B-splines

B-spline is a particular kind of spline function. A spline is a real function composed by piecewise polynomial functions.

More formally speaking, a spline may be defined as


The interval is divided into “knots” as


In each subinterval , the spline is given by the polynomial . The order of the spline corresponds to the highest order of polynomials. Each polynomial is called a basis of the spline function.

A B-spline is a particular category of splines characterized by minimum support (number of points where the function has value different of zero). Each B-spline basis is defined through


Finally, the B-spline curve is given by


where is the degree of the basis and corresponds to the knots.

4 Proposed Method

Beyond its importance as a statistical analysis tool, FDA has demonstrated to be an efficient technique to extract relevant information from a large data set. For example, in [40], a large amount of data respect to the water quality is collected in a specific local. Thus, the FDA approximating function is obtained from each curve of observed values and coefficients are used to extract important characteristics from data.

In [28], FDA is used to reduce the dimensionality and extract utile information from fractal descriptors used in a task of shape analysis. In that case, instead of using directly coefficients, it is employed a transform of that coefficients which takes into account the contribution of the function basis space used.

This transform is performed by the canonical transform matrix


where are the basis functions. Besides, to simplify the notation, we use , corresponding to the set of coefficients from the basis functions. Thus, the transformed coefficients are given through


where is the result from the matrix decomposed by the Choleski method [30], that is, is a unique lower triangular decomposition matrix of , such that , where is the conjugate transpose of .

The Figure 7 shows the discriminative power of FDA. We observe a data represented in two curves with similar visual aspect and the discrimination of FDA coefficients without and with transform.

Figure 7: Discriminative property of FDA. (a) Original data curve. (b) FDA coefficients. (c) Transformed coefficients.

The present work extends the application in [28] to the analysis of volumetric Bouligand-Minkowski descriptors, applied to texture classification. The motivation for this application comes from the fact that volumetric Bouligand-Minkowski descriptors corresponds to a typical case of data whose FDA representation is interesting, according to [29, 39]. In fact, the descriptors present a global smooth aspect, being therefore analytical. Besides, they are extracted from a nonlinear space (log-log curve) and then are provided by a domain irregularly spaced. Another motivation is the fact that fractal descriptors may involve a derivative operation which becomes more intuitive by the handling of descriptors as a function and not only as a simple set of non related values.

Unlike the situation in [28], the objective of using FDA here is not the simple dimensionality reduction, even because volumetric descriptors are easily treated by traditional classification methods. A problem with volumetric descriptors is that, although they allow for the achievemet of good results, they present a high level of correlation, that is, the original descriptors have a high dependence among themselves. This fact implies in difficulties for discrimination tasks envolving a large number of samples and classes. Our purpose is to evidentiate patterns in the global structure of descriptors which turn possible the enhancement of the discrimination power of original volumetric descriptors.

For this goal, we propose the use of direct coefficients or transformed coefficients replacing conventional fractal descriptors in a classification method. We call this representation form as FDA transform. In fact, we have a typical transform, in which the data is mapped from the log-log space of volumetric Bouligand-Minkowski descriptors onto the space of coefficients of functional data. The Figure 8 illustrates the FDA transform steps.

Figure 8: Functional data representation. A diagram illustrating the steps in the FDA transform applied to fractal descriptors.

5 Experiments

The performance of the FDA transform on volumetric Bouligand-Minkowski fractal descriptors is tested in an application to the classification of textures from two different datasets.

The first is the classical Brodatz texture dataset [31], composed by 111 classes, each one with 10 samples (images) corresponding to photographies of real world textures. The second analyzed dataset is the also well known Outex dataset [32], composed by 68 classes with 20 images in each class.

Figure 9: Some examples of images from the Brodatz dataset.
Figure 10: Some examples of images from the OuTex dataset.

The steps in the experiments may be summarized through the following items:

  1. Extraction of Volumetric Bouligand-Minkowski descriptors from each image in the dataset;

  2. Computation of coefficients of the approximating analytical function (functional data);

  3. Obtainment of the coefficients after the transform described in the Section 4;

  4. Use of coefficients and as input to different classification methods;

  5. Comparison, in terms of classification performance, among the proposed approach and the direct use of volumetric Bouligand-Minkowski descriptors.

The performance of the FDA transform is verified in direct approach (using ) and transformed approach (using ). The basis used was B-spline. For the classification process we use classical methods from the literature [33], that is, Bayesian, K-Nearest Neighbor (KNN) and Linear Discriminant Analysis (LDA).

6 Results

The results are showed in graphs and tables which represent the different ways for the use of the FDA transform combined to Bouligand-Minkowksi descriptors in the datasets analyzed. Empirically, we found an optimal interval for the number of descriptors used, that is, between 60 and 100 for direct FDA coefficients and between 10 and 50 for transformed FDA fractal descriptors.

Firstly, the Figure 11 shows the correctness rate for the use of FDA fractal descriptors in the classification of Brodatz data set. At left, we show the results for normal FDA. At right, for transformed FDA. From above to below, we use Bayesian, KNN and LDA classifier. Initially, we cannot notice any direct relation among number of descriptors, basis order and correctness rate. The exception occurs with the use of LDA with transformed FDA descriptors. In this case, it is clear that the correctness rate increases with the number of descriptors. In the most of cases, however, it is noticeable that higher order basis yield greater correctness. This is explained by the fact that those basis are capable of capture more details from the conventional VBFD descriptors. Relative to the number of descriptors, the graph shows that each specific combination of FDA descriptors and classifier provides a different pattern for the correctness rate results. This is also waited due to the fact that each classifier has a particular way of dealing with correlation and irregularity information.

Figure 11: Correctness rate for Brodatz data set. At left, using normal FDA coefficients. At right, using transformed FDA descriptors. From above to below using Bayesian, KNN and LDA classifier.

The Figure 2 shows the correctness rate in Outex data set. The graphs are organized in the same way as in the Figure 11. The observations from Brodatz data set are also valid in this case. Particularly, an interesting observation is that the aspect of the graph of each combination descriptor/classifier is similar in both data sets. The unique significant difference is the global correctness that is smaller in Outex, due to its greater difficulty level when compare to Brodatz data set.

Figure 12: Correctness rate for Outex data set. At left, using normal FDA coefficients. At right, using transformed FDA descriptors. From above to below using Bayesian, KNN and LDA classifier.

Now we show the best results achieved by each combination of descriptors and classifiers and the number of used descriptors. In the Table 1 we show the correctness rate for Brodatz data set. We observe that even using a reduced set of descriptors, FDA achieved results sensibly more precise than classical VB fractal descriptors. This advantage is more notable in KNN and Bayesian classifier. In Bayesian, FDA presented an advantage of 42% while in KNN this advantage was 27%. Another important observation from the table is that in this specific application the use of normal FDA coefficients demonstrated to be the better solution. This is very encouraging since this FDA approach is computationally simple and allows an easy statistical interpretation of the analyzed data.

Descriptors Classifier Number of descriptors Correctness rate (%)
Original Bayesian 86 40.0 0.2
Normal FDA Bayesian 80 56.8 0.1
Transformed FDA Bayesian 10 43.2 0.1
Original KNN 86 54.8 0.1
Normal FDA KNN 75 69.8 0.2
Transformed FDA KNN 20 56.8 0.1
Original LDA 86 98.6 0.1
Normal FDA LDA 70 99.0 0.1
Transformed FDA LDA 50 98.2 0.0
Table 1: Correctness rate for the use of fractal descriptors enhanced by FDA normal and transformed coefficients in Brodatz dataset.

Concluding, we present the results of FDA descriptors in OuTex dataset. Again, the performance of FDA descriptors is very good. A highlight must be given to the Bayesian result. FDA provided a correctness rate 123% greater than classical fractal descriptors.

Descriptors Classifier Number of descriptors Correctness rate (%)
Original Bayesian 86 23.0 0.1
Normal FDA Bayesian 80 51.4 0.2
Transformed FDA Bayesian 25 24.9 0.1
Original KNN 86 47.3 0.2
Normal FDA KNN 65 57.0 0.1
Transformed FDA KNN 15 48.1 0.1
Original LDA 86 92.0 0.0
Normal FDA LDA 80 92.5 0.1
Transformed FDA LDA 50 91.8 0.2
Table 2: Correctness rate for the use of fractal descriptors enhanced by FDA normal and transformed coefficients in Outex dataset.

From the previous results, we observe that we cannot extract an exact relation between the number of FDA basis (and, consequently, descriptors), the order of used basis and the correctness rate results. However, analyzing without excessive severity, we observe that generally the use of higher order basis increases the classification performance. Nevertheless, it is always important to verify the combinations for each different application.

Analyzing more globally the results, we observe initially that the FDA transform provides a significant increasing in the performance of volumetric Bouligand-Minkowski descriptors, mainly when we used KNN and Bayesian classifier. This fact attests that the FDA transform is capable of extract relevant features from the set of descriptors, allowing for the classifiers to provide a more precise classification result. As discussed in the Section 4

, the good performance of the FDA transform was expected due to the smooth analytical and irregularly spaced nature of Bouligand-Minkowski descriptors. The smaller efficiency in LDA classifier is easily explained by the fact that one step in the LDA method involves a correlation space transform (Principal Component Analysis). So, features extracted by the FDA transform do not necessarily to have the same correlation properties as the original descriptors and this fact prejudices the performance of the whole classification process.

7 Conclusions

This work proposed and analyzed the use of the FDA transform aiming at enhancing the performance of volumetric Bouligand-Minkowski fractal descriptors, applied to texture classification. The transform consists in the use of coefficients from the functional data representation replacing the original descriptors.

Results demonstrated that the FDA transform increased significantly the accuracy of classification process, mainly when using Bayesian and KNN classifiers. Results confirmed what is expected from the theory, once FDA is a powerful statistical tool for the representation of smooth analytical data, like fractal descriptors. The FDA transform extracts important features and patterns from the original descriptors set yielding a better classification performance.

Results suggest strongly that FDA must be considered as an auxiliary tool for other methods shown in the literature for obtaining fractal descriptors or even other techniques in texture analysis that generate a set of values which may be handled as an analytical function.

8 Acknowledgements

Odemir M. Bruno gratefully acknowledges the financial support of CNPq (National Council for Scientific and Technological Development, Brazil) (Grant #308449/2010-0 and #473893/2010-0) and FAPESP (The State of São Paulo Research Foundation) (Grant # 2011/01523-1). João B. Florindo is grateful to CNPq (National Council for Scientific and Technological Development, Brazil) for his doctorate grant.


  • [1] R. Quevedo, F. Mendoza, J. M. Aguilera, J. Chanona, G. Gutierrez-Lopez, Determination of Senescent Spotting in Banana (Musa cavendish) Using Fractal Texture Fourier Image, Journal of Food Engineering 84 (4) (2008) 509–515.
  • [2] H. Wang, J. Siopongco, L. J. Wade, A. Yamauchi, Fractal analysis on root systems of rice plants in response to drought stress, ENVIRONMENTAL AND EXPERIMENTAL BOTANY 65 (2-3) (2009) 338–344.
  • [3] R. Quevedo, M. Jaramillo, O. Diaz, F. Pedreschi, J. Miguel Aguilera, Quantification of enzymatic browning in apple slices applying the fractal texture Fourier image, JOURNAL OF FOOD ENGINEERING 95 (2) (2009) 285–290.
  • [4] L. Tian-Gang, S. Wang, N. Zhao, Fractal Research of Pathological Tissue Images, Computerized Medical Imaging and Graphics 31 (8) (2007) 665–671.
  • [5] S. Lorthois, F. Cassot, Fractal analysis of vascular networks: Insights from morphogenesis, JOURNAL OF THEORETICAL BIOLOGY 262 (4) (2010) 614–633.
  • [6] R. Lopes, M. Steinling, W. Szurhaj, S. Maouche, P. Dubois, N. Betrouni, Fractal features for localization of temporal lobe epileptic foci using SPECT imaging, COMPUTERS IN BIOLOGY AND MEDICINE 40 (5) (2010) 469–477.
  • [7] H. Millan, M. Gonzalez-Posada, Modelling Soil Water Retention Scaling. Comparison of a Classical Fractal Model with a Piecewise Approach, GEODERMA 125 (1-2) (2005) 25–38.
  • [8] N. Bird, M. Diaz, A. Saa, A. Tarquis, Fractal and multifractal analysis of pore-scale images of soil, JOURNAL OF HYDROLOGY 322 (1-4) (2006) 211–219, International Conference on Fractals in the Hydrosciences (Hydrofractals 03), Ascona, SWITZERLAND, AUG, 2003.
  • [9] G. Di Martino, A. Iodice, D. Riccio, G. Ruello, A novel approach for disaster monitoring: Fractal models and tools, IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING 45 (6, Part 1) (2007) 1559–1570.
  • [10] E. I. Scarlat, M. Mihailescu, A. Sobetkii, Spatial frequency and fractal complexity in single-to-triple beam holograms, Journal of Optoelectronics and Advanced Materials 12 (1) (2010) 105–109.
  • [11] D. Han, M. Wang, J. Zhou, Fractal analysis of self-mixing speckle signal in velocity sensing, OPTICS EXPRESS 16 (5) (2008) 3204–3211.
  • [12] M. Hiltunen, L. Dal Negro, N.-N. Feng, L. C. Kimerling, J. Michel, Modeling of aperiodic fractal waveguide structures for multifrequency light transport, JOURNAL OF LIGHTWAVE TECHNOLOGY 25 (7) (2007) 1841–1847.
  • [13] D. Chappard, I. Degasne, G. Hure, E. Legrand, M. Audran, M. Basle, Image analysis measurements of roughness by texture and fractal analysis correlate with contact profilometry, Biomaterials 24 (8) (2003) 1399–1407.
  • [14] R. P. Wool, Twinkling Fractal Theory of the Glass Transition, JOURNAL OF POLYMER SCIENCE PART B-POLYMER PHYSICS 46 (24) (2008) 2765–2778, Annual Meeting of the American-Physical-Society, New Orleans, LA, MAR 10, 2008.
  • [15] I. Das, N. R. Agrawal, S. K. Gupta, S. K. Gupta, R. P. Rastogi, Fractal Growth Kinetics and Electric Potential Oscillations during Electropolymerization of Pyrrole, JOURNAL OF PHYSICAL CHEMISTRY A 113 (18) (2009) 5296–5301.
  • [16] W. Y. Chen, S. J. Chang, M. H. Weng, C. Y. Hung, Design of the fractal-based dual-mode bandpass filter on ultra thin liquid-crystal-polymer substrate, Journal of electromagnetic waves and applications 24 (2-3) (2010) 391–399.
  • [17] M. Salmasi, M. Modarres-Hashemi, Design and analysis of fractal detector for high resolution radars, CHAOS SOLITONS & FRACTALS 40 (5) (2009) 2133–2145.
  • [18] K. Vinoy, J. Abraham, V. Varadan, On the relationship between fractal dimension and the performance of multi-resonant dipole antennas using Koch curves, IEEE TRANSACTIONS ON ANTENNAS AND PROPAGATION 51 (9) (2003) 2296–2303.
  • [19] B. B. Mandelbrot, The Fractal Geometry of Nature, Freeman, NY, USA, 1975.
  • [20] D. Harte, Multifractals: theory and applications, Chapman and Hall/CRC, 2001.
  • [21] B. Lashermes, S. G. Roux, P. Abry, S. Jaffard, Comprehensive multifractal analysis of turbulent velocity using the wavelet leaders, EUROPEAN PHYSICAL JOURNAL B 61 (2) (2008) 201–215.
  • [22] S. Lovejoy, P. Garrido, D. Schertzer, Multifractal absolute galactic luminosity distributions and the multifractal Hubble 3/2 law, PHYSICA A-STATISTICAL MECHANICS AND ITS APPLICATIONS 287 (1-2) (2000) 49–82.
  • [23] E. T. M. Manoel, L. da Fontoura Costa, J. Streicher, G. B. Müller, Multiscale fractal characterization of three-dimensional gene expression data, in: SIBGRAPI, IEEE Computer Society, 2002, pp. 269–274.
  • [24] L. da F. Costa, R. M. Cesar, Jr., Shape Analysis and Classification: Theory and Practice, CRC Press, 2000.
  • [25] O. M. Bruno, R. de Oliveira Plotze, M. Falvo, M. de Castro, Fractal dimension applied to plant identification, Information Sciences 178 (12) (2008) 2722–2733.
  • [26]

    A. R. Backes, D. Casanova, O. M. Bruno, Plant leaf identification based on volumetric fractal dimension, International Journal of Pattern Recognition and Artificial Intelligence (IJPRAI) 23 (6) (2009) 1145–1160.

  • [27] R. O. Plotze, J. G. Padua, M. Falvo, M. L. C. Vieira, G. C. X. Oliveira, O. M. Bruno, Leaf shape analysis by the multiscale minkowski fractal dimension, a new morphometric method: a study in passiflora l. (passifloraceae), Canadian Journal of Botany-Revue Canadienne de Botanique 83 (3) (2005) 287–301.
  • [28] J. B. Florindo, M. De Castro, O. M. Bruno, Enhancing Multiscale Fractal Descriptors Using Functional Data Analysis, International Journal of Bifurcation and Chaos 20 (11) (2010) 3443–3460.
  • [29] J. O. Ramsay, B. W. Silverman, Applied Functional Data Analysis: Methods and Case Studies, Springer-Verlag, New York, 2002.
  • [30]

    F. Rossi, N. Delannay, B. Conan-Guez, M. Verleysen, Representation of functional data in neural networks, Neurocomputing 64 (1) (2005) 183–210.

  • [31] P. Brodatz, Textures: A photographic album for artists and designers, Dover Publications, New York, 1966.
  • [32] T. Ojala, T. Maenpaa, M. Pietikainen, J. Viertola, J. Kyllonen, S. Huovinen, Outex - New framework for empirical evaluation of texture analysis algorithms, in: Kasturi, R and Laurendeau, D and Suen, C (Ed.), 16TH INTERNATIONAL CONFERENCE ON PATTERN RECOGNITION, VOL I, PROCEEDINGS, INTERNATIONAL CONFERENCE ON PATTERN RECOGNITION, IEEE COMPUTER SOC, 10662 LOS VAQUEROS CIRCLE, PO BOX 3014, LOS ALAMITOS, CA 90720-1264 USA, 2002, pp. 701–706.
  • [33] R. O. Duda, P. E. Hart, Pattern Classification, John Wiley and Sons, New York, 2000.
  • [34]

    C. M. Bishop, Pattern Recognition and Machine Learning (Information Science and Statistics), 1st Edition, Springer, 2007.

  • [35] K. Fukunaga, Introduction to Statistical Pattern Recognition, 2nd Edition, Academic Press, 1990.
  • [36] K. J. Falconer, The Geometry of Fractal Sets, Cambridge University Press, New York, NY, USA, 1986.
  • [37] A. P. Witkin, Scale space filtering: a new approach to multi-scale descriptions, in: Proceedings…, ICASSP - IEEE International Conference on Acoustics, Speech, and Signal Processing, GRETSI, Saint Martin d’Hères, France, 2003, pp. 79–95.
  • [38] R. Fabbri, L. da Fontoura Costa, J. C. Torelli, O. M. Bruno, 2d euclidean distance transforms: A comparative survey, ACM Computing Surveys 40 (1) (2008) 2:1–2:44.
  • [39] J. O. Ramsay, B. W. Silverman, Functional Data Analysis, Springer-Verlag, New York, 2005.
  • [40] B. Henderson, Exploring between site differences in water quality trends: a functional data analysis approach, Environmetrics 17 (1) (2005) 65–80.