Available online at www.sciencedirect.

com

ScienceDirect
Journal of Electrical Systems and Information Technology 4 (2017) 185–197

Texture classification using Steerable Pyramid based Laws’ Masks

Sonali Dash ∗ , Uma Ranjan Jena
Department of Electronics and Tele-communication Engineering, Veer Surendra Sai University of Technology (VSSUT), Burla 768018, Odisha,
India
Received 11 January 2016; received in revised form 7 September 2016; accepted 1 October 2016
Available online 11 October 2016

Abstract
This paper progress towards a new feature extraction technique by combining the two existing methods named as Laws’ mask and
steerable pyramid for texture classification. Texture parameters are derived and classified for accepted Laws’ mask method. In this
paper texture features are extracted and classified using new approaches, which are carried out by integrating both steerable pyramid
and Laws’ mask (SPLM) methods. The comparison of the methods yields that the Steerable Pyramid based Laws’ Mask (SPLM)
texture feature extraction technique using fifth level of image decomposition level resulted in the best classification accuracy. We
use simple k-NN classifier for classification purpose. Our proposed approaches are tested on Brodatz database. Experimental results
on fused features established the combination of two feature sets always outperform the conventional Laws’ mask method.
© 2016 Electronics Research Institute (ERI). Production and hosting by Elsevier B.V. This is an open access article under the CC
BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).

Keywords: Steerable pyramid; Laws’ mask; Feature extraction; Texture classification

1. Introduction

In the last decade’s texture analysis, which plays an important role in image processing has been studied thoroughly
(Zhang and Tan, 2002). Initially texture classification methods focus on statistical analysis of texture from the grey
level co-occurrence matrix (GLCM) for texture images and the filtering based methods (Haralick et al., 1973; Randen
and Husy, 1999). To describe texture, Local binary pattern (LBP) is also a commonly used method (Ojala et al., 2002).
In the above-mentioned methods, there is no control over orientation and scale. Several multiresolution and multi-
channel transform algorithms have been used for texture classification such as dyadic wavelet transform (Arivazhagan
and Ganesan, 2003; Mallat, 1989a; Mallat, 1989b), wavelet frame transform (Unser, 1995), Gabor filters (Jain and
Farrokhnia, 1991; Manjunath and Ma, 1996) and steerable pyramids (Freeman William and Simoncelli Eero, 1995)
for controlling the scale and orientation.

∗ Corresponding author.
E-mail addresses: [email protected] (S. Dash), [email protected] (U.R. Jena).
Peer review under the responsibility of Electronics Research Institute (ERI).

http://dx.doi.org/10.1016/j.jesit.2016.10.001
2314-7172/© 2016 Electronics Research Institute (ERI). Production and hosting by Elsevier B.V. This is an open access article under the CC
BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
186 S. Dash, U.R. Jena / Journal of Electrical Systems and Information Technology 4 (2017) 185–197

Image information can be represented by an effectual method known as pyramid for analysis. Among all the pyramids
steerable pyramid retains some of the advantages of orthonormal transforms (Freeman William and Simoncelli Eero,
1995). Pyramid filtering indeed is faster than the equivalent filtering done with a fast Fourier transform. The nodes
in each level represent information that includes in both space and spatial frequency, so convenient to use (Adelson
et al., 1984). Recursive multi-scale transforms (e.g., wavelets) are now a standard tool in signal and image processing.
The steerable pyramid is a particular variant of this type of transform, which has found use full in a variety of image
processing applications. In this linear decomposition, an image is subdivided into a collection of sub-bands localized
in scale (they have octave bandwidth) and orientation. The transform is computed recursively using convolution
and decimation operations, and it is “self-inverting”. The advantages of this representation are that the sub-bands
are translation and rotation invariant (Karasaridis and Simoncelli, 1996). The steerable pyramid decomposition is
an invertible representation similar to the two-dimensional discrete wavelet transform, but with interesting shift- and
rotation-invariance properties. It is slightly over complete and responsive to a filter bank implementation via convolution
and down- and up sampling operations (Castleman et al., 1998).
The Laws’ masks awareness conveyed by K.I. Laws’ where he approached towards the filtering of images with
specific masks created from the combination of one-dimensional kernel vector in order to assess the texture properties.
Laws’ inferred some of the significant properties for defining texture, which are uniformity, density, coarseness,
roughness, regularity, linearity, directionality, direction, phase and frequency. Analysis of the frequency component
of the image in spatial domain and capture high texture energy points of the windows (Laws, 1979; Laws, 1980).
Recently some hybrid models also have been developed by combining the traditional texture descriptors to obtain
better classification accuracy (Yadav Arvind et al., 2015; Yadav et al., 2015; Qi et al., 2015; Zhe et al., 2015).
In the medical image analysis Law’s masks has received wide acceptance (Rachidi et al., 2008; Elnemr Heba, 2013).
In mammographic image analysis, Arden obtained a classification accuracy of 93.90% for Laws’ masks method by
using ANN as classifier in 2015 (Sagiterry Arden Setiawan et al., 2015). Pietikainen found that the texture energy
measures performance depends on the general forms of the masks rather than on their specific values (Pietikainen
et al., 1983) in 1983. For appropriate image and mask sizes, for finding local rank order correlations of images Laws’
masks could achieve better result than the basic convolutions achieved by Harwood in 1985 (Harwood et al., 1985).
Different methodologies associated with laws measures have been considered by Ade and Unser in 1983 and 1986
(Ade, 1983; Unser, 1986). These kinds of texture measures have executed finely and comparative studies have been
carried out by Du Buf and Ng in 1990 and 1992 (Du Buf et al., 1990; Ng et al., 1992). In 1995 Harwood, Ojala and
Pietikainen proposed a new method of texture analysis based on abstract, local auto-covariance measures of texture
with classifier as nearest neighbour texture classification using Kullback’s log-likelihood discrimination of sample and
prototype distributions of these measures (Harwood et al., 1995). This method can be viewed as a generalization of
Laws’ approach and the results are compared with Laws’ measures. They did their final experiments with 24 different
Laws’ 5 × 5 masks and achieved classification error rates of 25.9% and 39.2% for 64 × 64 and 32 × 32 samples of
Brodatz database. However, in texture classification analysis a detailed study of research shows that Law’s masks
provided poor classification accuracy. In addition, the study reveals that there has not been any further research for the
improvement of texture classification accuracy by using Laws’ masks method (Vidya et al., 2000; Timo et al., 1996a,b).
In virtue of above, in this paper, we propose a new method of texture feature extraction technique by combining
the Laws’ mask and Steerable pyramid named as Steerable Pyramid based Laws’ Mask method (SPLM). Firstly,
we carry out the texture classification problem with conventional Laws’ mask method, which is used as texture
descriptor. By using Laws’ mask descriptor, texture features like mean square, absolute mean and entropy are extracted.
Then, the textures are classified by using k-NN classifier using these features. We have used k-nearest neighbour
approach as a classifier in this work because the classification method is a simplest technique both in conceptually and
computationally, but still it provides good classification accuracy. The motivation for this classifier is that patterns,
which are close to each other in the feature space, are likely to belong to the same pattern class. The neighbours are
taken from a set of samples for which the correct classification is known. K-nearest neighbour is a supervised learning
algorithm. The k-NN classification is based on a majority vote of k-nearest neighbour classes. k-NN classifier takes
only k-nearest neighbour classes. So that majority vote is then taken to predict the best fit class for a point. That means,
in a k = 5 nearest neighbour classifier, the algorithm will take majority vote of its 5 nearest neighbours. These k-nearest
neighbours are selected by using a predefined distance metric (correlation, Euclidian distance) to compute and select
the nearest training patterns that are neighbour to input sample in the sense of the selected metric. The k-NN classifiers
do not use any model to fit and only based on memory. Afterwards the experiments are executed for the proposed
 S. Dash, U.R. Jena / Journal of Electrical Systems and Information Technology 4 (2017) 185–197 187

Fig. 1. System diagram for a kth derivative steerable pyramid.

method (SPLM) for different level and domain. Here in this proposed method we integrate three different types of
steerable pyramid individually with Laws’ mask descriptor as follows:

(1) The textural images are passed over Steerable pyramid through convolution for level 2 by using decomposition
filters sp1, sp3 and sp5. Subsequently for level 3 and 5 by using decomposition filter sp5. Then output images of
steerable pyramid are distributed over traditional Laws’ mask descriptor for feature extraction.
(2) The textural images are passed through Steerable pyramid in frequency domain for level 2, 3, 5 and 7 for four
sub bands. Then output images of steerable pyramid distributed over traditional Laws’ mask technique for feature
extraction.
(3) The textural images are passed through complex Steerable pyramid in frequency domain for level 2, 3, 5 and 7 for
four sub bands. Then output images of steerable pyramid distributed over traditional Laws’ mask technique for
feature extraction.

The experimental results show that the integrated SPLM method gives better classification accuracy than the tra-
ditional Laws’ mask method. Since steerable pyramid has nice reconstruction properties in addition to translation
invariance and rotation invariance, we select steerable pyramid.
The rest of this paper is arranged as follows. Brief introduction to steerable pyramid and Laws’ mask methods are
in Section 2. Section 3 describes the proposed method. Experiments and results are discussed in Section 4. Section 5
gives some concluding remarks.

2. Review of steerable pyramid and Laws’ masks method

The Steerable pyramid generates multi-scale and multidirectional representation of the image. It is a wavelet-like
representation whose analysis functions are dilated and rotated versions of a single directional wavelet. Steerability
refers to the property that the underlying wavelets can be rotated to any orientation by forming suitable linear com-
binations of a primary set of equiangular directional wavelet components. This provides a powerful mechanism for
adapting the transform to the local characteristics of the image by steering the basis functions in the direction of
maximal response (Mythili, 2011). The system diagram of steerable pyramid for a kth stage is shown in Fig. 1. The
system is divided into two parts, analysis and synthesis. On the left side of the diagram is analysis part. The image
is decomposed into lowpass and highpass subbands, using steerable filters L0 and H0 . The lowpass band continues to
break down into a set of bandpass subbands B0 , . . ., Bk and lower lowpass subband H1 . The lower lowpass subband is
subsampled by a factor of 2 along the x and y directions. Repeating the shaded area provides the recursive structure. On
188 S. Dash, U.R. Jena / Journal of Electrical Systems and Information Technology 4 (2017) 185–197

the other hand, the right side of the diagram is synthesis part. The synthesized image is reconstructed by upsampling the
lower lowpass subband by the factor of 2 and adding up with the set of bandpass subbands and the highpass subbands
(Unser et al., 2011).
The Laws’ masks method of feature extraction was given by Kenneth Ivan Laws in 1980, where the main idea
behind it is the filtering of images with specific masks produced from the combination of one-dimensional Kernel
vector in order to assess the texture properties. The Laws’ masks match the pixel neighbourhood to the set of standard
masks to compute the texture properties of images. In addition, this image processing technique could be further
divided into 5 × 5 dimensions and 3 × 3 dimensions. Laws’ properties, which he called “texture energy measures”, are
extracted from three simple vectors of length 3, L3 (1, 2, 1), E3 (−1, 0, 1), and S3 (−1, 2, −1), which represent
the one-dimensional operations of centre-weighted local averaging, symmetric first differencing (edge detection), and
second differencing (spot detection). If we convolve these vectors with themselves or each other, we obtain five vectors
of length 5:
L5 (1,4,6,4,1) = L3 × L3, S5 (−1,0,2,0,−1) = −E3 × E3 = L3 × S3, R5 (1,−4,6,−4,1) = S3 × S3,
E5 (−1,−2,0,2,1) = L3 × E3 and W5 (−1,2,0,−2,1) = −E3 × S3. Sets of more number of vectors can be
defined by repeating this convolution process. Therefore, there are five types of masks namely Edge (E), Level (L),
Spot (S), Ripple (R) and Wave (W). From these set of masks 25 different masks are produced. These masks are
subsequently convolved with a texture image to accentuate its microstructure giving an image from which the energy
of the microstructure arrays is measured together with other statistics. The level vector determines the average grey
level, edge vector bear a resemblance to gradient operator that responds to the column and row stepped edges in an
image, spot vector represents the spot extraction, ripple vector detects ripple from the image.

3. Proposed Steerable Pyramid based Laws’ Masks (SPLM) method

The key purpose of this work is to propose the Steerable Pyramid based Laws’ Masks (SPLM) method based
feature extraction technique and to verify their effectiveness in classification. The simplest K-NN classifier is applied
for verification of classification. Steerable Pyramid based Laws’ Masks (SPLM) is introduced for further enhancement
of classification accuracy than the traditional Laws’ masks method. The steps carried in present work for texture
classification are as follows

Step 1: It starts with creating the database by splitting the original texture images into subimages.
Step 2: This step is to decompose the subimages into multiscale and multiorientation by using steerable pyramid.

The filter bank decomposition can be implemented in either the spatial or the frequency domain. We construct
Steerable pyramids in three different ways, steerable pyramid through convolution, steerable pyramid in frequency
domain and complex steerable pyramid implemented in frequency domain. For steerable pyramid through convolution
the decomposition filters sp1, sp3 and sp5 are used. Different levels for which these decomposition filters used are
(sp1, level = 2), (sp3, level = 2), (sp5, level = 2), (sp5, level = 3) and (sp5, level = 5). For steerable pyramid in frequency
domain four subbands and level two, three, five and seven are considered. Similarly, for complex steerable pyramid
implemented in frequency domain also have four subbands and level considered are two, three, five and seven.

Step 3: The resulting decomposed images are convolved using Laws’ mask texture descriptor for three energy
measurement filters. The convolved outputs are passed through three energy measurement filters (Elnemr Heba,
2013), which are described as follows:

N Neighbouringpixels
Mean = (1)
N

N abs (Neighbouringpixels)
Absolute mean = (2)
N


N (Neighbouringpixels − mean)2
Standard deviation = (3)
N
 S. Dash, U.R. Jena / Journal of Electrical Systems and Information Technology 4 (2017) 185–197 189

Fig. 2. Block diagram of proposed method.

where N represents the window size.

Step 4: Texture features are normalized using max min normalization because a wide range of numerical values are
present in the texture features. For each normalized feature vector three most important statistics are evaluated. Those
are absolute mean (ABSM), mean square or energy (MS) and entropy represented as textural parameters.
Step 5: In the concluding step, k-NN classifier classifies the texture feature vectors.

Accordingly, the best combination of steerable pyramid and Laws’ mask with k-NN as classifier is achieved.
The procedural steps are shown in block diagram Fig. 2.
Algorithm for the proposed method:

1. Input images are read.

2. Each image is divided into non-overlapped subimages of pixel size 128 × 128.
3. Subimages are decomposed by using three individual steerable pyramids (SP for convolution, SP for frequency
domain and Complex SP for frequency domain) for different domain and level.
4. Three different types of decomposed steerable images are convolved with Laws’ mask descriptor separately.
5. Again the convolved outputs are passed individually through three energy measurement filters and features like
mean square, entropy and absolute mean are calculated.
6. Above features are normalized by using min-max normalization.
7. Finally classification is done by applying the k-NN classifier and classification accuracies are calculated.

4. Experimental results, comparison and discussion

The key motivation of this section is to investigate our proposed method with conventional Laws’ masks method to
derive texture features and classification. Results are compared for both the methods. The performances of conventional
Laws’ masks and our proposed method SPLM for texture features extraction have examined using randomly selecting
textural images from Brodatz database (Brodatz, 1999).
190 S. Dash, U.R. Jena / Journal of Electrical Systems and Information Technology 4 (2017) 185–197

Fig. 3. 24 classes of Brodatz texture.

Table 1
Classification results for conventional Laws’ masks (5 × 5).
Different size of masks No. of features Classification accuracy

Mean Absolute mean Standard deviation

15 number of masks(without wave) 45 93.27 86.22 89.42

25 number of masks 75 95.83 91.35 91.6

In this experiment, we use only Brodatz database. The database shown in Fig. 3 has 24 texture images, each of
size 640 × 640. There are substantial variations within some textures and some of them are very alike to each other
that is why this Brodatz dataset is very challenging. For the dataset the original texture images are subdivided into
non-overlapping samples of size 128 × 128. Therefore, there are 600 sub images derived from selected 24 images as
shown in Fig. 3. Out of 25 sub images of each texture, we have taken 12 sub images for training and the remaining 13
sub images for testing. The favourable outcome of classification is verified by using k-NN classifier.
In the first experiment, we use the conventional Laws’ masks method in two ways. Initially only four masks Level (L),
Edge (E), Spot (S) and Ripple (R) are used. Fifteen different masks are produced by using these four masks. Secondly,
we use all the five masks Level (L), Edge (E), Spot (S), Ripple (R) and Wave (W) by which twenty-five different
masks are produced by using these five masks. Both training and testing images are convolved with these fifteen and
twenty five numbers of masks corresponding to our two methods. The convolved outputs are passed through three
energy measurement filters known as Mean, Absolute mean and Standard deviation of window size 15 × 15, which are
normalized by min-max normalization method and then the statistical features Absolute mean, Mean square or energy
and entropy are calculated, which are used as texture feature parameters. For fifteen number of masks 45 no’s and for
twenty five number of masks 75 no’s of features are extracted. The classification results for two sets of Laws’ masks
are presented in Table 1. The performance results are shown in Fig. 4.
 S. Dash, U.R. Jena / Journal of Electrical Systems and Information Technology 4 (2017) 185–197 191

Fig. 4. Performance results for only Law’s mask with two different sets of masks.

Table 2
Performance evaluation of Steerable Pyramid through convolution based Laws’ Mask (SPLM) for 15 numbers of masks.
Feature extraction technique Number of features Classification accuracy

Mean Absolute mean Standard deviation

SPLM (sp1, L = 2) 45 91.03 83.01 88.14

SPLM (sp3, L = 2) 45 91.35 83.65 88.14
SPLM (sp5, L = 2) 45 91.03 84.29 89.42
SPLM (sp5, L = 3) 45 95.13 86.86 90.38
SPLM (sp5, L = 5) 45 94.23 86.86 90.71

The bold value represent the classification accuracy achieved from Steerable Pyramid based Laws’ mask method (SPLM) method which are higher
than the traditional Laws’ mask descriptor.

For second experiment, we use steerable pyramid through convolution with decomposition filters sp1, sp3 and
sp5 to generate filtered images combined with Laws’ masks. Different levels for which these decomposition filters
used are (sp1, level = 2), (sp3, level = 2), (sp5, level = 2), (sp5, level = 3) and (sp5, level = 5). These filtered images are
promoted to Laws’ masks for convolution and to extract texture features with an aim of classification. Features are
derived by using both fifteen and twenty five numbers of Laws’ masks. Classification accuracies are verified for both
15 and 25 numbers of different masks. The number of features for fifteen and twenty five Laws’ masks are 45 and 75
respectively. The classification results for 15 numbers of masks are shown in Table 2. From Table 2, we can see that
Steerable Pyramid Laws’ masks (SPLM) with decomposition filter sp5 at level three achieves highest classification
accuracy of 95.13% for mean energy filter. The second highest accuracy of 94.30% is achieved from Steerable Pyramid
Laws’ Masks (SPLM) with decomposition filter sp5 at level five for mean energy filter. The classification results for
25 numbers of masks are shown in Table 3 from which we observed that SPLM with decomposition filter sp5 at level
three has the best classification accuracy of 96.47% for mean energy filter. SPLM with decomposition filter sp5 at
level five gives the second highest classification accuracy of 96.15% for the same mean energy filter. The performance
results are shown in Figs. 5 and 6.
Third experiment is the combination of Steerable Pyramid in frequency domain with Laws’ mask having two
different set masks of fifteen and twenty five numbers. The Steerable pyramid used for investigation is of four levels
and four sub bands. Different levels for which the decomposed images are achieved are level 2, level 3, level 5 and
level 7. The decomposed images are then convolved with Laws’ masks of two different sets having fifteen and twenty
five number of masks for feature extraction and classification. The Classification results of combined sets are shown
192 S. Dash, U.R. Jena / Journal of Electrical Systems and Information Technology 4 (2017) 185–197

Table 3
Performance evaluation of Steerable Pyramid through convolution based Laws’ Mask (SPLM) for 25 numbers of masks.
Feature extraction technique Number of features Classification accuracy

Mean Absolute mean Standard deviation

SPLM (sp1, L = 2) 75 92.95 86.22 89.10

SPLM (sp3, L = 2) 75 93.27 87.81 90.71
SPLM (sp5, L = 2) 75 93.27 87.81 90.71
SPLM (sp5, L = 3) 75 96.47 91.03 91.35
SPLM (sp5, L = 5) 75 96.15 91.99 91.35

The bold value represent the classification accuracy achieved from Steerable Pyramid based Laws’ mask method (SPLM) method which are higher
than the traditional Laws’ mask descriptor.

Fig. 5. Performance results for proposed SPLM method with fifteen numbers of masks.

Table 4
Performance evaluation of Steerable Pyramid in frequency domain based Laws’ Mask (SPLM) for 15 numbers of masks.
Feature extraction technique Number of features Classification accuracy

Mean Absolute mean Standard deviation

SP (4 subbands, L = 2) 45 93.27 86.22 89.42

SP (4 subbands, L = 3) 45 93.91 87.18 89.74
SP (4 subbands, L = 5) 45 95.19 87.18 90.06
SP (4 subbands, L = 7) 45 94.55 81.41 88.14

The bold value represent the classification accuracy achieved from Steerable Pyramid based Laws’ mask method (SPLM) method which are higher
than the traditional Laws’ mask descriptor.

in Tables 4 and 5. From Table 4, we can observe that Steerable Pyramid Laws’ masks (SPLM) at level five achieve
highest classification accuracy of 95.19% for mean energy filter. The second highest accuracy of 94.55% is achieved
from Steerable Pyramid Laws’ masks (SPLM) at level seven for mean energy filter. Whereas for Steerable Pyramid
Laws’ masks (SPLM) having twenty five number of masks provides highest classification accuracy simultaneously at
level five and seven of 96.15%. The second best classification is achieved at level three of 96.00%. The performance
results are shown in Figs. 7 and 8.
 S. Dash, U.R. Jena / Journal of Electrical Systems and Information Technology 4 (2017) 185–197 193

Fig. 6. Performance results for proposed SPLM method with twenty-five numbers of masks.

Table 5
Performance evaluation of Steerable Pyramid in frequency domain based Laws’ Mask (SPLM) for 25 numbers of masks.
Feature extraction technique Number of features Classification accuracy

Mean Absolute mean Standard deviation

SP (4 subbands, L = 2) 75 95.83 91.35 91.67

SP (4 subbands, L = 3) 75 96.00 91.35 91.35
SP (4 subbands, L = 5) 75 96.15 91.67 91.03
SP (4 subbands, L = 7) 75 96.15 88.78 91.67

The bold values represent the classification accuracy achieved from Steerable Pyramid based Laws’ mask method (SPLM) method which are higher
than the traditional Laws’ mask descriptor.

Table 6
Performance evaluation of complex Steerable Pyramid implemented in frequency domain based Laws’ Mask (SPLM) for 15 numbers of masks.
Feature extraction technique Number of features Classification accuracy

Mean Absolute mean Standard deviation

SP (4 subbands, L = 2) 45 93.27 86.22 89.42

SP (4 subbands, L = 3) 45 93.91 85.26 89.42
SP (4 subbands, L = 5) 45 94.55 85.90 89.42
SP (4 subbands, L = 7) 45 95.83 81.09 87.82

The bold value represent the classification accuracy achieved from Steerable Pyramid based Laws’ mask method (SPLM) method which are higher
than the traditional Laws’ mask descriptor.

Fourth experiment is the integration of complex Steerable pyramid implemented in frequency domain with Laws’
mask. The levels and sub bands are same used for third experiment. The process for obtaining the texture features and
classification are exactly same as we did in third experiment. The Classification results for fifteen number of Steerable
Laws’ masks (SPLM) are shown in Table 6 in which it is observed that the highest classification accuracy of 95.83%
is obtained at level seven from mean energy filter. The Classification results for twenty five number of Steerable Laws’
194 S. Dash, U.R. Jena / Journal of Electrical Systems and Information Technology 4 (2017) 185–197

Fig. 7. Performance results for proposed SPLM method with fifteen numbers of masks.

Fig. 8. Performance results for proposed SPLM method with twenty-five numbers of masks.

masks (SPLM) are shown in Table 7 which represents that the highest classification accuracy of 96.79% is obtained at
level five and seven concurrently from mean energy filter. The performance results are shown in Figs. 9 and 10.
Among all the above proposed methods, we observe that the highest classification accuracy of 96.79% is obtained
from complex Steerable Pyramid implemented in frequency domain with Laws’ Masks (SPLM). All the experimental
results show that out of three-energy measurement filter of Steerable Pyramid Laws’ Masks (SPLM) method the mean
 S. Dash, U.R. Jena / Journal of Electrical Systems and Information Technology 4 (2017) 185–197 195

Table 7
Performance evaluation of complex Steerable pyramid implemented in frequency domain based Laws’ mask (SPLM) for 25 numbers of masks.
Feature extraction technique Number of features Classification accuracy

Mean Absolute mean Standard deviation

SP (4 subbands, L = 2) 75 95.83 91.35 91.67

SP (4 subbands, L = 3) 75 96.47 91.03 91.99
SP (4 subbands, L = 5) 75 96.79 92.31 92.31
SP (4 subbands, L = 7) 75 96.79 86.22 91.35

The bold values represent the classification accuracy achieved from Steerable Pyramid based Laws’ mask method (SPLM) method which are higher
than the traditional Laws’ mask descriptor.

Fig. 9. Performance results for proposed SPLM method with fifteen numbers of masks.

energy filter always gives the best classification accuracy. Furthermore we compiled some other hybrid models used for
texture classification for different databases proposed by other researchers. Fernando, William, Schwartz implemented
a new hybrid model of pyramid decomposition with GLCM and Gaussian smoothing for the GLCM for combined
feature descriptors. For pyramid decomposition with GLCM highest accuracy of 68.27% for Brodatz database at
scale 2 for local and normalization+ global normalization combination with nearest neighbour as classifier. When the
Gaussian smoothing is considered for the GLCM, highest classification accuracy achieved was of 87.19% at scale 4 for
corresponding approach with nearest neighbour as classifier (Siqueira et al., 2013). Combination of Gabor filter and
GLCM for feature extraction with MLP-BP-NN as classifier produced recognition accuracy of 90.33% and 90.20%
(Yusof et al., 2010). Yadav, Anand, Dewal and Gupta proposed a new hybrid model by combining DWT and LBP with
LBP extensions. The highest classification accuracy of 97.40% ± 1.06 was obtained for DWTCLBPu2 on full feature
for hard wood species by using linear SVM as classifier. For reduced features by PCA they achieved accuracy of
97.87 ± 0.82% for DWTCLBPu2 (Yadav Arvind et al., 2015). In another paper of same authors where they integrated
Gaussian Pyramid with LBP, LCP, LPQ, and achieved highest classification accuracy of 92.30% at level 1, 2, 3 for
GPLBP, highest accuracy of 96.10% for GPLPQ at level 2 and 98.00% for GPLCPriu2 at level 3 by using UIUC texture
database (Yadav et al., 2015).
196 S. Dash, U.R. Jena / Journal of Electrical Systems and Information Technology 4 (2017) 185–197

Fig. 10. Performance results for proposed SPLM method with twenty-five numbers of masks.

5. Conclusions

The Steerable image pyramid based generation of texture feature extraction with integration of existing Laws’
masks method is proposed here. The multi resolution capability of steerable pyramid is employed to extract features
at different levels and different orientations. For evaluation of efficiency of the proposed method, k-NN classifier is
used for classification. The best classification accuracy of 96.79% for brodatz database has been accomplished here,
which shows a better classification than individual Laws’ mask method. For future work, different databases can be
implemented on the proposed method to verify the further improvement.

References

Zhang, J., Tan, T., 2002. Brief review of invariant texture analysis methods. Pattern Recognit. 35 (3), 735–747.
Haralick, R.M., Shanmugam, K., Dinstein, I., 1973. Textural features for image classification. IEEE Trans. Syst. Man Cybern. 3 (6), 610–621.
Randen, T., Husy, J.H., 1999. Filtering for texture classification: a comparative study. IEEE Trans. Pattern Anal. Mach. Intell. 21 (4), 291–310.
Ojala, T., Pietikainen, M., Maenpaa, T.T., 2002. Multiresolution gray-scale and rotation invariant texture classification with Local Binary Pattern.
IEEE Trans. Pattern Anal. Mach. Intell. 24 (7), 971–987.
Arivazhagan, S., Ganesan, L., 2003. Texture classification using wavelet transform. Pattern Recognit. Lett. 24 (9–10), 1513–1521.
Mallat, S.G., 1989a. A theory for multiresolution signal decomposition: the wavelet representation. IEEE Trans. Pattern Anal. Mach. Intell. 11 (7),
674–693.
Mallat, S.G., 1989b. Multifrequency channel decomposition of images and wavelet models. IEEE Trans. Acoustic Speech Signal Process. 37 (120),
2091–2110.
Unser, M., 1995. Texture classification and segmentation using wavelet frames. IEEE Trans. Image Process. 4 (11), 1549–1560.
Jain, A.K., Farrokhnia, F., 1991. Unsupervised segmentation using Gabor filters. Pattern Recognit. 24 (12), 1167–1186.
Manjunath, B.S., Ma, W.Y., 1996. Texture features for browsing and retrieval of large image data. IEEE Trans. Pattern Anal. Mach. Intell. 18 (8),
837–849.
Freeman William, T., Simoncelli Eero, P., 1995, October. The Steerable Pyramid: a flexible architecture for multi-scale derivative computation. In:
Presented at 2nd Annual IEEE International Conference on Image Processing, Washington, DC.
Adelson, E.H., Anderson, C.H., Bergen, J.R., Burt, P.J., Ogden, J.M., 1984. Pyramid methods in image processing. RCA Eng. November/December,
29–36.
Karasaridis, A., Simoncelli, E.P., 1996, May. A Filter Design Technique for Steerable Pyramid Image Transforms. In: Proc. ICASSP-96, Atlanta.
Castleman, K.R., Schulze, M., Wu, Q., 1998. Simplified Design of Steerable Pyramid Filters. Proc. IEEE Int. Sympos. Circ. Syst. 5, 329–332.
 S. Dash, U.R. Jena / Journal of Electrical Systems and Information Technology 4 (2017) 185–197 197

Laws, K.I., 1979. Texture energy measures. Proc. Image Underst. Workshop, 47–51.
Laws, K.I., 1980. Textured image segmentation. In: Image Processing Institute, Univ. of Southern California, Report 940.
Yadav Arvind, R., Anand, R.S., Dewal, M.I., Gupta Sangeeta, 2015. Multiresolution local binary pattern variants based texture feature extraction
technique for efficient classification of microscopic images of hard wood species. Appl. Soft Comput. 32, 101–112.
Yadav, A.R., Anand, R.S., Dewal, M.I., Gupta, S., 2015. Gaussian image pyramid based texture features for classification of microscopic images of
hardwood species. Optik 126 (24), 5570–5578.
Qi, X., Shen, L., Zhao, G., Li, Q., Pietikainen, M., 2015. Globally rotation invariant multi-scale co-occurrence local binary pattern. Image Vis.
Comput. 43, 16–26.
Zhe, T., Sua, Y., Erb, M.J., Qia, F., Li, Z., Zhouc, J., 2015. A local binary pattern based texture descriptors for classification of tea leaves.
NeuroComputing 168, 1011–1023.
Rachidi, M., Marchadier, A., Gadois, C., Lespessailles, E., Chappard, C., Benhamou, C.L., 2008. Laws’ masks descriptors applied to bone texture
analysis: an innovative and discriminant tool in osteoporosis. Skeletal Radiol. 37 (June (6)), 541–548.
Elnemr Heba, A., 2013. Statistical analysis of law’s masks texture features for cancer and water lung detection. IJCSI Int. J. Comput. Sci. Issues 10
(November (6, No. 2)).
Sagiterry Arden Setiawan, E., Julian, W., Yudy, P., 2015. Mammogram classification using law’s energy measure and neural network. Proc. Comput.
Sci. (59), 92–97.
Pietikainen, M., Rosenfled, A., Davis, L.S., 1983. Experiments with texture classification using averages of local pattern matches. IEEE Trans. Syst.
Man Cybern. SMC-13 (3), 42–426.
Harwood, D., Subbarao, M., Davis, L.S., 1985. Texture classification by local rank correlation. Graph. Image Process. 32, 404–411.
Ade, F., 1983. Characterization of textures by eigenfilter. Signal Process. 5, 451–457.
Unser, M., 1986. Local linear transforms for texture measurements. Signal Process. 11, 61–79.
Du Buf, H.M.J., Kardan, J.M.H., Spann, M.M., 1990. Texture feature performance for image segmentation. Pattern Recognit. 23 (3/4), 291–309.
Ng, I., Tan, T., Kittler, J., 1992. On local linear transform and Gabor filter representation of texture. In: Proc. 11th International Conference Pattern
Recognition, III, The Hague, The Netherlands, pp. 627–631.
Harwood, D., Timo, O., Pietikäinen, M., Shalom, K., Larry, D., 1995. classification by centre-symmetric auto-correlation using Kullback discrimi-
nation of distributions. Pattern Recognit. Lett. 16 (January (1)), 1–10.
Timo, O., Matti, P., David, H., 1996a. A comparative study of texture measures with classification based on feature distributions. Pattern Recognit.
29 (l), 51–59.
Vidya, M., Ramón, V., Praveen, K., 2000. Texture Classification Using Logical Operators. IEEE Trans. Image Process. 9 (October (10)).
Timo, O., Matti, P., Jarkko, N., 1996b. Determining composition of grain mixtures by texture classification based on feature distributions. Pattern
Recognit. Artif. Intell. 10 (73).
Mythili, S., 2011. Fabric defect detection using Steerable pyramid. ICTACT J. Image Video Process. 01 (04 (May)).
Unser, M., Chenouard, N., Ville, D.V.D., 2011. Steerable pyramids and tight wavelet frames in L2 (Rd ). IEEE Trans. Image Process. 20 (October
(10)).
Brodatz, P., 1999. Textures: “A Photographic Album for Artists and Designers”. Dover & Greer Publishing Company.
Siqueira, F.R., Schwartz, D., Pedrini, W.R.H., 2013. Multi-scale gray level co-occurrence matrices for texture description. Neurocomputing (120),
336–345.
Yusof, R., Rosli, N.R., Khalid, M., 2010. Using gabor filters as image multiplier for tropical wood species recognition system. In: 12th IEEE
International Conference on Modelling and Simulation (UKSim), pp. 289–294.