2015 IEEE International Conference on Computational Intelligence & Communication Technology

ICA in Image Processing: A Survey

Swati Goel Akhilesh Verma Savita Goel Komal Juneja

M.Tech Scholar Assistant Professor Senior System M.Tech Scholar
Department of CSE Department of CSE Programmer Department of CSE
AKGEC, Ghaziabad AKGEC, Ghaziabad CSC, IIT Delhi AKGEC, Ghaziabad
[email protected] [email protected] [email protected] [email protected]

Abstract-Source separation is a problem in which signals are Here we are going to introduce several variants of ICA for
mixed together. It is becoming a tedious task to recuperate extracting interesting batch of sources that are independent
original components signal from the signal mixture. Blind Source from its linear mixture.
Separation (BSS) is suggested as a key to the problem aiming at Aapo Hyvarien et al. [2] proposed FastICA algorithm.It is
finding the linear representation in such a way that the
highly efficient computable method performing an ICA
components are statistically (stochastically) independent.
Independent Component Analysis (ICA) is an approach that estimation for one unit or multi-unit output. ICA using filter
attained a wider attention and a growing significance in a diverse bank is an approach proposed by R. Jenessen et al. [3] used for
range of research fields for accomplishing Blind Source texture segmentation. The basis filters calculated using this
Separation. This paper includes preface of ICA, its variants and approach capture the textures inherent structure and hence it
their list of applications in brief. becomes possible to create features used for fruitful
segmenting multitextured image [3]. Kernel Independent
Keywords- ICA; Image model; Filter bank; Global features; Component Analysis (KICA) [4] assimilate kernel trick in ICA
Texture segmentation. that upgrade the rate of performance of ICA for face
I. INTRODUCTION recognition [4]. ICA with Reference (ICA-R) [5] is an
approach for extracting interesting independent sources from
Signals are the function of independent variables such as their mixture, given a pre information of the original sources
distance, time, position and temperature that carry some available as templates [5]. Regardless of wide applications,
meaningful data. Depending on independent variables nature ICA often confronts two exigent problems : one is smaller size
and the function value for defining the signal, signals are of sample and the other is the selection of the basis functions.
classified as 1-D signal and other is the multidimensional M-D Jian Cheng et al. [6] gave an enhanced ICA algorithm by
signal. Cocktail Party Problem detailed in [1], [8] embellishes ensemble learning approach, Random Independent Subspace
the concept very well. In such a situation ICA works as a tool (RIS) [6] that aims at solving the two major problems of ICA.
for separating the original components from its mixture. The Weibau Zou et al. [7] suggested combination of wavelet and
concept of ICA is revealed through Fig.1. Independent Component Analysis for image classification.
This approach aimed at finding components that are Xiaoyan Ziang et al. [8] suggested to combine ICA and
independent at maximum level and non-gaussian [2]. There are watershed algorithm for segmenting medical images efficiently
several classical methods that are proposed for classifying reducing the noise effect and keeping the images with clear
multivariate data such as Statistical methods often based on co- texture simultaneously. Sindhumol S. et al. [9] introduced an
occurrence matrix [34]-[37], model based method [38] algorithm Spectral Clustering Independent Component
including Gaussian Markov random fields [39], Gibbs random Analysis (SC-ICA) [9] used for brain tissue classification.
fields [40], Wold models [41] but the assumption of non- Originally ICA was organized to handle the problems
gaussianity in ICA enables it to identify original, underlyning related to cocktail party problem. The increasing interest in
components making it different from other methods. ICA made the model useful in diverse areas like texture
The input data taken as a single matrix of data and no segmentation [3], [4], Face recognition [6], Network Intrusion
assumption of desired output make ICA an unsupervised Detection [10], Digital Image Forgery Detection [11],
method. Seperation of mixed images [12], Image Encryption [13],
Medical Imaging [14], [15], [16], [17], Classificaion of
globular clusters [18], 3D statistical shape analysis [19], Blind
identification of damage in time varying syatems [20].
II. ICA BASIC THEORY
Independent Component Analysis (ICA) [31, 49] is an
approach of finding the linear transform of the mixtures in
such a way that the recovered signals are as independent as
possible. Consider a dataset of n training samples x1, x2......xn
as X={x1.....xt......xn}, where n is the quantity of samples and
xi is an n dimensional vector. The signals observed xi can be
Fig 1: Example of ICA [27]. moulded as a combination of latent variables sj, j=1....m with
some unknown coefficients aij presented as

978-1-4799-6023-1/15 $31.00 © 2015 IEEE 144

DOI 10.1109/CICT.2015.91
Authorized licensed use limited to: ULAKBIM UASL - ISTANBUL KULTUR UNIVERSITESI. Downloaded on April 19,2025 at 12:00:08 UTC from IEEE Xplore. Restrictions apply.
 m
It is used over an artificial mixture of speech recordings
xi = ¦ aij si , for all i=1....n (1)
j =1
where the sources are already known.The input applied to
method is a data matrix where each column belongs to
individual signal. The data is preprocessed using centering
Here si= group of independent components and aij= mixing
process and whitening process. Centering is done basically to
coefficients. The values of xi and si are composed into a vector
simplify the FastICA algorithm by making each component
x and s respectively of n dimension and aij are composed into a
mean zero. Whitening is a procedure to make the components
mixing matrix A. The model can be expressed as
mixture uncorrelated giving their variance unit value. It is
done by applying a whitening matrix to the centered data. The
x = As (2) whitening matrix given as:

where A = mixing matrix , s(t) = source signals. W = D− 1/ 2

ET (4)
Considering that the source signals are mutually
statistically (stochastically) independent [2], unsupervised where D= diagonal matrix of eigen values, D-1/2 =diag(d1-
1/2
process is hence used to get the solution by finding the .......dn-1/2) and ET= transpose of the orthogonal matrix of
demixing matrix (W) , such that eigenvectors.
FastICA has many advantages such as it is
y = Wx (3) fast,computationally simple, parallel, distributed and requires
slight memory space. The methods performance can be
is an estimate of s. The y components should be as independent improved by choosing a suitable nonlinearity.
as possible. Equation (1) expresses the basic model of ICA B. ICA using filter bank
illustrated through Fig.2.
R. Jennesen et al. [3] presented a method of generating a
III. VARIANTS OF ICA new filter bank that are data dependent useful for texture
segmentation. These filters are capable of capturing the
A. FastICA
texture inherent structure enabling to create features for
Hyvarinen et al. [2] proposed that their are number of segmenting composite input textured images taken from
different measures of non gaussianity that are deployed as an Brodatz album [28].
ICA objective function including kurtosis and negentropy. ICA has been introduced as a generic model for images as
FastICA [2] is a highly efficient computable method that shown in Fig.3. An image x can be formulated as
performs an ICA. Fast ICA is named as “Fast” ICA as it has
high capability of fast convergence which means faster N
estimation and refinement of the independent components. It x = ¦ si ai = As (5)
i =1
can efficiently perform ICA for single unit or multi unit output,
overcoming the limitations of using higher order cumulants.
The algorithm uses a fixed point iterative scheme [21], [29] to where ai; i=1.....N are referred to as ICA image basis filters
obtain a subspace of ICs. that are represented as columns of mixing matrix A, si;
i=1.....N are their respective weighting components that are
statistically (stocastically) independent. The si’s are the
independent components.
The experimental setup discussed in [3] is presented as
Fig.4.

Fig 2: ICA Basic Model [8].

Fig.3. Image modelled as weighted summation of basis images [3].

Fig 4. Experimental setup [3].

145

Authorized licensed use limited to: ULAKBIM UASL - ISTANBUL KULTUR UNIVERSITESI. Downloaded on April 19,2025 at 12:00:08 UTC from IEEE Xplore. Restrictions apply.
 The ICA filter bank performance depands on the number degrade the quality of the signals. Thus ICA-R removes the
of filter banks used that outputs the misclassified rate of the drawback of basic ICA technique mentioned above.
pixels at the true boundary of the adjacent regions.
E. Enhanced ICA algorithm
C. Kernel Independent Component Analysis (KICA) Independent Component Analysis, applicable to high
Jian Cheng et al. [4] proposed a innovative method named dimensional tasks such as face recognition affected from two
as Kernel Independent Component Analysis (KICA) meant problems: one is smaller size of sample and other is the
for texture features extraction.The algorithm incorporates ICA selection of basis functions.
and kernel trick to improve face recognition. Kernel trick [34] Jian Cheng et al. [6] introduced an enhanced ICA
is a pivotal trick of machine learning. The objective is to carry algorithm by ensemble learning approach called random
out a non linear mapping by projecting the data into a higher independent subspace (RIS) [6].The smaller size of sample
dimensional implicit feature space F using SVM (Support means that the samples used for training are small in number
Vector Machine) as resulting in biased classifiers having a variance with higher
value and unstable performance. Ensemble learning
φ : x ∈ RN →φ(x) ∈ F (6) techniques such as Random Subspace method [6] gives a way
of combining multiple classifiers into ensembled (group)
classifier resulting in improved performance than its base
Then ICA is applied to generate a group of nonlinear features
classifier overcoming the two limitations of ICA algorithm.
of input data. There are a various kernel functions used in
Let X=[x1, x2.....xn] be a training set matrix consisting of n
kernel approach. It includes polynomial kernel
training samples. Each column is an N-dimensional feature.
From this N-dimensional feature space; K r-dimensional
k ( x, y) = ( x. y)d (7) feature subspaces are resampled. This sampling is repeated
with replacement producing training set Xk= [x1k,
Gaussian kernel : x2k...,xnk].Then one ICA classifier Ck can be constructed from
each subspace and gives a prediction Ck(y). Aggregation of all
|| x − y ||2 predictions using decision rule produces a final prediction
k(x, y) = exp(− ) (8)
2σ 2
C(y) = aggre{C1(y),C2(y).......,Ck (y)} (11)
Sigmoid kernel :
There are basically two ways of resampling. One is the
k ( x, y) = tanh(k ( x< y) + θ ) (9) resampling in original (unmodified) feature space and second
is the resampling in the feature space that has been whitened
before. Based on these two ways, the three schemes for RIS
For improving the performance,cosine kernel function is are developed.
introduced given as:
F. ICA with wavelet
~ k(x, y) Day by day increasing size of image database leads to the
k(x, y) = (10)
k(x, x)k( y, y) image classification problem. Weibao Zu et al. [7] proposed a
method based on combination of wavelet and Independent
Component Analysis for classification of images. Using
where k is a polynomial kernel. wavelet, an image can be disintegrated into low frequency
KICA tends to be used for extracting features of texture band and high frequency band reflecting the image features
pixels used for classifying these pixels. KICA outperforms like edges of an object.
ICA and Gabor and proved to be most prominent method to Characterization of images is effective using histograms of
represent texture feature. low frequency bands .Because different images can have
D. ICA with Reference (ICA-R) similar histograms of high frequency bands and hence they
Wei Lu et al. [5] presented ICA-R with a motivation of cannot be used as features. ICA is used for extracting features
seperating components that are independent and selecting from high frequency band thereby improving image
desired sources in a single stage when the rough templates of classification. Extraction of features using wavelet includes
the preferred ICs are available from a given information following steps:
referring to as reference signals. The components extracted 1. Image decomposition using wavelet.
using classical ICA algorithms are same as the amount of the 2. Structural representation of image.
mixtures hence, additional pre processing is required to 3. Description of histogram of wavelet coefficients.
choose for the preferred sources from the set of recovered Combining wavelet and ICA leads to an improvement in
sources. This two stage process in high dimensional the accuracy in classification as well as it become more robust
applications may involve redundant computation, large and accurate.
enough memory for estimating unnecessary signals and may G. Watershed algorithm using ICA

146

Authorized licensed use limited to: ULAKBIM UASL - ISTANBUL KULTUR UNIVERSITESI. Downloaded on April 19,2025 at 12:00:08 UTC from IEEE Xplore. Restrictions apply.
 Segmenting medical image is a complicated task and there Computational reliability is itself a major problem that the
is a risk of misclassifying due to presence of noise, uniform computationally intensive and complex estimation methods
gray intensity, and many low contrast areas adjacently causing face. The problem arises when the records are perfect for
images to be over segmented automatically using watershed statistical inference but the algorithm may get trapped in local
algorithm alone. minima and fails to produce meaningful results. There exist
Xiaoyan Zhang et al. [8] attempted to introduce ICA as several ICA algorithms based on local optimization methods
pre-process step of watershed algorithm for segmenting at that start from a arbitrary initial point and attempt to increase
clear boundary and also reduce the noise affection during the objective function at each iteration. Hence it shall be
image segmentation by eliminating wavebands redundancy possible that ICA may not provide the real (global) optimum
and extracting the image signal components from source of the objective function [44]. This becomes an additional
signal as shown in Fig. 5. source of randomness and produce results with error. In order
to access the reliability of the data, we can perform the
H. Spectral Clustering Independent Component Analysis (SC-
randomization of the data [45] or can execute the algorithm on
ICA)
different initial points.
S. Sindumol et al. [9] proposed an ICA variant SC-ICA
that provides an identical priority to global and local features V. ICA APPLICATIONS
improving the classification of tissue of brain from MRI. It is
A. Texture Segmentation
a two stage process. In the first stage, Spectral distance based
clustering is used to split multispectral MRI into clusters and ICA filter bank serves as a basis of capturing intrinsic
in second stage ICA in conjunction with SVM is used over the properties of texture images used for feature extraction of
clustered data for brain tissue analysis to obtain least textured images and segment multi textured image [3].
frequently occurred objects such as small lesions from large B. Network Intrusion Detection
multi spectral data.
The Illegal activities on computer system can be detected
ICA as a global issue is good at extracting repeatedly
using Intrusion Detection System. It includes feature
occurring features but fails at preserving local features. SC-
extraction and feature selection. ICA serves as a tool for
ICA is an expansion to ICA aimed at improving the concept
extracting features deriving best features with online
of extracting features from multispectral MRI. SVM using
computation with less computation power, high accuracy and
RBF kernel [26] is used as a classifier for analysing brain
fast real time delivery [10].
tissues.
C. Forgery Detection
IV. TESTING INDEPENDENT COMPONENTS
Forgery of digital images is the task of eluding the viewer
After estimating independent components using any by inserting some object of other image or pasting part of
variant of ICA, it is required to test the reliability or statistical same image to cover the useful information. Watermarking
significance of the components [43]. The basic ethics of serves as a traditional technique to detect such fraud with
scientific publication states that the components estimated image. The step is processed only at the moment of capturing
using various scientific data should be validated. Inputting the image. ICA eliminates this issue by its capability of
pure noise to the ICA algorithm and then observing that separating the statistically (stochastically) independent
whether ICA is able to estimate original components is a sources [11].
calculation of reliability itself.
D. Separation of Mixed Images
ICA proposed a FastICA algorithm with modified contrast
function that can be used to obtain high quality images with
high PSNR (Peak Signal to Noise Ratio), SNR (Signal to Noise
Ratio), ISNR (Improved Signal to Noise Ratio) and low RMSE
(Root Mean Square Error) value [12].
E. Image Encryption
Transmitting useful information over internet is a high risk
task. In order to prevent the data to be hacked, at the sender
side the target images are covered using an insignificant image
to hide them and send over the channel. The mixing signals
generated are close to Gaussian signals. At the receiver side,
FastICA is applied over the mixed signals to obtain original
signals without any familiarity of mixing matrix or original
images [13].
F. X-Ray coronary digital subtraction angiography
Fig 5: Framework for denoising image [8].

147

Authorized licensed use limited to: ULAKBIM UASL - ISTANBUL KULTUR UNIVERSITESI. Downloaded on April 19,2025 at 12:00:08 UTC from IEEE Xplore. Restrictions apply.
 Digital subtraction angiography (DSA) is the technique of [6] Jian Cheng, Qingshan Liu, Hanqing Lu and Yen-Wei Chen,
“Ensemble learning for independent component analysis,” Pattern
visualizing and examining blood vessels in the human body. Recognition 39, 2006, pp. 81 – 88,
ICA using fixed point algorithm [21, 28] is proposed as an doi:10.1016/j.patcog.2005.06.018.
effective method of removing tube-like background from the [7] Weibao Zou, Yan Li, King Chuen Lo and Zheru Chi, “Improvement
coronary vessels [14], [21], [30]. of image classification with Wavelet and Independent Component
Analysis (ICA) based on Structured Neural Networks,” IEEE,
G. Automatic separation of breast masses International Conference on Neural Networks, 2006, pp. 3949-3954.
[8] Xiaoyan Zhang, Lichao Chen, Lihu Pan and Lizhi Xiong, “Study on
ICA assumes that the data generating sources should be the image segmentation based on ICA and Watershed algorithm,”
independent. Enhanced ICA mixture model (EICAMM) IEEE, Fifth International Conference on Intelligent Computation
relaxes the limitation by using the idea of image model. X-ray Technology and Automation, 2012, pp. 505-508,
mammography is considered as a technique of analysing and doi:10.1109/ICICTA.2012.132.
[9] Sindhumol S., Anil Kumar and Kannan Balakrishnan, “Spectral
diagnosing breast cancer. EICAMM can be imposed over the clustering independent component analysis for tissue classification
breast images for segmenting masses in dense breast region from brain MRI,” Biomedical Signal Processing and Control 8,
[15]. 2013, pp. 667– 674, doi:10.1016/j.bspc.2013.06.007.
[10] Dayu Yang and Hairong Qi, “A network intrusion detection method
H. Electrocardiogram beat detection enhancement using independent component analysis,” IEEE, 2008.
[11] Sunil Kumar, J.V. Desai, S. Mukherjee and P. K. Das, “Suitability of
Electrocardiogram (ECG) is used to plot the beat rate of independent component analysis in digital image forgery detection,”
heart. When using ECG as a diagnosis tool, the investigator International Journal of Engineering and Technology (IJET), vol. 5,
may face a problem of extracting essential beat in the existence no 1, 2013, pp. 226 – 231.
of artifacts. The original ECG signals in addition with artifacts [12] Minal Barhate, “Seperation of mixed images using independent
component analysis on the basis of blind Source Seperation,”
are transformed to the components using ICA. Joint International Journal of Advanced Electronics & Communication
diagonalization based ICA algorithm called JADE algorithm Systems, vol 1, issue 3, 2013.
can be used to spot beats in presence of artifacts and noises [13] A. Alfalou and A.Mansour, “NEW IMAGE ENCRYPTION
[16], [17]. METHOD BASED ON ICA,” IAPR Conference on Machine Vision
Applications, 2007.
I. 3D Statistical shape analysis [14] Songyuan Tang, YongtianWang , Yen-WeiChen, “Application of
ICA to Xray coronary digital subtraction angiography,”
Statistical shape analysis is a powerful means of providing Neurocomputing 79, 2012, pp.168–172,
shape based features used for diagnosing and classifying doi:10.1016/j.neucom.2011.10.012.
medicine. ICA is used to extract surface features closely [15] Patricia B. Ribeiro , Roseli A.F. Romero ,Patrícia R. Oliveira ,
Homero Schiabel and Luciana B. Verçosa, “Automatic segmentation
resembling the original basis surfaces for the right ventricle of breast masses using enhanced ICA mixture model,”
endocardial surfaces (RVES) capturing more localized Neurocomputing 120, 2013, pp. 61–71,
variations in the shape [19]. doi:10.1016/j.neucom.2012.08.062.
[16] Jakub Kuzilek and Lenka Lhotska, “Electrocardiogram beat detection
VI. CONCLUSIONS enhancement using Independent Component Analysis,” Medical
Engineering & Physics 35, 2013, pp. 704– 711,
ICA serves as a general purpose statistical technique of doi:10.1016/j.medengphy.2012.07.017.
linearly transforming observed random data into components [17] Ricardo Vigário, Jaakko Särelä, Veikko Jousmäki, Matti
having more Gaussian nature than originals. ICA is a global Hämäläinen, and Erkki Oja, “Independent Component Approach to
the Analysis of EEG and MEG Recordings,” IEEE transaction on
technique that works on the image as whole and it is used for Biomedical Engineering, vol. 47, no. 5, 2000.
extracting image features that serve as a basis for segmenting [18] Asis Kumar Chattopadhyay, Saptarshi Mondal and Tanuka
images. It is a highly computational technique and has various Chattopadhyay, “Independent Component Analysis for the objective
classification of globular clusters of the galaxy NGC5128,”
fields of application such as Texture segmentation, Face
Computational Statistics and Data Analysis 57, 2013, pp. 17–32,
recognition, Cyber security, Data mining, Medical imaging etc. doi:10.1016/j.csda.2012.06.008.
Each application field is dependent on specific variant of ICA. [19] Jia Wu, Katharine G. Brigham, Marc A. Simon and John C. Brigham,
The future development of ICA is hence driven by particular “An implementation of independent component analysis for 3D
needs of the application fields and may be specific to each field. statistical shape analysis,” Biomedical Signal Processing and
Control 13, 2013, pp. 345–356.
[20] Yongchao Yang and Satish Nagarajaiah, “Blind identification of
REFERENCES damage in time-varying systems using independent component
[1] Yoel Ainhoren, Shlomo Engelberg, and Shai Friedman, analysis with wavelet transform,” Mechanical Systems and Signal
instrumentationnotes, IEEE Instrumentation & Measurement Processing 47, 2014, pp. 3–20, doi:10.1016/j.bspc.2014.06.003.
Magazine, pp. 45-48, 2008. [21] A. Hyvarinen, “Fast and robust fixed point algorithms for
[2] A. Hyvarinen and E. Oja, “Independent component analysis: independent component analysis,” IEEE Trans. Neural Networks,
algorithms and applications,” Neural Networks 13, pp. 411–430, vol.10, 1999, pp. 626-634.
2000. [22] S. Theodoridis and K. Koutroumbas, Pattern Recognition, Academic
[3] R. Jenssen and T. Eltoft, “Independent component analysis for Press, New York, 1999.
texture segmentation,” Pattern Recognition 36, 2003, pp. 2301 – [23] M.Tuceryan, A.K.Jain, in: C.H.Chen, L.F.Pau, P.S.P.Wang (Eds.),
2315, doi: 10.1016/S0031-3203(03)00131-6. the Handbook of Pattern Recognition and Computer Vision, 2nd
[4] Jian Cheng, Qingshan Liu and Hanqing Lu, “Texture classification Edition, World Scientific, Singapore, 1998, pp. 207-248.
using Kernel Independent Component Analysis,” IEEE, 2004. [24] J. Karhunen and J. Joustsensalo, “Representation and separation of
[5] Wei Lu and Jagath C. Rajapakse, “ICA with Reference,” signals using nonlinear PCA type learning,” Neural Networks vol.7,
Neurocomputing 69, 2006, pp. 2244–2257, 1999, pp. 113-127.
doi:10.1016/j.neucom.2005.06.021.

148

Authorized licensed use limited to: ULAKBIM UASL - ISTANBUL KULTUR UNIVERSITESI. Downloaded on April 19,2025 at 12:00:08 UTC from IEEE Xplore. Restrictions apply.
 [25] Vincent L and Soille P., “Watersheds in digital spaces: an efficient
algorithm based on immersion simulations,” J.IEEE Trans Pattern
Analysis and Machine Intelligence, vol. 13, 1991, pp. 583-598.
[26] C.M. Bishop, Pattern Recognition and Machine Learning, Springer,
2006.
[27] Dominic Langlois, Sylvain Chartier and Dominique Gosselin, “An
Introduction to Independent Component Analysis: Infomax and
FastICA algorithms,” Tutorials in Quantitative Methods for
Psychology, vol. 6 (1), 2010, pp. 31-38.
[28] P. Brodatz, Texture: a photographic album for artists and designers,
Dover, New York, 1996.
[29] E. Bingham and A. Hyvarinen, “A fast fixed-point algorithm for
independent component analysis of complex valued signals,” Int. J.
Neural Syst. vol.10, 2000, pp. 1–8.
[30] A. Frangi, W. Niessen, K. Vincken and M. Viergever, “Multiscale
vessel enhance- ment filtering,” Proceedings of Medical Image
Computing and Computer- Assisted Interventation (MICCAI98),
1998, pp. 130–137.
[31] Comon P., “Independent component analysis, a new concept?”
Signal Processing 36, 1994, pp. 287–314.
[32] Lijuan Duan, Jicai Ma, Jun Miao and Yuanhua Qiao, “A texture
images segmentation method based on ICA filters,” IEEE, fifth
International Conference on Natural Computation, 2009, pp. 484-
487, doi:10.1109/ICNC.2009.400.
[33] Qingshan Liu, Jian Cheng, Hanqing Lu and Songde Ma, “Modelling
Face Appearance with Kernel Independent Component Analysis,” to
appear in FG2004, Seoul, Korea, 2004.
[34] R. M. Haralick, K. Shanmugam and I. Distein, “Textural features for
image classification,” IEEE Trans. Syst. ManCybern. , 3 (6), 1973,
pp. 610-621.
[35] S. Theodoridis and K. Koutroumbas, Pattern Recognition, Academic
Press, New York, 1999.
[36] L. Davis, S. Johns, J. K. Aggrawal, “Texture analysis using
generalized co-occurrence matrices,” IEEE Trans. Pattern Anal.
Mach. Intell, , 1 (3), 1979, pp. 251-259.
[37] H. Tamura, S. Mori and T. Yamawaki, “Textural features
corresponding to visual perceptions,” IEEE Trans. Syst. ManCybern,
8 (6), 1978, pp. 460-473.
[38] J. Jhang and T. Tan, “Brief review of invariant texture analysis
methods,” Pattern Recognition, 35, 2002, pp. 735-747, PII: S0031-
3203(01)00074-7.
[39] F. Cohen, “Classification of rotated and scaled texture images using
gaussian markov random field models,” IEEE Trans. Pattern Anal.
Mach. Intell., 21 (2), 1999, pp. 148-153.
[40] K. Sivakumar, “Morphologically constrained GRF’s: applications to
texture synthesis and analysis,” IEEE Trans. Pattern Anal. Mach.
Intell. , 21 (2), 1999, pp. 148-153.
[41] L. Fang, R. W. Picard, “Periodicity directionality and randomness:
Wold features for image modelling and retrieval,” IEEE Trans. On
Pattern Anal. And Mach. Intell., 18 (7) , 1996, pp. 722-733.
[42] L. Fang and R. W. Picard, “Periodicity, directionality, and
randomness: Wold features for image modeling and retrieval,” IEEE,
Pattern Analysis and Machine Intelligence, vol. 18, 1996, pp. 722-
733, doi: 10.1109/34.506794.
[43] Aapo Hyvarinen, “Independent component analysis: recent
advances,” Phil. Trans. R. Soc. A, 2012.
[44] Meinecke F, Ziehe A, Kawanabe M and Müller K-R.,”A resampling
approach to estimate the stability of one-dimensional or
multidimensional independent components,” IEEE Trans. Biomed.
Eng. 49, 2002, pp. 1514–1525, doi: 10.1109/TBME.2002.805480.
[45] Hyvärinen A., “Testing the ICA mixing matrix based on inter-subject
or inter-session consistency,” NeuroImage, 58, 2011, pp. 122–136,
doi: 10.1016/j.neuroimage.2011.05.086.
[46] Gonzales, R. and Wintz, “Digital Image Processing,” Addison-
Wesley, 1987.
[47] James V. Stone, “Independent component analysis: an introduction,”
TRENDS in Cognitive Sciences, vol. 6, no. 2, 2002, PII: S1364-
6613(00)01813-1.

149

Authorized licensed use limited to: ULAKBIM UASL - ISTANBUL KULTUR UNIVERSITESI. Downloaded on April 19,2025 at 12:00:08 UTC from IEEE Xplore. Restrictions apply.