Texture Classification with Fisher Kernel Extracted from the Continuous
Models of RBM
Tayyaba Azim and Mahesan Niranjan
School of Electronics and Computer Science, University of Southampton, Southampton, U.K.
Keywords:
Fisher Kernel, Factored 3-way RBM, Gaussian Bernoulli RBM, Texture Classification, Brodatz, Emphysema.
Abstract:
In this paper, we introduce a novel technique of deriving Fisher kernels from the Gaussian Bernoulli restricted
Boltzmann machine (GBRBM) and factored 3-way restricted Boltzmann machine (FRBM) to yield better
texture classification results. GBRBM and FRBM, both, are stochastic probabilistic models that have already
shown their suitability for modelling real valued continuous data, however, they are not efficient models for
classification based on their likelihood performances (Jaakkola and Haussler, 1999; Azim and Niranjan, 2013).
We induce discrimination in these models with the help of Fisher kernel that is constructed from the gradients
of the parameters of the generative model. From the empirical results shown on two different texture data
sets, i.e. Emphysema and Brodatz, we demonstrate how a useful texture classifier could be built from a very
compact generative model that represents the data in the Fisher score space discriminately. The proposed
discriminative technique allows us to achieve competitive classification performance on texture data sets,
without expanding the size of the generative model with large number of hidden units. Also, comparative
analysis shows that factored 3-way RBM is a good representative model of textures, giving rise to a Fisher
score space that is less sparse and efficient for classification.
1 INTRODUCTION
Texture analysis and classification is one of the most
widely explored research problems in computer vi-
sion. Since textures form an important feature of
the objects in an image, and the perception of tex-
tures has played an important role in the human vi-
sual system for recognition and interpretation, there
has been a great interest in developing artificial recog-
nition systems that deploy texture based features for
classification (Hangarge et al., 2013; Zhang et al.,
2005). There are a number of different approaches:
statistical, geometrical and model based, that have
become state of the art techniques for texture classi-
fication. The statistical techniques compute the lo-
cal features at each point in an image and derive
a set of statistics from the distribution of local fea-
tures, for example, the co-occurrence features (Haral-
ick et al., 1973) or the gray level differences (Weszka
et al., 1976). The geometrical methods assume that
the building blocks of textures are textons that gov-
ern the spatial organization of the textures. These
primitives are usually extracted by edge detection fil-
ters such as Laplacian-of-Gaussian or difference-of-
Gaussian (Marr and Vaina, 1982; Poggio et al., 1988;
Tuceryan and Jain, 1998), by adaptive region extrac-
tion (Tomita and Tsuji, 1990) or mathematical mor-
phology (Serra, 1983; Matheron, 1967). After the
primitives have been identified, some statistics of the
primitives such as intensity and area are computed
for analysis. Model-based methods hypothesize the
underlying texture process, constructing a parametric
generative model, which can create the observed in-
tensity distribution of the image. Some of the exam-
ples of this method are pixel based models and region
based models (Cristani et al., 2002).
In this work, we introduce a novel approach for
texture classification that models the textures through
a probabilistic statistical model and then uses the gra-
dients of the model parameters as features for classi-
fication. The energy based probabilistic models used
for capturing the pixel intensity variations present
in the continuous images are Gaussian Bernoulli re-
stricted Boltzmann machine (GBRBM) and factored
3-way restricted Boltzmann machine. Both the mod-
els have been successfully used previously to capture
the pixel intensity variations in real valued images like
natural scenes and textures (Kivinen and Williams,
2012; Ranzato et al., 2010; Cho et al., 2011). These
variants of RBM provide a solution to the poor mod-
684
Azim T. and Niranjan M..
Texture Classification with Fisher Kernel Extracted from the Continuous Models of RBM.
DOI: 10.5220/0004857506840690
In Proceedings of the 9th International Conference on Computer Vision Theory and Applications (VISAPP-2014), pages 684-690
ISBN: 978-989-758-004-8
Copyright
c
2014 SCITEPRESS (Science and Technology Publications, Lda.)
elling strength of the classical RBM for real valued
data. The GBRBM, although useful, is much slower
to train (Krizhevsky, 2009) and is not a good model of
the covariance structure of an image because it does
not capture the fact that the intensity of a pixel is al-
most exactly the average of its neighbours. Also it
lacks a type of structure that has proven very effective
in vision applications. These challenges have been
addressed in factored 3-way RBM that uses the states
of its hidden units to represent the abnormalities in
the local covariance structure of an image.
To the best of our knowledge, Fisher kernel has
not been extracted from the GBRBM and factored
3-way RBM before. In this work, we’ll first show
how this extraction is possible, and then discuss the
advantages one may get for texture classification, if
the probabilistic models used for Fisher kernel is fac-
tored 3-way RBM rather than a binary-binary RBM
(BBRBM) or GBRBM. We observed that the tex-
ture classification accuracies are better if the under-
lying probabilistic model represents the data in gra-
dient space discriminatively and with less sparsity.
This manuscript is organized as follows: Section 2 de-
scribes the Fisher kernel framework, Sections 3 and 4
explain the generative models used for deriving the
Fisher score space, Section 5 discusses the experi-
mental design and the results obtained on benchmark
texture data sets. We then conclude with a discussion
on the obtained results and future work.
2 THE FISHER KERNEL
The Fisher kernel provides a generic framework for
deriving a kernel from a generative probability model,
p(x|θ) by computing Fisher scores which are the gra-
dients of the log likelihood of the data with respect
to the model parameters, θ. Thus, the Fisher kernel
function is mathematically expressed as:
K(x
i
, x
j
) = [
θ
log p(x
i
|θ)]
T
U
1
[
θ
log p(x
j
|θ)],
where
θ
log p(x|θ) = φ
x
and U = E
x
[φ
x
φ
x
T
].
The magnitude of the Fisher score, φ
x
specifies the ex-
tent to which each parameter contributes in generating
the data. The Fisher scores obtained for each indepen-
dent parameter are arranged in a vector of fixed di-
mension. The Fisher information matrix, U is the co-
variance matrix of the score vectors φ
x
and is approxi-
mated as an identity matrix in this work to avoid com-
putational complexity (Taylor and Cristianini, 2004).
Once a kernel function is derived from a generative
probability model, it could be embedded into any dis-
criminative classifier such as support vector machines
(SVM), linear discriminant analysis (LDA), etc. We
have used SVM as a classifier to classify the images in
benchmark data sets. The two probabilistic generative
models from which the Fisher kernel has been derived
in this work are Gaussian-Bernoulli restricted Boltz-
mann machine and factored 3-way restricted Boltz-
mann machine, both discussed below.
3 GAUSSIAN BERNOULLI
RESTRICTED BOLTZMANN
MACHINE (GBRBM)
An RBM is a bipartite graph in which the visible units
that represent the observations are connected to bi-
nary stochastic hidden units using undirected weight
connections. The hidden units allow the network to
discover interesting features that represent complex
regularities in the observations fed to the visible layer
during training. The connectivity of the units is re-
stricted with no visible-visible or hidden-hidden con-
nections, thus allowing us to update all the units in
the same layer in parallel. Moreover, biases are con-
nected as an external input to each of the unit in the
network. In a GBRBM, the visible units have a Gaus-
sian activation function that modifies the energy of the
RBM in the following way:
E(v, h) =
V
i=1
(v
i
b
v
i
)
2
2σ
2
i
H
j=1
b
h
j
h
j
V
i=1
H
j=1
v
i
σ
i
h
j
w
i j
,
where b
v
i
and b
h
j
are the biases attached to the visi-
ble, v and hidden units, h and w
i j
refers to the weight
interaction between the visible unit i and hidden unit
j. Since, there is no direct connection of the units of
each layer with each other, it is easy to infer samples
via the following conditional distributions:
p(v|h) =
V
i=1
N
b
v
i
+
K
j=1
h
j
W
i j
, σ
2
i
!
,
p(h|v) =
H
j=1
sigmoid
b
h
j
+
V
i=1
W
i j
v
i
σ
i
!
,
where N (., σ
2
) denotes the probability density func-
tion of the Gaussian distribution with mean, µ and
variance, σ
2
and sigmoid(x)=
1
1+exp(x)
. The gradient
to update the model parameters are:
L
W
=
1
σ
vh
data
1
σ
vh
model
,
L
b
v
=
1
σ
2
(v b
v
)
data
1
σ
2
(v b
v
)
model
,
L
b
h
= hhi
data
hhi
model
.
TextureClassificationwithFisherKernelExtractedfromtheContinuousModelsofRBM
685
The angle brackets represent the expectation un-
der the probability distribution specified by the sub-
script. The first expectation under the data is calcu-
lable whereas the second expectation over the model
distribution is intractable and can be approximated
by drawing samples through a Markov chain Monte
Carlo algorithm running for a very short time, i.e. 1
step, as proposed in Contrastive Divergence 1 (CD-
1) algorithm (Hinton, 2002). GBRBM in general is
known as difficult to train and this difficulty arises
from learning standard deviations σ
i
of the visible
neurons. Unlike other parameters, the standard de-
viations are constrained to be positive. However, with
an inappropriate learning rate, it is possible for the ob-
tained gradient update rule to result in a non-positive
standard deviation which may result in an infinite en-
ergy of the model ( in case of σ
i
=0) or to an ill-
defined conditional distribution of the visible neuron
(in case of σ
i
=0). Since, all gradients other than that
of the hidden biases are scaled by the standard devi-
ation, inappropriate learning of it affects the learning
of other parameters too. In this context, (Krizhevsky,
2009) suggested using a separate learning rate for
the standard deviations which should be 100 to 1000
times smaller than that of the other parameters. How-
ever, there has been a general consensus to update
the weights and the biases only, and use fixed, pos-
sibly unit standard deviation that results in impres-
sive performances (Hinton and Salakhutdinov, 2009;
Krizhevsky, 2009; Mohamed et al., 2010). For this
reason, we have taken σ = 1 in our work. The Fisher
score is derived from the log likelihood of the model
as:
θ
log p(x
n
|θ) =
θ
L =
L
W
.
.
.
L
b
v
.
.
.
L
b
h
,
where θ = {W, b
v
, b
h
}.
4 FACTORED 3-WAY
RESTRICTED BOLTZMANN
MACHINE (FRBM)
Ranzato et al. (Ranzato et al., 2010) proposed that
an RBM’s visible and hidden units can be modified to
incorporate three-way interactions so that the covari-
ance of the visible units is captured. This modified
RBM which allows the hidden units to modulate pair-
wise interactions between the visible units is called
three-way RBM shown in Figure 1.
Capturing the interactions between the visible
units has far too many parameters, therefore, to keep
their count under control and make learning efficient
Figure 1: A graphical representation of the factored 3-way
RBM in which the triangular symbol represents a factor that
computes the projection of the input image whose pixels are
denoted by v
i
with a set of filters (columns of matrix C). The
square outputs of the visible units are sent to the binary hid-
den units after projection with a second layer matrix (matrix
P) that pools similar filters.
in practice, it is necessary to factorize these 3-way in-
teractions. The energy function is redefined in terms
of the three-way multiplicative interactions between
the two visible binary units, v
i
, v
j
and one hidden bi-
nary unit, h
k
as:
E(v, h) =
i, j,k
v
i
v
j
h
k
W
i jk
. (1)
For real images, we expect the lateral interactions in
the visible layer to have a lot of regular structure,
therefore the three-way tensor can be approximated
as a sum of factors:
W
i jk
=
f
B
i f
C
j f
P
k f
, (2)
where i refers to the number of visible units, f refers
to the number of factors and k refers to the number of
hidden units. The matrix P is regarded as the factor-
hidden or pooling matrix and the matrix C
i f
is known
as visible to factor matrix. It is sensible to assume that
matrix B = C in Eq. 2 and the final approximation
(Eq. 1) becomes :
E(v, h) =
f
i
v
i
(C
i f
)
!
2
k
h
k
P
k f
!
.
The hidden units of the model are conditionally inde-
pendent given the states of the visible units and their
binary states are sampled as:
p(h
k
= 1|v) = σ
f
P
k f
i
v
i
C
i f
!
2
+ b
k
,
where σ is a logistic function and b
k
is the bias
of the k-th hidden unit. In contrast, given the hidden
states, the visible units are not independent and there-
fore it is much more difficult to compute the recon-
struction of the data from the hidden states. To resolve
this, in practice, the hidden units are integrated out by
VISAPP2014-InternationalConferenceonComputerVisionTheoryandApplications
686
(a) Normal Tissue
(NT)
(b) Centrilobular Em-
physema (CLE)
(c) Paraseptal
Emphysema (PSE)
Figure 2: Examples of different lung tissue patterns ex-
tracted through computed tomography are shown. NT rep-
resents the sample of a healthy tissue, CLE reveals a healthy
smokers tissue and PSE shows the distorted tissue of a per-
son suffering from chronic obstructive pulmonary disease
(COPD).
calculating free energy function of the model, and the
visible samples are inferred using the Hybrid Monte
Carlo (HMC) sampling technique (Neal, 1996) that
calculates gradient of the free energy w.r.t the visible
vector as:
F(v) =
k
log
1 +exp
1
2
f
P
k f
i
C
i f
v
i
!
2
+ b
k
i
b
i
v
i
,
F(v)
v
=
f
C
i f
k
P
k f
i
C
i f
v
i
1 + exp(0.5
f
P
k f
(
i
C
i f
v
i
)
2
b
k
)
.
The gradients of the log likelihood function of free en-
ergy w.r.t each model parameter are given as (Ranzato
et al., 2010):
L1 =
1
2
i
C
i f
v
i
!
2
1
1 + exp(0.5
f
P
k f
(
i
C
i f
v
i
)
2
b
k
)
,
L2 = v
i
k
P
k f
i
C
i f
v
i
1 + exp(0.5
f
P
k f
(
i
C
i f
v
i
)
2
b
k
)
,
L3 =
1
2
1
1 + exp(0.5
f
P
k f
(
i
C
i f
v
i
)
2
b
k
)
,
L4 =
n
i
v
i
n
, where
L1 =
F(v)
P
, L2 =
F(v)
C
, L3 =
F(v)
b
h
and L4 =
F(v)
b
v
.
The F0isher score is derived from the log likelihood
of this model as:
θ
log p(x
n
|θ) =
F(v)
P
.
.
.
F(v)
C
.
.
.
F(v)
b
h
.
.
.
F(v)
b
v
,
where θ = {P, C, b
v
, b
h
}.
5 EXPERIMENTS AND RESULTS
We have carried out the classification experiments on
two different kinds of texture data sets: a medical im-
age database called Emphysema, and a famous texture
data set called Brodatz. The experimental design and
the results obtained on the two data sets are described
as follows:
5.1 Emphysema Data Set
The Emphysema database (Sørensen et al., 2010)
consists of 115 high-resolution computed tomogra-
phy (CT) slices as well as 168 61 × 61 dimensional
patches extracted from the subset of slices, and man-
ually annotated for texture analysis techniques. Em-
physema is a disease characterised by a loss of lung
tissue and is one of the main reasons of chronic ob-
structive pulmonary disease (COPD). A proper clas-
sification of emphysematous - and healthy - lung tis-
sue is useful for a more detailed analysis of the dis-
ease. The 61 × 61 pixel patches
1
are from three dif-
ferent classes: normal tissue (NT) with 59 observa-
tions, Centrilobular Emphysema (CLE) with 50 ob-
servations, and Paraseptal Emphysema (PSE) with
59 observations. The NT patches were annotated as
never smokers, while the CLE and PSE region of in-
terests were annotated as healthy smokers and smok-
ers with COPD. These texture patterns serve as a good
basis for assessing the modelling power of RBMs de-
signed specifically for capturing pixel intensity varia-
tions present in the textures. As a preprocessing step,
we crop 31 × 31 dimensional patch from the center of
each 61 × 61 patch and threshold the pixel values in
the dynamic range [-1000, 500]. The thresholding is
based on the knowledge that the CT density values
of lung parenchyma pixels are usually between the
Hounsefield unit range [-1000HU, 500HU]. In order
to classify these patches into 3 different classes, we
have used Fisher kernel derived from three different
probabilistic models: binary-binary RBM, Gaussian-
Bernoulli RBM and factored 3-way RBM that model
the data representations through different distribu-
tions. Once each of the generative model is trained,
we calculate the gradients of the log likelihood func-
tion to form Fisher scores for the Fisher kernel. The
Fisher kernel is then embedded into the SVM clas-
sifier that finally performs multi-class classification
through one versus one training technique. The op-
timal value for hyperparameter C in SVM is decided
via grid search method. In factored 3-way RBM, we
maintained an average rejection rate of 6% with HMC
sampling that used an adaptive step size to control the
average acceptance rate of the drawn samples, thus
yielding fast mixing rate. The summary of the classi-
fication results of Fisher kernel derived from different
probabilistic models is shown in Table 1 (second col-
1
http://image.diku.dk/emphysema database/,
http://www.ee.oulu.fi/research/imag/texture/image data/
Brodatz32.html
TextureClassificationwithFisherKernelExtractedfromtheContinuousModelsofRBM
687
Table 1: Summary of classification results attained by different classifiers on the Emphysema and Brodatz texture data sets.
Classifier Emphysema Performance (Acc) Brodatz Performance (Acc)
k-Nearest Neighbour [Input=Image pixels, k=1] 46.04 ± 5.27% 29.06 ± 1.66%
Condensed Nearest Neighbour [Input=Image Pixels, 45% Data Retrieved | 25% Data Retrieved] 46.06 ± 5.19% 28.11 ± 2.01%
FK (Binary Binary RBM) [5 hid units] 47.31 ± 5.54% 16.81 ± 2.008%
FK (GaussianBinary RBM) [5 hidden units, σ = 1] 47.85 ± 4.83% 16.96 ± 2.40%
FK (Factored 3-Way RBM ) [5 hid units, 32 factors] 86.97 ± 5.54% 65 ± 4.6%
k-Nearest Neighbour [Input=Local Binary Pattern features, k=1] 95.2%(Sørensen et al., 2010) 91.4%(?)
1 2 3 4 5 6 7 8 9 10
4.469
4.47
4.471
4.472
4.473
4.474
4.475
x 10
10
No. of Epochs
Reconstruction Error
(a) RBM-BB(Emphysema)
1 2 3 4 5 6 7 8 9 10
4.3723
4.3724
4.3725
4.3726
4.3727
4.3728
4.3729
x 10
10
No. of Epochs
Reconstruction Error
(b) RBM-GB(Emphysema)
1 2 3 4 5 6 7 8 9 10
4.42
4.422
4.424
4.426
4.428
4.43
4.432
x 10
10
No. of Epochs
Reconstruction Error
(c) Factored 3-Way RBM (Emphy-
sema)
1 2 3 4 5 6 7 8 9 10
2.81
2.815
2.82
2.825
2.83
2.835
2.84
x 10
10
No. of Epochs
Reconstruction Error
(d) Factored 3-Way RBM (Brodatz)
Figure 3: The reconstruction error shown after training different variants of RBM generative model on the Emphysema and
Brodatz data set for 10 epochs. The error for each of these models drops after several epochs; for factored 3-way model on
Emphysema, it first rises, stabilises and then drops.
umn) and the reconstruction error for each model is
shown in Figure 3. Note that the best known perfor-
mance on Emphysema data set has been achieved by
(Sørensen et al., 2010), in which he used the leave one
subject out methodology to test the classifier. Such
a partitioning scheme did not reveal discriminative
Fisher score space in our case, due to which we chose
holdout estimation method to train models and draw
Fisher scores. Consequently, the Fisher kernel de-
rived from factored 3-way RBM does give competi-
tive classification performance in the same league as
shown by (Sørensen et al., 2010).
5.2 Brodatz Texture Data Set
The Brodatz textures (Valkealahti and Oja, 1998)
data set consists of a subset of 32 different classes
chosen randomly from the main Brodatz data set.
These textures are histogram equalized and then 20
patches
1
of size 64 × 64 are drawn from random lo-
cations of each class database for further experimen-
tation. Table 1 shows the classification performance
of distance based approaches, i.e. k-NN and con-
densed NN on these preprocessed patches. The same
patches are also fed to the generative probability mod-
els for representational learning. Once the models
(binary-binary, gaussian-binary and factored 3-way)
are trained, a Fisher kernel is extracted from them and
then embedded into the SVM classifier. The SVM
classifies these textures using one versus one train-
ing of the gradients learnt by different models. The
hyperparameter C in SVM is once again decided via
Figure 4: Samples of texture images from the Brodatz data
set.
Figure 5: The visual factor filters, C
i f
learnt from the 64 ×
64 size patches of Brodatz data set (1
st
row) and 31 × 31
size patches of Emphysema data set (2
nd
row).
grid search method. From the results obtained, we
observe that the Fisher kernel derived from a fac-
tored 3-way RBM gives better classification perfor-
mance in comparison to the other Fisher kernel based
approaches and distance based classifiers on prepro-
cessed images. The best performance on the data set
is once again shown by local binary pattern features
classified through k-NN and shown in Table 1.
VISAPP2014-InternationalConferenceonComputerVisionTheoryandApplications
688
6 DISCUSSION AND FUTURE
WORK
In this paper, we present a novel approach of deriv-
ing a suitable classifier for texture classification that
uses the gradients of the generative model to differ-
entiate between different categories of textures. From
the experiments conducted above, we observed that
the performance of the Fisher kernel approach re-
lies on the discriminative quality of the Fisher score
space attained via maximum likelihood training of
the generative models. On a comparative scale, the
factored 3-way RBM proves better than the GBRBM
and BBRBM since it was able to provide less sparse
Fisher vectors that makes them suitable for discrimi-
nation in the dot product space. The dot product space
is not suitable for learning distance metric similarities
over sparse data, therefore Fisher vectors with zero or
very small gradients donot provide a space discrimi-
nant enough for texture classification, as revealed for
FK-BBRBM and FK-GBRBM in Table 1. It is also
important to note that despite the availability of less
sparse Fisher vectors, the Fisher kernel classification
performance still does not beat the best known clas-
sification performance on the Brodatz data set. This
follows us to the conclusion that a generative model
which is trained well via maximum likelihood learn-
ing does not necessarily give rise to a representation
that is well suited for classification tasks. In practice,
the Fisher vectors for objects that have high proba-
bility under the model, will comprise of very small
gradients that are less likely to form a discriminative
basis for kernel functions. We would like to explore
this in more detail by overcoming the gradient scaling
problem through kernel normalization techniques in
the future. Such a kernel should satisfy the rationale
of achieving a discriminant Fisher score space by as-
signing similar gradients to two similar objects, and
maintaining inter-class separability too. The impact
of generative model’s scale on the Fisher score space
is also worth studying and will be pursued in future.
REFERENCES
Azim, T. and Niranjan, M. (2013). Inducing Discrimina-
tion in Biologically Inspired Models of Visual Scene
Recognition. In MLSP.
Chen, J., Kellokumpu, V., and Pietikinen, G. Z. . M. (2013).
RLBP: Robust Local Binary Pattern. In BMVC 2013,
Bristol, UK.
Cho, K., Alexander, A., and R.Tapani (2011). Improved
Learning of Gaussian-Bernoulli Restricted Boltzmann
Machines. In ANN - Volume Part I, ICANN, pages
1017, Berlin, Heidelberg. Springer-Verlag.
Cristani, M., Bicego, M., and Murino, V. (2002). Inte-
grated Region and Pixel Based Approach to Back-
ground Modelling. In Workshop on MVC, pages 3–8.
Hangarge, M., Santosh, K., Doddamani, S., and Pardeshi,
R. (2013). Statistical Texture Features Based Hand-
written and Printed Text Classification in South Indian
Documents. CoRR, 1(32).
Haralick, R., Shanmugam, K., and Dinstein, I. (1973).
Textural Features for Image Classification. SMC,
3(6):610–621.
Hinton, G. (2002). Training Products of Experts by Mini-
mizing Contrastive Divergence. Neural Computation,
14:1771–1800.
Hinton, G. and Salakhutdinov, R. (2009). Semantic Hash-
ing. IJAR, 50(7):969–978.
Jaakkola, T. and Haussler, D. (1999). Exploiting Genera-
tive Models in Discriminative Classifiers. NIPS, pages
487–493.
Kivinen, J. and Williams, C. (2012). Multiple Texture
Boltzmann Machines. In AISTATS, volume 22, pages
638–646.
Krizhevsky, A. (2009). Learning Multiple Layers of Fea-
tures from Tiny Images. Master’s thesis.
Marr, D. and Vaina, L. (1982). Representation and Recog-
nition of the Movements of Shapes. Proceedings of
the Royal Society of London. Series B. Biological Sci-
ences, 214(1197):501524.
Matheron, G. (1967). Representation and Recognition
of the Movements of Shapes. Proceedings of the
Royal Society of London. Series B. Biological Sci-
ences, 214(1197):501–524.
Mohamed, A., Dahl, G., and Hinton, G. (2010). Deep Belief
Networks for Phone Recognition. In NIPS.
Neal, R. (1996). Bayesian Learning for Neural Networks.
Springer-Verlag New York, Inc., Secaucus, NJ, USA.
Poggio, T., Voorhees, H., and Yuille, A. (1988). A Regular-
ized Solution to Edge Detection. Journal of Complex-
ity, 4(2):106 123
Ranzato, M., Krizhevsky, A., and Hinton, G. (2010).
Factored 3-Way Restricted Boltzmann Machines for
Modeling Natural Images. In AISTATS.
Serra, J. (1983). Image Analysis and Mathematical Mor-
phology. Academic Press, Inc., Orlando, FL, USA.
Sørensen, L., Shaker, S., and de Bruijne, M. (2010). Quan-
titative Analysis of Pulmonary Emphysema Using Lo-
cal Binary Patterns. Medical Imaging, 29(2):559–569.
Taylor, J. and Cristianini, N. (2004). Kernel Methods for
Pattern Analysis. Cambridge University Press, New
York, USA.
Tomita, F. and Tsuji, S. (1990). Computer Analysis of Vi-
sual Textures. Kluwer Academic Publishers, Norwell,
MA, USA.
Tuceryan, M. and Jain, A. (1998). Handbook of Pattern
Recognition & Computer Vision. Chapter Texture
Analysis, pages 235276. River Edge, NJ, USA
Valkealahti, K. and Oja, E. (1998). Reduced Multidimen-
sional Co-occurrence Histograms in Texture Classifi-
cation. PAMI, 20(1):90–94.
TextureClassificationwithFisherKernelExtractedfromtheContinuousModelsofRBM
689
Weszka, J., Dyer, C., and Rosenfeld, A. (1976). A Compar-
ative Study of Texture Measures for Terrain Classifi-
cation. SMC, 6(4):269–285.
Zhang, Y., X. Jian and J. Han (2005). Texture Feature-based
Image Classification Using Wavelet Package Trans-
form. In AIC, volume 3644 of LNCS, pages 165 173.
Springer Berlin Heidelberg.
VISAPP2014-InternationalConferenceonComputerVisionTheoryandApplications
690