![](bg6.png)
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.
TextureClassificationwithFisherKernelExtractedfromtheContinuousModelsofRBM
689