VOL. 11, NO. 15, AUGUST 2016 ISSN 1819-6608
ARPN Journal of Engineering and Applied Sciences
©2006-2016 Asian Research Publishing Network (ARPN). All rights reserved.
www.arpnjournals.com
9177
A NOVEL TEMPLATE MATCHING IMPLEMENTATION OF OBJECT
BASED IMAGE CLASSIFICATION BASED ON MULTIKERNEL
FUSION SPARSE REPRESENTATION
Shivakumar G. S.
1
, S. Natarajan
2
and K. Srikanta Murthy
3
1
Department of Computer Science and Engineering, Srinivas Institute of Technology, Mangalore, India
2
Department of Information Science, PESIT, Bangalore, India
3
Department of Computer Science and Engineering, PESSE, Bangalore, India
E-Mail: shivakumar_gs1@yahoo.co.in
ABSTRACT
This paper introduces and implements a novel object based image classification method on remote sensing
images. The novelty introduced in this implementation is the application of a Multikernel Sparse Representation method on
the object based image classification. The template-matching algorithm inspired from the object tracking implementation
replaces the process of segmentation usually applied in object based image classification. The Multikernel fusion sparse
representation based learning and prediction method is developed for remote sensing image classification. A particle filter
framework for the sample template selection with the Multikernel Fusion Sparse Representation optimization technique is
used to develop the image classification algorithm. The particle filter will act as the template-matching framework for our
classification algorithm and the optimization of the observation model of this framework is carried out using the
Multikernel Fusion Sparse Representation. Multikernel implementation has been proved to be more accurate than the
feature extraction techniques since it extracts the internal intricacies of the image vector. The Kernels consume lesser
memory space and lesser computational complexity compared to the traditional feature extracting methods. Multikernel
Sparse representation has been proved to be more accurate and less computationally complex while implemented in other
applications like the video object tracking. Affine transform based templates are extracted from the image which have to be
trained and the kernel matrix is generated which is used for comparison with the templates extracted from the test images.
Kernel Coordinate Descent (KCD) algorithm is used to find the similarity measure between the database kernel and the
testing kernel. The weight values updated using the observation likelihood method that would indicate whether the test
template matches with the database templates. The comparison is carried out with the multikernel method using the SVM
classifier. The results that are observed are kappa coefficient and overall accuracy, which measure the classification
accuracy, for images with higher and lower illumination and also the images are analysed for robustness to direction
change and the classification performance for two different hyperspectral images.
Keywords: active learning, sparse representation, remote sensing, classification.
INTRODUCTION
Conditional Probability is the probability of
occurrence of an event provided another event had
occurred. Posterior Probability is the conditional
probability, which considers the relevance between the
events that occurred due to the conditional probability.
Particle framework defined in the literature [18] is based
on the posterior probability method on the video tracking
implementation. Posterior probability defines the most
certain occurrence of an event. The recent literature used
the particle filter framework on image classification where
the filtering, labeling and statistics are carried out using
particle filter framework [17]. Sparse representation is
castoff to symbolize the data under study in a comparable
and expressive manner in a smaller memory space. SR-
based algorithm has a drawback of high computation cost
and less accuracy which is overcome by using kernel
sparse representation (KSR)-based algorithm using single
feature kernel and it also avoids introducing large number
of trivial templates which will speed up compared to SR-
based methods. Furthermore to overcome the weakness of
single feature in object description, multikernel fusion
method is proposed for multiple features integration.
Furthermore to improve the method to have higher
classification accuracy we utilize a prediction strategy
called the adaptive multikernel fusion. Sparse
representation would be a signal specific basis pursuit
method that would be a more accurate representation for
any non-linear signal. The deeper insight of sparse
representation into the data structure would make it
eligible for the classification algorithms of remote sensing.
The sparse representation is the most compact
representation by using the linear combination of the
building blocks of the data. Remote sensing image
classification using sparse implementation and active
learning method is taken into account for implementation.
The active learning method is the process where the data
under learning has to be decided by the human
intervention in order to have higher redundancy [1]. SVMs
can be considered as techniques which use hypothesis
space of linear separators in a high dimensional feature
space, trained with a learning algorithm from optimization
theory that makes a learning bias derived from statistical
learning theory [2, 3]. Support vector machine is one of
the important methods for use in the active learning
method. The classification was binary but now the
multiclass Support Vector Machine has been introduced
for better classification.