A Review of Image Classification Techniques

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 04 Issue: 11 | Nov -2017 www.irjet.net p-ISSN: 2395-0072
© 2017, IRJET | Impact Factor value: 6.171 | ISO 9001:2008 Certified Journal | Page 1588
A REVIEW OF IMAGE CLASSIFICATION TECHNIQUES
Nupur Thakur1, Deepa Maheshwari2
1,2Electronics and Telecommunication, Pune Institute of Computer Technology, Pune-Satara Road, Behind Bharati
Vidyapeeth College, Dhankawadi, Pune, Maharashtra, India.
------------------------------------------------------------------------***-------------------------------------------------------------------------
Abstract - Image classification is an important tool for
extracting information from digital images. The aim of this
paper is to summarize information about few image
classification techniques. The paper also elaborates different
categories of image classification techniques. The image
classification techniques considered in this paper are
Parallelpiped Technique, Minimum Distance Technique,
Maximum Likelihood (ML) Technique, Artificial Neural
Networks (ANN) and Support Vector Machine (SVM).
Key Words: classifier, image, supervised, classification,
class, pixel
1. INTRODUCTION
The process of classifying pixels into finite set of individual
classes based on their data values is known as image
classification. The pixel is assigned to a particular class if it
satisfies the certain set of rules to fit in a particularclass.The
classes can be known or unknown. If the user is able to
separate the classes based on the training data, then the
classes are known, else they are unknown.
In general, the image classification techniques can be
categorized as parametric and non-parametric, or
supervised and unsupervised, or hard and soft classifiers.
Depending on the whether there is prior knowledge of the
classes, the techniques are divided into two groups,
supervised and unsupervised classification techniques.
Supervised classification technique requires the training
data set in order to teach the classifier to define the decision
boundary. It recognizes the instance of the necessary
information in the image, which are known as training sites.
This is then used to expand a statistical description of the
reflectance of information for each class, which is known as
‘Signature Analysis’. The last step is to classify the image by
searching the reflectance for each pixel and evaluating the
resemblance to the signatures. [1] The data provided during
the signature analysis, also known as training phase, is
stored in a file called training data file. The classification
phase uses this information for classifying the input images.
[2] The advantage of this kind of technique is that the errors
can be easily identified and solved. The only disadvantage is
the large time required for training phase. [3]
[4] states that unsupervised learning explores the
underlying structure of thedata andautomaticallypartitions
them based on it. It produces a set of centroids which
represent the prototypes of the classes. These are used for
further classification. [5] divides unsupervised, also known
as clustering method, into two groups, namely Hierarchical
clustering and Partitioningclustering.Theformergroups the
data with a sequence of partitions, while the later one
divides the data into pre-specified number of clusters.
According to [6], the algorithm starts with initialization that
is done by executing an initial segmentation rule. Next, the
classification is done using differentstrategies.The resultsof
these techniques is physically explicable, but the accuracy
highly depends on the design of the algorithm. This
technique is fast and fully automated.
The parametric and non-parametric image classification
techniques come under the supervised learning. The
parametric classifiers use the algebraic possibility for
allocation to each class. Some of the parametric classifiers
are Bayesian Classifier, Naïve Bayes classifier and decision
tree. [7] The parameters required are taken from the
training data. Parameters like mean and co-variance are
used in these classifiers. [8]
The non-parametric classifiers are used when there is no
density function available. It approximates the probability
density function for further use. Some of thenon-parametric
classifiers are K-Nearest Neighbor, Logical Regression and
Multilayer Perceptron. [7]
The hard classifiers develop classes by combining the
spectra of all the pixels in a training set from a given feature.
The classes contains the contributions of all the pixels in the
training set. [9] The hard classifiers assume that the pixels
are pure and categorize them into one and only one class.
This makes them inefficient in handling with the problem of
mixed pixel. It treats the mixed pixels as noise, uncertainty
or error. [10]
The soft classifiers work at the sub-pixel level but they
cannot deal with pure pixels accurately. They treat them as
mixed pixels, which produces huge error. [9] Theygroupthe
mixed pixels into multiple classes and handles this problem
very well. There are two soft classifiers namely FCM (Fuzzy
c-means) and PCM (Possibilitic c-means). The FCM has a
probabilistic constraint and PCM is based on modified
version of PCM. [10] This classifierassignsthepixel fractions
according to the area it represents inside a pixel, which
makes it efficient in handling mixed pixels. [11]
The paper is organized as follows. Section 2 presents the
literature survey. Section 3 illustrates few image

classification techniques. Section 4 provides the conclusion
of the paper.
2. LITERATURE SURVEY
Image classification is an important step in the object
detection and image analysis. The output of the image
classification step can be the final outputortheintermediate
output. A lot of image classification techniques have been
proposed till date. Various studies have been conducted in
order to conclude aboutthe bestsatelliteimageclassification
technique. It is hard to decide any one technique as the best
technique among all, because the results and its accuracy
depend on a number of factors [2].
Over the last few decades, there is a constant modificationin
the conventional methods as well as invention of new image
classification techniques in order to get maximum accurate
results. Each of the classification technique has its own
advantages and disadvantages. The research now
concentrates on combining the desired features of these
techniques in order to increase the efficiency.
As the hard classifiers cannot handle the problem of mixed
pixels, the soft classifiers are used. But, soft classifiers have
their own disadvantages. [9] presents a study that combines
the desirable features of a soft classifieranda hardclassifier.
It makes the use of LSMM, a soft classifier and SVM, a hard
classifier and compared the results with the ones produced
by LSMM and SVM separately. The results showed that the
combination of bothclassifiersproducedbetterresultswhen
compared to either of them.
Another study was conducted on PoISAR data combined the
Markov randomfieldsmoothness constraintwithsupervised
Softmax regression model [6]. The experiments conducted
during this study proved that the combination of supervised
and unsupervised algorithms produced better results as
compared to the ones producedby eitherofthesetechniques
separately.
[16] presents the combination of fuzzy logic and neural
networks in order to design a system that can detecttheface
and fingerprints of the person. This is done in order to
determine the authenticity of the person. This systemcan be
used for various security purposes.
3. IMAGE CLASSIFICATION TECHNIQUES
This section delineates the supervised image classification
techniques that are used recently.
3.1 Parallelpiped technique
This method of classification is used by determining the
parallelpiped- shaped boxes for each pre-defined class [12].
The parallelpiped boundaries for the classesaredetermined
by the minimum and maximum of pixels ina particularclass.
These boundaries help in assigning a pixel to a given class.
The classifier is trained by analyzing the histograms of the
individual spectral components of the trainingsamples.This
technique has many advantages like:
• It is easy to understand and implement.
• The speed of this classifier is high.
Though the above advantages are significant, it has a lot of
drawbacks due to which it is not used practically. These
disadvantages are:
• There can be significant gaps between the
parallelpipeds and the pixels within this region will remain
unclassified.
• Another drawback is that the prior probabilities of
the class memberships are not taken into consideration.
3.2 Minimum distance classifier
Minimum distance classifier is a supervised image
classification technique, in which the pixels are classified
based on their distance from the mean spectra of the pre-
defined classes [12]. In this method, first themeanvectorfor
each class is calculated based on the training dataset. Next,
using the Minimum distance algorithm, the Euclidean
distance of every unclassified pixel from the mean vector is
calculated. The pixel is then assigned to the class with the
minimum distance.
The distance (dcx) of a particular pixel from different class
mean vectors (x) is usually calculated using Euclidean
distance:
This type of classifier is mathematicallysimpleandtherefore
computationally less complex. It requires the least time for
computation among all the other supervised classification
techniques. The disadvantage of this technique is that it
takes into account only the mean value, and so it is less
efficient than maximum likelihood technique.
3.3 Maximum likelihood classifier
Maximum Likelihood is a supervised image classification
technique in which the probability value of pixels is taken
into consideration for classifying the pixels [14]. In this
method, the probability of each pixel belonging to a class is
calculated. These values are then compared. The pixel is
assigned to the class where the probability value is highest.
In this method, it is assumed that all the input bands have
normal distribution.
where is the likelihood membership function ofxbelonging
to class k,

is the vector of a pixel with n bands
is the mean of the kth class
is the variance-covariance matrix
This method is highly efficient when it comes to classifying
the satellite images, especially the multi-spectral images.
Though this is an efficient technique, it requires large
computational time.
3.4 Artificial Neural Network
Artificial neural networksarenon-parametric classifiers. The
structure of the artificial neural networks is inspired from
the human nervous system. The basic unit of this type of
network is unified processing rudiment known as neuron.
Each neuron has two stages- training and using phase [1]. In
the training phase, the neuron learn to performanoperation
while in the testing phase, they use the training information
to predict the output. Generally, these neural networks are
used in order to detect specific trends or patterns in the
given data.
The artificial neural networks have several advantages:
• It has high computation rate.
• It deals with the noisy inputs efficiently.
• This technique is data driven as it learns from the
training data.
Though these artificial neural networks are popular, they
have some disadvantages:
• As it requires prior training, it is time-consuming.
• It is considered to be semantically poor.
• It encounters the problem of overfitting.
3.5 Support Vector Machine
Support Vector Machine, also known as SVM is a non-
parametric classifier. Support Vector Machine is a binary
classifier and separates the classes using a linear boundary.
This classifier assumes that there is no prior information on
how to classify the data. This optimizes the use of training
data, which is the biggest advantage of this classifier over
other classifiers like Maximum Likelihood Classifier [15].
The real power of SVM lies in the kernel illustration as it
facilitates the non-linear mapping of the input space to the
feature space [1]. Therefore, the choice of kernel function is
the most significant step in Support VectorMachine.Someof
the commonly used kernel functions are:
• Linear Kernel
• Polynomial Kernel
• Gaussian Kernel
As SVM optimizes the training data use, it increases the
speed of the classifier to a great extent. It also minimizes the
classification errors that would occur due to prior
assumptions on the unsupervised data [15]. The major
advantages of SVM are:
• It has excellent generalization capacity.
• It does not face the problem of overfitting.
• It makes use of the non-linear transformation.
The disadvantages of this classifier are:
• The structure of the algorithm is complex and
therefore difficult to understand.
• Optimal parameters cannot be defined easily.
CONCLUSION
There are many ways in which these techniques are
classified and categorizing these techniques into supervised
and unsupervised is the most common way.Thecomparison
of these techniques on the basis of efficiencylargelydepends
on the type of data they are being used for. This paper
summarizes the information about commonly used image
classification techniques. This will help the researchers to
select the most appropriate classification technique
according to their requirements.
REFERENCES
[1] Rajesh Sharma R Beaula A, Marikkannu P, Akey
Sungheeth, C. Sahana, “Comparative Study of Distinctive
Image Classification Techniques”, 10th International
Conference on Intelligent Systems and Control (ISCO), 2016
[2] Kalra K., Goswami A.K., Gupta R.,“AComparativeStudy of
Supervised Image Classification Algorithms for Satellite
Images”, International Journal of Electrical, Electronics and
Data Communication, ISSN: 2320-2084, Volume-1,Issue-10,
Dec-2013
[3] Kurian J., Karunakaran V., “A Survey on Image
Classification Methods”, International Journal of Advanced
Research in Electronics and Communication Engineering
(IJARECE) Volume 1, Issue 4, October 2012
[4] Y. Hu and K. Ashenayi, R. Veltri, G. O'Dowd and G. Miller,
R. Hurst and R. Bonner, “A Comparison of Neural Network
and Fuzzy c-Means Methods in Bladder Cancer Cell
Classification”,1994.IEEEWorldCongressonComputational
Intelligence
[5] Yun Yang and Ke Chen, “Unsupervised Learning via
Iteratively Constructed Clustering Ensemble”, The 2010
International Joint Conference on Neural Networks (IJCNN)

[6] Haixia Bi, Jian Sun, Zongben Xu, “Unsupervised PolSAR
Image Classiﬁcation Using Discriminative Clustering”, IEEE
Transactions on Geoscience and RemoteSensing,Volume55
Issue 6
[7] M. S. Sonawane, C. A. Dhawale, “Evaluation and Analysis
of few Parametric and Nonparametric Classification
Methods”, 2016 Second International Conference on
Computational Intelligence & Communication Technology
[8] M. P. Sampat, A. C. Bovik, J. K. Aggarwal, K. R. Castleman,
“SupervisedParametric andNon-Parametric Classiﬁcationof
Chromosome Images”, Pattern Recognition 38(8): 1209-
1223 (2005)
[9] Tangao Hu, Wenyuan Wu, Lijuan Liu, “Combination of
Hard and Soft Classification Method Based on Adaptive
Threshold”, IGARSS 2014
[10] Ranjana Sharma, R. K. Dwivedi, Achal Kumar Goyal,
“Review of soft classification approaches on satellite Image
and accuracy assessment”, 4th International Conference on
System Modeling & Advancement in Research Trends
(SMART) College of Computing Sciences and Information
Technology (CCSIT), 2015
[11] Amit Masoud Chegoonian, Mehdi Mokhtarzade,
Mohammad Javad Valadan Zoej, Maryam Salehi, “Soft
Supervised Classification and Improved Method for Coral
Reef Classification Using Medium Resolution Satellite
Images”, IGARSS 2016
[12] Sunitha Abburu and Suresh Babu Golla, “SatelliteImage
Classification Methods and Techniques: A Review”,
International Journal of Computer Applications (0975 –
8887) Volume 119 – No.8, June 2015
[13] Aykut AKGÜN, A.Hüsnü ERONAT and Necdet TÜRK,
“Comparing Different SatelliteImageClassificationMethods:
An Application in Ayvalik District, Western Turkey”, 4th
International Congress for Photogrammetry and Remote
Sensing, Istanbul, Turkey
[14] F. S. Al-Ahmadi and A. S. Hames, “Comparison of Four
Classification Methods to Extract Land Use and Land Cover
from Raw Satellite Images for Some Remote Arid Areas,
Kingdom of Saudi Arabia”, JKAU; Earth Sci., Vol. 20 No.1, pp:
167-191 (2009 A.D./1430 A.H.)
[15] Maryam Niknejad, Vahid MirzaeiZadeh,MehdiHeydari,
“Comparing different classifications of satellite imagery in
forest mapping”, International Research Journal of Applied
and Basic Sciences, 2014
[16] G. Prabhakar Reddya, Y. Deepika, K. Sai Prasad, Dr. G.
Kiran Kumar, “Fuzzy Logics associated with Neural
Networks in the Real Time for Better World”, International
Conference on Advancements in Aeromechanical Materials
for Manufacturing (ICAAMM-2016)

A Review of Image Classification Techniques

More Related Content

What's hot (17)

Similar to A Review of Image Classification Techniques (20)

More from IRJET Journal (20)

Recently uploaded (20)

A Review of Image Classification Techniques