Proposed algorithm for image classification using regression-based pre-processing and recognition models

International Journal of Electrical and Computer Engineering (IJECE)
Vol. 9, No. 2, April 2019, pp. 1021~1027
ISSN: 2088-8708, DOI: 10.11591/ijece.v9i2.pp1021-1027  1021
Journal homepage: http://iaescore.com/journals/index.php/IJECE
Proposed algorithm for image classification using
regression-based pre-processing and recognition models
Chanintorn Jittawiriyanukoon
Assumption University, Thailand
Article Info ABSTRACT
Article history:
Received Apr 27, 2018
Revised Sep 30, 2018
Accepted Oct 15, 2018
Image classification algorithms can categorize pixels regarding image
attributes with the pre-processing of learner’s trained samples. The precision
and classification accuracy are complex to compute due to the variable size
of pixels (different image width and height) and numerous characteristics of
image per se. This research proposes an image classification algorithm based
on regression-based pre-processing and the recognition models. The
proposed algorithm focuses on an optimization of pre-processing results such
as accuracy and precision. To evaluate and validate, the recognition model is
mapped in order to cluster the digital images which are developing the
problem of a multidimensional state space. Simulation results show that
compared to existing algorithms, the proposed method outperforms with the
optimal amount of precision and accuracy in classification as well as results
higher matching percentage based upon image analytics.
Keywords:
Classification
Image embedder
Recognition model
Regression-based approach
Simulation
Copyright © 2019 Institute of Advanced Engineering and Science.
All rights reserved.
Corresponding Author:
Chanintorn Jittawiriyanukoon,
Assumption University,
Samut Prakan, Thailand.
Email: chanintornjtt@au.edu
1. INTRODUCTION
An image does not contain a structured form of digital data but rather an unstructured pattern. In
order to analyze an image and curate such data however the unstructured form of data needs to be converted
to a vector representation accordingly. Deep learning and embedding methods help recover the numeric
pattern from unstructured data. The quality and accuracy of clustering images depend on firstly the accuracy
of pre-processing and secondly the subsequent embedding algorithm [1]. An image is segmented in order to
classify the pixels correctly in a decision-making application. This approach is beneficial in the field of the
complicated image such as pattern recognition, medical image processing, or traffic image. Different
approaches for image segmentation described in [1] opt clustering, training, threshold-based value for the
computation. Clustering techniques include fuzzy c-means (FCM) and K-means method.
Farhang [2] has introduced a new K-means clustering algorithm as the implementation of K-means
algorithm per se is simple. The proposed algorithm is applied to face image mining. Experimental images are
chosen from the database system. The experimental results improve accuracy rate, reduce processing time
due to a reducible number of iterations and smaller cluster distance. However, image classification and
recognition model are not cited. These interesting occurrences need extra investigation.
Xin and Sagan [3] have proposed an image clustering algorithm using multiple agents for
optimizing cluster center. The algorithm based on fuzzy logic maps the image clusters as intelligent agent
moving in the state space. Simulation results support the proposed algorithm compared to existing methods
can optimize a number of cluster centers, the number of categories and the classification accuracy (CA) is
higher. The size of big data state space is reduced by optimizing the collection of center. The paper somehow
needs to additionally consider the processing time.

 ISSN: 2088-8708
Int J Elec & Comp Eng, Vol. 9, No. 2, April 2019 : 1021 - 1027
1022
This paper presents the proposed algorithm for classifying digital images based upon pre-process
and recognition models. Firstly, in this regards, widely used three image classifications are examined. The
proposed algorithm characterized by regression-based pre-processing and recognition models is introduced.
Secondly, the proposed algorithm compared to three traditional approaches is evaluated using the
classification accuracy (CA) and precision metrics. Thirdly, several algorithms used for recognition model
and their evaluation are scrutinized. Lastly, experimental results and analysis are discussed in order to state
the further research investigation.
2. IMAGE CLASSIFICATION
Several approaches have been presented to improve the classification quality of digital images. The
algorithm developed by fuzzy logic to deal with ambiguity in digital images [4], [5]. The classification and
space vector relationship have been inspected based on Markovian models as introduced in [6], [7]. The
hierarchical classification has been also employed for image classification. Artificial Intelligence (AI)
technology has been opted to choose the variables in order to increase the exclusive classification quality.
The neighborhood decision is introduced by cellular network reconfiguration in order to improve judgment
classification quality. In this section, pre-processing approaches are presented. Our proposed method which is
applicable for image classification and the algorithm is discussed.
2.1. Preprocessing approaches
The objective is to ensure classification accuracy (CA), precision and to accelerate the pre-
processing time. K-means, Naïve Bayes and Ada Boost and the proposed classification models are presented
in this section. To excel the post-processing computation, the recognition model as presented in [8] is used.
The MOA simulation [9] results and their performance are evaluated.
2.1.1. K-Means classification (KM)
K-means classification [10] is a kind of unsupervised classifier, which is employed as data is not yet
labeled (i.e., data without groups). The algorithm’s objective is to classify defined K groups for data. The
algorithm calculates repeatedly to allocate each data to one of variable K groups based on data
characteristics. Data is classified based upon the similarity of their characteristics. Closest with K groups (K-
means) employed in classification has different but unique functions which differ from other algorithms. It is
unsupervised which requires no input probability density function. This K-means is a lagging learning
algorithm, which computes data during the testing period, rather than in the learning phase. A benefit of K-
means is that it quickly adapts any alterations. But a drawback is the computational cost due to state space
complexity.
2.1.2. Naïve Bayes classification (NB)
The Naive Bayes Classification [11] based on the Bayesian theory is appropriate for autonomous
input variables. Regardless of its incomplexity and low computational cost, NB can outperform more
advanced classification. NB classifiers can lever a number of independent variables whether classified or
repeated. Given a set of dimensional attribute vector of X = {x1,x2,x3,...,xd}, and the subsequent probability
for the event Cj among all possible outcomes C = {c1,c2,c3,...,cd}, in a conventional language, X is called the
predictors and C is the set of different classes presented in the dependent variable. Assume xd can take
different Cj values, namely, P(Cj/X) > P(Ck/X) for 1 ≤ k ≤ d and k ≠ j. The NB classiﬁer computes a
probability of Cj as following P(Cj/X) = P(X/Cj) P(Cj) / P(X). The values P(X/Cj) and P(X) are estimated
from the training. The NB algorithm is shown in Figure 1.
Algorithm: NB
Require: Data matrix [D]xy with x rows and y columns
for p= 1 to x do
for k = 1 to y do
Construct a frequency table for all characteristics for Cp
Build the prospect table for all characteristics for Cp
Calculate the conditional probability for Cp
Calculate the maximum probability for Cp
end for
end for
Figure 1. NB algorithm

Int J Elec & Comp Eng ISSN: 2088-8708 
Proposed algorithm for image classification using regression-based… (Chanintorn Jittawiriyanukoon)
1023
2.1.3. Ada Boost classification (AB)
Ada Boost (AB) algorithm [12] repairs delicate to a tough learning environment. The weight divides
the data matrix Dxy into 2 parts symmetrically. First tough part of the weight is set to be the perfect classified
part, and the delicate part is allocated to the non-classified part. The Poisson distribution function for
calculating the random probability in order to classify the data model has opted. The idea of AB is to agree
on a series of delicate learners. The weighted variable is designed to a data model which is misclassified in
the previous repetition. Only the present the weighting variable changes according to the AB weight as
proceeding through each iteration of calculation. The approximation moves on with iteratively computing
through the weighted classification until the terminal round. The algorithm is depicted in Figure 2.
Algorithm: AB
Ensure:[D]xy → [D1] and [D2], Q = dimension of [D]
Set:Weight variable is initially set to be wq (=1/Q)
for i = 1 to x do
for j = 1 to y do
for k = 1 to K do
Take Ck (a) after minimizing error of weight variable Ek
Calculate Ek = ∑
( )
( )
Calculate αk = ∑ ( )
( ) ∑ ( )
Calculate βk = loge ( )
Using Poisson distribution function to randomize and update the weight
variable
( )
=
( )
* ( )+
end for
Approximate through final round YK(a) = sgn ∑ ( ) {-1, 0, 1}
end for
end for
Figure 2. AB algorithm
2.2. Proposed algorithm
The proposed algorithm is based upon a logistic regression-based learning environment which
integrates multiple classifications in order to maximize the observed probability figure. At the low level of
computation, there are miscellaneous learning algorithms which are trained independently. This is not similar
to other algorithms which take the sample values that minimize the total of mean squared errors. The
proposed method deals with the grouping of pre-processing techniques for the post-processing of the
outcome at higher learning level. Notice that the original learning environment is not tailored while the
proposed algorithm targets at achievable higher accuracy (as well as precision) in the classification of data.
The proposed model is iteratively trained through the outcomes from the low level of computation. The
proposed algorithm is listed in Figure 3.
Proposed Algorithm
Ensure:[D]xy , Z = number of classifiers, Q = dimension of [D]
for I = 1 to x do
for j = 1 to y do
for k = 1 to P do /** Low level computation **/
Learner Zk applying for data matrix D
end for
for n = 1 to Q do /** Regression-based computation to maximize
probability**/
Dz = {x’n,yn}, in which x’n= z0 + z1 xn + z2 xn+ ... + zP xn
end for
Apply learner Z with Dz /** High level computation **/
Return Z
end for
end for
Figure 3. Proposed algorithm

 ISSN: 2088-8708
1024
3. RECOGNITION MODELS
To search databases of digital images is a provoking job particularly for image data retrieval. Search
engines compute the match between the query image and all images stored in the database and classify the
sequence of images due to their matches. One drawback is that time to sort all matched images requires
O(log d!) in case of d data images. To train to recognize images requires numerous images. This includes
inputting several million images into deep learning level [13]. In between, the process produces
characteristics from the digital image. These characteristics are just the outlines or image extraction. The idea
is that those characteristics can dig deeper down to another layer, for instance, the distance vector from each
class or the density of each characteristic. Once these characteristics are saturated then other recognition jobs
or the image matching requires only machine learning level. It is positive to make use of an algorithm which
has already been trained on lots of images while a collection and a training of several million images is next
to impossible. For instance, ResNet [14] is a deep learning model trained for image recognition in computer
vision. With built-in function, the software can generate a model with a vector representation of
4,096 characteristics between zero and one from input images. The image recognition model based upon
these characteristic vectors is developed without the workload of training at deep learning level onto massive
data images. In this research, the open-source based simulation tool, MOA is used for the image analytics.
Twenty digital images in the database have been chosen. The experiment has been executed on an Asus
Windows 7 with Intel® Core ™ i7 CPU, 2.2 GHz Processor and 8 GB RAM on board. The images have been
chosen in order that they are all different in size, number of attributes, instances, and contents. The
experimental model is given in Figure 4. In order to evaluate the performance of the proposed algorithm for
an improvement of CA and precision is also denoted by the pre-processing results from randomly selected
four images as presented in Table 1.
Figure 4. Experimental model
Table 1. Results from Pre-processing Evaluation
Image 2
Classifier Type CA(%) Precision(%)
KM 52.2 47.2
NB 50.2 74.3
AB 73.9 75.2
Proposed 82.6 84.3
Image 6
KM 41.9 37.2
NB 22.6 38.3
AB 51.6 42.2
Proposed 58.4 62.4
Image 11
KM 53.6 43
NB 17.9 3.2
AB 46.4 35.9
Proposed 60.7 46.9
Image 14
KM 48 50.5
NB 36 36.9
AB 44 67
Proposed 68 84.8

1025
In the recognition model, the speed of matching calculation depends on its P(n) which is the nth
pixel
moves with velocity (v). In Figure 5, one pixel moves at a velocity of v at t. The position of the matching
pixel in the state space at t + 1 can be given by Equation (1).
Figure 5. Matching check of the pixels in the state space
Pn+1(t+1) = Pn(t) + v (1)
Let ῡ be the average velocity and q be the dimension of a digital image then the computational cost
for recognizing an image is O(q). Twenty digital images in the database as listed in Table 2 have been
employed by post-process with different embedding algorithms (GLN, ILSVRC and CPVR) in order to
recognize the matches. A processing for reducing of state space problem and time is demonstrated in [15].
Table 2. Feature of twenty images
Image Dimension (pixels) Size (KB)
1 255x198 9.025
2 275x183 8.142
3 200x231 7.223
4 275x183 5.906
5 400x325 80.420
6 400x294 76.870
7 259x194 9.984
8 290x178 12.959
9 400x226 88.894
10 400x326 117.366
11 400x317 103.177
12 1024x849 726.251
13 335x151 8.908
14 400x286 102.382
15 318x159 10.995
16 400x261 96.226
17 384x247 17.360
18 400x269 41.717
19 225x225 6.089
20 229x220 10.135
4. RESULTS AND ANALYSIS
The computation of the image matching based on the input parameters from pre-processing
algorithms and their final results including match percentage and acceptance range of high (hi: above 51),
medium (med: between 25 and 50) and low (lo: lower than 24). Image number one is used as a reference
image in the recognition model and in finding for another two matches. The retrieval similarities (matching
percentage) for the GLN embedding algorithms with the four pre-processing methods are summarised in
Table 3. As seen the matching percentage for the proposed method is higher than others, but a few percentage
points. The difference in this percentage helps reduce a clarification in decision making for image similarity.
Results from the ILSVRC-2014 algorithm are also seen in the identical direction. The results obtained by the
proposed method are higher in matching percentage. A more thorough evaluation using CPVR-2015
algorithm confirms the results obtained from the proposed method still outperform. The performances
depicted in Table 3 are similar to the corresponding results in Table 1 demonstrating that there is no
degradation in the retrieval performance when the pre-process is chosen in order to classify images from the
database. Not to mention for all three different investigations, the proposed method results in a mid-range-
acceptance level while others give only low-range-level.

 ISSN: 2088-8708
1026
Table 3. Results of twenty images matching
Preprocess Embedding Image No. Matching (%) Acceptance
KM
GLN
19
2
13.49
18.22
Lo
Lo
NB
10
19
13.8
17.06
Lo
Lo
AB
15
19
23.32
26.85
Lo
Med
Proposed
20
2
28.9
29.45
Med
Med
KM
ILSVRC-2014
15
16
14.5
23.7
Lo
Lo
NB
15
16
15.51
18.07
Lo
Lo
AB
11
16
18.07
23.5
Lo
Lo
Proposed
5
9
27.0
27.62
Med
Med
KM
CPVR-2015
20
15
14.94
16.6
Lo
Lo
NB
2
9
15.51
16.47
Lo
Lo
AB
2
4
18.31
19.96
Lo
Lo
Proposed
5
9
27.04
30.1
Med
Med
5. CONCLUSIONS AND FUTURE WORK
The results obtained with the proposed algorithm demonstrate that pre-process increases the
classification accuracy and precision without sacrificing the amount of required matching computation. The
proposed technique can be used for a scalable digital image from large databases. The proposed algorithm
outperforms other three algorithms in the pre-processing phase, even though only marginally in all cases. In
the post-processing phase, three algorithms namely (GLN, ILSVRC and CPVR) are applied for recognizing
image similarity. In the post-processing phase, results from proposed method also marginally improve the
matching percentage. More sophisticated similarity measures [16] which have been used in the video stream
are being currently investigated and these results will be presented in the near future. Another direction of
future work is to identify a benchmark against which the different matching ranges can be set. The execution
time in each phase will be taken into account as well.
REFERENCES
[1] N. Dhanachandra, Y. J. Chanu and K. Manglem, “Image Segmentation Using K -means Clustering Algorithm and
Subtractive Clustering Algorithm,” Procedia Computer Science, vol. 54, pp. 764-771, 2015.
[2] Y. Farhang, “Face Extraction from Image based on K-Means Clustering Algorithms,” International Journal of
Advanced Computer Science and Applications (IJACSA), vol. 8, no. 9, pp. 96-107, 2017.
[3] P. Xin and H. Sagan, “Digital Image Clustering Algorithm based on Multi-agent Center Optimization,” Journal of
Digital Information Management (IJDIM), vol. 14, no. 1, pp. 8-14, 2016.
[4] X. Zhao, Y. Li and Q. Zhao, “Mahalanobis distance based on fuzzy clustering algorithm for image segmentation,”
Digital Signal Processing, vol. 43, no. 1, pp. 8-16, 2015.
[5] N. S. Mishra, S. Ghosh and A. Ghosh, “Fuzzy clustering algorithms incorporating local information for change
detection in remotely sensed images,” Applied Soft Computing, vol. 12, no. 8, pp. 2683-2692, 2012.
[6] B. N. Subudhi, F. Bovolo, A. Ghosh and L. Bruzzone, “Spatio-contextual fuzzy clustering with Markov random
field model for change detection in remotely sensed images,” Optics and Laser Technology, vol. 57, no. 1,
pp. 284-292, 2014.
[7] C. Benedek, M. Shadaydeh, Z. Kato, T. Szirányi and Z. Zerubia, “Multilayer Markov Random Field models for
change detection in optical remote sensing images,” Journal of Photogrammetry and Remote Sensing, vol. 107,
no. 1, pp. 22-37, 2015.
[8] S.K. Dash and M. Panda, “Image Classification using Data Mining Techniques,” Advances in Computer Science
and Information Technology (ACSIT), vol. 3, no. 3, pp. 157-162, 2016.
[9] A. Bifet, R. Kirkby, G. Holmes and B. Pfahringer, “MOA: Massive Online Analysis,” Journal of Machine
Learning Research, vol. 11, pp. 1601-1604, 2010.
[10] R. Thirumahal and P. A. Deepali, “KNN and ARL Based Imputation to Estimate Missing Values,” Indonesian
Journal of Electrical Engineering and Informatics, vol. 2, pp. 119-124, 2014.

1027
[11] H. Y. Mussa, J. B. Mitchell and R. C. Glen, “Full Laplacianised Posterior Naive Bayesian Algorithm,” Journal of
Cheminformatics, pp. 1-6, 2013.
[12] W. Hu, W. Hu and S. Maybank, “Ada Boost-Based Algorithm for Network Intrusion Detection,” IEEE
Transactions on Systems, Man and Cybernetics, vol. 38, no. 2, pp. 577-582, 2008.
[13] A. Ucar, Y. Demir and C. Guzelis, “Object recognition and detection with deep learning for autonomous driving
applications,” Transactions of the Society for Modelling and Simulation International, vol. 93, no. 9, pp. 759-769,
2017.
[14] K. He, X. Zhang, S. Ren and J. Sun, “Deep Residual Learning for Image Recognition,” Computer Vision and
Pattern Recognition, pp. 1-12, 2015.
[15] S. Krishnamurthy and R. Tzoneva, “Decomposition-Coordinating Method for Parallel Solution of a Multi Area
Combined Economic Emission Dispatch Problem,” International Journal of Electrical and Computer Engineering,
vol. 6, no. 5, pp. 2048-2063, 2016.
[16] S. F. C. Haviana and M. Taufik, “Comparison of Various Similarity Measures for Average Image Hash in Mobile
Phone Application,” Proceeding of the 3rd International Conference on Electrical Engineering, Computer Science
and Informatics (EECSI), vol. 3, pp.1-4, 2016.

Proposed algorithm for image classification using regression-based pre-processing and recognition models

More Related Content

What's hot (20)

Similar to Proposed algorithm for image classification using regression-based pre-processing and recognition models (20)

More from IJECEIAES (20)

Recently uploaded (20)

Proposed algorithm for image classification using regression-based pre-processing and recognition models