Volume 9, Issue 1, January 2024 International Journal of Innovative Science and Research Technology

ISSN No:-2456-2165

Performance Evaluation of Selected

Feature Extraction Techniques in
Digital Face Image Processing
Ismaila Folasade. M. Afolabi O. Adeolu
Computer Science Department Department of Cybersecurity Science,
Osun State Polytechnic, Iree, Nigeria Ladoke Akintola University of Tech Ogbomoso, Nigeria

Ismaila W. Oladimeji, Alo Oluwaseun O.

Computer Science Department, Department of Information System,
Ladoke Akintola University of Technology, Ladoke Akintola University of Tech,
Ogbomoso, Nigeria Ogbomoso, Nigeria

Abstract:- Digital image processing is the use of Feature extraction is related to dimensionality
computer algorithms to analyze digital images. Digital reduction (also named a feature vector) [40], [3]. Feature
image processing, involves many processing stages of extraction method can be either holistic or local. The holistic
which feature extraction stage is important. Feature feature extraction techniques are principle component
extraction involves reducing the number of resources analysis [39], [33], linear discriminate analysis [28], [29],
required to describe a large set of data. However, independent component analyse (ICA) [37] are used to
choosing a feature extraction techniques is a problem extracted the features of the whole facial images or sub-
because of their deficiencies. Thus, this paper presents a images. Although considerably good recognition accuracy
comparative performance analysis of selected feature can be obtained using holistic-based techniques but the
extraction techniques in human face images. 90 face limitation is that they cannot control illumination and local
images were acquired with three different poses viz: variations are difficult to describe. Using holistic
normal, angry and laughing. The face images were first algorithms, the global structure of the image is pin-pointed
pre-processed and then subjected to selected feature and the correlation between the images is computed. Many
extraction techniques (Local binary pattern, Principal scholars have employed local-based feature extraction
component analysis, Gabor filter and Linear techniques to overcome the drawbacks of holistic-based
discriminant analysis). The extracted features were then feature extraction techniques. The most local feature
classified using Backpropagation neural network. The methods that are used in face recognition systems are: local
results of recognition accuracy produced by Gabor filter, binary patterns (LBP) method [4], discrete wavelet
PCA, LDA and LBP at 0.76 threshold are 76.7%, 72.2%. transform (DWT) method [24] and Gabor filter [27], [20].
78.9% and 85.6%. Hence, it can be deduced that LBP
performed the best among the four selected feature However all these techniques have pros and cons. LBP
extraction techniques. is an appearance-based method which produce relatively
good results even at variability of the images collection.
Keywords:- Digital Image Processing, Feature Extraction, LBP is relatively insensitive to illumination changes, and its
Local Binary Pattern, Principal Component Analysis, computational simplicity. However, It is sensitive to noisy
Gabor Filter, Linear Discriminant Analysis. pixels wherein the value of the pixels can be easily affected
by the erroneous surrounding pixels. The PCA technique18
I. INTRODUCTION converts each two dimensional image into a one
dimensional vector. This vector is then decomposed into
Digital image processing is a field that explains the use orthogonal (uncorrelated) principle components (known as
of computer algorithms to analyze digital images. The action eigenfaces). PCA improves the performance of the machine
of retrieving an image from source, usually a hardware- learning algorithm as it eliminates correlated variables that
based source for processing (webcam or camera). Several don't contribute in any decision making.; helps in
images have been subjected to digital classifications like overcoming data over fitting issues by decreasing the
trauma-related data [15]; hypertensive data [41]; cancer data number of features; and produces results in high variance
[36]; [39] used Baboon, House, Lena and Peppers; [41] which improves visualization. However, PCA produces
plant leaf diseases; [19] concrete surface images; [46] breast low interpretability of principal components [2]. Gabor
cancer, facial images [9] etc. Digital image processing filtering is a popular approach to extract features from the
involves four stages viz; image acquisition, pre-processing, face images with different orientations by convolving the
feature extraction and classification. [12], [43]. image with the Gabor wavelet basis function. However,
Every image produces a large number of Gabor filtered
images depending upon the number of scale and orientation

IJISRT24JAN1306 www.ijisrt.com 2292

Volume 9, Issue 1, January 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
[25]. The most prominent shortcomings of Gabor filter are developed system produced 88.46% accuracy. [14] proposed
the orientation selectivity and spatial locality. LDA is a local binary pattern variant based on gray level and
dimensionality reduction and classification technique structural information (CLBP_GLSI). Firstly, the authors
commonly used in machine learning and pattern recognition. proposed an structural texture operator called gray level and
In the context of classification it aims to find a linear structural information (GLSI), which adopts the average
combination of features that best separates different classes gray level of image to make image converted into binary
or categories of data. It seeks to reduce the dimensionality of images, and binary images are encoded as border or interior
the feature space while preserving as much of the class pixels image by Border/Interior Pixel Classification (BIC).
separability information as possible. Hence the aim of this Secondly, by combining with CLBP_M, CLBP_S and GLSI
work is to compare the performances of some selected into joint or hybrid distributions, the CLBP_GLSI are
feature extraction techniques on image classification system. obtained. Experimental results obtained from two databases
show that the best recognition accuracy for
II. RELATED WORKS CLBP_S/M/GLSI is 95.95%, while the best results for
CLBP_S/M/C is 95.47%.
Reference [44] designed a face recognition system
using gabor filter based elastic bunch graph matching Reference [23] presented a novel two-layer back
technique and the result is 94.29% accuracy for seventy propagation based neural network algorithm that is cost
input images. However, there is degradation in recognition effective, to classify the normal, malignant and benign
rate due to increased in number of images.. The other tumor dataset.. The acquired dataset were preprocessed to
drawback is that it produces poor recognition accuracy if the remove irrelevant features. Then, backpropagation neural
lighting conditions arenot constant during the enrolment and network algorithm was employed to train the two-layered
verification process [21]. Reference [9] developed an network. The results of correct classification rate gave by
effective and real time face recognition system. The system two layer neural networks and single layered
was tested on YALE Face database B and ORL Face backpropagation neural network are 97.12% and 96.63%
Database. The acquired facial database were extracted with respectively. [32], in their research, applied
PCA. The correct recognition rate achieved using the Backpropagation Neural Network to classify books using
Mahalanobis distance is 92.3% in comparison to the 73.1% its cover. The data was in form of scanned images with three
for the normal PCA with euclidean distance. [17] sizes and image resolution. The features extraction was done
investigated the effectiveness of combining Independent using maximally stable externally egions to identify the
Component Analysis (ICA) with Gabor algorithm (called I- region of book heading, and optical character recognition to
Gabor) as feature extractor from eye and nose of known detect the heading. Then, the features extracted were
faces. The resultant features matrices were trained with transformed into a numerical matrix and then passed as
support vector machine for classification and performances input to the BPNN. The accuracies obtained using one
were evaluated with false acceptance rate, false rejection hidden layer and 2 hidden layers were 63.31% and 79.89%
rate and accuracy. respectively.

Reference [25] performed feature extraction on face III. MATERIALS

images by regulating the scale and orientation parameters of
Gabor Filters. In order to select significant features only, The section entailed the techniques to be evaluated and
different properties of the features like entropy, correlation compared in the work. It also revealed the dataset to be used
coefficient and variance are analysed. Removal of irrelevant for implementing the techniques.
features, effectively by Gabor filter reduced dimensionality
of the feature space without sacrificing accuracy which is A. Dataset
94.5%. [35] developed an approach to solve this limitation In the literature, many data have been used like
using salient distance features, which are obtained by trauma-related data, hypertensive data, cancer data, Baboon,
extracting patch-based 3D Gabor features, selecting the House, Lena, Peppers, plant leaf diseases, concrete surface
salient patches, and performing patch matching operations. images, breast cancer, facial images etc. however, human
The simulation results produced high accuracy, significant faces were used for experimenting the techniques.
performance improvements due to the consideration of
facial element. B. LBP method
LBP technique was founded by [30]. This method
Reference [1] proposed a novel approach in classifying makes it possible to explain the shape and texture of a
gesture into single or double handed subcategories. The digital image. This is executed by splitting the image into
approach employed morphological operation, filtered binary several small regions from which the features are extracted
images to extract the geometric features. The obtained (figure 1).
values were used to classify the single or double handed
gestures. Histogram of Oriented Gradients techniques are
used to extract local orientations and intensity distribution
for detecting the shapes of the object. Classification was
done by applying K- Nearest Neighboring algorithm on the
geometric features extracted.. The result showed that the

IJISRT24JAN1306 www.ijisrt.com 2293

Volume 9, Issue 1, January 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
this problem, the solution is to calculate eigenvectors from
the covariance matrix with reduced dimensionality. The
eigenvectors will be sorted according to their corresponding
eigenvalues from high to low. Then, the eigenvectors
corresponding to zero eigen values are discarded while those
associated with non-zero eigen values are kept. PCA
improves the performance of the machine learning algorithm
Fig. 1: The Original LBP Operator [16][30] as it eliminates correlated variables that don't contribute in
any decision making.
This operator works with the eight neighbors of a
pixel, using the value of this center pixel as a threshold. If a E. Linear Discriminant Analysis
neighbor pixel has a higher gray value than the center pixel LDA is one of the supervised approach to Dimension
(or the same gray value) than a one is assigned to that pixel, reduction. The LDA technique is developed to transform the
else it gets a zero. The LBP code for the center pixel is then features into a lower dimensional space, which maximizes
produced by joining the eight ones or zeros to a binary code. the ratio of the between-class variance to the within-class
However, LBP has advantage of relatively insensitive to variance, thereby guaranteeing maximum class separability
illumination changes wherein the value of the pixels can be [31]. There are two types of LDA technique to deal with
easily affected by the erroneous surrounding pixels and also classes: class-dependent and class-independent. In the class-
it’s computational simplicity [16]. dependent LDA, one separate lower dimensional space is
calculated for each class to project its data on it whereas, in
C. Gabor filter the class independent LDA, each class will be considered as
Hungarian-born electrical engineer Dennis Gabor a separate class against the other classes [6], [47]. In this
developed a Gabor wavelet in 1946. It was created from one type, there is just one lower dimensional space for all classes
particular atom by dilation and rotation in a two- to project their data on it. Although the LDA technique is
dimensional case and provides a complete image considered the most well-used data reduction techniques, it
representation. Recently, Gabor functions are frequently suffers from a number of problems. In the first problem,
used for classification, segmentation, or edge detection) and, LDA fails to find the lower dimensional space if the
more practically, pattern recognition. An image processing dimensions are much higher than the number of samples in
task can be seen in the form of a wavelet transform. The the data matrix. Thus, the within-class matrix becomes
scale of a wavelet provides a magnification that allows one singular, which is known as the small sample problem. This
to see the image through the lens. Gabor functions are type of feature selection techniques are applicable
particularly useful in texture-based image analysis due to biometrics [22], [8]; Bioinformatics [45] and chemistry [7].
their ability to capture both spatial and frequency
information. By applying Gabor filters at different scales F. Performance Metrics
and orientations, the wavelet transform can effectively In [16], it was stated that the performance metrics to
extract features from an image, enabling tasks such as evaluated biometric systems include recognition accuracy,
classification, segmentation, edge detection, and pattern false match rate, false non-match rate, Specificity,
recognition. This approach allows for a comprehensive Sensitivity and receiver operating characteristic (ROC).
representation of the image, revealing intricate details and Confusion matrix was employed to estimate the values of
enhancing its visual interpretation [44]. each of the metrics.
𝑇𝑟𝑢𝑒 𝑁𝑒𝑔𝑎𝑡𝑖𝑣𝑒+𝑇𝑟𝑢𝑒 𝑝𝑜𝑠𝑖𝑡𝑖𝑣𝑒
D. PCA 𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 𝑖𝑠 𝑒𝑥𝑝𝑟𝑒𝑠𝑠𝑒𝑑 𝑎𝑠
𝑇𝑜𝑡𝑎𝑙 𝑁𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝐼𝑚𝑎𝑔𝑒𝑠
PCA is a statistical approach used for reducing the
number of variables in image classification. In PCA, every 𝐹𝑎𝑙𝑠𝑒 𝑝𝑜𝑠𝑖𝑡𝑖𝑣𝑒
image in the training set is represented as a linear 𝐹𝑎𝑙𝑠𝑒 𝑀𝑎𝑡𝑐ℎ 𝑅𝑎𝑡𝑒 𝑖𝑠 𝑒𝑠𝑡𝑖𝑚𝑎𝑡𝑒𝑑 𝑎𝑠
𝑇𝑟𝑢𝑒 𝑛𝑒𝑔𝑎𝑡𝑖𝑣𝑒+𝐹𝑎𝑙𝑠𝑒 𝑝𝑜𝑠𝑖𝑡𝑖𝑣𝑒
combination of weighted eigenvectors. These eigenvectors
are obtained from covariance matrix of a training image set. 𝐹𝑎𝑙𝑠𝑒 𝑁𝑜𝑛𝑚𝑎𝑡𝑐ℎ 𝑅𝑎𝑡𝑒 𝑖𝑠
𝐹𝑎𝑙𝑠𝑒 𝑁𝑒𝑔𝑎𝑡𝑖𝑣𝑒
The weights are found out after selecting a set of most 𝑇𝑟𝑢𝑒 𝑃𝑜𝑠𝑖𝑡𝑖𝑣𝑒+𝐹𝑎𝑙𝑠𝑒 𝑁𝑒𝑔𝑎𝑡𝑖𝑣𝑒
relevant eigenvectors [2]. PCA eigenvector method
True Negative
considers each pixel in an image as a separate dimension, Specificity is termed
True Negative+False Positive
that is, N by N image has N2 pixels or N2 dimensions. To
calculate eigenvector, there is a need to calculate the
covariance matric C as shown in Equation 1. True Positive
Recall is represented by
True Positive+False Negative
C = A.AT where A = N2 by M (1)
IV. METHODOLOGY
where N is the dimension of the image, M is the
number of column vector. If eigenvector will be calculated The work flow of the image classification system is
from a covariance matrix before dimension reduction, the shown in figure 2 and explained in the subsections.
system would slow down terribly or run the system out of
memory, due to huge computations. In order to overcome

IJISRT24JAN1306 www.ijisrt.com 2294

Volume 9, Issue 1, January 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
A. Data Acquisition The face poses of the 30 subjects comprises of three poses
Faces of thirty subjects (sample shown in figure 3) viz: normal pose, angry pose and laughing pose which made
were acquired with a good sony camera of high resolution. a total of 90 faces stored in the database.

Fig. 3: Sample of face subjects

 Pre-processing Stage the accuracy of systems. In this work, it involves image

This section entails the steps to be adopted in preparing resizing (figure 4a), pose orientation (figure 4b), and colour
the subject images for face recognition system to increase transformation (figure 4c).

Face database

Pre-processing Steps
Image Resize

Pose Orientation

Colour Transformation

LBP Gabor LDA PCA

Feature extraction stage

Classification

yes < threshold no

Fig. 2: Flow diagram of the Image Classification System

(a) (b) (c)

Fig. 4: Sample of pre processing outputs (a) Image resize (b) pose orientation (c) Colour transformation

IJISRT24JAN1306 www.ijisrt.com 2295

Volume 9, Issue 1, January 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
B. Feature Extraction max Current_Img(x, y then
This stage involves the techniques used for dimension (x,y)ϵBlock
reduction of preprocessed images. If Current_Tmg (x, y )>
𝟏
∑𝒓𝒐𝒘 ∑𝒄𝒐𝒍 𝐶𝑢𝑟𝑟𝑒𝑛𝑡_𝑇𝑚𝑔 (𝑥, 𝑦 )𝐭𝐡𝐞𝐧
𝒓𝒐𝒘.𝒄𝒐𝒍 𝒙=𝟏 𝒚=𝟏
 Technique One: LBP algorithm is shown in Algorithm 1.
While Filter_No ≤ Total_Filter do
Algorithm 1: The LBP algorithm [30]
VALUE (Filter_No) := GABOR_M AG{I, j}(x, y)
 Step 1: Set 𝒈𝒄 which corresponds to the gray value of
end
the center pixel.
end
 Step 2: Set 𝒈𝒑 as the gray values of the “n” neighbour end
pixels. FEATURE_VECTOR=Value;
 Step 3: Set end
𝟏, 𝒊𝒇𝒈𝒄 ≥ 𝟎 end
𝑴={ } (2)
𝟎, 𝒊𝒇𝒈𝒄 < 0
 Step 4: Compute LBP features as described thus;  Technique 4: LDA algorithm is shown in algorithm 4
𝑳𝑩𝑷𝒑,𝒓 (𝒙𝒄 , 𝒚𝒄 ) = ∑𝒏−𝟏
𝒑=𝟎 𝑴(𝒈𝒑 − 𝒈𝒄 ) ∗ 𝟐
𝒑
(3) Algorithm 4. : LDA. [13]
 Read the training images (X = {x1;x2…..xN}), where xi(ro
Where 𝒙𝒄 and 𝒚𝒄 represent the horizontal and vertical x co) represents the ith training image, ro and co
component of the image; 𝑴𝒈𝒑 and 𝑴𝒈𝒄 are neighborhood represent the rows (height) and columns (width) of xi,
patterns, P represent the bit binary number resulting in respectively, N represents the total number of training
𝟐𝒑 distinct values for the LBP code. images.

 Step 5: Output selected LBP features  Convert all images in vector representation Z = {z1; z2…..
zN}, where the dimension of Z is M x_1;M = ro x_co.
 Technique Two: PCA algorithm is shown in algorithm  Calculate the mean of each class μi, total mean of all data
2. μ, between-class matrix SB(M _M), and within-class
Algorithm 2: PCA [18] matrix SW(M_M) of X
Following are steps involve;  Find the k eigenvectors of SB with nonzero eigenvalues,
 Step 1: Column or row vector of size N2 represents the and denote them as U =[u1;u2……uk}(i.e. UT SbU > 0).
set of M images (B1, B2, B3…BM) with size N*N (9)
 Step 2: The training set image average (μ) is described as  Calculate the eigenvalues and eigenvectors of UT SWU,
1 and then sort the eigenvalues and discard the
𝜇 = 𝑚 ∑𝑀 𝑛=1 𝐵𝑛 (4)
eigenvectors which have high values. The selected
 Step 3: the average image by vector (W) is different for eigenvectors are denoted by V; thus, V represents the null
each trainee image space of SW.
Wi = Bi - μ (5)  The final LDA matrix (ᴪ) consists of the range SB and
 Step 4: Total Scatter Matrix or Covariance Matrix is the null space of SW as follows, ᴪ =UV.
calculated from Φ as shown below:  The original data are projected on the LDA space as
𝐶 = ∑𝑀 𝑛=1 𝑤𝑛 𝑤𝑛 𝑡 = 𝐴𝐴𝑇 (6)
follows,
where A= [W1W2W3…Wn] Y = X ᴪ = XUV (10)
 Step 5: Measure the eigenvectors UL and eigenvalues λL
of the covariance matrix C. C. Image Matching
 Step6: For image classification, this feature space can be The image classification was done by Back-
utilized. Measure the vectors of weights propagation neural network. The algorithm of BNN is given
below. The back propagation algorithm is implemented
ΩT = [w1, w2, …, wM'], (7)
using following steps.
 Initialize weights to small random values Wih ,Whj.
whereby,  Feed input vector X1,X2....through Network and
Hk = UkT (B - μ), k = 1, 2, …, M' (8) computing weighting sum coming into node and then
apply the sigmoid function
 Technique 3: Gabor Filters algorithm is shown in
algorithm 3
Algorithm 3: Gabor filter [43] (11)
Gabor Filters: Algorithm 3: Algorithm to Extract Feature
Vector Where Wih is weights between input nodes and hidden
Data: GABOR_MAG{No_of_Scale, No_of_Orient} nodes,
Result: FEATURE_VECTOR
Block_Size :=32 x 32, Total_Filter := 40; Sh is weighting sum coming into hidden node
While i ≤ No_of_Scale do
While i ≤ No_of_Orient do
Current_Tmg := GABOR_M AG{i, j} if (12)
GABOR_M AG{I, j}(x,y) =

IJISRT24JAN1306 www.ijisrt.com 2296

Volume 9, Issue 1, January 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
Where Yh is probability of each hidden node for each Repeated all the steps except 1 till MSE is within
pixel reasonable limits.
(13)
7. After training neural network using Training pixels, find
Yh and Yj for each pixel using weights Whj, Wih which is
Where Whj is weights between hidden nodes and
obtain from training of neural network.
output nodes,
8. Pixel goes in Yj class if Yj have maximum probability for
Sj is weighting sum coming into output node
this pixel. According to this all pixels of image are
classified.
(14)
 Calculate error term for each output unit V. DISCUSSION OF RESULTS
(15)
 Mean square error(MSE) of output node The simulation tool used was MATLAB R2012a
version on Windows 7. The dataset contains 90 Face
images consisting of 30 normal face, 30 angry face and 30
(16)
laughing face. Sixty of the face images were stored and are
 Calculate error term of each of hidden node
labeled known faces while remaining 30 were not stored
(17) which are labeled as unknown faces. These face images
 Adjust weights to minimize mean square error were trained and tested using 50-50 cross-validation
method. The image classification system was executed with
(18) thresholds of 0.22, 0.35, 0.5 and 0.76. The results of the
execution are shown in Table 1, Table 2, Table 3 and Table
(19) 4 with the performance metrics measured in percentages.

Table 1: Results of LBP-BPNN system Table 2: Results of PCA-BPNN system

Thresholds RECALL SPEC FMR FNMR ACC Thresholds RECALL SPEC FMR FNMR ACC
0.22 71.7 60.0 40.0 28.3 67.8 0.22 65.0 50.0 50.0 35.0 60.0
0.35 76.7 70.0 30.0 23.3 74.4 0.35 66.7 56.7 43.3 33.3 63.3
0.5 85.0 73.3 26.7 15.0 81.1 0.5 71.7 60.0 40.0 28.3 67.8
0.76 88.3 80.0 20.0 11.7 85.6 0.76 75.0 66.7 33.3 25.0 72.2

Table 3: Results of Gabor-BPNN system Table 4: Results of LDA-BPNN system

Threshold RECALL SPEC FMR FNMR ACC Threshold RECALL SPEC FMR FNMR ACC
0.22 66.7 53.3 46.7 33.3 62.2 0.22 68.3 56.7 43.3 31.7 64.4
0.35 68.3 56.7 43.3 31.7 64.4 0.35 71.7 60.0 40.0 28.3 67.8
0.5 73.3 63.3 36.7 26.7 70.0 0.5 76.7 66.7 33.3 23.3 73.3
0.76 78.3 73.3 26.7 21.7 76.7 0.76 80.0 76.7 23.3 20.0 78.9

The results obtained from the tables showed that obtained for in Table 1 for LBP at 0.76 threshold are REC of
maximum values were got at threshold of 0.76. The results 88.3%, SPEC of 80.0%, FMR of 20.0%, FNMR of 11.7%
obtained for in Table 3 Gabor filter at 0.76 threshold are and ACC of 85.6%.
Recall of 78.3% , SPEC of 73.3%, FMR of 26.7%, FNMR
of 21.7% and ACC of 76.7%. The results obtained for in Hence, it can be deduced that LBP performed the best
Table 2 for PCA at 0.76 threshold are REC of 75.0%, SPEC among the four selected feature extraction techniques. The
of 66.7%, FMR of 33.3%, FNMR of 25.0% and ACC of ROC graphs of Specificity vs thresholds, Recall vs
72.2%. The results obtained for in Table 4 for LDA at 0.76 thresholds, and Accuracy vs thresholds are shown in figure
threshold are REC of 80.0%, SPEC of 76.7%, FMR of 6, 7 and 8 respectively.
23.3%, FNMR of 20.0% and ACC of 78.9%. The results

IJISRT24JAN1306 www.ijisrt.com 2297

Volume 9, Issue 1, January 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165

Fig. 6: ROC graph of Specificity vs Thresholds

Fig. 7: ROC graph of Recall vs Thresholds

Fig. 8: ROC graph of Accuracy vs Thresholds

IJISRT24JAN1306 www.ijisrt.com 2298

Volume 9, Issue 1, January 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
VI. CONCLUSION [9]. R. Deepesh (2011). A Realtime Face Recognition
system using PCA and various Distance Classifiers,
This paper presented an evaluation of performances of Computer vision and processing. 5(3), 15-22.
four selected feature extraction techniques viz LBP, PCA, [10]. F. Ehsan et al., (2004). "Robust Face Recognition
Gabor filter and LDA). These techniques were selected through Local GraphMatching," Journal of
based on their notable performances in image classification Multimedia, vol. 2, 31-37.
problems. The evaluation procedure employed four stages [11]. H. R. Ekenel and R. Stiefelhagen (2006) "Analysis of
which are image collection, pre-processing, feature Local Appearance-BasedFace Recognition: Effects of
extraction and images matching. The face images were first Feature Selection and Feature Normalization,"in
pre-processed and then subjected to selected feature Conference on Computer Vision and Pattern
extraction techniques (LBP, PCA, Gabor filter and LDA). Recognition Workshop,CVPRW,, 34-34.
The extracted features were then matched with BPNN. The [12]. P. Franti, (2001). Digital Image Processing, Univ. of
results of recognition accuracy produced by Gabor filter, Joensuu, Dept. of Computer Science, Lecture notes
PCA, LDA and LBP at 0.76 threshold are 76.7%, 72.2%. [13]. T. Feng .and G. Wu. (2014) A theoretical contribution
78.9% and 85.6%. Hence, it can be deduced that LBP to the fast implementation of null linear discriminant
performed the best among the four selected feature analysis method using random matrix multiplication
extraction techniques. However, this work can be improved with scatter matrices. arXiv preprint arXiv:1409.2579.
on by considering other notable feature extraction [14]. T. Huawei, W. Rugang and Li Z. (2016). An
techniques for evaluation and also possibly improve the Enhanced Local Binary Pattern for Texture
efficiency of any of these techniques through cascade or Classification, Advances in Computer Science
optimization methods. Research, volume 58, Modeling, Simulation and
Optimization Technologies and Applications (MSOTA
ACKNOWLEDGEMENT 2016).
[15]. S. Hassanipour, H. Ghaem, M. Arab-Zozani, M. Seif,,
I appreciate TETFUND for her financial support Fararouei, M., Abdzadeh, E., Paydar, S. (2019).
toward the successful completion of this project. Comparison of artificial neural network and logistic
regression models for prediction of outcomes in trauma
REFERENCES patients: A systematic review and meta-analysis.
Injury, 50(2), 244–250.
[1]. S. Akanksha, A. Saloni. S. Pushkar. M. Ankush, http://doi.org/10.1016/j.injury.2019.01.007.
(2015). “Indian Sign Language Gesture Classification [16]. F. M Ismaila. (2022). Hybridized Local Binay Pattern-
as Single or Double Handed Gestures”, International Levyflight Firefly Algorithm based Human face
Conference on Image Information Processing, 2015. Recognition Access Control System, PhD Thesis,
[2]. R. Abbassi, A. Abbassi, A. Heidari., S. Mirjalili, Unpublished, Ladoke Akintola University of
(2019). An efficient salp swarm-inspired Algorithm Technology, Nigeria.
for parameters identification of photovoltaic cell [17]. W. O. Ismaila, A . Adetunji, A. Falohun and
models. Energy Convers Manag 179:362–372. Iwashokun G. B. (2012). A Study of Features
[3]. E. Alpaydin, (2010). Introduction to Machine Extraction Algorithms for Human Face Recognition.
Learning. London: The MIT Press. Transnational Journal of Science and Technology,
[4]. T. Ahonen, et al., (2004)."Face recognition with local 2(6), 14-22.
binary patterns," Lecture Notes in Computer Science [18]. S. Karamizadeh, S, Abdullah, A. Manaf., Mazdak Z.,
(including subseries Lecture Notes in Artificial Hooman A. (2013). An Overview of Principal
Intelligence and Lecture Notes in Bioinformatics), vol. Component Analysis, Journal of Signal and
3021, 469-481. Information Processing, 2013, 4, 173-175
[5]. A. Bansal et al., (2012). "Face Recognition Using [19]. H. E. Kim,. M. Ahn, M. Shin, and S.-H. Sim. (2019).
PCA and LDA Algorithm," in Second International Crack and non crack classification from concrete
Conference on advanced Computing & surface images using machine learning. Structural
Communication Technologies (ACCT), 2012, 251- Health Monitoring, 18(3), 725–738.
254. [20]. C. Ki-Chung., et al., (1999). "Face recognition using
[6]. S. Balakrishnama and A. Ganapathiraju (1998). principal component analysis of Gabor filter
Linear discriminant analysis-a brief tutorial. Institute responses," in International Workshop on Recognition,
for Signal and information Processing. Analysis, and Tracking of Faces and Gestures in Real-
[7]. D. Coomans,, Massart D., and Kaufman. L. (1979). Time Systems, Proceedings, 53-57.
Optimization by statistical linear discriminant analysis [21]. P. Lo (2007). "Dynamic local feature analysis,"
in analytical chemistry. Analytica Chimica Acta, Biometric Technology Today, vol. 15, 8-8.
112(2), 97–122.. [22]. J. Lu,, K. Plataniotis. and Venetsanopoulos. A. N.
[8]. J. Cui and Y.Xu.. (2011). Three dimensional palmprint (2003). Face recognition using lda-based algorithms.
recognition using linear discriminant analysis method. IEEE Transactions on Neural Networks, 14(1):195–
In Proceedings of the second International Conference 200.
on Innovations in Bioinspired Computing and
Applications (IBICA),, 107–111

IJISRT24JAN1306 www.ijisrt.com 2299

Volume 9, Issue 1, January 2024 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
[23]. J. A. Mohamed (2016). Classification Using Two [38]. M. Shen, et al., (2012). "Independent component
Layer Neural Network Back Propagation Algorithm, analysis for face recognition basedon two dimension
Circuits and Systems, 7, 1207-1212. symmetrical image matrix," in 24th Chines Control
[24]. D. Murugan, et al. (2010) "Performance Evaluation of and Decision Conference (CCDC, pp. 4145-4149.
Face Recognition UsingGabor Filter, Log Gabor filter [39]. P. B. Saurabh. and D. S. Chaudhari (2012). "Principal
and Discrete Wavelet Transform,"International Component Analysis for Face Recognition,"
Journal of Computer Science & Information International Journal of Engineering and Advanced
Technology, vol. 2, pp. 125-133. Technology, vol. 1, pp. 91-94.
[25]. A. Majumder, Chowdhury S, Sil J. (2015).Gabor Filter [40]. A.Singh, N. Thakur, and A. Sharma. (2016). A review
Based Face Recognition Using Non-frontal Face of supervised machine learning algorithms. In: 3rd
Images, Conference Paper · August 2015, WCI ’15, International Conference on Computing for
August 10 - 13, 2015, Kochi, India. 419-425. Sustainable Global Development (INDIACom), 1310–
[26]. D. .Murugan et al. (2010) "Performance Evaluation 1315. New Delhi, India.
of Face Recognition UsingGabor Filter, Log Gabor [41]. Sarangi, S, Sahidullah, M,; Saha, G. (2020).
filter and Discrete Wavelet Transform,"International "Optimization of data-driven filterbank for automatic
Journal of Computer Science & Information speaker verification". Digital Signal Processing.
Technology, vol. 2, 125-133. 104. arXiv:2007.10729. doi:10.1016/j.dsp.2020.10279
[27]. S., Muhammad et al (2007)., "Face Recognition using 5.
Gabor Filters," Journal of Applied Computer Science [42]. P.Sharma, P. Hans, and S. C. Gupta. (2020).
& Mathematics, vol. 5, 53-57.. Classification of plant leaf diseases using machine
[28]. J. Mazanec, et al., (2008). "Support vector machines, learning and image preprocessing techniques. In: 10th
PCA and LDA in face recognition," Journal of International Conference on Cloud Computing, Data
electrical engineering -electrotechnickycasopis, vol. Science Engineering (Confluence), 480–484. Noida,
59, pp. 203-209. India.
[29]. J. Mazanec et al., (2008). "Support vector machines, [43]. M. Ture, I. Kurt, Turhan Kürüm, A., and Ozdamar, K.
PCA and LDA in face recognition," Journal of (2005). Comparing classification techniques for
electrical engineering -electrotechnickycasopis, vol. predicting essential hypertension. Expert Systems with
59, pp. 203-209. Applications, 29(3), 583–588.
[30]. A. Martinez . and A. Kak (2001). "PCA versus LDA," http://doi.org/10.1016/j.eswa.2005.04.014.
Pattern Analysis and Machine Intelligence, IEEE [44]. V. Tyagi (2013). Understanding Digital Image
Transactions on, vol. 23, pp. 228-233. Processing, CRC Press Taylor & Francis Group,.
[31]. T. Ojala, M. Pietikainen and Harwood D., (1996). “A [45]. L. Wiskott, et al (1997). "Face recognition by elastic
comparative study of texture measures with bunch graph matching,"Pattern Analysis and Machine
classification based on feature distributions” Pattern ntelligence, IEEE Transactions on, vol. 19, 775-779.
Recognition vol. 29, 1996. [46]. Wu, M. C, Zhang, L., Wang, Z.,. Christiani, D. C and
[32]. F. Pan,, Song, G. Gan, X. and Gu. Q. (2014). Lin. X. (2009). Sparse linear discriminant analysis for
Consistent feature selection and its application to face simultaneous testing for the significance of a gene
recognition. Journal of Intelligent Information set/pathway and gene selection.Bioinformatics,
Systems, 43(2):307–321. 25(9):1145–1151.
[33]. P. B. Purwanta, Adi C. K., Dewi N. P.(2020). [47]. N. Yassin, S.. Omran, E. El Houby, and H. Allam.
Backpropagation Neural Network for Book (2017). Machine learning techniques for breast cancer
Classification Using the Image Cover, International computer aided diagnosis using different image
Journal of Applied Sciences and Smart Technologies, modalities: A systematic review. Computer Methods
2(2), 179–196. and Programs in Biomedicine, 156, 25–45.
[34]. S. M. Prakash. and Kalpna C. J., (2011). "Face [48]. Viszlay, M. Lojka, and Juh´ar J. (2014). . Class-
Recognition Using PCA," International Journal of dependent two dimensional linear discriminant
Artificial Intelligence & Knowledge Discovery, vol. 1, analysis using two-pass recognition strategy. In
pp. 25-28. Proceedings of the 22nd European Signal Processing
[35]. P. J. Phillips., et al., (1998). "The FERET database and Conference (EUSIPCO), 1796–1800. IEEE.
evaluation procedure for face recognitionalgorithms,"
Image and Vision Computing, vol. 16, pp. 295-306.
[36]. C. Ruchi (2015). Facial Expression Recognition Using
Facial Movement Features, Coputer Science &
Engineering, India, vol. 4 Issue 4, 2318 – 2324
[37]. M. Remzi, T. Anagnostou, V. Ravery, Zlotta, A.,
Stephan, C., and Marberger, M. (2003). An artificial
neural network to predict the outcome of repeat
prostate biopsies. Urology, 62(3), 456–460.
http://doi.org/10.1016/s0090-4295(03)00409-6.

IJISRT24JAN1306 www.ijisrt.com 2300