a case study on various preprocessing methods and their impact on face recognition using randomforest

International Journal of Computer Applications (0975 – 8887)
Volume 175 – No.9, October 2017
5
A Case Study on Various Preprocessing Methods and
their Impact on Face Recognition using Random Forest
Hanan M. S. Algharib
Computer Engineering
Ankara Yıldırım Beyazıt university
Hanan.m.algharib@gmail.com
Shafqat Ur Rehman
Computer Engineering
Ankara Yıldırım Beyazıt university
shafqat.rehman@gmail.com
ABSTRACT
The Random forest is a well-known powerful classifier, that
used to classify a wide range of patterns in our daily life for
different purposes, it enters into many fields such as images
and objects classification. In this paper, we studied the impact
of a five-common preprocessing method in face recognition
on the random forest performance, the study included
applying five different pre-processing methods (Single Scale
Retinex, Discreet Cosine Transform, wavelet Denoising,
Gradient faces, and the method proposed by tan and et Known
as pp chain or TT), each one has applied separately with a
general random forest as a classifier, we computed the error
rate for each method. The study was conducted on a face
recognition system under occlusion and illumination variation.
All experiments were done using MATLAB and Extended
Yale B database.
General Terms
Face Recognition, Occlusion, Illumination Variation.
Keywords
Random Forest, Gradient Faces, Wavelet Denoising, Discrete
Cosine Transform, Single Scale Retinex, TT (tan and et).
1. INTRODUCTION
Classification of Patterns is a very important term and an
essential matter in the machine learning, the urgent need for
strong Classifiers leads us to think about the possibility of
searching for irreplaceable classifiers, Testing the strength of
classifiers under different challenges is the focus of our goal.
In this regard, we have conducted an actual study on a face
recognition system includes human faces with frontal pose
under different illumination and occlusion. The preprocessing
phase includes applying five preprocessing separately each
method has its impact on the random forest. The choice of the
best preprocessing method that increase the efficiency of a
random forest is the focus of our work. Unstructured forests
have a very high velocity, especially in the classification task
that require speed such as face recognition at airports and this
observation should be taken into consideration. Random trees
have been applied in many fields and have been adopted as a
strong classifier. Based on the rationale the combination of
learning models increases the classification accuracy which is
called bagging, the main goal of bagging is to average noisy
unbiased models in order to create a model with low
invariance.
2. PREVIOUS WORKS
A Survey of some previous researches, [1] Ghoshal (2009)
have used random forests to classify faces with occlusion, it
also used LBP and Gabor filter as preprocessing. In [2]
presents an algorithm for face recognition using modified
local binary pattern (MLBP) and random forest (RF). In [3]
suggested to detect and recognize human faces using random
forests for mobile robots. In [4] proposed a Gabor wavelet
transform to extract features from a small dataset of images,
then applied the Random Forest algorithm for classification
based on the regression technique. In [5] compute the
performance of Random Forest (RF) and Support Vector
Machine (SVM) to classify faces. In [6] proposed Principal
Component Analysis (PCA) and Eigenface to extract features,
then applied classifiers K-Nearest Neighbors (K-NN),
Random Forest and K-Star for recognition. In [7] suggested
combination of Principal Component Analysis (PCA),
Discrete Cosine Transform (DCT), Discrete Wavelet
Transform (DWT) as a feature extractors and random forest as
classifier. In [8] proposed algorithm to image fusion between
the visible and thermal IR face images based on a-torus
wavelet transform and random forest (RF). In [9] face
recognition system was introduced using random forests as a
classifier with Histograms of oriented gradients as feature
representation. In [11] it focuses to enhance illumination
variation using adaptive multiscale retinex as a preprocessing
to overcome the lighting condition effects. In [11] proposed to
combine two methods to image enhancement Based on
variants of Retinex algorithms and image epitome. In [12]
applied the retinex model on color constancy and lightness
and test in an image to study the Performance of a
Center/Surround Retinex and its Properties. [13] In this paper
14 normalization algorithm were applied on a face verification
system as preprocessing and compared under the illumination
condition through 3 verification methods. In [14] a book by
Gonzales and earth, explained the Retinex algorithm and
discuss its work and advantages from different wide aspects.
While in [15] presented a new approach based on discrete
cosine transform (DCT) used for image normalization
illumination under different lighting conditions. In [16]
Robust approach for illumination preprocessing based on
Discrete cosine transform (DCT) applied on sample of images
with proper scaling of low frequency DCT coefficients. [17]
This paper proposed algorithm to extract local information
depending on block-based discrete cosine transform. [18] it
used DCT as a feature extractor from the images that suffer
from poor illumination. In [19] (DCT) is employed to address
the human face images in difficult lighting conditions. In
Reference [21] a paper, explain a method of a wavelet-based
illumination normalization for human face recognition under
different directions and strength of light. In [21] The human
face images are preprocessed and normalized by a group of
combination of preprocessing normalization method (DCT-
WD-Gradient face-LCE-Weber faces) to eliminate the
negative impacts of illumination. In [22] book discusses in

6
detail Photometric Normalization Techniques for inconstant
Illumination Examines the impacts of preprocessing (WD-
DCT-SSR) and other methods on the images that Distorted
images by illumination. In [23] this paper addressed the
problem of illumination based on method called
“Gradientfaces” The proposed method has the ability to
extract illumination insensitive. [24] comparative study, it
discussed a group of illumination variation preprocessing
(HE, LT, GIC, DGD, LoG, SSR, GHP, SQI, LDCT, LTV, LN
and TT) with different three classes: gradient and edge
extraction, gray-level transformation and face reflection field
estimation. In [25] face image with different illumination
solved using DCT coefficients in logarithm domain. In [26]
Discussed a lot of normalization preprocessing methods for
practical FRS, in local face representations, one of presented
methods is a preprocessing chain which is consist of a series
of sequential processing (Gamma Correction, Difference of
Gaussian (DoG), Filtering Masking, Contrast Equalization),
that eliminates uncontrolled lighting conditions.
2.1 The Aim of study
The main goal is to measure the impact of five suggested
preprocessing methods on the random forest performance.
The preprocessing method that improves the performance of
Preprocessing
SSR-DCT- WD-GF-TT
Feature Extraction
LDA
Classification
Random Forest
Faces under Random
occlusion and
uncontrolled lighting.
random forests will be considered as the best preprocessing
method suitable to apply with the random Forest classifier in
case of occlusion and illumination variation. Random forest
will also be applied without any of these preprocessing
methods to demonstrate their performance when it standalone.
The study will be conducted on a face recognition system
under difficult conditions of illumination variation and
occlusion. The main motive of this paper is to introduce a new
and unique case study of random forest efficiency as a strong
classifier with different preprocessing methods and examine
the impacts. The processing and classification stages are
important stages in any classification system where they are
complementary processes in any recognition system. This
study could provide a Foundation stone in face recognition
system that uses the random forest classifier, it focused on two
Common problems which are occlusion and illumination
variation, also Constitute a basis for information that can be
used as a reference to future studies.
3. THE PROPOSED METHODS
Five different powerful preprocessing methods that are used
to solve illumination variation problem were applied in this
study.
- Single Scale Retinex (SSR)
- Discreet Cosine Transform (DCT)
- Wavelet Denoising (WD)
- Gradient Faces (GF)
- PP Chain (TT)
The following figure illustrates two important stages in the
recognition system and the impact between them. The pre-
processing stage where applied five different pre-processing,
which they are a Collection of powerful pre-processing
methods used in illumination varying for recognition purpose,
and the classification stage where the random forest as
classification algorithm used to make classification decision
based on a number of random decision trees.
Fig 1: Diagram to explain the proposed Techniques
The figure above explains the five suggested preprocessing
methods in the preprocessing stage, and the random forest in
the classification stage. In our face recognition system, we use
the Linear Discriminant Analysis (LDA) as a necessary step
for feature extraction and dimension reduction and to reduce
the huge image dimensions.
3.1 The imaging model and retinex theory
Illumination model or lighting model is a well-known model
in image processing, many of illumination preprocessing
methods were derived from this model, it based on two factors
L (luminance) and R (Reflectance).
I (x,y)= R(x,y)*L(x,y) (1)
L(x,y) is the luminance (amount of lighting ) is a dynamic
(changes slowly with the spatial position of the image I )
Contains a blurred image , represents Low frequency
phenomenon. R(x,y) is the reflectance at each point
(Structure of object surface) illumination invariant or
insensitive measure (Key facial features), the edges in the
image I(x,y) also represent the edges in the reflectance. R(x,y)
represents high frequency phenomenon. Retinex stands for
Retina and cortex, Proposed by Edwin land in 1986 ,it is a
model of lightness and human vision perception. Retinex
improves visual rendering of an image when lighting
conditions are not good. To bridge the gap between images
and human observation of scenes. A lot of illumination
enhancement preprocessing methods developed based on this
theory.
3.1 Single Scale Retinex
Single scale retinex also called as center/surround retinex it is
straightforward method, it is photometric based techniques
derived from imaging model. offered by Edwin Land in 1986
based on his experimentation involving color constancy.
Retinex algorithm, is used to enhance images based on
Retinex Technique, which mainly suggested to be used in
image improving with regard to the brightness, contrast and
sharpness of an image mostly through compression in the

7
MN
dynamic range [11]. The impact of SSR on the image can be
illustrates in the figure (2). SSR were applied on a set of
images belongs to the same person under different
illumination.
Fig 2: The Impact of SSR
Based on the imaging model in eq (1) that mentioned above,
Where I (x, y) represents the 2D input image to be enhanced.
we convert the image into logarithmic mode as in eq 2, to
separate between R and L.
Fig 3: The Impact of DCT
The following steps explain how DCT works:
1- First step, separates between the reflectance (R) and
luminance (L) by apply the logarithm (log) to the input image
I (x, y), next step, the all image is converted to frequency
domain by apply the DCT transform, where the coefficients of
DCT manipulated until, reaching the goal of obtaining a
constant illumination, the initial DCT coefficient C (0,0) is set
to as the eq (6) :
logI (x, y)  logR(x, y)  logL(x, y)
R' (x, y)  logI(x, y)  log[I(x, y) * k(x, y)]
(2)
(3)
C(0,0)  log. (6)
MN refers to the dimensions of the input image I (x, y),
μ is the average value of I (x, y). The number of DCT
coefficients is defined previously from the DCT coefficient
where “*” denotes the convolution operator, K(x,y) is a
smoothing kernel, which can be computed by eq (4). R’(x,y)
refer to the illumination invariant reflectance output of the
SSR algorithm.
matrix in a snaky way. encoding the low-frequency
information of the image is then set to zero. In the final step,
the modified matrix of DCT coefficients is converted again to
the local domain by taking the inverse of DCT to generate
K(x, y)k exp(
r 2
)
c 2
(4)
the representation of the facial image under illumination
varying.
3.3 Wavelet Denoising (WD)
C can be compute by eq(3.5) It is calculated in a manner that
it is equal to 1.
 K (x, y)dxdy 1 (5)
While , r  (x2
 y 2
)1/ 2
, I(x,y) is orginal image, R(x,y)
This technique presented by (Zhang et al., 2009), it is one of
the photometric normalization technique, the wavelet-
denoising approach is exploited to get a face representation
under illumination invariant [22] . To clarify the WD impact
on illumination variation we applied it on a set of images
belong to the same person of class 8 under different
illumination as shown in the following figure.
is reflectance and L(x,y) represent illumination .
3.2 Discreet Cosine Transform (DCT)
The discrete-cosine-transform-(DCT) is a photometric
normalization technique recently introduced by Chen et al in
2006. The technique also relies on the retinex theory and
imaging model [22] it supposes a similar assumption about
illumination characteristics in varying appearance. The
technique assume that the variations of illumination are
closely related, to the low frequency coefficients of the DCT
transform. these low frequency coefficients are discarded
before converting the image back to the spatial domain by
using the inverse DCT to obtain face representation under
constant illumination, the following figure shows the impact
of DCT on the human face Under different possibilities of
illumination belongs to the same person, where the upper
images represent the images under illumination and the
images on the lower rows represent the images that has been
processed by DCT.
Fig 4: The Impact of WD
The wavelet denoising depends on four major steps
here is an important operation applied during denoising
model they are: (2D DWT), thresholding, 2DIDWT. The
following figure in the next page summarizes the general
steps of wavelet denoising, where it starts with input image in
logarithm domain then decompose the image and apply
thresholding then reconstruct the image to obtain the output
image in local domain. (see the figure 5)

8
The wavelet denoising steps during applied on the image
Apply Thresholding
Where : Gx (x,y,σ), Gy (x,y,σ) are the derivation of Gaussian
kernal function in the x,y directions Respectively 3-compute
the illumination in insensitive measure by following
G  arctan(Ix/Iy)[0,2 ] (10)
Reconstruct
apply
2DIDWT
using synthesis
Filter
3.4 Gradient Faces (GF)
Gradientfaces is also known as an illumination insensitive
measure this technique proposes the principle that based on
extract illumination insensitive facial feature under different
lighting situations in gradient domain. There are many ways
to compute the gradient of image. However, the numerical
calculation of derivative directly without smoothing (gradient)
is typically not the ideal computation. to compute the gradient
in more stable way, we firstly smooth the image by using
Gaussian kernel function with a convolution type smoothing,
the numerical calculation of gradient will be more stable in
calculation [22]. moreover, the main feature for using
Gaussian kernel is multiplied, also robustness of
Gradientfaces is very high to deal with the image noise in
satisfied, it can eliminate the effects of shadows. The steps of
applying the Gradientfaces illustrated below, the following
figure illustrates the Gradientfaces impact.
Fig 6: The Impact of Gradient face
1- The first step starts with Smoothing operation which
applied on the input image with convolving Gaussian kernal
function
I’=I*G (x ,y,σ) (7)
Where * is a convolution operator a ,while G(x,y,σ) Contains
the following equation
G (x, y ,σ) = 1/2nσ2 exp (-x2 + y2) (8)
Where it represents the Gaussian kernal function with
standard deviation σ.
2- Compute the Gradient of image I by feeding the soothed
image through convolution operation with the derivation of
Gaussian kernal function in the x, y direction
Ix= I’*Gx (x, y, σ) and I =I* Gy (x, y, σ) (9)
3.6 PP chain (TT)
also known as (TT) Proportion of its inventors (Tan and
Triggs) [22], as shown by the name it is a chain of
preprocessing steps , one of illumination normalization
method , it consists of a series Sequential stages (Gamma
Correction, Difference of Gaussian (DoG), Filtering Masking,
Contrast Equalization). This chain is suggested to be applyed
like any preprocessing before feature extraction, the main goal
is to eliminate the effects of illumination, local shadowing and
highlights with keeping the essential parts of facial
appearance.
1- Gamma Correction  : the first step starts withapplying
the Gamma correction, which is a nonlinear transformation
method in gray-level it replaces the gray-level of image I by
I
 .
Where   [0,1] the parameter can be defined by the user to
enhances the darkness or shadowed regions and press it in the
bright regions and at highlights in the local dynamic range of
the image, the default value of  usually equal to 0.2.
2- Difference of Gaussian (DoG) Filtering:
Gamma correction couldn't eliminate the impact of all
intensity gradients such as shading influences, Shading comes
from the surface structure that effected by illumination
gradients. this make the spatial information in frequency
domain hard to separate from those effects. To solve this
problem, we using the Filtering process but, there are some
factors must be considered when selecting filters, for example
using High pass filtering may removes both the useful and
unwanted information, this make the recognition task more
complex. In contrast suppress the high frequencies reduces the
noise without effecting on the important recognition aspects.
DoG filtering is a proper filter to keep the fine details. Dog
filter has two filters to deals with high frequency and low
frequencies at the same time so this make it as a strong filter
that gives the important details that needed in recognition
[22,26].
3-Masking
Masking is an optional step, we need to apply the mask in
case of getting facial regions such as hair, beard that may
considered as irrelevant or too variable, so this area needed to
be mask out and the mask should be applied on this point [26].
4-Contrast Equalization
It is the final step of preprocessing chain, it basically proposed
to deal with any extreme values may be generated by the
filtering process, it takes extreme noise in the dominant noise
source and makes it approximately uniform [26]. The figure
(7) shows the impact of TT on the person belongs to class 8.
Fig 7: The Impact of TT
Fig 5: The Steps of WD
Decompose
apply 2DDWT
using Analysis
Filter
Input image
in logarithm
domain
Obtain log(R)
=Log(I)-Log(L)
Log R
Log (L)
Denoised
Image

9
12
25 50 77 90
4. FEATURE EXTRACTION
Feature extraction is the task of reducing the high dimension
of training data to a set of useful features. we applied LDA as
a necessary step before the classification step using random
forest. In our face recognition system we used the Linear
Discriminant Analysis (LDA) to reduce the huge image
dimensions, thus restricting the classification of selected
features in a limited number of samples.
6. IMPLEMENTATION
The focus of our study on the random forests efficiency and
the impact of different preprocessing on random forest
performance, we divided the experiments into 3 as following:
1. Experiment 1: Compute the error rate for each
preprocessing with Random forest (Applied on
subset 1 (faces with illumination variation).
• Single Scale Retinex (SSR).
• Discreet Cosine Transform (DCT).
• wavelet Denoising (WD).
Features
Fig 8: Feature extraction
New Features • Gradient Faces (GF).
• PP Chain (TT).
2. Experiment 2: Compute the error rate for each
preprocessing with RF under different occlusion
(Applied on subset 1+2).
3. Experiment 3: compute the Preprocessing time
and classification time.
5. RANDOM FOREST ALGORITHM
It is an ensemble learning method based on generate many
classifiers and aggregate their results. Random forest (RF) is
a supervised classification algorithm and a strong modern
approach to data classification and modeling. As illustrated
from its name “forest” it’s a group of trees, each tree consists
of splitting node, internal node or decision node and leaf
node. The image is classified by sending it down to the
terminal nodes of trees and aggregating the reached leaf
distributions based on voting. Randomness can be at two
points during training: in subsampling the training data so
that each tree is grown using a different subset. each subset
represents sub node that may selected to be split more based
on GiNi index, this process repeated in recursive manner.
The classification results from each tree are collected for an
input image, then the majority voting is gathered to give the
resulting class label. random forest is grown using
randomization strategy. The leaf nodes are the final node that
used to make the final decision. Each internal node contains
a test to select the best splits of data to be classified.
6.1 Database
Experiments will be conducted on Extended Yale B database
which contains a lot of frontal human faces under strong
effects for lighting. The number of stored images is about
2432 image of 38 persons with frontal pose, each person has
64 images were captured under different illumination
conditions. We have added a block of random occlusion on
frontal faces from 10% to 50%, to measure the efficiency of
the random forest as classifier even under Random occlusion.
The database is divided into five subsets based on the degree
of the light source direction and the central camera axis (
o
,
o
,
o
,
o
,
o
) . Fig.10, the image size is resized to
128 x 128. Subset 1, subset 2 are images with simple
illumination effects, the effects are gradually increased until
subset 5 which is contain the strong illumination effects, the
subset 3 are images with different occlusion.
e.g. (Lighting 50 <  >77)
(Occlusion=50%)
Fig 10: Samples of challenging faces
6.2 Computation the Error rate
We compute the error rate to refer the number of images that
classified incorrectly and usually calculate as shown in
equation below.
ER =
Number of Classified images
incorrectly
Total number of testing images
* 100
Fig 9: The Flow chart of Random forest
We dependent the equation above to calculate the accuracy of
random forest performance as a classifier with each
preprocessing. In contrast, the recognition rate (RR) can be
calculated instead of the error rate as found in some research
Start For each Tree
Select Randomly any Training
sub set
Stop condition at each node
Calculate predication error Build next
split
End
Build a model
Original image
Extract key Features
Reduced
data

10
Fig 11: Bar chart of experiment 1 results
but in this case the equation will vary, it will have computed as
the recognized images correctly divided by total images.
However, in our study, we calculated the error Rate not the
Recognition Rate, since the focus of our study is to measure
performance accurately to determine the rate of error and to
know which of the preprocessing give a large error rate.
7. EXPERIMENTS
7.1 Impact of Preprocessing Methods
We computed the Error Rate (ER) or accuracy for Five
preprocessing methods, also we used Subset 1 (samples with
illumination variation) as training set and the whole database
as testing set. we used five readings for each preprocessing for
each different number of trees and computed the average for
each trial to get more accurate results. The table below shows
the average results of each preprocessing, from the table we
can note that the ER increases in case of no preprocessing and
in case of using minimum number of trees, in our experiment
the minimum number of trees was equal to 10 The WD an TT
shows a good result with minimum number of trees, while the
other methods shown a good result from 20 number of trees to
above. The results of this experiment is explained in table 1 in
the next page.
There is no one static reading every time, but there is a certain
reading range, due to the randomized strategy of the random
forests in every classification process.
7.1.1 Evaluation of Experiment 1 In Table (1)
The Error rate of random forest is increased without any
preprocessing, also increased when the number of trees
reaches to the minimum, while the error rate Significantly
decreased when the number of trees >= 500. All the
preprocessing methods (TT, WD, GF, SSR,DCT) Sequentially
, give a good result when the number of trees is equal to 50 or
more. The PP chain (TT) and WD gives a good impact on
random forest even in little number of trees = 10. In case the
number of trees = 10, and without any preprocessing the AVG
of ER almost equal to 36 % . The following bar chart explains
the error rate for each preprocessing.
As shown in the figure above , no preprocessing gives thehigh
error rate , while the (TT,WD) shown an excellent resultunder
minimum number of trees , while (GF,SSR,DCT) did not give
a good results with minimum trees =10 .
7.2 Preprocessing methods under occlusion
we Computed the error rate for each preprocessing under
different occlusion and static number of trees = 500, we used
Subset 1 and subset 2 (samples with illumination variation)
as training set while, subset 3 (samples with different
occlusion)
as testing. The goal to show the performance of RF under
occlusion and how preprocessing may be affected by
occlusion, as we know the suggested pre-processing methods
is mainly applied for the illumination variation problem not
for occlusion, but some methods may deal with occlusion as a
noise and can enhances the result. The table (2) in the page 8
shows the results of each preprocessing under different
occlusion.
The Random Forest is able to classify the occluded faces with
different ratios of occlusion even without pre-processing, but
the error rate increases when there is no preprocessing, also
the number of trees that used in classification is another
factor, since the minimum number of trees gives the higher
error rate even if there a preprocessing.
7.2.1 Evaluation of Experiment 2 In Table (2)
As we seen in the table (2) the Random forest does not affect
by occlusion, no matter how much the occlusion ratio is, RF
shows a good classification even under 50 % occlusion, but
the problem arises in the case of the absence of preprocessing,
as seen from the results table (2) the higher error rate in case
the absence of preprocessing. Another important problem can
arise in case of choice a sensitive preprocessing for occlusion
such as SSR it gives a negative impact on classification
process. SSR gives a negative impact that are greater than the
impact in case of absence of preprocessing itself. DCT and GF
had shown the best results under occlusion and classification
with random forest, this is due to its powerful nature to resist
the occlusion. The TT and WD had shown also a positive
impact.
Fig 12: Bar chart experiment 2 results
As shown in the figure above, the SSR is a case sensitive to
occlusion, the DCT and GF had given the best result with RF,
also the TT and WD had given a positive impact with RF,
while in case of absence the preprocessing the RF gives a high
error rate of classification.

11
Preprocessing Trees Number Trial 1 Trial 2 Trial 3 Trial 4 Trial 5 AVG
No pre-
processing
10 37.39 37.58 35.64 35.22 34.62 36.09
50 27.14 24.05 26.77 27.42 25.57 26.19
100 23.68 24.33 24.14 24 24.06 24.04
500 21.51 22.16 21.09 21.23 23.08 21.81
SSR
10 9.56 9.23 9.37 8.68 8.67 9.10
50 1.94 1.66 2.17 1.34 1.29 1.68
100 1.06 0.92 1.34 0.97 1.15 1.08
500 1.25 1.20 0.78 0.74 0.88 0.97
DCT
10 14.04 15.05 12.60 14.13 13.67 13.89
50 3.23 3.97 3.23 3.42 3.51 3.47
100 2.49 2.68 2.49 2.86 2.17 2.53
500 1.71 1.52 1.39 1.38 1.20 1.44
WD
10 3.46 3.23 3.09 2.35 3.19 3.06
50 0.42 0.32 0.60 0.74 0.60 0.53
100 0.37 0.42 0.46 0.32 0.37 0.38
500 0.32 0.37 0.37 0.32 0.23 0.32
GF
10 8.86 6.37 6.42 6.14 6.83 6.92
50 1.20 1.66 1.11 2.21 0.92 1.42
100 0.97 1.29 1.06 0.78 1.20 1.06
500 1.02 0.69 0.83 0.60 1.11 0.85
PP-Chain (TT)
10 4.34 3.83 3.28 3.60 3.19 3.64
50 0.55 0.69 0.60 0.41 0.46 0.54
100 0.32 0.32 3.37 3.32 0.18 1.50
500 0.23 0.32 0.27 0.41 0.32 0.31
Table 1. Impact of preprocessing methods on ER (Experiment 1)
In this Experiment, we applied the techniques below as a preprocessing and the random forest as a classifier with number of trees
from 10 to 500, applied on the subset 1 as training and the all database as testing.

12
In this Experiment, we applied the techniques below as a preprocessing and the random forest as a classifier with number of trees = 500,
applied on the subset 1 and subset 2 as training and the subset 3 as testing under different occlusion from 10% to 50% .
Table 2. preprocessing methods Under Different Occlusion (Experiment 2)
Preprocessing Occlusion
Ratio (%)
Trial 1 Trial 2 Trial 3 Trial 4 Trial 5 AVG
No
preprocessing
10 8.57 7.47 8.13 9.01 5.71 7.77
20 9.45 7.69 7.03 7.91 6.59 7.73
30 8.13 7.69 4.17 6.15 7.25 6.67
40 7.47 8.13 5.71 6.81 6.59 6.94
50 7.91 6.59 8.13 5.71 6.15 6.89
SSR
10 7 9 7 11.8 9 8.76
20 7.47 7.91 8.35 7.47 10.76 8.39
30 7.25 7.91 8.35 8.79 7.69 7.99
40 8.35 7.25 9.23 9.23 8.79 8.57
50 9.01 8.35 9.89 8.79 8.57 8.92
DCT
10 1.31 1.31 0.87 0.65 0.87 1.00
20 1.97 2.19 1.09 1.97 1.09 1.66
30 1.09 0.65 2.85 1.09 1.09 1.35
40 1.31 1.97 2.19 1.75 1.75 1.79
50 1.53 1.31 1.97 1.09 0.21 1.22
WD
10 3.07 3.29 3.51 2.19 2.63 2.93
20 2.41 1.31 2.19 3.29 2.41 2.32
30 2.41 1.75 1.97 3.07 2.85 2.41
40 1.53 2.85 3.07 2.41 2.85 2.54
50 2.85 4.61 2.41 2.63 3.73 3.24
GF
10 1.31 1.97 1.75 1.09 1.75 1.57
20 1.97 1.75 1.09 0.87 1.75 1.48
30 2.41 0.43 0.31 0.43 1.31 0.97
40 1.09 1.31 1.09 0.44 1.09 1.00
50 1.97 1.53 1.31 1.09 1.31 1.44
TT
10 1.97 2.63 2.41 2.63 2.41 2.41
20 2.41 3.29 1.31 2.19 3.07 2.45
30 2.63 2.41 2.85 2.63 1.09 2.32
40 3.07 2.85 2.19 3.07 2.85 2.80
50 1.53 3.29 1.97 2.63 0.65 2.01

13
7.3 TIME COMPUTATION
In the experiment, we computed the preprocessing time and
classification time. in order to compute the total recognition
time.
• Software & Hardware Specification
We applied on computer with Core i7, windows 7, RAM 6
Giga. MATLAB 2012
7.3.1 Preprocessing Time
The time taken to apply the preprocessing on the whole
Database images, as we see from the table below the system is
very fast without a Preprocessing it takes about 8 seconds,
while the DCT has consumed the longest preprocessing time,
while the Gradient faces (GF) has consumed the shortest
preprocessing time compared with the other methods, this is
maybe depending on the algorithm complexity that are
adopted by each preprocessing.
Table 3. Preprocessing Time for each preprocessing
PREPROCESSING NAME PREPROCESSING TIME
Without Preprocessing 8 Sec
SSR 22 Sec
DCT 70 Sec
WD 37.5 Sec
GF 16.9 Sec
TT 25 Sec
From the table above, we note that DCT and WD we note that
DCT and WD have taken a long time in preprocessing, this
depends on the algorithm complexity which contains a lot of
steps includes decompose and reconstruct the image again
(see the DCT and WD steps on page 3 ). The following Bar
chart shows the preprocessing time for each preprocessing
applied, we can note that DCT is the highest Preprocessing
Time while without preprocessing is the fast preprocessing,
GF is the lowest preprocessing Time.
Fig13: Bar chart for preprocessing time
7.3.2 Classification Time
Is the time taken by the random forest classifier to train and
classify the images in the database, the random forest is
known as a fast and strong classifier, it classifies faces even
under occlusion. we noted this from the results, it works
efficiently even under 50% occlusion. The Classification Time
and the error rate are affected by the number of trees that used
in the classification process, when the number of trees is low
the classification time is decreased while, the error rate is
increased.
Table 4. Classification Time for each number of trees
NUMBER OF TREES CLASSIFICATION TIME
10 0.06 Sec
50 0.25 Sec
100 0.47 Sec
500 2.28 Sec
1000 4.47 Sec
As shown in the table above when the number of trees =10 the
classification time becomes very fast, while the number of
trees = 1000 the time increases until 4.47. As we have seen
from the previous experiments İt is important to note that the
minimum number of trees gives a fast classification time with
a high error rate and Vice versa.
The following Bar chart shows the Classification time for
each Number of Trees, we can note that when Number of
trees =1000 we get a slow Classification Time, while when
number of trees =10 we get a fast Classification Time, but
with a higher error rate as we mentioned before.
Fig 14: Bar chart for classification time
• Total Time
The Total Time is computed by Adding the training time to
classification time.

IJCATM : www.ijcaonline.org
14
8. CONCLUSTION
In this paper, we studied the impact of five common
preprocessing methods in face recognition on the random
forest performance. The study was conducted on Extended
Yale B database which contains faces under occlusion and
different illumination. We applied three different experiments.
Based on the practical experiments we got the following
results: In illumination variation experiment, we noticed that
TT and WD have shown an excellent result with random
forest even under using a little number of trees =10.
The other methods like GF and SSR have shown a good result
when using number of trees equal to 50 or above, while DCT
has shown the lowest performance with random forest
compared with the other methods, also in case of absence
preprocessing method with random forest we get the highest
error rate. In occlusion experiment, GF and DCT have shown
the best results with random forest even under 50% occlusion,
TT and WD have shown a good result with random forest,
while SSR and absence preprocessing method
, we get the highest error rate. In time experiment, we noticed
in case of there is no preprocessing we get the fastest
preprocessing time, the GF has shown the fastest
preprocessing time compared with the other methods, while
the DCT has shown the slowest preprocessing time.
9. FUTURE WORK
The pattern recognition is a very wide area, where the human
faces are considered as a special case of pattern recognition.
In the future work, we plan to conduct a methodology based
on several classifiers to explain how to choose the appropriate
preprocessing for each classifier, in other areas of patterns
classification.
10. REFERENCES
[1] V. Ghosal, “Efficient Face Recognition System Using
Random Forests,” Work, 2009.
[2] B. O. Connor and K. Roy, “Facial Recognition using
Modified Local Binary Pattern and Random Forest,” vol.
4, no. 6, pp. 25–33, 2013.
[3] H. O. A. Achen and P. R. O. F. G. E. L. Akemeyer,
“Detection and Recognition of Human Faces using
Random Forests for a Mobile Robot.”
[4] V. Vishwakarma and A. K. Srivastava, “Performance
Improvements in Face Classification using Random
Forest,” vol. 2, no. 3, pp. 2384–2388, 2012.
[5] E. Kremic and A. Subasi, “Performance of Random
Forest and SVM in Face Recognition,” vol. 13, no. 2, pp.
287–293, 2016.
[6] F. Classifiers et al., “A Facial Recognition System Based
on Techniques of Principal Component Analysis and
Auto faces with K-NN , K-Star and Random,” vol. 12,
no. April, pp. 7–14, 2013.
[7] V. W. Parate, M. T. Scholar, and V. Branch, “PCA ,
DCT and DWT based Face Recognition System using
Random Forest Classifier,” vol. 3, no. 6, 2015.
[8] A. Seal, D. Bhattacharjee, and M. Nasipuri, “Human
Face Recognition using Random Forest based Fusion of
A-trous Wavelet Transform Coefficients from Thermal
and Visible ,” AEUE - Int. J. Electron. Common., no.
May, 2016.
[9] A. I. Salhi, M. Kardouchi, and N. Belacel, “Fast and
efficient face recognition system using RandomForest
and Histograms of Oriented Gradients,” pp. 293–303,
2012.
[10] K. A. Ishak, S. A. Samad, M. A. Hannan, and M. M.
Sani, “Image Enhancement Technique using Adaptive
Multiscale Retinex for Face Recognition Systems 2 The
Original Multiscale Retinex.
[11] B. Academy, O. F. Sciences, and I. Technologies,
“Image Enhancement Using Retinex Algorithms and
Epitomic Representation Violeta Bogdanova,” vol. 10,
no. 3, pp. 10–19, 2010.
[12] D. J. Jobson, Z. Rahman, and G. A. Woodell, “Properties
and Performance of a Center / Surround Retinex,” vol. 6,
no. 3, pp. 451–462, 1997.
[13] M. Leszczyński, “Image Preprocessing for Illumination
Invariant Face Verification,” pp. 19–25
[14] C. Constancy, “Fast RETINEX for Color Image
Enhancement: Methods and Algorithms,” pp. 1–23.
[15] V. S. Manjula, “Image Normalization Robust using
Histogram Equalization and Logarithm Transform
Frequency DCT Coefficients for Illumination in Facial
Images,” vol. 3, no. 11, pp. 1891–1895, 2014.
[16] V. P. Vishwakarma, S. Pandey, and M. N. Gupta,
“Adaptive Histogram Equalization and Logarithm
Transform with Rescaled Low Frequency DCT
Coefficients for Illumination Normalization,” vol. 1, no.
1, pp. 1–5, 2009.
[17] H. K. Ekenel and R. Stiefelhagen, “local appearance
based face recognition using discrete cosine transform.”
[18] Z. Sufyanu, F. S. Mohamad, A. A. Yusuf, and B.
Mustafa, “discreet Cosine Transform,” no. February,
2016.
[19] W. Chen, M. J. Er, and S. Wu, “Transform in Logarithm
Domain,” vol. 36, no. 2, pp. 458–466, 2006.
[20] O. Of, “A Wavelet-Based Image Preprocessing Method
for Illumination Insensitive Face Recognition,” vol.
1731, pp. 1711–1731, 2015.
[21] A. P. C. S. Devi and V. Mahesh, “performance
evaluation of illumination normalization techniques for
face recognition,” vol. 1, no. 2, pp. 11–20, 2013.
[22] “Photometric Normalization Techniques for Illumination
Invariance AVTOR : Vitomir Štruc INTERNAL
REPORT : LUKS Photometric Normalization
Techniques for Illumination.”
[23] C. N. Raghu, “Illumination Insensitive Face Recognition
Using Gradientfaces,” no. 1, pp. 38–44, 2012.
[24] H. Han, S. Shan, X. Chen, and W. Gao, “A comparative
study on illumination preprocessing in face recognition,”
[25] Z. Yang, X. He, W. Xiong, and X. Nie, “Face
Recognition under Varying Illumination Using
[26] X. Tan and B. Triggs, “Enhanced Local Texture Feature
Sets for Face Recognition under Difficult Lighting
Conditions,” pp. 1–16, 2009.

a case study on various preprocessing methods and their impact on face recognition using randomforest

More Related Content

What's hot (18)

Similar to a case study on various preprocessing methods and their impact on face recognition using randomforest (20)

Recently uploaded (20)

a case study on various preprocessing methods and their impact on face recognition using randomforest