A Modified CNN-Based Face Recognition System

International Journal of Artificial Intelligence and Applications (IJAIA), Vol.12, No.2, March 2021
DOI: 10.5121/ijaia.2021.12201 1
A MODIFIED CNN-BASED FACE
RECOGNITION SYSTEM
Jayanthi Raghavan and Majid Ahmadi
Department of Electrical and Computer Engineering,
University of Windsor, Windsor, Canada
ABSTRACT
In this work, deep CNN based model have been suggested for face recognition. CNN is employed to extract
unique facial features and softmax classifier is applied to classify facial images in a fully connected layer
of CNN. The experiments conducted in Extended YALE B and FERET databases for smaller batch sizes
and low value of learning rate, showed that the proposed model has improved the face recognition
accuracy. Accuracy rates of up to 96.2% is achieved using the proposed model in Extended Yale B
database. To improve the accuracy rate further, preprocessing techniques like SQI, HE, LTISN, GIC and
DoG are applied to the CNN model. After the application of preprocessing techniques, the improved
accuracy of 99.8% is achieved with deep CNN model for the YALE B Extended Database. In FERET
Database with frontal face, before the application of preprocessing techniques, CNN model yields the
maximum accuracy of 71.4%. After applying the above-mentioned preprocessing techniques, the accuracy
is improved to 76.3%
KEYWORDS
CNN, ANN, GPU
1. INTRODUCTION
For human brain, recognizing face is a very simple task and can be performed fast. In computer
vision, face recognition is a very challenging task. Even though the face recognition research is in
an advanced state [7], till now it is not possible to obtain results on par with humans.
To date, many approaches have been suggested for facial recognition. Holistic method works by
projecting facial images onto a low-dimensional space, which neglects surplus details and
variations that are not needed for the facial recognition [11]. One of the methods under this
category is PCA [2]. The holistic methods are sensitive to local distortions like facial expression
or illumination variation.
Subsequently, progress in the field of computer vision led to the growth of feature-based method
in which features are extracted from various parts of a face image. Feature-based methods are
robust to local variations such as intensity variation and face expression changes [22]. With the
development of the local feature descriptors, feature based methods gained popularity. Local
Binary Pattern (LBP) [6] is an extensively applied local feature descriptor in face recognition.
The recent trend is towards neural network-based approach [30]. Deep learning-based methods
achieve excellent results in many fields like robotics and autonomous driving cars [3]. Deep
learning methods are based on convolutional neural networks (CNNs). CNNs are slightlydifferent
from normal neural network. In CNN, neurons in convolutional layer are thinly connected to the

2
neurons in the next layer based on their relative location. CNNs are multilayer network trained
from end to end with raw image pixel values assigned to classifier outputs.
The main advantage of deep learning methods is that they can be trained with very big datasets to
learn the vital features to represent the input data. The main issue with deep learning method is
models trained with small datasets are having the problem of poor generalization, which results in
over-fitting.
Generalization term indicates the performance difference of a network model when assessed on
earlier viewed training data against the testing data, the network has never viewed before [8].
Models with poor generalizability have overfitted the training data.
Overfitting is a term used when the network model functions extremely good with the training
data, but could not work well with the test data. In the overfitted network validation error goes up
while the training error comes down [21].
To reduce the overfitting, regularization process is employed on the network model.
Regularization is a method of making minute changes to the actual network model and the
learning algorithm, so that the model functions better in both training and testing data set.
Regularization is defined as “Allowing to generalize well to unseen data even when training on a
finite training set or with an imperfect optimization procedure” [4]. There are various
regularization techniques available in machine learning like Dropout, Data Augmentation, Early
stopping, batch normalization etc. [10]. Some of the regularization techniques are explained
below.
Dropout means removing units temporarily in a neural network, together with all the incoming
and outgoing links [23]. Dropout can be explained as the regularization technique by including
noise to the hidden units of the network.
Another popular technique is batch normalization. Batch Normalization operates by deducting
the batch mean from each activation and dividing by the standard deviation of the batch [8]. The
normalization technique together with standardization is used as a typical combination in the
preprocessing of pixel values. Batch normalization technique can be employed to any individual
layer within the network. Hence it is powerful [5].
Data augmentation is an extensively applied technique in deep learning. The performance of the
deep neural network is based on the size of the dataset [45]. Deep learning network needs large
dataset to avoid the problem of overfitting. Sometimes it is difficult to get quality and huge
database especially in medical field. Data augmentation helps to artificially inflate the size of the
training dataset by the methods called data warping or oversampling. The augmented data is a
representation of detailed set of feasible data points, thus minimizing the distance between the
training and the validation set.
Data augmentation techniques are useful to provide powerful regularization in terms of improved
generalization which in turn yields better network performance [46]. Bengio et al [48]
demonstrate that the data augmentation techniques are very effective in deep networks compared
to shallow networks. In addition to flipping and cropping various techniques like color casting,
vignetting, rotation, horizontal and vertical stretching proven to be effective data augmentation
methods [49].
The main aim of augmentation technique is to diminish the effect of overfitting on models using
traditional transformations to manipulate the training data. As an example, a labeled image in the

3
dataset can be increased by various processes like flipping, rotation, morphing and zooming.
After some time, trained network gets exposure to such modifications and the it can identify the
same object with different variations.
There are some limitations associated with the application of data augmentation techniques. One
of the main disadvantages is it demands extra memory, computational expenses involved with
implementing augmentation methods and longer training time. Some geometric transformations
such as translation or random cropping must be monitored carefully to ensure that the labels of
the images are not altered during the transformation process [8].
Optimizers update the weight parameters to minimize the cost function. Cost function is defined
as the difference between predicted and actual output. One of the popular optimizers is Adam
optimizer [1].
Adam is an adaptive learning rate optimization algorithm, which calculates learning rates for
different parameters individually. Adam uses computations of first and second moments of
gradient to adjust the learning rate for each weight of the network.
Application of regularization methods help to improve the accuracy rates. To increase the
accuracy rates further, preprocessing techniques are applied to deep CNN architecture.
2. PREPROCESSING METHODS
Face recognition task becomes challenging due to illumination conditions, occlusion, pose and
facial expression variations [33]. The variation in illumination is one of the main challenging
problems which affects the performance of the face recognition system. Among all, shadowing
effect, underexposure, and overexposure conditions are challenging problems that need to be
addressed in the face recognition process [35]. If the lighting conditions present in the gallery
image is different from the probe image, then the process of face recognition may completely fail
[34]. A good face recognition system should be able to give accurate recognition rate under the
different illumination conditions between images of the same face [32]. Image enhancement
algorithms play a great role in handling the illumination variation.
The main aim of preprocessing is to remove features that obstruct the process of classifying the
images of the same person (within-class differences), thereby boosting the difference of them
with others (between-class differences) [36].
For better face recognition under uncontrolled and illumination variation conditions, the vital
features responsible for differentiating two different faces require to be retained. The shadows
produced in facial images due to variation in lighting directions may cause loss of important
facial features which are helpful for recognition. A preprocessing method must enhance the
intensity in the regions of inadequately illuminated and decrease the intensity in the densely
illuminated regions while retaining the intensity in the fairly illuminated portions [37]. Few
important preprocessing techniques are discussed below.
2.1. Histogram Equalization (HE)
HE [38] flattens the histogram of the image and expands the dynamic range of the pixel intensity
values by employing cumulative density function. The Histogram Equalization is a global
preprocessing technique.

4
For image I(x, y) with discrete k. gray values histogram is defined by the probability of
occurrence of the gray level I [39] is given by equation (1) as follows.
p(i) = 𝑛𝑖 /N (1)
Where i∈ 0, 1…k -1 grey level and N is total number of pixels in the image.
2.2. Self Quotient Image (SQI)
SQI [40] is one of the illumination invariant algorithm suggested for handling both shadow and
lighting changes. It is defined as the ratio of the intensity of the input image to its smooth version
as given in equations (2) and (3).
Q(x,y)=I(x,y)/S(x,y) (2)
= I(x,y) /(F(x,y) *I(x,y)) (3)
where I (x, y) is the face image and S (x, y) is a smoothed version of the image and ∗ is the
convolution operation. F is the smoothing kernel which in this case is a weighted Gaussian filter
and Q is the Self Quotient Image since it is derived from one image and has the same quotient
form as that in the quotient image method.
2.3. Locally Tuned Inverse Sine Nonlinear (LTISN)
LTISN [41] is a nonlinear and pixel by pixel approach, where the improved intensity values are
calculated by applying the inverse sine function with a tunable parameter based on the nearby
pixel values given in the equations (4), (5), (6), (7) and (8). The intensity range of the image is
rescaled to [0 1] followed by a nonlinear transfer function.
2.4.Gamma Intensity Correction (GIC)
GIC [42] is a nonlinear gray-level transformation that substitutes gray-level I with the gray level
𝐼
1
ϒ, given by the equation (9).
𝐼 = 𝐼
1
ϒ(9)

5
As shown in the (Fig.1), for the values of gamma less than 1.0 darkens the image and for the
values of gamma greater than 1.0 lightens the image. When gamma value is 1.0, it does not
produce any effect.
Fig.1. Gamma Intensity Correction
2.5. Difference of Gaussian (DoG)
DoG is a grayscale image enhancement algorithm [42] that involves the subtraction of one
blurred version of an original grayscale image from another less blurred version of the original.
The blurred images are obtained by convolving the original grayscale image with Gaussian
kernels having differing standard deviations [34], which is given in the equation (10).
𝜎1, 𝜎2 are Gaussian kernel widths.
2.6. Contrast-Limited Adaptive Histogram Equalization (CLAHE)
CLAHE works on small areas in the image, known as tiles, rather than the whole image [43].
Individual tile's contrast is improved. Hence histogram of the output area is roughly matching
with the histogram specified by the distribution parameter. The tiles present in the neighborhood
regions are then joined by applying bilinear interpolation to minimize the effect of artificially
induced border line [39]. In CLAHE, the image is divided into a limited number of regions and
the same histogram equalization technique is applied to pixels in each region [44].
3. CONVOLUTIONAL NEURAL NETWORK
Deep learning-based methods have shown better performances in terms of accuracy and speed of
processing in image recognition
3.1. CNN Basics
CNN is biologically inspired by visual cortex in brain [20]. Different phases of learning
procedure in CNN is similar to the visual cortex. The visual cortex has small group of cells that
are reactive to particular regions of the visual field. Hubel and Wiesel [17] experimentally
showed that particular group of neurons in the cat’s brain reacted to the appearance of edges of a
certain orientation as shown in the (Fig.2). Further they illustrated that one specific set of neurons

6
were activated, when exhibited to vertical edges and also another set of neurons fired when
exposed to horizontal or diagonal edges. All of these categories of neurons are arranged in a
columnar configuration and collectively these neurons can produce visual perception.
Different portions of the visual cortex are classified as V1, V2, V3, and V4. In general, V1 and
V2 regions of visual cortex are having close resemblance with convolutional and subsampling
layers, whereas inferior temporal region resembles the higher layers of CNN [16]. During
theprocess of training, CNN learns with the help of backpropagation algorithm by making
adjustments in weights with respect to the target.
Fig.2. Hubel and Wiesel experiment
3.2. CNN Architecture
In CNN architecture, network layers are divided into three types: the convolutional, pooling and
fully connected layers [20]. The architecture of CNN is as shown in Fig 3.
3.2.1. Convolutional Layer
In CNN, every neuron in the convolutional layer is linked only to a small portion of the neurons
in the preceding layer, which is in square shape area across the height and width dimensions. The
size of this square is a hyperparameter (controllable parameter) and called as Receptive Field. For
the depth dimension, there is no hyperparameter available since the convolution operations are
normally carried out for the whole depth. Generally, the depth dimension of the input describes
the various colors of the image. Hence it is usually required to link them in order to bring out
necessary information.
Fig.3.The CNN architecture

7
Neurons present in the convolution operator can recognize certain local patterns of the previous
layer’s output. Once features are obtained, its actual position is not important and all the neurons
are expected to identify the same pattern. This is achieved by forcing all the neurons to have a
common single set of parameters known as Parameter Sharing [9].
In order to identify various unique features within one layer, it is necessary to have multiple
filters, where each filter is a group of neurons that recognize a certain pattern at different
locations in the image.
3.2.2. Pooling Layer
The main aim of pooling layer is to reduce the complexity of CNNs. The neurons present in the
pooling layer form a square shaped area across the width and height dimensions of the preceding
layer. Even though it is very similar to the convolutional layer, it is different from convolutional
layer because the pooling layer is Non-Parametrized Layer.
The function carried out by this layer is known as subsampling or down sampling. During this
process, contraction in size results in concurrent loss of data. On the other hand, such a loss is
helpful to the network because the reduction in size not only reduces the computational burden
for the succeeding layers of the network and also it reduces the effects of overfitting [12].
Max pooling and average pooling are the generally used techniques shown in (Fig.4). Max
Pooling chooses the largest element within each receptive field [14] whereas Average Pooling
computes the average among the output neurons within the pooling window. Max-pooling
chooses the most prominent feature in a pooling window. On the other hand, average-pooling
method selects whole features into consideration. Thus, max-pooling method keeps texture
related information, while average pooling method retains the background related data [24].
Pooling operation does not combine neurons with different depth values. Instead, the resulting
pooling layer will have the uniform depth as the previous layer and it will only combine local
areas within a filter.
Fig.4.Max Pooling and Average Pooling Operations Illustration.
3.2.3. Fully Connected Layer
The filters and neurons present in this layer are connected to all the activation in the preceding
layers resulting in a completely connected structure. Hence the name. The output feature maps of
the final convolution or pooling layer is converted into a one-dimensional (1D) array of numbers
[18]. High-level reasoning in the network is carried out via fully connected layers [19].

8
The final fully connected layer has the same number of output nodes as the number of classes.
Each fully connected layer is followed by a nonlinear function, such as ReLU (Rectified Linear
Units). ReLU is an activation function operates by thresholding values at 0, i.e. f (x) = max (0, x).
In other words, it outputs 0 when x < 0, and contrarily, it outputs a linear function with a slope of
1 when x ≥ 0 [15] as shown in the (Fig.5).
Fig.5. The Rectified Linear Unit (ReLU)
3.3. CNN Operation
Based on local receptive field, each component in a convolutional layer accepts inputs from a set
of adjacent units belonging to the preceding layer layer. This way neurons are proficient in
extracting rudimentary features like edges or corners. These features are then linked by the
succeeding convolutional layers in order to further extract high level features.
The components of a convolutional layer are arranged in planes. All units of a plane share the
same set of weights. Thus, each plane is in charge for building a particular feature. The results
obtained from the plane is termed as feature maps. Each convolutional layer consists of several
planes, so that multiple feature maps can be constructed at each location. The most significant
features derived are passed from initial layers to higher layers. As the features are passed to the
higher layer, there is a dimensionality reduction in features determined by kernel size of the
convolutional and max-pooling layers. On the other hand, there is an increase in number of
feature maps for representing better features of the input images for ensuring classification
accuracy [13]. The derived feature vector either could be an input for classification task or could
be treated as a feature vector for next level processing.
The network designed to do the classification task consists of several convolution pooling layers
followed by some fully-connected layers. The first two layers mentioned above perform
convolution and pooling operation in order to extract high-level features. The final layer’s output
of CNN is applied as the input to a fully connected network, which does the task of classification.
The output layer for classification task consists of one neuron for each class and the values of
these neurons describes the score of each class. If score distribution is chosen, the score range
will be between zero and one. The summation of class scores is one. The values of each neuron
can be presumed as the probability of occurrence of the class.

9
Fig.6. A visual representation of the various hyperparameters of convolutional layers: receptive field,
stride and padding
4. EXPERIMENTAL SETUP
The implementations are carried out in MATLAB 2019b on a workstation with windows 10 OS,
AMD processor with 2.0 GHz, hard drive with 8 GB RAM. The experiments are conducted in
Extended Yale B and FERET Database with deep CNN model.
4.1. Extended Yale B Database
The extended Yale Face Database B contains 16128 images of 28 human subjects under 9 poses
and 64 illumination conditions as shown in (Fig.7). The data format of this database is the same
as that of the Yale Face Database B.
Fig.7. Images from extended Yale B database.
4.2. FERET Database
The FERET database was collected in 15 sessions between August 1993 and July 1996. The
database contains 1564 sets of images for a total of 14,126 images that includes 1199 individuals
and 365 duplicate sets of images as shown in (Fig.8). A duplicate set is a second set of images of
a person already in the database and was usually taken on a different day.

10
Fig.8. Images from FERET database
4.3. Experiment
The model uses a deep CNN network with six convolution layers (with size 3x3), two
maxpooling and two fully convolutional layers. The first two convolutions use 8 filters, next two
uses 16 and the last convolution layers uses 32 each in order to extract complex features.
The model uses batch normalization and dropouts in all convolutional layers along with ReLU
activation function. In order to reduce the overfitting, the dropout rate is kept higher on the final
convolutional layers. Two maxpooling layers are introduced after second and fourth
convolutional layers to reduce the feature space dimension. The two fully connected layers use
128 and 28 neurons respectively. The final classification layer uses softmax activation with
categorical cross entropy loss function. The softmax layer is used to produce the classification
scores, in which each score is the probability of a particular class for a given instance [31].
The dropout principle is employed on the convolutional neural networks model and the value of
drop out is chosen by trial and error method. The (Fig.9) shows the network model with drop out
and the values of the drop out is different for different layers. The model also uses Adam
Optimizer with a learning rate of 0.001 which was found empirically after trying different
combinations.
The feature maps are subsampled with maxpooling layers with a stride of 2x2. Stride is the
number of pixels which shifts over the input matrix. When the stride is 2, it means the filter is
moved by 2 pixels at a time. The number of neurons in this output layer is limited to 28 in
Extended Yale B database. In FERET database, the number of neurons in the output layer is 994.
Data Augmentation is implemented in the form of simple methods like resizing, rotation, and
reflection. The batch size of 128 and number of epochs 250 are chosen to achieve best results.
The training is designed to use different batch sizes to do a fair trade off between accuracy and
training time.
(a) (b)
Fig.9. (a) The neural networks model (b) The model after applying drop out

11
5. RESULTS
The CNN trained with back-propagation algorithm in batch mode with a various batch sizes of 4,
8, 16 and 32 for different epochs are computed for Extended YALE B with an image size of
128×128 for 50 epochs. The same experiment is conducted on FERET database for the frontal
face image size of 128×128 for 300 epochs.
For the Extended YALE B Database, the maximum accuracy rate of 97.2% is achieved for the
batch size of 4 without applying preprocessing technique as shown in (Fig.10). After applying the
preprocessing techniques, the maximum accuracy is improved to 99.8% as shown in (Fig.11).
For the batch size of 8, the maximum accuracy rate of 97.0% is achieved without applying any
preprocessing technique. After applying the preprocessing techniques, the maximum accuracy is
improved to 99.4%.
For the batch size of 8, SQI, LTISN and HE perform very equally well and yield the maximum
accuracy of 99.4%, 99.3% and 99.1%.
For the batch size of 16, the maximum accuracy rate of 96.8% is achieved without applying
improved to 99.1%.
For the batch size of 16 HE, SQI and GIC perform equally well and yield the maximum accuracy
of 99.1%. 98.8% and 98.8% respectively.
improved to 98.7% for the same batch size.
For the batch size of 32, GIC, SQI, and HE perform equally well and yield the maximum
accuracy of 98.7%, 98.3% and 98.3% respectively. Table 1 shows the accuracy rates of Extended
Yale B database for various batch sizes before and after application of preprocessing techniques.
For the FERET Database, the maximum accuracy rate of 71.4% is achieved without applying
preprocessing technique, for the batch size of 4, as shown in (Fig.12). After applying the
preprocessing techniques, the maximum accuracy improved to 76.6% as shown in (Fig.13).
DoG, HE and LTISN perform equally well and yield the maximum accuracy of 76.6%, 76.3%
and 75.6% respectively.
improved to 76.4%.
For the batch size of 8, HE, SQI and LTISN perform very equally well and yield the maximum
accuracy of 76.4%, 74.6% and 72.5% respectively.
improved to 76.1%.

12
For the batch size of 16 HE, LTISN and SQI perform very equally well and yield the maximum
accuracy of 76.1%, 72.0% and 71.5% respectively.
improved to 71.0%.
For the batch size of 32, SQI, GIC, LTISN and DoG perform equally well and yield the
maximum accuracy of 71%, 70.9%, 70.8% and 69.9% respectively.
SQI and HE perform equally well and provide best results for all the batch sizes in both Extended
Yale B and FERET database.
Fig.10.Accuracy using Extended YALE B Database before applying of preprocessing techniques
Fig.11.Accuracy using Extended YALE B Database after application of preprocessing techniques

13
Fig.12.Accuracy using FERET Database before the application of preprocessing techniques.
Fig.13.Accuracy using FERET Database after application of preprocessing techniques.
Table 1 Accuracy rates of Extended Yale B database for various batch sizes before and after application
of preprocessing techniques.
Database Accuracy Rates
Batch Size
4 8 16 32
Extended Yale B (WOAPP) 97.2 97.0 96.8 96.2
DoG 99.1 99.0 98.7 97.9
SQI 99.8 99.4 98.8 98.3
LTISN 99.2 99.3 98.4 98.1
HE 99.7 99.1 99.1 98.3
GIC 99.6 98.9 98.8 98.7

14
Table 2 Accuracy rates of FERET database for various batch sizes before and after application
of preprocessing techniques
Database AccuracyRates
Batch Size
4 8 16 32
FERET(WOAPP) 71.4 71.2 71 68.7
DoG 76.6 72.6 71.3 69.9
SQI 74.9 74.6 71.5 71.0
LTISN 75.6 72.5 72.0 70.8
HE 76.3 76.4 76.1 69.2
GIC 74.2 71.3 71.0 70,9
Data augmentation with the batch size of 128 and number of epochs 250 yields best results,
especially with the preprocessing technique Histogram Equalization. In FERET Database, the
maximum accuracy of 79.6% is achieved.
For the same batch size and number of epochs, LTISN provides almost the same accuracy rate of
78.9%.
DoG, SQI and GIC performs almost equally and provides the accuracy enhancement of 78.2%,
78.3% and 78.1%
Data augmentation Methods does not have much effect in Extended Yale B database. With the
batch size of 128 and number of epochs 250 yields best results, especially with the preprocessing
technique of Histogram Equalization. In Extended Yale B Database, the maximum accuracy of
99.86% is achieved.
For the same batch size and number of epochs, SQI and LTISN provides almost the same
accuracy rate of 99.91% and 99.31%
DoG, HE and GIC performs almost equally and provides the accuracy enhancement of 99.21%,
99.78% and 99.71%
5.1. Number of trainable parameters calculation in Model
The number of trainable parameters is computed for each layer as below. The deep architecture is
as shown in (Fig.14).
C1 Layer
Input image 128×128×1 and having one channel. Number of weights per filter=3×3×1
There are totally eight filters in C1 layer and there is one bias parameter for each filter.
Total number of trainable parameters in C1 Layer is = (3×3×1+1) × 8= 80.
C2 Layer
Input feature maps to C2 layer having 8 channels. Number of weights per filter is 3×3×8.

15
There are 8 such filters in the layer. Hence total number of weights = ((3×3×8) +1) ×8 and there
is one bias for each filter.
Total weights in C2 = ((3×3×8) +1) ×12=584.
S1 Layer
S1 layer is using max pooling operation. Hence no trainable parameter available for this layer.
C3 Layer
Input to C3 has eight feature channels. Number of weights per filter is 3×3×8.
There are 16 such filters in the layer. Hence total number of weights = ((3×3) ×8+1) ×16 and
there is one bias for each filter.
Total weights in C3 = (3×3×8+1) ×16=1168.
C4 Layer
Input to C4 has sixteen feature channels. Number of weights per filter is 3×3×16.
S2 Layer
S2 layer is using max pooling operation. Hence no trainable parameter available for this layer.
C5 Layer
Input to C5 has sixteen feature channels. Number of weights per filter is 3×3×16.
Fig.14. Deep CNN Model Architecture (BN-Batch Normalization)

16
C6 Layer
Input to C6 has 32 feature channels. Number of weights per filter is 3×3×32
There are 32 such filters in the layer. Hence total number of weights = (3×3 ×32+1) ×32 and
F1 Layer
In the fully connected layer, the input is having total feature space of size 32×32 and 32. Input
samples to F1 layer=32×32×32
There are 128 neurons in F1 layer and each have a bias.
Total Number of trainable parameters for F1 layer is = (32×32×2+1) ×128=4194432.
Fig.15.Feature map obtained for the batch size 32 and epoch 2

17
F2 Layer
The input to F2 layer is the output from F1 layer. It is connected to 28 neurons (since there are 28
classes). So, total number of trainable parameters = (128 +1) ×28= 16512.
Feature maps extracted for different batch size and epochs are as shown in the (Fig.15,16,17).
6. COMPARISON WITH OTHER APPROACHES
In this section, the proposed approach is compared with the various approaches.
Hybrid system, which combines convolutional neural network (CNN) and a Logistic regression
classifier (LRC) yields the accuracy of 80%. A CNN is trained to recognize facial images and
LRC is used to classify the features learned by the convolutional network [26]. The neural
network with the nonlinear theories, such as wavelet theory and fuzzy set is a new research idea,
when applied in face recognition, gives accuracy around 93% in Wavelet-BP [25]. CNN
architecture along with Max-Feature-Map (MFM) could extract compact vital information yields
the accuracy rate of 98% [28]. The popular LeNet-5 architecture in FERET database yields
accuracy rate is 81.25% [27]. The pre-trained CNN model along with the VGG- Face provides
face verification rate accuracy of 86.85% and 83.32% on FRGC and LFW databases respectively
[29].
Table 3. Comparison of Accuracy rates of different approaches.
Sno Architectureused Dataset Accuracy
1 Proposedmethod(C
NN)
Extended Yale
BFERET
99.8%
76.3%
2 VGG+CNN[29] FRGC,LFW 86.85%,
83.32%
3 CNN withMAF
activationfunction
[28]
LFW 98%
4 Wavelet-BP[25] AT&T 93.0%
5 CNN-LRC
[26]
Yale 80%
6 CNN-LENET[27] FERET 81.25%
The proposed method compared with existing approaches are given in Table 3. The proposed
method performs better in comparison to all other approaches in Extended Yale B. In FERET
database, CNN-LENET method performs better than the proposed method.
7. CONCLUSION
In this paper, deep convolution neural network (CNN) is applied for extracting features and
classification. The performance of the network is assessed before and after applying
preprocessing techniques for various batch sizes and epochs on YALE B and FERET dataset.
Adam optimizers with a learning rate of 0.001 is applied in this experiment. The batch sizes used
in this experiment are [4, 8, 16,32];
The optimum batch sizes are chosen using trial and error method. Initially, multiple batch sizes of
[4,8,16,32,64,128] are selected for this experiment. Batch size of 4 with learning rate of0.001 and

18
number of epochs 50 give the best results. Batch sizes of [4,8,16,32] provide consistent and best
results. The batch sizes are normally selected in power of 2. Choosing smaller batch size with
low value of learning rate yields best result.
The results showed the maximum accuracy rate of 97.2% of achieved without using
preprocessing techniques. When the preprocessing techniques are applied, the improved accuracy
rates are achieved up to 99.8%.
The same experiments are conducted in FERET. dataset also. For the frontal face, the maximum
71.4% is achieved without optimization. After applying the preprocessing techniques, the
accuracy rate increased to 76.3%.
The proposed approach performs very well as compared to existing methods.
REFERENCES
[1] S. Setiowati, Zulfanahri, E. L. Franita and I. Ardiyanto, "A review of optimization method in face
recognition: Comparison deep learning and non-deep learning methods” 9th International Conference
on Information Technology and Electrical Engineering (ICITEE), Phuket, pp. 1-6, 2017.
[2] Aleix M. Marttinez and Avinash C. Kak, “PCA versus LDA”. IEEE Transactions on Pattern Analysis
and Machine Intelligence, Vol. 23, No.2, pp. 228-233, 2001.
[3] Rahul Haridas, Jyothi R L “Convolutional Neural Networks: A Comprehensive Survey” International
Journal of Applied Engineering Research Volume 14, Number 3, pp. 780-789, (2019).
[4] Goodfellow, I. J., Bengio, Y., and Courville, A. Deep Learning MIT Press, (2016).
[5] Jason Wang, Luis Perez The Effectiveness of Data Augmentation in Image Classification using Deep
Learning. In: Stanford University research report, 2017.
[6] Di Huang, Caifeng Shan, Mohsen Ardebilian, Yunhong Wang, and Liming Chen “Local Binary
Patterns and Its Application to Facial Image Analysis: A Survey” IEEE Transactions on Systems,
Man and Cybernetics, part C, volume 41, pp 765-781, 2011.
[7] Shailaja A Patil and Dr. P. J. Deore “Face Recognition: A Survey” Informatics Engineering, an
International Journal (IEIJ), Vol.1, No.1, December 2013.
[8] Connor Shorten and Taghi M. Khoshgoftaar “A survey on Image Data Augmentation for Deep
Learning” Jounal of Big Data, - Springer, 2019.
[9] Felix Ellenberger, Claus Lenz “A Non-Technical Survey on Deep Convolutional
Neural Network Architectures” Computer Vision and Pattern Recognition, 2018.
[10] S.V.G. Reddy, K. Thammi Reddy, V. ValliKumari “Optimization of Deep Learning using various
Optimizers, Loss functions and Dropout” International Journal of Recent Technology and
Engineering (IJRTE) Volume-7 Issue-4S2, pp 448-455, December 2018.
[11] SaezTrigueros, D.; Meng, L.; Hartnett, M. Face Recognition: From Traditional to Deep Learning
Methods. J. Theory. Appl. Inf. Technol. 2018, 97, 3332–3342.
[12] Athanasios Voulodimos, Nikolaos Doulamis, AnastasiosDoulamis, and EftychiosProtopapadakis
“Deep Learning for Computer Vision: A Brief Review” Computational Intelligence and
Neuroscience, pp 1-13, 2018.
[13] MdZahangirAlom, Tarek M. Taha, Chris Yakopcic, Stefan Westberg, PahedingSidike,
MstShamimaNasrin, Mahmudul Hasan, Brian C. van Essen, Abdul A. S. Awwal and Vijayan K.
Asari “A State-of-the-Art Survey on Deep Learning Theory and Architectures”. Electronics 2019, 8,
292
[14] WaseemRawat and Zenghui Wang “Deep Convolutional Neural Networks for Image Classification:
A Comprehensive Review” Volume 29, pp 2352-2449 September 2017.
[15] A. F. Agarap. Deep learning using rectified linear units (relu). CoRR, abs/1803.08375, URL
http://arxiv.org/abs/1803.08375 , 2018.
[16] Khan, Asifullah&Sohail, Anabia&Zahoora, Umme& Saeed, Aqsa “A Survey of the Recent
Architectures of Deep Convolutional Neural Networks”, published in Artificial Intelligence Review
(2019).

19
[17] D. H. Hubel and T. N. Wiesel. “Receptive fields of single neurons in the cat’s striate cortex” The
Journal of Physiology, 148(3):574–91, 1959.
[18] Yamashita, R., Nishio, M., Do, R.K.G. et al. “Convolutional neural networks: an overview and
application in radiology” Insights Imaging 9, pp 611–629, 2018.
[19] MamounAlazab, MingJian Tang “Deep Learning Applications for Cyber Security”, Springer, ISBN
978-3- 030-13057-2, May 2019.
[20] N. Aloysius and M. Geetha, "A review on deep convolutional neural networks," 2017 International
Conference on Communication and Signal Processing (ICCSP), Chennai, pp. 0588-0592, doi:
10.1109/ICCSP.2017.8286426, 2017.
[21] Nusrat, Ismoilov and Sung-Bong Jang. “A Comparison of Regularization Techniques in Deep Neural
Networks.” Symmetry 10 (2018): 648.
[22] B. S. Manjunath, R. Chellappa and C. von der Malsburg, "A feature-based approach to face
recognition," Proceedings 1992 IEEE Computer Society Conference on Computer Vision and Pattern
Recognition, Champaign, IL, USA, pp. 373-378, 1992.
[23] Srivastava, Nitish et al. “Dropout: a simple way to prevent neural networks from overfitting.” J.
Mach. Learn. Res. 15, pp 1929-1958, (2014).
[24] Zhao, Qi et al. “Multiactivation Pooling Method in Convolutional Neural Networks for Image
Recognition.” Wireless Communications and Mobile Computing (2018): n. page. Wireless
Communications and Mobile Computing. Web, 2018.
[25] J. Kittler, P. Koppen, et. al, "Conformal Mapping of a 3D Face Representation onto a 2D Image for
CNN Based Face Recognition," 2018 International Conference on Biometrics (ICB), Gold Coast,
QLD, pp. 124-131, 2018.
[26] H. Khalajzadeh, M. Manthouri and M. Teshnehlab, ’Face recognition using convolutional neural
networks and simple logistic classifier’, Advances in intelligent systems and computing, 223, pp 197-
207, (2014).
[27] A.R. syafeeza, M.K. Hani, S.S. Uiw and R. Bakhteri,’ Convolutional neural network for face
recognition with pose and illumination variation’, International journal of engineering and
Technology, 6(1), pp 44-57, (2014).
[28] Y. Xu, F. Liang, G. Zhang and H. Xu, ‘Image intelligent detection based on the Gabor wavelet and
the neural network’, Symmetry, 8(130), (2016).
[29] Wu, Xiang et al. “A Lightened CNN for Deep Face Representation.” ArXiv abs/1511.02683
(2015):n. pag.
[30] Z. Lu, X. Jiang and A. Kot, "Enhance deep learning performance in face recognition,” 2nd
International Conference on Image, Vision and Computing (ICIVC), Chengdu, pp. 244-248, 2017.
[31] Pouyanfar, S., Sadiq, S., Yan, Y., Tian, H., Tao, Y., Reyes, M. P, Iyengar, S. S. A survey on deep
learning: Algorithms, techniques, and applications. ACM Computing Surveys, 51(5), (2018).
[32] Y. Admi, Y. Moses and S. Ulman “Face Recognition: The problem of compensating for changes in
illumination direction”, IEEE Transaction. Pattern Anal. Mach. Intell., vol. 19, pp 721–732, Jul.
1997.
[33] SoodehNikan and M. Ahmadi. “Human Face Recognition under Occlusion using LBP and Entropy
Weighted Voting” 21st International Conference on Pattern Recognition ICPR pp. 1699-1702, 2012.
[34] S. Anila, K. Arulsukanya, K. Divya, C. Kanimozhi, and D. Kavitha “An efficient Preprocessing
Technique for Face Recognition under difficult Lighting Conditions,” in Proc. National Conference
on Emerging Trends in Computer Communication and Informatics (ETCCI-2011), March 10-11,
2011.
[35] Hu Han, Shiguang Shan, Xilin Chen, Wen Gao “A comparative study on illumination preprocessing
in face recognition” Pattern Recognition., Volume 46, Issue 6, Pages 1691-1699, June 2013.
[36] Calvo G., Baruque B., Corchado E. (2013) Study of the Pre-processing Impact in a Facial
Recognition System. In: Pan JS., Polycarpou M.M., Woźniak M., de Carvalho A.C.P.L.F., Quintián
H., Corchado E. (eds) Hybrid Artificial Intelligent Systems. HAIS, Lecture Notes in Computer
Science, vol 8073. Springer, Berlin, Heidelberg, 2013.
[37] Koringa P.A., Mitra S.K., Asari, V.K. Handling Illumination Variation: A Challenge for Face
Recognition. In: Raman B., Kumar S., Roy P., Sen D. (eds) Proceedings of International Conference
on Computer Vision and Image Processing. Advances in Intelligent Systems and Computing, vol 460.
Springer, Singapore, 2017.

20
[38] Dr. A. Sri Krishna, G. Srinivasa Rao and M. Sravya “Contrast Enhancement Techniques using
Histogram Equalization Methods on Color Images with Poor Lightning”, International Journal of
Computer Science, Engineering and Applications (IJCSEA) Vol.3, No.4, August 2013.
[39] P. Suganya, S. Gayathri, N. Mohanapriya “Survey on Image Enhancement Techniques” International
Journal of Computer Applications Technology and Research, Volume 2– Issue 5, 623 - 627, 2013.
[40] Biglari M, Mirzaei F, Ebrahimpour-Komeh H. Illumination invariant face recognition using SQI and
weighted LBP histogram[C]. Pattern Recognition and Image Analysis (PRIA), 2013 First Iranian
Conference on. IEEE, pp 1-7, 2013.
[41] E Krieger, VK Asari, and S Arigela. Color image enhancement of low-resolution images captured in
extreme lighting conditions. In SPIE Sensing Technology and Applications, pages 91200Q– 91200Q.
International Society for Optics and Photonics, 2014.
[42] Anila, S., and N. Devarajan. "Preprocessing technique for face recognition applications under varying
illumination conditions." Global Journal of Computer Science and Technology Volume 12, issue 11,
Version 1, 2012.
[43] NungsanginlaLongkumer, Mukesh Kumar and RohiniSaxena “Contrast Enhancement Techniques
using Histogram Equalization: A Survey” International Journal of Current Engineering and
Technology, Vol.4, No.3, E-ISSN 2277 – 4106, P-ISSN 2347 – 5161 (June 2014).
[44] Ali M. Reza. "Realization of the Contrast Limited Adaptive Histogram Equalization (CLAHE) for
Real-Time Image Enhancement." Journal of VLSI Signal Processing, vol. 38, pp 35-44, 2004.
[45] Porcu, S.; Floris, A.; Atzori, L. Evaluation of Data Augmentation Techniques for Facial Expression
Recognition Systems. Electronics 2020, 9, 1892. https://doi.org/10.3390/electronics9111892.
[46] Zagoruyko, K. Wide residual networks. In BMVC, 2016.
[47] DeVries, T. and Taylor, G. W. Improved regularization of convolutional neural networks with cutout.
arXiv preprint arXiv:1708.04552, 2017.
[48] Y. Bengio, A. Bergeron, N. Boulanger-Lewandowski, T. Breuel, Y. Chherawala, et al. Deep learners
benefit more from out-of-distribution examples. In Proceedings of the Fourteenth International
Conference on Artificial Intelligence and Statistics, pages 164–172, 2011.
[49] R. Wu, S. Yan, Y. Shan, Q. Dang, and G. Sun. Deep image: Scaling up image recognition. arXiv
preprint arXiv:1501.02876, 7(8), 2015.

A Modified CNN-Based Face Recognition System

More Related Content

What's hot (19)

Similar to A Modified CNN-Based Face Recognition System (20)

More from gerogepatton (20)

Recently uploaded (20)

A Modified CNN-Based Face Recognition System