https://doi.org/10.31449/inf.v44i4.3390 Informatica 44 (2020) 491 –495 491 
 
Face Recognition Based on Deep Learning Under the Background of Big 
Data 
Hongbiao Ni 
Department of Information Engineering, Jilin Police College, Changchun 130117, Jilin, China 
E-mail: nhbhongb@yeah.net 
Keywords: big data, deep learning, face recognition, CNN, loss function 
Received: December 8, 2020 
Face recognition has important value in real life. In this study, the application of the deep learning method 
in the field of face recognition was studied. The structure of LeNet-5 in convolutional neural network 
(CNN) was selected and improved; based on it, a face recognition method was designed. The performance 
of the method was analyzed taking CelebA as training set and LEW as testing set. The results showed that 
the improved LeNet-5 model which took A-softmax Loss as loss function not only had shorter training 
time, but also had higher recognition accuracy, its accuracy increased with the increase of sample size, 
and the highest accuracy rate reached 97.9%. The experimental results showed that the face recognition 
method designed in this study had good performance in large data background as it could effectively 
reduce the running time of the algorithm and improve the recognition accuracy. This study proves the 
reliability of deep learning methods such as CNN in face recognition, which is conducive to the further 
development of face recognition technology. 
Povzetek: Opisano je prepoznavanje obrazov z metodami globokih nevronskih mrež in z velikimi podatki. 
 
1 Introduction 
With the development of computer technology and in the 
context of big data, people pay more attention to issues 
such as data security and personal privacy, and the social 
requirements for human identification are also increasing. 
Traditional identification methods based on identity cards 
and passwords have low reliability because they are easy 
to be counterfeited and lost. Therefore, biometric 
identification technologies such as fingerprints and voices 
have been widely recognized [1]. Face recognition is a 
kind of biometric recognition, which has attracted more 
and more research and attention. However, due to the 
difference of face pose and illumination, face recognition 
is  difficult [2]. The deep learning method has excellent 
performance in face recognition, especially in big data 
processing [3], and relevant research is also deepening. 
Ding et al. [4] studied the recognition of face images with 
severe noise. Based on the deep neural network, an anti-
noise network was designed, and the reliability of the 
network in face recognition with noise was proved by 
experiments. Lu et al. [5] proposed a deeply coupled 
ResNet model, which was composed of a relay network 
and two branch networks. It could extract various possible 
resolutions of images, and the reality of the model was 
proved by experiments in LFW and SCface databases. 
Jiang et al. [6] designed an unsupervised deep learning 
network by combining 2-D Gabor filter with PCA to 
improve the computing speed through short binary 
hashing and then proved the excellent performance of this 
method by testing in face database. Singh et al. [7] applied 
convolutional neural network (CNN) to neonatal 
recognition and found that  CNN had a good accuracy in 
neonatal recognition compared with conventional 
technology and CNN with two convolution layers and one 
hidden layer had the highest accuracy. In this study, deep 
learning was analyzed. Based on CNN in deep learning, a 
face recognition method was designed. The reliability of 
the method was proved by LFW data set, which provides 
some theoretical support for the further application of 
deep learning in face recognition. 
2 Face recognition 
Face recognition refers to extracting feature information 
from static or dynamic images collected by computer and 
then analyzing and matching to realize identity 
recognition. Compared with other biometric methods, face 
recognition image acquisition is more convenient, with 
rich personal characteristics, high recognition degree and 
good interaction. It has been widely used in surveillance 
video, intelligent consumption [8], criminal investigation 
[9] and so on. 
Traditional face recognition methods include 
geometric features, template matching and so on, but there 
are also some shortcomings. Face feature extraction is a 
very important step in recognition, which has a great 
impact on the final results. In traditional recognition 
methods, feature extraction is mostly based on manual 
method. Under the background of massive data, the 
traditional recognition methods not only take a lot of time 
and energy, but also are difficult to recognize images 
because they are easily affected by illumination, occlusion 
and other factors. Deep learning can automatically extract 

492 Informatica 44 (2020) 491 –495 H. Ni 
 
features, which is less affected by external factors, and it 
has been proved to have good recognition effect. 
3 CNN 
3.1 Overview of CNN algorithm 
CNN is a common model of deep learning. Its basic 
structure is shown in Figure 1. 
(1) Convolutional layer 
Convolution layer is the core component of CNN. It 
extracts image features by convolution operation, 
generates different feature maps by different 
convolution kernels and superimposes them to obtain 
various features of input image. Its output calculation 
method is: 
𝑦 𝑗 𝑙 = 𝑓 ( ∑ 𝑤 𝑖 ,𝑗 ⊗ 𝑥 𝑖 𝑙 −1
+ 𝑏 𝑗 𝑙 𝑁 𝑗 𝑙 −1
𝑖 =1
) ,𝑗 = 1,2,⋯,𝑚 
where 𝑙 represents the current number of 
layer, 𝑤 represents the convolution kernel weight 
matrix, 𝑥 𝑖 𝑙 −1
 represents the output characteristic 
pattern matrix, 𝑓 represents an activation function,  
⊗ represents convolution operations, and 𝑏 𝑗 𝑙 
represents the offset of the 𝑗 -th characteristic pattern 
of the 𝑙 -th layer. 
(2) Pooling layer 
The role of the pooling layer is to compress data and 
reduce the amount of computation. There are two 
common methods, average pooling and maximum 
pooling. Figure 2 shows an example of maximum 
pooling. The size of image is 4×4, the size of pooling 
window is 2×2, and the step length of maximum 
pooling operation is 2. In the first pooling window, 
the values are 5, 7, 9 and 2 and the maximum value is 
9; thus the maximum pooling result can be obtained 
by traversing the whole image. 
(3) Fully connected layer 
Fully connected layer plays the role of classification, 
and its calculation formula is: 
𝛿 𝑗 𝑙 = 𝑓 ( ∑𝑥 𝑖 𝑙 −1
𝑛 𝑖 =1
𝑤 𝑖𝑗
𝑙 + 𝑏 𝑗 𝑙 ) 
where 𝑙 represents the current level, 𝑛 represents the 
number of neurons, 𝑤 represents weights, 𝑏 𝑗 𝑙 
represents offset, and 𝑓 represents an activation 
function. 
3.2 Training process of CNN 
The training process of CNN can be divided into two 
stages: 
(1) Forward propagation 
A sample (𝑋 ,𝑌 𝑝 ) is selected from the sample set, and 
𝑋 is input into CNN. 
Actual output 𝑂 𝑝 of CNN is calculated. 
(2) Reverse propagation 
(1) The error between actual output 𝑂 𝑝 and expected 
output 𝑌 𝑝 is calculated. 
(2) The error is reversely propagated, weight matrix 
is adjusted, and parameters are optimized. 
4 Face recognition based on deep 
learning 
4.1 Experimental environment 
The experiment was carried out on Ubuntu 16.04 
operating system. The program was written in C++ 
language and Python language. The training and testing of 
CNN model was realized by Caffe framework, which 
supports GPU acceleration, runs faster and operates more 
simply. 
4.2 Experimental data set 
At present, data sets commonly used in face recognition 
include CAS-PEAL, CASIA-WebFace, LFW, MSCeleb, 
CelebA and so on. In this study, CelebA was selected as 
the experimental training set, and LEW was used as the 
testing set. CelebA can train the model well as it includes 
200,000 face images of 10,177 people and there are 
changes in expression, posture, occlusion and 
illumination. LFW which has been widely used in the 
performance analysis of face recognition algorithms 
includes 13,233 images, a total of 6000 face combinations. 
4.3 Data preprocessing 
The main task of data preprocessing is face alignment. As 
the face image is partly inclined (Figure 3), the difficulty 
of recognition increases. Therefore, in order to obtain 
better recognition effect, image alignment is needed. The 
face images obtained after alignment are shown in 
Figure 4. 
 
Figure 1: The structure of CNN. 
 
Figure 2: Maximum pooling. 

Face Recognition Based on Deep Learning... Informatica 44 (2020) 491 –495 493 
 
4.4 Improved LeNet-5 
LeNet-5 is one of the most representative structures in 
CNN [10]. In order to improve the recognition 
performance of the network, the structure of LeNet-5 was 
improved in this study. Five convolution layers, four 
pooling layers and one fully connected layer were used. 
The specific parameters of each layer are shown in 
Table 1. 
In order to improve the training speed of the 
algorithm, an improved ReLU function, LReLU, was used 
as the activation function of the model: 
LReLU(𝑦 ) = {
𝑦 ,𝑖𝑓 (𝑦 > 0)
𝑎𝑦 ,𝑖𝑓 (𝑦 ≤ 0)
 
where 
a
 represents a small constant, so that the 
function is not zero when the input is negative, preventing 
neuron necrosis. 
There were two choices of loss function for the model: 
Softmax and A-softmax Loss: 
(1) Softmax: For input 
x
, it is divided into
k
 classes, then 
the probability of sample belonging to class 
i
 can be 
expressed as: 
𝑔 𝜃 (𝑥 (𝑥 )
) =
[
 
 
 
 
𝑝 (𝑦 (𝑖 )
= 1|𝑥 (𝑖 )
;𝜃 )
𝑝 (𝑦 (𝑖 )
= 2|𝑥 (𝑖 )
;𝜃 )
⋯
𝑝 (𝑦 (𝑖 )
= 𝑘 |𝑥 (𝑖 )
;𝜃 )
]
 
 
 
 
=
1
∑ 𝑒 𝜃 𝑗 𝑇 𝑥 (𝑖 )
𝑘 𝑗 =1
[
 
 
 
 
𝑒 𝜃 1
𝑇 𝑥 (𝑖 )
𝑒 𝜃 2
𝑇 𝑥 (𝑖 )
⋯
𝑒 𝜃 𝑘 𝑇 𝑥 (𝑖 )
]
 
 
 
 
 
where 𝑔 𝑒 (𝑥 ) is a hypothetical functions and 𝜃 𝑖 is a model 
parameter. 
(2) A-softmax Loss: A-softmax Loss is an improvement 
of Softmax, which introduces angular distance and 
angular margin, and its expression is: 












+
− =

i
y j
i m x i m x
i m x
ang
i
j i
i
y i
i
y i
e e
e
N
L
, cos , cos
, cos
log
1
 

,
 
where 𝑚 represents an integer, which is used for 
controlling the angular distance. 
5 Experimental results 
Images of 100 people were selected from CelebA to train 
the model, ten images each people. The training time of 
different models is shown in Table 2. 
It was found from Figure 2 that the training time of 
LeNet-5 model was longer than that of the improved 
LeNet-5 model when using the same samples. In the same 
CNN model, the training time of the model which used A-
softmax Loss as the loss function was shorter than that of 
the model which used Softmax function, and the training 
time of the improved LeNet-5 model with A-softmax Loss 
as the loss function was the least. 
Taking A-softmax Loss as the loss function, two CNN 
models were tested using LFW data sets. 100 pairs, 500 
pairs, 1000 pairs and 2000 pairs of matched face images 
were taken as positive samples; as shown in Figure 5, the 
two images matched each other, which was called a pair 
of positive samples. Mismatched face images were taken 
as negative samples; as shown in Figure 6, the two images 
did not match, which was called a pair of negative 
samples.The recognition results of the model can be 
divided into four cases, as shown in Table 3. 
The recognition accuracy of the model = (TP+TN)/ 
the total number of samples. 
Under different number of samples, the recognition 
accuracy of the two models is shown in Figure 7. 
It was found from Figure 7 that the recognition 
accuracy of the model increased with the increase f the 
 
Figure 3: Face images. 
 
Figure 4: Face images after preprocessing. 
Type 
Convolution 
kernel 
Number of 
characteristic 
patterns 
Number of 
neurons 
Convolution 
layer 1 
5×5 16 16128 
Pooling  
layer 1 
2×2 16 4032 
Convolution 
layer 2 
4×4 32 1536 
Pooling  
layer 2 
2×2 32 384 
Convolution 
layer 3 
3×3 64 128 
Convolution 
layer 4 
6×6 16 4032 
Pooling  
layer 3 
2×2 16 1008 
Convolution 
layer 5 
5×5 32 480 
Pooling  
layer 4 
2×2 32 256 
Fully 
connected 
layer 
- - 192 
Table 1: Improved LeNet-5. 
CNN model Loss function Training time 
LeNet-5 Softmax 59.27 s 
LeNet-5 A-softmax Loss 57.46 s 
Improved 
LeNet-5 
Softmax 45.39 s 
Improved 
LeNet-5 
A-softmax Loss 42.18 s 
Table 2: Comparison of training time of models. 

494 Informatica 44 (2020) 491 –495 H. Ni 
 
sample size, which showed that CNN model had excellent 
performance in recognizing massive face data and could 
accurately recognize large-scale data. From the 
comparison of the two models, it was found that the 
accuracy of the improved LeNet-5 was higher than that of 
LeNet-5. When the sample size was 4000, the recognition 
accuracy of LeNet-5 was 87.1%, while that of the 
improved LeNet-5 was 97.9%. The results showed that the 
improved CNN model could extract face features more 
comprehensively and obtain better recognition effect. 
6 Discussion and conclusion 
Deep learning is an important part of machine learning. It 
is based on big data and can automatically extract feature 
information from massive data by certain algorithms 
instead of traditional manual feature acquisition. It has 
higher accuracy than shallow learning and better 
performance in dealing with non-linear problems. It has 
shown great advantages in fields such as computer vision 
and semantic analysis. CNN is one of the deep learning 
methods, which has been widely used in object 
recognition and detection. With the support of massive 
data, face recognition based on CNN has excellent 
performance [11]. 
In this study, CNN was analyzed firstly. Traditional 
recognition methods, such SVM [12], can only extract 
shallow features when extracting image features, which is 
easily affected by other factors, and the recognition rate is 
not high. Deep learning methods such as CNN can extract 
abstract and conceptual features in depth [13], which is 
less disturbed by illumination, gesture and expression. 
CNN can extract multiple image features by convolution 
operation, then reduce the dimension by pooling layer to 
reduce the amount of calculation, and finally classify 
them. Based on LeNet-5 in CNN, the network structure 
was improved to make it more suitable for face image 
processing. Then, the improved ReLU function, LReLU, 
was used as activation function, and the influence of loss 
function on the performance of the model was analyzed. 
In the experiment, CelebA was used as training set to train 
the model, and then LEW was used as testing set to test 
the performance. The results showed that the improved 
LeNet-5 model using A-softmax Loss had shorter training 
time among LeNet-5 models using softmax and A-
software Loss as the loss function and the improved 
LeNet-5 models, which showed that it had higher 
convergence speed. Then in the processing of LFW testing 
set, A-softmax Loss was used as the loss function, and the 
recognition accuracy of the improved LeNet-5 was 
significantly higher than that of LeNet-5. The recognition 
rate of the two models increased with the increase of 
sample size, and the gap between the two models 
increased as well. When the sample size was 4000, the 
recognition accuracy of LeNet-5 was 87.1%, while that of 
the improved LeNet-5 was 97.9%. 
In summary, the face recognition method designed in 
this study has short training time and high recognition 
accuracy. It has excellent performance when facing a large 
number of face images. The reliability of deep learning 
methods such as CNN is proved, which makes some 
contributions to their further application. 
7 References 
[1] Galbally J, Marcel S, Fierrez J (2014). Biometric 
Antispoofing Methods: A Survey in Face 
Recognition. IEEE Access, 2, pp. 1530-1552. 
https://doi.org/10.1109/ACCESS.2014.2381273. 
[2] Zhang K, Zhang Z, Li Z, et al (2016). Joint Face 
Detection and Alignment Using Multitask Cascaded 
Convolutional Networks. IEEE Signal Processing 
Letters, 23, pp. 1499-1503. 
https://doi.org/10.1109/LSP.2016.2603342. 
[3] Pang SC, Yu Z (2015). Face recognition: a novel deep 
learning approach. Journal of Optical Technology C/c 
of Opticheskii Zhurnal, 82, pp. 237. 
[4] Ding Y, Cheng Y, Cheng X, et al (2017). Noise-
resistant network: a deep-learning method for face 
 
Figure 5: An example of positive sample. 
 
Figure 6: An example of negative sample. 
  Identified as 
positive 
samples 
Identified as 
negative 
samples 
Actual positive 
sample 
True Positive 
(TP) 
False Positive 
(FP) 
Actual negative 
sample 
False Negative 
(FN) 
True Negative 
(TN) 
Table 3: Classification of recognition results. 
 
Figure 7: Comparison of recognition accuracy 
between models. 

Face Recognition Based on Deep Learning... Informatica 44 (2020) 491 –495 495 
 
recognition under noise. Eurasip Journal on Image & 
Video Processing, 2017, pp. 43. 
[5] Lu Z, Jiang X, Kot C (2018). Deep Coupled ResNet 
for Low-Resolution Face Recognition. IEEE Signal 
Processing Letters, pp. 1-1. 
https://doi.org/10.1109/LSP.2018.2810121. 
[6] Jiang M, Lu R, Kong J, et al (2017). GB (2D) 2 PCA-
based convolutional network for face recognition. 
Neuroreha, 06, pp. 131-135. 
[7] Singh R, Om H (2017). Newborn face recognition 
using deep convolutional neural network. Multimedia 
Tools & Applications, 76, pp. 1-11. 
https://doi.org/10.1007/s11042-016-4342-x. 
[8] Smith DF, Wiliem A, Lovell BC (2015). Face 
Recognition on Consumer Devices: Reflections on 
Replay Attacks. IEEE Transactions on Information 
Forensics and Security, 10, pp. 736-745. 
https://doi.org/10.1109/TIFS.2015.2398819. 
[9] Ghiass RS, Arandjelovic O, Bendada H, et al (2014). 
Infrared face recognition: a comprehensive review of 
methodologies and databases. Pattern Recognition, 
47, pp. 2807-2824. 
https://doi.org/10.1016/j.patcog.2014.03.015. 
[10] Zhao ZH, Yang SP, Ma ZQ (2010). License Plate 
Character Recognition Based on Convolutional 
Neural Network LeNet-5. Journal of System 
Simulation, 22, pp. 638-641. 
https://doi.org/10.3724/SP.J.1187.2010.00953. 
[11] Wu W, Yin Y, Wang X, et al (2018). Face Detection 
With Different Scales Based on Faster R-CNN. IEEE 
Transactions on Cybernetics, PP, pp. 1-12. 
https://doi.org/10.1109/TCYB.2018.2859482. 
[12] Zhang L, Zhou WD, Li FZ (2015). Kernel sparse 
representation-based classifier ensemble for face 
recognition. Multimedia Tools & Applications, 74, 
pp. 123-137. 
https://doi.org/10.1007/s11042-013-1457-1. 
[13] Li Y, Lu Z, Jing L, et al (2018). Improving Deep 
Learning Feature with Facial Texture Feature for 
Face Recognition. Wireless Personal 
Communications, pp. 1-12. 
  

496 Informatica 44 (2020) 491 –495 H. Ni