Exeley Inc.

Zhu, Lingli

Wang, Tingzhong

International Journal of Advanced Network, Monitoring and Controls            Volume 03, No.04, 2018 
83 
Application of Improved BP Neural Network in Hybrid Control Model of Lime 
Quality 
Lingli Zhu *  
College of Information Technology 
Luoyang Normal University 
Luoyang, 471934, China 
e-mail: zhulinglils@163.com 
Tingzhong Wang 
College of Information Technology 
Luoyang Normal University 
Luoyang, 471934, China 
 
 
 
Abstract—In this paper, a method of combining neural network 
with expert system is proposed, which can realize the control 
model of real time feedback and control parameters. The model 
is based on the parameters of production condition, and the 
current lime quality is predicted. Through the quality of lime 
and control parameters, using association rule base of expert 
system, reasoning that lime control parameter adjustment, 
timely feedback to the lime production control system, to 
achieve the purpose of real-time control of the quality of lime. 
The paper presents application of improved BP neural network 
in hybrid control model of lime quality.  
Keywords-Improved Neural Network; Lime; Expert System; 
BP; Hybrid Control 
I. INTRODUCTION 
The production control process of the rotary kiln is the 
production of high quality lime under the premise of stable 
output. The automatic control system of lime rotary kiln can 
control the quality of the product by controlling the operation 
of the fuel quantity, air quantity, and equipment. However, the 
quality of lime can not be obtained in real time, and the 
production control system can not get the product quality 
information in time [1]. At present, the control system of lime 
rotary kiln is the main technology workers. Through the 
process system on-line monitoring state parameters to predict 
the current product quality, and it is as the basis for the 
adjustment of control parameters. The quality of products is 
greatly influenced by artificial judgment or off-line detection. 
In this paper, the ability of the neural network is used to 
deal with the nonlinear mapping problem, and the quality of 
the product is predicted based on the on-line monitoring state 
parameters. Combined with experience of experts in the field 
and the production of large amounts of data knowledge 
mining association rules from database and matching 
products quality prediction results and adjusting control 
parameters, obtained the lime rotary kiln control parameter 
correction information, lime rotary kiln control system for 
product quality of real time pre measurement and accurate 
feedback adjustment of intelligent control. 
II. MODEL AND KEY TECHNOLOGY OF PRODUCT QUALITY 
CONTROL OF LIME ROTARY KILN 
The product quality control model flow chart of lime 
rotary kiln is shown in figure 1. The data preprocessing 
module receives data from the automatic control system of the 
lime rotary kiln and the data stored in the database. And then 
the product quality prediction model is activated according to 
the preset time interval. Forecast to the product quality and 
the quality of the pre - set quality information. To maintain the 
original control parameter information and you meet the 
requirements. If not up to meet the requirements, combined 
with the control parameters and product quality information, 
through association rules base matching control to adjust the 
parameters of, and feedback to the automatic control system 
of lime rotary kiln. 
The activity of the product can be reduced by the lime 
burning or over burning, but the burning and burning of the 
lime is completely different from the operation, so it is 
necessary to distinguish the state of the lime. Secondly, the 
different chemical composition of raw material production of 
lime activity difference is very big, the different raw material 
must have the corresponding product quality requirements. 
Therefore, in order to make a correct judgment, the quality of 
the product must be compared with the corresponding product 
quality setting. 
BP neural network is composed of input layer (X), hidden 
layer (H) and output layer (Z). The hidden layer (H) can be 
one or more layers, and the article chooses only one layer of 
hidden layer of neural network [2]. For the three layer BP 
neural network, the hidden layer node number can be 
determined by using the following formula. 
p m n 

Where: m is the input layer node number, n is the output 
layer node number, P is the hidden layer node number. 
BP neural network is a kind of iterative algorithm for the 
forward propagation of data and the error back propagation. 
The input signal propagates through the activation function to 
the hidden layer, and the hidden layer nodes are transmitted to 
the output layer by the activation function. By calculating 
output between neurons and the expected output error, in 
accordance with the negative gradient of error modify the 
weight of each layer, error signal along the path to the original 
return, to calculate the output, as is shown by figure1, until the 
error achieve ideal value. 
DOI: 10.21307/ijanmc-2019-025
International Journal of Advanced Network, Monitoring and Controls            Volume 03, No.04, 2018 
84 
 
Figure 1. Flow chart of control mode 
Lime rotary kiln control system online condition 
monitoring parameters (x) many, some reflect the process of 
production condition parameters of the system, some safety 
protection monitoring parameters, more two features in both. 
According to the actual production experience, exhaust gas 
temperature (x1), the tail of the kiln temperature (x2), kiln 
pressure (x3), secondary air temperature (x4), feed 
temperature (x5) are reflected lime quality is the most 
significant factor. 
The control parameters (Y) of the lime rotary kiln are 
essentially determined by the quality of the product. Including 
fuel consumption (Y1), rotary kiln rotation speed (Y2), the 
main exhaust fan (Y3) and other control parameters is the 
product quality of the decisive parameters. Product quality 
information (Z) contains the lime activity (z1) and the (z2) 
[3][4][5]. First, it is the higher the activity of lime, the better 
the quality of the product. In the ideal state, there is no 
maximum level of activity of the lime, which has the highest 
degree of burning or burning. 
In the process of production of lime rotary kiln, there are 
inevitable abnormal conditions, such as mechanical 
equipment failure, monitoring equipment abnormal and so on 
will cause the distortion of online detection data. If you do not 
filter out the data, it will cause the system model to send out 
the wrong instruction, affecting the normal operation of the 
system. For this kind of abnormal data points to take filtering, 
mean filling and other means to ensure the accuracy and 
integrity of the data. 
III. CONSTRUCTION OF EXPERT SYSTEM RULE BASE 
Prediction to the quality of the product is not ideal, the 
need to adjust the control parameters to achieve the purpose 
of product quality control. Through the neural network 
prediction period of product quality information model is 
obtained, and presupposition information after comparing the 
parameters (
1 1 1
j jz z z  
, 2
jz ) and control parameters 
( 1 2 3, ,
j j jy y y ) as the association rule premise, matching the 
amount of correction of the control parameters as output item 
( 1 2 3, ,
j j jy y y   ), then optimized control parameters in the next 
time interval for: 
1j j j
n n ny y y
   
（，，）
BP neural network algorithm is as follows: 
(1)Initialization. In the form (2), the value of the license is 
randomly assigned to the initial value of the network, setting 
the number of learning times N and calculating precision. 
(2)A study sample K input and the corresponding 
expected output. 
(3)The output value of I node is calculated. 
   
m
ai
a 1
k k
i a ih f x 

 
  
 


(4) J node output value of the output layer. 
   
j i
1
o
p
k k
i j
i
z f h 

 
  
 


(5) Calculating the weights between the output layer and 
the hidden layer. 
     1 +1 1 +1i j i1 +
k kN N N N N N
i ib h b    
    

Which is the learning rate, B is the additional momentum. 
The value of the judgment is determined by the formula (3) 
and (4) [6]. 
(6) Calculating the weights between the hidden layer and 
the input layer. 
International Journal of Advanced Network, Monitoring and Controls            Volume 03, No.04, 2018 
85 
     1 +1 1 +1ai i a1 +
k kN N N N N N
ai aib x b    
    

(7)Judgment. Sample not read, jump second steps. 
Otherwise the next step. 
(8)According to the corrected formula weights (5) (6), 
fixed weights. The output of each layer is calculated by the 
formula (7). 
(9)To calculate the global error. 
    
2K n
1 j 1
1
2K
k k
j j
k
E z zo
 
 

(10)Judgment. Does the error reach the design accuracy or 
exceed the maximum number of training sessions, and then 
the error is. No, then make N=N++, return second steps. 
The model association rule mining algorithm using OLAP 
and Apriori algorithm, to find the relationship between 
product quality information and control parameters 
adjustment. Product quality information and control 
parameters of the association rules mining, with the help of 
the production process by the control parameters to adjust the 
information on the quality of the product data. 
On-line analytical processing technology is the main tool 
of data analysis in data warehouse. The OLAP system uses the 
data storage scheme of materialized data cube, in exchange 
for the space cost for time efficiency. For association rule 
mining, a large amount of intermediate data which is used in 
the mining process may already be materialized in the data 
cube and not need to be re calculated. 
IV. RUNNING EXAMPLES OF NEURAL NETWORK FORECAST 
MODEL 
Sample data is from a steel mill Nissan 800 tons of lime 
rotary kiln production line. The enterprise product quality 
detection is per hour from the sampling hole sampling time, 
artificial naked eye observation quality of products; sampling 
every 3 hours on the quality of the products by experimental 
method determination of a, so the sample selection 
experimental measurement data shall prevail. Select the 
normal operation of the equipment in the case of 30 days and 
240 data for the sample, of which 120 were used as training 
samples, 40 as a test sample, as is shown by table1. 
TABLE I. SAMPLE 
NO. The 
temperat
ure of 
exhaust 
gas 
（℃） 
The 
temperat
ure of 
inlet 
（℃） 
The 
pressure 
of inlet 
（Pa） 
The 
temperat
ure of the 
second 
air 
（℃） 
The 
temperat
ure of 
product 
（℃） 
Activity 
（ml） 
Product 
state 
1 
280 
1009 -57.78 536 171 
360 
Over-burn
ed 
2 
286 
1014 -23.64 625 143 
351 
Over-burn
ed 
3 
273 
983 -25.54 630 95 
335 
under-bur
ned 
… … … … … … … … 
116 
270 
949 -39.46 578 155 
329 
under-bur
ned 
117 
261 
954 7.42 571 134 
353 
under-bur
ned 
118 
282 
1002 -51.04 630 127 
365 
Over-burn
ed 
119 
267 
967 -20.32 547 109 
344 
under-bur
ned 
120 
282 
1006 -55.68 628 152 
357 
Over-burn
ed 
… … … … … … … … 
159 
286 
996 -12.5 627 141 
354 
under-bur
ned 
160 
274 
949 -12.01 597 112 
341 
under-bur
ned 
 
Definition1: The sum rule for the attribute values of the 
same attribute in the 1 N dimensional data is as follows: 
Let SUM=Attribute.Value1+Attribute.Value2 
If Attribute.Value1 is not empty and 
Attribute.Value1==Attribute.Value2, then SUM=N+1; 
If Attribute.Value1 is null and Attribute.Values2 is not 
empty, or Attribute.Value1 is not empty and Attribute.Value2 
is empty, then SUM=1; 
In other cases, SUM=0. 
The rule can be used to determine whether the 2 frequent 
attribute sets satisfy the connection conditions. To judge two 
k-item frequent attribute set Trans1 and Trans2 whether that 
International Journal of Advanced Network, Monitoring and Controls            Volume 03, No.04, 2018 
86 
satisfies the join condition, and Trans2 Trans1 the 
corresponding added and all the sum of the results 
accumulated the resulting value to result. 
Write n for the number of fields in the data table, if 
Result= (n + 1) (k-1) + 2, and Trans2 Trans1 a (k-1) of non 
null attribute values are the same, two attribute values are 
different, and different two attribute values and are 1. That 
was non null attribute value and attribute values are added. At 
this point, Trans1 and Trans2 meet the connection conditions, 
can be connected operation. Otherwise, Trans1 and Trans2 
can not be connected operation. 
The 2 meet the connection conditions of frequent sets 
Trans1 and Trans2, can be achieved by the "add" connection 
operation. 
In order to carry out the operation of the frequent attribute 
set, the "or" rule of the value of the attribute is defined first. 
From the production site, the use of the quality control 
system before and after 1 month of the 240 measured lime 
activity contrast as shown in Figure 2. 
 
320
325
330
335
340
345
350
355
360
365
370
375
380
0 14 2
8 42 56 70 84 9
8
11
2
12
6
14
0
15
4
16
8
18
2
19
6
21
0
22
4
23
8
The sample
Ac
ti
vi
ty
After
Before
 
Figure 2. The contrast of product quality before and after used the system 
From Figure 2 it can be seen that the use of the quality 
control system, the average lime activity is about 370ml, 
compared with the previous product, the average 340ml level 
has greatly improved. In addition to the specific time period, 
the change of the activity of the lime is kept within 10ml, 
which indicates that the stability of the lime production 
system is greatly improved. 
V. CONCLUSION 
This paper presents a combination of neural network and 
association rule library based on lime rotary kiln products 
quality prediction and control parameters feedback control 
model can be in line completed the accurate prediction of the 
quality of the products and to achieve timely control 
parameter adjusting information feedback. In the control 
model, the process of reading and processing the data of the 
state parameter is added to improve the robustness of the 
control model, and reduce the influence of equipment failure 
on the system stability. 
ACKNOWLEDGEMENTS 
This paper is supported by Science and technology project 
of Henan Province (152102210123, 142102210475). 
 
REFERENCES 
[1] Liu Xiao-li, Wang Shen-tao, Dai Rui, Wei Shi-feng, “On fault 
diagnosis of net work based on    improved particle swarm 
optimization”, Computer applications and software, vol.28, no.1,   
(2011),pp.207-209. 
[2] Wang Cheng-liang,Pang Xu,Lu ZHi-jian, “Research on hybrid control 
methods of aluminium reduction on neural  network”, Application 
Research of Computer, vol.27,no.7, (2010),pp.2536-2539. 
[3] Li Dao-zhong,ZHang Lei,Wu ZHao-hui, “Test method of physical 
metallurgical lime, Bei Jing:Metallurgical Industry 
Press,(2005),pp.1-5. 
[4] Chu Hui, Lai Hui-cheng, “An improved back-propagation N 
algoriythm and its application”, Computer simulation,vol.24, no.4, 
(2007), pp.75-111. 
[5] Shi ZHong-zhi, “Knowledge discovery”,Bei jing:Tsinghua University 
press, (2002),pp.8-9. 
[6] Zhang Lei, Hu Chun, Qian Feng, “Research progress of BP algorithm 
for local minimum problem”, The industrial control computer, 
vol.17,no.9, (2004),pp.33-50. 


English

Application of Improved BP Neural Network in Hybrid Control Model of Lime Quality

Yubian, Wang

Shebzukhov, Yuri

International Journal of Advanced Network, Monitoring and Controls          Volume 03, No.04, 2018 
97 
Improvement and Realization of CamShift Algorithm Based MotionImage Tracking
Wang Yubian*  
Department of Railway Transportation Control 
Belarusian State University of Transport, 
34, Kirova street, Gomel,246653, Republic of Belarus 
*is the communication author. 
e-mail: alika_wang@mail.ru 
 
Yuri Shebzukhov 
Department of the International Relations 
Belarusian State University of Transport,  
Republic of Belarus  
34, Kirova street, Gomel, 246653, Republic of Belarus 
e-mail: oms@bsut.by 
 
Abstract—The detection and tracking technology of moving 
object image is one of the key technologies of computer vision, 
and is widely used in many fields such as industry, 
transportation, and military. The detection and tracing of 
moving object in the motion scenes based on the UAV 
platform is a technical difficulty in this field. In practical 
applications, the characteristics of complicated environment, 
small target object and moving platform require higher 
real-time performance and reliability of the algorithm. If it is 
possible to add some other features of the target to the 
tracking process, it will be possible to improve the 
shortcomings of Camshift itself. Based on the Camshift 
tracking algorithm, this paper integrates SURF feature 
detection into the algorithm, which greatly improves the 
tracking accuracy of the target object, and also has better 
real-time performance which can achieve better tracking 
performance of the object. 
Keywords-Motion Image; CamShift Algorithm; SURF; 
Feature Detection 
I. INTRODUCTION 
Video tracking is the key technology for motion target 
object detection in dynamic sceneson UAV platform. It can 
be realized by two methods: one is based on target 
recognition technology, of which the core concept is 
frame-by-frame recognition algorithm for motion video to 
identify the target object and determine the target matching. 
The other is based on the detection technology of the 
moving object, of which the core concept is the active 
detection of the moving object, and the position of the 
moving object is determined in accordance with the 
detection result to realize the tracking.This method can 
achieve the tracking of any moving object without the need 
of complicated priori information for detection, such 
ascharacteristics of object shape, object sizes. However, the 
tracking effect of various tracking algorithms also depends 
on the background migration of the object, the unpredictable 
tracking path, the unpredictable target motion path and 
mode, the scene switch, the target object movement does not 
have an analyzable pattern, the change of camera model, the 
camera shift, and the change  of illumination condition. 
And the causes of the changes in the color and shape of the 
moving object are very different [1]. The current 
mainstream motion object tracking methods with good 
tracking performance mainly include feature-based tracking 
methods, region-based tracking methods, model-based 
tracking methods, motion estimation-based tracking 
methods, and contour-based tracking methods. 
The detection and tracking algorithm of the conventional 
motion object is only suitable for the scene with static or 
almost static background, but it is not suitable for the 
detection and tracking of moving objects in the UAV video. 
Therefore, the digital sequence information of the video 
images acquired by the drone should meet the real-time, 
Accurate, robust and other requirements [2]. The current 
research shows that the target object tracking method based 
on Camshift algorithm can meet the requirements of drone 
video target tracking. 
The CamShift algorithm performs target object 
recognition and tracking by analyzing the Hue component 
DOI: 10.21307/ijanmc-2019-028
International Journal of Advanced Network, Monitoring and Controls          Volume 03, No.04, 2018 
98 
information of the target region in the HSV color space. The 
target has low deformation sensitivity. The algorithm has 
good real-time performance, little computationand low 
complexity. Therefore it has been extensively studied. 
In the video target tracking by drone, the Camshift 
algorithm [3],comparing to other tracking methods, has 
many advantages and disadvantages. Under the condition 
such as the similar color setting to the target object in the 
background or in the complicated scene, the Camshift 
algorithm may have tracking error or fail to track because of 
the tracking characteristicsthat the algorithm is based on the 
color information of the moving object. If other auxiliary 
features of the object can be obtained during the search 
process and input conditionsof algorithm are specified, it is 
possible to make up for the problems caused by Camshift in 
such scenes. Because the SURF algorithm has the advantage 
of good object recognition, the implementation of the 
SpeedUp Robust Features (SURF) in the Camshift 
algorithm will greatly improve the tracking accuracy and the 
tracking reliability of the object. The improved method 
ensures the good real-time performance of the tracking, 
greatly improves the tracking accuracy of the object by 
UAV, and eventually achieves a better tracking effect of the 
moving object. 
II. THE CHARACTERISTICS OF SURF 
The tracking of the characteristics of the target object 
isbased on the tracking of point of interest, which is often 
used in engineering applications and it works well. The 
difficulty of this method lies in the selection and extraction 
of features [4]. The selection of features should be able to 
fully cover the key features of the target object in different 
scene settings, and such features shall also be convenientfor 
extraction. In general, if the number of sampling points is 
insufficient in the process of extracting features, it is easy to 
lose tracking of the object and the tracking effectwould 
deteriorate. On the contrary, the calculation amount and 
complexity will be greatly increased, and unableto satisfy 
the actual application. 
Although the Harris corner detection method is a 
traditional point of interest detection method, due to its 
fixed scale characteristics, it is becoming difficult to 
determine the change ofposition of the target object between 
image frames when the target object is deformed or there is 
change in dimensions. Prof. David G. Lowe of the 
University of British Columbia in Canada first proposed the 
Scale Invariant Feature Transform (SIFT) [5]. The SIFT 
algorithm for querying feature key points is to use feature 
detection within the constructed scale. The orientational 
selection of point of interest is calculated according to the 
gradient of its neighborhood, thus achieving the goal that 
the feature in the scale does not change in accordance with 
the orientation. However, the algorithm requires high 
computational complexity, and has high requirements for 
hardware devices, and a shortcoming of poor real-time 
computation. On the basis of this algorithm, Yan Ke et al. 
introduced the Principal Components Analysis(PCA) in the 
SIFT system and proposed the PCA-SIFT algorithm, which 
profoundly improved the matching inefficiency of the 
common SIFT algorithm. However, this method will lead to 
the failure of feature extraction in the later stage as well as 
the deterioration of the characteristic distinctiveness. 
On this basis, some people have proposed a point of 
interest algorithm based on SURF [6]. The extracted 
features are used as the key local features of the images. The 
speed of the fast calculation is improved by the integral 
image method. Then the point of interest obtained by Haar 
wavelet transform are used to obtain the main orientation 
and its feature vector [7]. Finally, the Euclidean metric is 
calculated to verify the matching effect between the images. 
The SURF feature is invariant in the presence of changes in 
brightness, translation, rotation, and scale. In addition, the 
method does not show any negative effect to its robustness 
under noise interference conditions even if the viewing 
angle changes. This method not only realizes theaccurate 
feature recognition, but also reduces the computational 
complexity, which greatly improves the inefficiency of the 
method in use, and has broad application. 
 
International Journal of Advanced Network, Monitoring and Controls          Volume 03, No.04, 2018 
99 
III. THE SURF ALGORITHM 
A. Constructing Hessian matrix 
SURF relies on approximation of the determinant 
images of Hessian matrix. The Hessian matrix is the core of 
the SURF algorithm [8]. First, the Hessian matrix of each 
pixel is calculated according to the equation, and then the 
Hessian matrix discriminant is used to determine whether 
the point is an extremum. Gaussian filtering is used to 
construct the Gaussian pyramid [9]. Compared with the 
Gaussian pyramid construction process of SIFT algorithm, 
the speed of SURF algorithm is improved. In the SURF 
algorithm, the image space occupancy of each set is 
different. The previous set of images is downsampled by 1/4 
to obtain the next scale. At the same time, the same set of 
images has the same size, and the difference is that different 
scales σ are used. In addition, in the process of blurring, the 
Gaussian template size is always constant, and only the 
scale changes σ. For the SURF algorithm, the image size 
remains the same, only the Gaussian blur template size and 
scale σ need to be changed. 
B. Preliminary determination of point of interestusing 
non-maximum suppression 
Each pixel processed by the hessian matrix is compared 
with its 26 points in the three-dimensional domain. If it is 
the extremum of these points, it shall beselected as a 
preliminary point of interest for the next process [10]. The 
detection process uses a filter corresponding to the size of 
the scale image to detect. In this paper, a 3×3 filter is used 
as an example for detection analysis. The point of interestfor 
detection which is one of the 9 pixel points in the scale and 
the remaining 8 points in the scale is compared with the 9 
pixel pointsin each of theadjacenttwo scale layers above and 
below it to complete the calculation of 26 points in the 
three-dimensional domain. 
C. Precisely locate extremum 
The three-dimensional linear interpolation method 
removes the points whose valuesare less than a certain 
threshold, but the sub-pixel point of interest can be obtained, 
and the extremumbe increased, the number of detected point 
of interest is therefore reduced, and finally several 
highly-featured points are detected and the amount of 
workis reduced [11]. 
D. Select the main orientation of the point of interest 
In order to ensure the rotational invariance, the gradient 
histogram is not countedin the SURF, but the Haar wavelet 
response around the point of interestregion is computed [12]. 
That is, centering on the point of interest, within the circular 
neighborhood of a radius of 6s (s is the scale of the point of 
interest), the Haar wavelet side length is 4s, and the Haar 
wavelet response of all points in the 60 degree fan in both x- 
and y-directions is computed. Sum, and assign a Gaussian 
weight coefficient to the Haar wavelet responses, to improve 
the contribution of the response close to the point of interest, 
suppress the contribution of the response away from the 
point of interest, and then add the response within the range 
of 60 degrees, eventually yielding a new vector, traversing 
the whole circular region. The main orientation of the point 
of interestwas defined by the orientation of the longest 
vector [13]. 
E. Construct descriptor of the SURF point of interest[14] 
Extract a square region around the point of interest. The 
size of the window is 20s. The orientation of the region is 
the main orientation detected in step 2.4. The region is then 
divided into 16 sub-regions, each of which the Haar wavelet 
responses of the x- and y-direction of 25 pixels are summed, 
where the x- and y-directions are relative to the main 
orientation. The Haar wavelet response is the sum of the 
horizontal direction, the sum of the absolute values in the 
horizontal direction, the sum of the vertical directions, and 
the sum of the absolute values in the vertical direction. 
IV. EXTRACT AND MATCH THE POINT OF INTEREST 
（1）Select a frame from the drone tracking video and 
extract the points of interest using SURF detection method, 
as shown in Figure 1. 
 
 
 
 
International Journal of Advanced Network, Monitoring and Controls          Volume 03, No.04, 2018 
100 
 
 
Figure 1. Extract points of interest 
（2）Matching the target region. After selecting the 
target region, extract the target regions in the adjacent two 
frames and match the points of interest of the two frames. In 
Figure 2, it can be seen that 6 points of interest are 
successfully matched. 
 
 
Figure 2. Matching the points of interest 
V. VERIFICATION 
After manually selecting the target window, the 
feedback mechanism is used to calculate the color similarity 
between the Camshift tracking window and the initial 
window, and the feature similarity between the SURF 
tracking window and the initial window. Suppress the major 
displacement, and assign the weight of displacement 
dynamically to the two tracking algorithms in accordance 
with Bhattacharyya distance [15]. The Camshift tracking 
algorithm is preferred when the motion tracking is stable, 
and otherwise, the SURF tracking method [16] is preferred. 
The experiments show that the tracking method is ideal for 
tracking, and it can solve the problem of tracking 
interference caused by background changes and object 
similarity. Takes a picture of tracking object every 15 
seconds, and an example is shown in Figures 3, 4, 5, and 6. 
 
 
Figure 3. Tracking image 1 
 
Figure 4. Tracking image 2 
 
Figure 5. Tracking image 3 
 
International Journal of Advanced Network, Monitoring and Controls          Volume 03, No.04, 2018 
101 
 
 
Figure 6. Tracking image 4 
As shown in the above figures 3, 4, 5, and 6, the drone 
has achieved good results in tracking the moving object. 
VI. CONCLUSION 
Based on the classic tracking algorithm Camshift and 
focused on the deficiencies of the algorithm in the motion 
object tracking of drone video, this paper proposes the 
combination of Camshift algorithm and SURF feature 
detection to realize the tracking of motion object in UAV 
video. The experimental results show that the proposed 
method can effectively track and locate the target object in 
the more complex aerial photography background. The 
experiment has achieved good results, basically realizes the 
tracking of the moving object, and the tracking speed is fast, 
thus the real-time performance is satisfactory and the 
timeconsumption is little.  
However, in practical applications, the environment of 
the object tracking is more complicated and diverse. The 
further study of this paper can be carried out in the 
following respects: This paper only studied the tracking of a 
single object, which does not involve the tracking of 
multiple similar objects. The multi-object tracking has 
practical significance in video surveillance, intelligent 
traffic detection, air formation, and geographic monitoring. 
Therefore, it is very necessary to further study the tracking 
of multiple objects; the object tracking system in this paper 
needs to be improved, including optimization of the logic 
system and add the feature of parallel processing to improve 
the tracking efficiency of the system. The Camshift 
algorithm in this paper is still based on color histogram, 
which is sensitive to changes in illuminance and the change 
in color of objects. When the resolution of the camera is not 
high and the ambient light is insufficient, the tracking effect 
is not good,therefore the study should focus on the tracking 
of physical features that are insensitive to illumination. 
 
REFERENCES 
[1] Liu yanli, tang xianqi, Chen yuedong. Application research of 
moving target tracking algorithm based on improved Camshift [J]. 
Journal of Anhui Polytechnic University,2012, 27(2):74-77. 
[2] Xiong TAN, Xuchu YU, Jingzheng LIU, Weijie HUANG. Object 
Fast Tracking Based on Unmanned Aerial Vehicle 
Video[C]//Proceedings of PACIIA, IEEE Presss,2010:244-248. 
[3] C. Harris, M. J. Stephens. A Combined Corner and  Edge 
Detector[C]. Prco of the 4th Alvey Vision Conf, 2013: 147-152. 
[4] D. G. Lowe. Distinctive Image Features from Scale-Invariant 
Keypoints[J]. International Journal of Computer Vision.2014, 60(2): 
91-110. 
[5] C. Harris, M. J. Stephens. A Combined Corner and  Edge 
Detector[C]. Prco of the 4th Alvey Vision Conf, 2015: 147-152. 
[6] D. G. Lowe. Distinctive Image Features from Scale-Invariant 
Keypoints[J]. International Journal of Computer Vision.2014, 60(2): 
91-110. 
[7] Liu yawei. Review of target detection and tracking methods in uav 
aerial photography video [J]. Airborne missile,2016.(9):53-56. 
[8] Yan K,Sukthankar:a more distinctive representation for local image 
descriptors[C]//ProceedingsofCVPR,Los Alamitos,IEEE 
press,2014:268-235. 
[9] Bay H,Ess A,Tuytelaars T.Speeded up robust 
features(SURF).[J]Computer Vision and Image 
Understanding,2007,110(3):346-359. 
[10] Leutenegger S,Chli M,Siegwart R.BRISK:binary robust invariant 
scalable keypoints.[C]//Proceedings of ICCV,IEEE 
Press,2013:326-329. 
[11] Cui zhe. Image feature point extraction and matching based on SIFT 
algorithm [D]. Xi 'an: XiDian: university.2016:38-46. 
[12] Yu huai, Yang wen. A fast feature extraction and matching algorithm 
for uav aerial images [J]. Journal of electronics and information 
technology, 2016, 38(3):509-516. 
[13] Wang jianxiong. Research on key technologies of low altitude 
photogrammetry of unmanned airship and practice of large scale 
map formation [D]. Xi 'an: chang 'an university, 2011:36-48. 
[14] Li yifei, research on PID control in four-rotor aircraft,[J] technology 
and market, 2016.07:90-91. 
[15] Li xiang, wang yongjun, li zhi, misalignment error and correction of 
attitude system vector sensor,[J] journal of sensor technology, 
2017.02:266-271. 
[16] Wang donghua, yue dawei. Design and implementation of large 
remote sensing image correction effect detection system,[J] 
computer programming skills and maintenance,2015.12:118-120. 
 


Improvement and Realization of CamShift Algorithm Based MotionImage Tracking

PARKHATS, Vadzim

KRZYWOŃ, Rafał

EVALUATION OF TIME DEPENDENT PRESTRESS LOSSES IN PRETENSIONED CONCRETE MEMBER WITH TOP AND BOTTOM TENDONS

Majkowska-Zwolińska, Beata

Jędrzejczak, Joanna

7
E
d
it
o
r
ia
l
In memoriam
ogy, Queen Square, London (1958), Montreal Neuro-
logical Institute (fellowship of Rockefellers Founda-
tion) (1959–1960), Brain Research Laboratories, New 
York Medical School (1964–1965), California Institute 
of Technology, Pasadena (1965–1966).
 In the years 1952–1975 he worked at the Depart-
ment of Neurology at the Medical University of War-
saw; he was the head of the EEG and Neurophysiolo-
gy Laboratory and between 1973–1990 he continued 
as the head of the Experimental Neurology Laboratory 
of the Medical Academy in Warsaw were he conduct-
ed experimental research in the field of neurophysiol-
ogy and learning processes.
In 1979–1986 he was the chairman of Department of 
Clinical and Experimental Neurophysiology and Sei-
zure Unit, New York State Institute for Basic Research 
in Developmental Disabilities, and visiting professor 
at New York University, Medical Center. After return 
to Poland in 1988 he created the first in Poland De-
partment of Neurology and Epileptology and was its 
head until 1998.
After retirement he worked at Foundation of Epilep-
tology and Epilepsy Diagnostic and Therapeutic Cent-
er, which he established in 1993. He focused his ener-
gy on medical, scientific, educational and most of all 
publishing work. From 1993 he published and was ed-
itor-in chief of the quarterly Epileptologia, which since 
2010 has evolved into an international journal in Eng-
lish, and then in 2013 to take the current form of the 
Journal of Epileptology.
Professor’s scientific research focused on the follow-
ing issues: electrophysiological exponents of learning 
and memory in healthy animals and with damage of 
various brain structures; electrophysiological studies 
of epileptogenesis of experimental epilepsy (aluminum 
and kindling models), protective role of glia in the pre-
vention of post-traumatic epilepsy, assessment of pro-
phylactic and therapeutic effect of antiepileptic drugs 
in animal models and in the clinical practice.
 Professor Majkowski was involved in organization-
al academic activities. In 1974–1979 he was the organ-
izer and the first Dean of the Second Faculty of Medi-
cine of the Medical Academy in Warsaw. He was also 
very active in many scientific societies. He was the or-
ganizer and the first president of the Polish Society of 
P rofessor Jerzy Majkowski teacher and mentor of many generations of Polish neurologists and epileptologists and editor-in-chief of Journal of 
Epileptology passed away on July 18, 2019 in Warsaw 
at the age of 91.
The professor’s years of youth fell during World War 
II. He was 11 years old when the war broke out and 
throughout the occupation, he took an active part in the 
fight. In the Warsaw Uprising he was one of the young-
est soldiers. After the uprising he was taken prisoner in 
a prisoner-of-war camp in Germany. After the libera-
tion, he served in the Polish Armed Forces in Hameln 
in Germany and there, he continued his interrupted 
high school education.
After returning to Poland in 1947, he studied at the 
Faculty of Medicine of the University of Warsaw. He 
obtained his medical diploma in 1952 and became 
a specialist in the field of neurology. He underwent 
also extended abroad training at Dep. of Neurology, 
Karol University, Prague (1954), Institute of Neurol-
Copyright and photocopying by Foundation of Epileptology, 2019
Journal of Epileptology • 2019 • 27 • 7–8 • 10.21307/jepil-2019-007
8
Electroencephalography and Clinical Neurophysiolo-
gy, First president of the Polish League Against Epilep-
sy and from 2001 – the first vice president of the Pol-
ish Society of Epileptology. Therefore he is considered 
a father of the Polish epileptology. In 2007 he organ-
ized and became the first President of the Federation 
of Polish Medical Societies. He felt strongly bound to 
organizational and scientific work for the societies, in 
which was active until the end of his life. At the inter-
national League Against Epilepsy, he was an extremely 
active member of the Commission for European Affairs, 
Executive Board of the European Academy of Epilep-
sy, Commission on Therapeutic Strategies, key mem-
ber, chairman of Subcommission on Polytherapy and 
Drug Interactions, Commission on Classification and 
Terminology, member of the Council of European An-
tiepileptic Leagues. he received the Ambassador for Ep-
ilepsy award of the ILAE.
 Professor Majkowski’s publications include more 
than 500 publications, 11 books mainly in experimen-
tal models, clinical epilepsy, clinical trials of AED, clin-
ical and experimental electroneurophysiology, epileptic 
seizures and cognitive functions impairment.
The professor received many awards and honors for 
scientific work and military awards as well, including 
the last received before his death – Honorary Citizen 
of the Capital City of Warsaw.
He will be remembered for his inexhaustible ener-
gy and passion in teaching and passing on his experi-
ence to all of us who had this privilege to know him.
Beata Majkowska-Zwolińska & Joanna Jędrzejczak
•
In memoriam


In memoriam

Bhoil, Rohit

Kaushal, Sushruti

Sharma, Reena

Kaur, Jaswinder

Sharma, Tanupriya

Thakur, Rohini

Sharma, Rakesh

Color Doppler ultrasound of spiral artery blood flow in mid first trimester (4–8 weeks) in cases of threatened abortion and in normal pregnancies

Idowu, Bukunmi Michael

Onigbinde, Stephen Olaoluwa

Ebie, Isaiah Uzezi

Adeyemi, Michael Temidayo

Gallbladder diseases in pregnancy: Sonographic findings in an indigenous African population

Emslie, Nicola

THE ELEVATED RISK OF MELANOMA AMONG PILOTS – COULD UVA BE IMPLICATED?

Beebeejaun, Adam

Roby, Howard

DANGERS OF DEFIBRILLATION IN FLIGHT 
Adam Beebeejaun 
Intensive Care and Aeromedical Retrieval Registrar 
Dunedin Hospital, New Zealand 
BMBS PGCertAeroRT (Correspondence Author) 
Howard Roby 
Consultant Anaesthetist and Aeromedical Retrieval 
University of Otago, New Zealand 
BSc (Med) MBBS FANZCA FCICOU D Au Med RAAF 
 
 
ABSTRACT 
Aim 
Defibrillation is a critical aspect of advanced life 
support, but the inherent risks in the procedure are 
increased when used during an aeromedical 
retrieval. The dangers of defibrillation in flight can 
be divided in to fire, electrical, avionic interference 
and physical carriage and packaging. A limited 
body of evidence exists concerning defibrillation 
in flight, in part, due to under-reporting. Changes in 
incident reporting, increased team based 
simulation training and awareness of the dangers 
of defibrillation should allow aeromedical teams 
to defibrillate patients safely and expediently. 
 
INTRODUCTION 
Defibrillation is the termination of ventricular 
fibrillation (VF) or ventricular tachycardia (VT) after 
the delivery of an electrical current across the 
myocardium enabling restoration of coordinated 
electrical activity.1 
The use of electricity to resuscitate a patient was first 
described in animal models in the 1770s, however it 
was not until 1960s that Zoll demonstrated the 
effective use of externalised electricity to resuscitate 
a patient.2 Due to severe pain and burns, several years 
passed before technology allowed for defibrillation to 
be carried out safely and effectively.  
Early defibrillation has now become a core aspect of 
advanced life support and is one of the few 
interventions proven to improve outcomes from 
cardiac arrest secondary to ventricular fibrillation 
(VF) or pulseless ventricular tachycardia (VT).3 
Since the probability of successful defibrillation is 
time dependent, advances in technology and 
reduction in cost has allowed the procedure to be  
 
conducted outside of hospital.4 The rate of cardiac 
arrest during aeromedical transport is between 3.4- 
5% with a requirement for defibrillation occurring in 
0.8% of all missions.5 
 
Defibrillation can cause significant harm to both 
patient and health practitioners unless appropriate 
measures are undertaken to reduce the risk. When 
defibrillation is attempted during an aeromedical 
transport the risk increases. The dangers of 
defibrillation can be divided into four main categories: 
fire, electrical, avionic interference and physical 
carriage and packaging. 
 
FIRE 
Fire is a rare but devastating consequence of 
defibrillation however this risk is not mentioned in 
standard critical care or cardiology literature.6,7 If fire 
were to occur during an aeromedical evacuation, it 
could result in the loss of the patient and crews’ life 
and loss of the aircraft.  
The causes of fire can be further divided into two 
categories: 
1. Electrical Sparks and Arcs 
 
The production of any spark in an oxygen rich, 
enclosed environment has the risk of igniting a 
catastrophic fire. A case report from 2003 
demonstrated a fire from a spark formation during 
defibrillation. The anterior-posterior placement of 
defibrillation resulted in a spark which set the 
bedding, cotton wool and wool cap on fire.8 
Sparks can form when pads are improperly placed 
resulting in a reduced electrode-chest wall interface. 
The spark produced can ‘jump’ between conductors. 
This is called an electrical arc. Arcing of electricity 
during defibrillation can occur from several 
scenarios: 
• Poor skin electrode interface 
• Excess conducting gel forming a conduit between 
the pads 
• Paddles/pads placed too close to ECG electrodes 
• Other conduits present – saline soaked pads 
(surgical/open wounds) 
 
The spark or electrical arc is a source of heat and the 
first part of the triangle of fire. 
Oxygen, the second part of the triangle of fire, is 
always in close proximity to the patient during 
defibrillation and all case reports of fires during  
 
 
 
defibrillation have occurred in oxygen rich 
settings.9,10,11 Within an aircraft cabin, space is limited 
and moving oxygen far enough away from the patient 
may not be enough to reduce the oxygen 
concentration around the defibrillation paddles. 
Lower barometric pressure may reduce the 
combustion potential; however, this is potentially 
offset by the presence of oxygen delivered to the 
patient.  
The final part of the triangle of fire is fuel which, in 
effect, is anything that can burn. Many potential fuel 
sources exist at the beside including cleaning agents, 
linens, dressings, ointments and patient body hair.12 
Specific fuel sources within the aeromedical 
environment include aviation fuel, medical dressings, 
patient hair/skin and a variety of rubber supplies.  
2. Lithium battery failure 
 
Lithium-ion batteries have revolutionised portable 
medical devices providing increased power and 
longer life with reduced size and weight. Fires occur 
when one of the battery cells are damaged, punctured 
or overheated by overcharging or external heat. Due 
to the configuration of the lithium-ion batteries, 
external damage can cause contact between the 
positive and negative electrodes leading to short 
circuit and rapid electrical discharge.  
When a single cell becomes unstable, it causes a 
cascade overheating of adjacent cells causing in a 
chain reaction known as thermal runaway which can 
become an explosive inferno.13  
113 air incidents involving smoke, fire, extreme heat 
or explosion have been reported to the Federal 
Aviation Administration between 1991 and 2011.14 
The highest profile of these incidents was the 
catastrophic lithium-ion battery fire which resulted in 
the crash of a cargo plane in Dubai in 2010, killing the 
2 pilots on board.15 More recently, during an 
aeromedical transport of a patient in Utah, smoke 
was identified originating from a bag containing an 
oxygen cylinder. Investigation identified a smoking 
spare ventilator lithium-ion battery. The battery was 
removed from the vehicle and the transport continued 
uneventfully; however it took over 60 minutes to fully 
extinguish the battery.16 
The risk of fire from lithium ion batteries is well 
documented and has become an accepted risk during 
aeromedical evacuation. Recommendations 
including appropriate charging and storage have 
been well documented by the manufacturers, 
however due to operational constraints these are 
colloquially known to be compromised.  
The paucity of evidence on fires associated with 
defibrillation does not reflect the absence of these 
events, in part due to the litigious nature of the  
 
 
 
event.6,9 No reports exist of fires due to defibrillation 
during aeromedical evacuations. It could be argued 
part of the reason for the lack of reported cases could 
stem from a variety of reasons including a lower 
incidence of defibrillation during flight. 
 
ELECTRICAL 
Medical practitioners suffering an electrical shock 
from defibrillation is a recognised if rare 
occurrence.17 Although several case reports describe 
accidental defibrillation of medical practitioners or 
bystanders, the UK Medicines Health Regulatory 
Agency (MRHA) has received only a single report of 
an electrical shock to a paramedic during patient 
transport.18 
The risk of electrical shock is increased during 
aeromedical transport due to the constricted 
environment within the cabin. Depending on the 
configuration of the aircraft it may be impossible to 
move or see around the entire stretcher increasing 
the risk of accidental defibrillation of crew-members. 
Getting clear of the metal stretcher, including but not 
limited to, IV poles and infusion pumps can also be a 
challenge. Flight can be unpredictable and 
unexpected turbulence can present a significant risk 
to health practitioners. All these risks make the 
requirement for clear, effective communication and 
teamwork paramount.  
The Medicines and Healthcare products Regulatory 
Agency (MHRA) suspects a significant amount of 
under-reporting of electrical events during 
defibrillation and to date, no accidental bystander 
defibrillation events have been reported during an 
aeromedical evacuation.18 
Automated external defibrillators (AEDs) are 
commonplace both inside and outside of hospital but 
their use when transporting a patient is not 
recommended in current resuscitation guidelines.19 
AEDs can interpret artefact rhythms arising from 
movement within the ambulance as VF resulting in an 
inappropriate shock.  
In 2011, Sang et al, demonstrated AEDs can correctly 
identify shock delivery for cardiac rhythms during 
helicopter flight.20 However, despite an increasing 
body of evidence, there is a level of disagreement on 
whether AEDs can be used safely during aeromedical 
retrieval. It is important to note AEDs have been 
adopted and used successfully in many commercial 
airline operations with no reported injuries or 
inappropriate shocks delivered.21 
 
 
 
 
AVIONIC INTERFERENCE 
All electronic devices carried aboard aeromedical 
aircraft are subject to approval by the Civil Aviation 
Authority within the UK. An approval assessment is 
carried out to demonstrate the equipment, 
installation and operation will not adversely affect the 
proper functioning of the aircraft.22 
When a defibrillator is activated, it discharges 5000 
volts of electrical energy over 1-2 seconds. During 
this discharge, a small but distinct electromagnetic 
pulse is emitted, similar to a lightning bolt. It is best 
practice to inform the pilot prior to activating the 
defibrillator in case of any interruption to the avionics.  
In most modern aeromedical aircraft, there is no 
effect on the avionics however, in the past, a ‘blip’ 
would occur in the cockpit depicting an unknown, 
unexpected signal. This would require the pilot to 
check all avionics if they had not been informed prior 
to defibrillation.  
Informing the pilot has the additional benefit of 
improving safety during the mission. During critical 
moments of the flight, the pilot can instruct the 
defibrillation to be paused until it is safe to continue. 
An example is during flight (particularly take-off and 
landing) using instrument flight rules (IFR). 
Discharging the defibrillator can interrupt the 
instruments and make landing/take-off significantly 
more dangerous. With knowledge of impending 
defibrillation, if safe to do so, the pilot can maintain 
straight and level flight rather than manoeuvring, 
increasing the safety for the patient and health 
practitioners. 
 
PHYSICAL CARRIAGE AND PACKAGING 
Despite all the safety testing undertaken prior to a 
defibrillator being certified fit for use within an 
aircraft, the physical carriage of a defibrillator is the 
most common cause of injury to patients and medical 
staff. Two main categories exist: 
1. Falling Equipment 
 
Within an aircraft, unless all equipment is secured 
there is a high propensity for equipment falling and 
causing injury. The most common time is during 
disembarkation from the helicopter or aeroplane. 
Additionally, turbulence in-flight can cause equipment 
to shift and fall causing injury to both patient and 
medical team. As mentioned before, lithium batteries 
are a significant source of risk and falling equipment 
can damage batteries resulting in a catastrophic fire. 
Most established aeromedical services have 
integrated monitors and diagnostic equipment which 
are housed either on the stretcher or permanently in 
the aircraft. This significantly reduces the risk of  
 
 
 
individual pieces of equipment becoming dislodged 
and falling. 
If a service does not have specific housings, extra 
care must be taken by the crew to ensure all 
equipment is accessible but appropriately secured to 
prevent injury.  
2. Equipment Failure 
 
Failure of defibrillators have reduced significantly 
since they have become smaller and single use pads 
have been introduced. This has reduced the risk of 
exposed wires and subsequent electrical accidents.  
The MRHA dictates all equipment should be serviced 
regularly and maintained to a high order.22 However it 
remains the responsibility of the flight team to remain 
vigilant and inspect defibrillators regularly to ensure 
they remain in good working order.  
 
CONCLUSION 
Defibrillation is a critical resuscitative tool that 
requires time efficient use to maximise patient 
survival. Multiple factors present risk during 
defibrillation in the aeromedical environment and 
therefore appropriate training and effective teamwork 
is imperative. Despite few to no reports of any 
adverse events from defibrillation, factors unique the 
aeromedical environment, including turbulence and a 
confined environment magnifies the risk. A limited 
body of evidence exists to recommend specific 
measures to be taken to prevent risk beyond a 
‘common-sense’ approach adopted by all health 
practitioners. Even more debate exists surrounding 
whether AEDs can be used safely during aeromedical 
retrieval and whether this will reduce the risk to 
patient and healthcare practitioner. Further research 
will be required to provide a definitive answer. 
With ever increasing aeromedical missions, 
defibrillation will occur in flight more often and it is 
crucial health regulatory agencies work with 
aeromedical teams to promote safe practice and 
blame-free incident reporting. 
 
REFERENCES 
1. Deakin CD, Nolan J, Sunde K et al. European 
Resuscitation Council Guidelines of Resuscitation 
2010 Section 3. Electrical Therapies: automated 
external defibrillators, defibrillation, cardioversion 
and packing. Resuscitation. 2010;82:1293-304. 
2. Beck CF, Pritchard WH, Feil HS. Ventricular 
Fibrillation of long duration abolished by 
electrical shock. JAMA. 1947;136:985-6. 
3. Eisenberg MS, Copass MK, Hallstrom AP et al. 
Treatment of out-of-hospital cardiac arrests with 
 
 
 
 
 
 rapid defibrillation by emergency medical 
technicians. N Engl J Med. 1980;302:1379-83. 
4. Edelson DP, Abella BS, Kramer-Johansen J, et al. 
Effects of compression depth and pre-shock 
pauses predict defibrillation failure during cardiac 
arrest. Resuscitation. 2006;71:137-45. 
5. Frakes M, O’Donnell S, Sanjian M. Need for 
defibrillation during cardiac arrest management  
by rotor wing critical care transport teams. Air 
Med J. 2006;25:211. 
6. Emergency Care Research Institute. Fires from 
defibrillation during oxygen administration. 
Health Devices. 1994;23:307-9. 
7. Bersten AD & Soni N. Oh’s Intensive Care 
Manual. 7th ed. Oxford, Butterworth-Heinemann 
Elsevier; 2014. 
8. Theodorou AA, Gutierrez JA & Berg RA. Fire 
attributable to a defibrillation attempt in a 
neonate. Paediatrics. 2003;112:677. 
9. Miller PH. Potential fire hazard in defibrillation 
[letter]. JAMA. 1972;221:192. 
10. Lefever J & Smith A. Risk of fire when using 
defibrillation in an oxygen enriched atmosphere. 
Med Devices Agency. 1995. Safety notice MDA 
SN 9503. 
11. Hummel RS III, Ornato JP, Weinberg SM et al. 
Spark-generating properties of electrode gels 
used during defibrillation. A potential fire hazard. 
JAMA. 1988;260:3021-4. 
12. Barker SJ & Polson JS. Fire in the operating 
room: a case report and laboratory study. Anesth 
Analg. 2001;93:960-5. 
13. Ben D, Ma B, Liu L et al. Unusual burns with 
combined injuries caused by mobile phone 
explosion: Watch out for the “Mini-bomb”! J Burn 
Care Res. 2009;30:1048. 
14. FAA Office of Security and Hazardous Materials. 
Batteries & battery-powered devices: Aviation 
incidents involving smoke, fire, extreme heat or 
explosion. 
http://www.faa.gov/about/office_org/headquarte
rs_offices/ash/ash_programs/ 
hazmat/aircarrier_info/media/Battery_incident_c
hart.pdf. Accessed 5th August 2017. 
15. GCAA Air Accident Investigation Department 
(AAI). Air Accident Preliminary Report: Boeing 
747-400F/N571 UP. GCAA Accident Report N 
13/2010. http://www.gcaa.gov. 
ae/en/ePublication/admin/iradmin/Lists/Incident
s%20Investigation%20Reports/At 
tachments/16/2010-
GCAA%20Accident%20Preliminary%20Report%20
B747% 20DXB%20.pdf. Accessed 5th August 
2017. 
16. Thomas F, Mills G, Howe R, et al. Lithium battery 
fires: Implications for air medical transport. Air 
Med J. 2012;31(5):242-8. 
17. Dickinson CL, Hall CR & Soar J. Accidental shock 
to rescuer during successful defibrillation  
 
 
 
 
 
of ventricular fibrillation – a case of human 
automaticity. Resuscitation. 2008;76(3):489. 
18. Personal Correspondence, MRHA. 
19. European Resuscitation Council. Part 4: the 
automated external defibrillator: key link in the 
chain of survival. Resuscitation. 2000;46:73-91. 
20. Sang MJ, Sung Y, Tae NC, et al. Performance of 
an automated external defibrillator during 
simulated rotor-wing critical care transports. 
Resuscitation 2011;82:454-8. 
21. Personal Correspondence, International SOS. 
22. CAP 562 – Civil Aircraft Airworthiness 
Information & Procedures. London: HMG 
Stationary Office; 2011. 


DANGERS OF DEFIBRILLATION IN FLIGHT

Yang, Haoqi

Yao, Hongge

DOI: 10.21307/ijanmc-2019-058
International Journal of Advanced Network, Monitoring and Controls          Volume 04, No.03, 2019 
47 
Street View House Number Identification Based on Deep 
Learning 
 
Yang Haoqi 
School of computer science and engineering 
Xi’an Technological University 
Xi’an, China 
e-mail: curioyhq@gmail.com 
Yao Hongge 
School of computer science and engineering 
Xi’an Technological University 
Xi’an, China 
e-mail: 835092445@qq.com 
 
 
Abstract—In this paper, the difficult problem of character 
recognition in natural scenes caused by many factors such as 
variability of light in the natural scene, background clutter and 
inaccurate viewing angle, and inconsistent resolution. Based on 
the deep learning framework PyTorch, a convolutional neural 
network is implemented. Based on the classic LeNet-5 network, 
the network optimizes the input layer to accept three-channel 
images, changes the pooling method to maximum pooling to 
simplify parameters, and the activation function is replaced by 
Rectified Linear Unit with faster convergence. The cross-
entropy loss is used instead of the minimum mean square error 
to mitigate the slow learning. Furthermore, we also enroll the 
gradient descent optimization algorithm RMSprop and L2 
regularization to improve the accuracy, speed up the 
convergence and suppress the over-fitting. The experiment 
results show that our model achieved an accuracy of 92.32% 
after training for 7h24min on the street view house 
number(SVHN) dataset, effectively improving the performance 
of LeNet-5. 
Keywords-House Number Identification; Convolutional 
Neural Network; Lenet-5 
I. INTRODUCTION 
The traditional method of classifying house 
numbers from natural scene images is usually to use 
manual feature extraction[1-2] and template 
matching[3-4]. In order to identify the house number of 
the corridor environment, Zhang Shuai et al. used the 
combination of Robert edge detection and 
morphological operation to locate the position of the 
house number image, and then divide the house 
number by horizontal and vertical projection method, 
tilt correction, etc., and finally use pattern recognition 
to identify the house number [5]. Ma Liling et al. used 
the linear discriminant linear local tangent space 
alignment algorithm (ODLLTSA) and the support 
vector machine (SVM) method to identify the house 
number, use the extracted features to train the SVM 
classifier, and use the SVM classifier to the new house 
number classification [6]. 
For these traditional methods, the key to 
determining their performance is to have a good 
classifier, and the features in the classifier are mostly 
designed manually (such as SIFT, SURF, HOG, etc.), 
and the features of the artificial design are well 
interpreted. However, in the face of complex 
backgrounds, changing fonts and various deformations, 
it is rather troublesome and difficult to extract more 
general features[7]. 
The Convolutional Neural Network (CNN) is a 
multi-layered supervised learning neural network. 
Although the training process requires a large amount 
of data compared with the traditional method, the 
convolutional neural network can automatically 
summarize the target feature from these data. Features 
do not require human intervention. Overcome the 
shortcomings of manual design features that are time-
consuming and labor-intensive, have poor general use 
and require high experience in the designer field. It is 
precise because of these advantages of convolutional 
neural networks that a large number of researchers 
have begun to apply it to solve character recognition 
problems. 
In response to this situation, we implemented a 
LeNet-5-based neural network based on the deep 
learning framework PyTorch and achieved an accuracy 
of 92.32% on the SVHN dataset at a time of 6 hours 
and 17 minutes. 
II. RELATED WORK 
A. Network structure 
The network used in this experiment is modified by 
LeNet-5 as shown in Figure 1. LeNet-5 appeared to 
solve the problem of recognition of handwritten 
characters. The data set used in the training process is 
the MNIST. The samples in the data set are single-
International Journal of Advanced Network, Monitoring and Controls          Volume 04, No.03, 2019 
48 
channel grayscale images, and the street view dataset is 
The three-channel color picture, to improve the 
robustness of the model, minimize the intervention on 
the original data set, we do not pre-process it, such as 
grayscale, but choose to adjust the input layer of 
LeNet-5 to three channels. 
 
Figure 1. Network structure 
The pooling layer in the original LeNet-5 network is 
very different from the currently recognized pooling 
layer operation, so we replace it directly with the max-
pooling layer, which on the other hand reduces the 
number of trainable parameters of the network. It is 
conducive to controlling the scale of the network and 
speeding up the training. In terms of the activation 
function, the activation function in the original LeNet-5 
is Sigmoid or TanH. Here we use a Rectified Linear 
Unit (ReLU) with faster convergence speed and no 
significant impact on the generalization accuracy of the 
model. LeNet-5's loss function is Minimum Mean 
Squared Error: 
  
Where   is the number of samples，   represents 
the predicted value of the ith sample, and    is the 
labelof the  th sample. In the case of back-propagation 
by the gradient descent method, the minimum mean 
square error is easy to occur when the neuron output is 
close to ‘1’ and the gradient is too small to learn slow. 
We use the cross-entropy loss function here: 
  
In addition to the above improvements, we will 
introduce four optimization algorithms, SGD (with 
momentum), Adam, Adamax, and RMSprop. 
B. Comparison  effects of different optimizers 
The package ‘torch.optim’ in PyTorch encapsulates 
a large number of optimization algorithms, which are 
often referred to as optimizers. In Figure 2, we take the 
more common SGD, Adam, Adamax and RMSprop 
optimizers according to the parameters listed in Table 
1.After 90 epochs training, compare their optimization 
effects on the SVHN dataset used in the improved 
LeNet-5 network proposed in this paper. It can be seen 
from Figure 2 that the network using the SGD 
optimizer has almost no improvement in the test set 
accuracy in the first 14 epoch, and the 14th epoch only 
starts to rise significantly; the network using the other 
three optimizers is in the toptenepochs, a good test set 
accuracy rate is obtained, and the test set accuracy of 
the network using the RMSprop optimizer is the fastest. 
So in the next experiment, we will use RMSprop as the 
default optimizer. 
TABLE I.  OPTIMIZER PARAMETER SETTING 
Optimizer parameter 
SGD lr=0.001, 
Adam lr=0.001, 
Adamax lr=0.002, 
RMSprop lr=0.001, 
 
 
 
Figure 2. Optimizer effect 
It can be observed from Figure 2 that the accuracy 
of the test set except the SGD optimizer is not 
improved much in the later stage of training, and the 
accuracy of the test set of the other three networks even 
shows a small downward trend. Table 2 shows the 
statistics in Figure 2. The data of test sets with high 
accuracy rate, the difference ofhighest accuracy rate on 
the SVHN test set is only 1.793816%, which can be 
considered as the optimization effect of the four 
optimizers on the accuracy rate; From the position of 
the highest test set accuracy, only the 7th epoch 
appears in the network using the SGD optimizer. The 
highest test set accuracy of the network using the other 
three optimizers is relatively high, indicating that 
theperformanceofthese three optimizers in the latter 
part of the trainingdecreased. It may be that the 
International Journal of Advanced Network, Monitoring and Controls          Volume 04, No.03, 2019 
49 
network has been over-fitting, and it is necessary to 
introduce evaluation and avoid over-fitting. 
TABLE II.  OPTIMIZERS TRAINING RESULTS 
optimizer Top Accuracy/% 
SGD 87.350184 
Adam 89.090000 
Adamax 88.955900 
RMSprop 88.676000 
 
C. Comparison of the application of L2 regularization 
with the appropriate weight attenuation coefficient 
In the face of possible over-fitting, one possible 
inhibitory measure is the introduction of regularization. 
We first use the L2 regularization and introduce the 
training set accuracy rate, training set loss, test set loss 
three indicators to enrich the evaluation results of the 
experimental results. The experimental design is shown 
in Table 3. The default optimizer is RMSprop 
(lr=0.001, alpha=0.9), the maximum iteration number 
is still set to 90epoch, and the L2 pooling 
corresponding weight attenuation coefficient 
(weight_decay) is the best and the best. The resulting 
position is shown in Table 3, and the corresponding 
accuracy and loss curves are shown in Figure 2-6. 
It can be seen from Table 3 and Figure 4 that the 
training set loss curves show a smooth downward trend 
under the four values of the weight attenuation 
coefficient. Comparing the training set accuracy rate of 
Figure 3 with the test set accuracy rate of Figure 5, it 
can be seen that the accuracy rate under the 
corresponding weight attenuation coefficient is about 5% 
up and down, within an acceptable range, but the 
weight attenuation coefficient in Figure 6 is The loss 
curve of the test set at 0.001 is firstly decreased and 
then increased. This indicates that the over-fitting 
phenomenon appears under this parameter, which 
indicates that the improved LeNet-5 network proposed 
in this paper is attenuated by the weight of 0.001 when 
training on the SVHN dataset. The coefficient can not 
suppress the over-fitting, we should choose a higher 
weight attenuation coefficient; from Table 3, we can 
see that the data with the weight attenuation coefficient 
of 0.0025 is better than the weight attenuation 
coefficient of 0.005 and 0.01. On the accuracy curve of 
the test set in Figure 5, the curve with the weight 
attenuation coefficient of 0.0025 is higher than the 
curve with the weight attenuation coefficient of 0.005 
and 0.01, and the weight attenuation coefficient is 
0.0025 in the later stage of the training process. More 
obvious and less shocking. The above analysis shows 
that among the selected four sets of weight attenuation 
coefficients, the weight attenuation coefficient of 
0.0025 can avoid over-fitting and achieve better 
training results. 
TABLE III.  RESULT OF DIFFERENT WEIGHT DECAY 
weight_decay Train Acc%(e) Test Acc%(e) 
0.01 89.01538(85) 87.14659(85) 
0.005 91.59397(57) 88.95590(57) 
0.0025 94.51470(88) 90.01229(88) 
0.001 97.21119(90) 89.70498(24) 
 
 
 
Figure 3. Training Accuracy of different regularization 
 
 
 
 
Figure 4. Training Loss of different regularization 
International Journal of Advanced Network, Monitoring and Controls          Volume 04, No.03, 2019 
50 
 
 
Figure 5. Test Accuracy of different regularization 
 
 
 
Figure 6. Tetst Loss of different regularization 
III. EXPERIMENT AND ANALYSIS  
A. SVHN(The Street View House Numbers) 
Currently, for the identification task of the street 
view number, the better public data set is the SVHN 
data set. The SVHN dataset is a real-world image 
dataset focused on the development of machine 
learning and target detection algorithms with minimal 
need for data preprocessing and format conversion. 
There are ten types of labels in the dataset. Each class 
represents 1 number. For example, the category label of 
the number "1" is 1, and so on. The label of "9" is 9, 
and the label of "0" is 10. In general, the SVHN dataset 
contains three subsets: training set, test set, and 
extended set; the data set is divided into two formats 
based on the difficulty of recognition: a character-level 
bounding box containing the entire house number and a 
small number of wall backgrounds. The full resolution 
image (Figure 7); a 32x32 pixel centered on a single 
number similar to the MNIST dataset format image 
(Figure 8). The latter style is highly similar to the 
classic MNIST dataset, but is larger and more difficult 
to identify: the training set consists of 73,257 hard-to-
recognize digital images, and the test set consists of 
26,032 digital images, with an additional set of 531131. 
A simpler digital image that can be used to extend the 
test set. Unless otherwise stated, the SVHNdataset 
mentioned later in this article refers to the dataset in the 
format after cropping. 
 
Figure 7. SVHN-Complete house number 
 
 
Figure 8. SVHN-Part number 
B. Data augmentation 
Augmenting the data set is also an effective means 
to improve the accuracy of the model. The size of each 
subset of the SVHN dataset is shown in Table 4, where 
the extension set official mentions that it can be used to 
extend the training data. Figure 9-11 shows a small 
number of samples and their labels in each subset. It 
International Journal of Advanced Network, Monitoring and Controls          Volume 04, No.03, 2019 
51 
can be seen that the resolution and brightness of the 
extended set are high and the background interference 
is small. The human eye recognition is indeed higher 
than the training set and test set, that is, It is said that 
the recognition of the extended set is relatively difficult, 
but the addition of the extended set can make the 
training set expand to 8.25 times, which is still 
expected to improve the accuracy of the model. Figure 
12 shows the distribution of the original training set, 
extension set and test set ("1" for the number 1, "2" for 
the number 2, ..., "10" for the number 0), which can be 
seen in the three sub-categories The proportional 
distribution of each category is approximated, so the 
distribution of the new training set incorporating the 
extended set is similar to the distribution of the original 
training set. This correlation helps to suppress the 
disadvantages brought by the introduction of the 
extended set to the model training. influences. 
TABLE IV.  AUGMENTATION RESULT 
Subset category Number of samples 
Training set 73257 
Extra set 531131 
Test set 26032 
 
 
 
Figure 9. Example of train set 
 
 
Figure 10. Example of extra set 
 
 
Figure 11. Example of test set 
 
Figure 12. Category distribution of SVHN 
The effect of 90 epoch training before and after the 
introduction of the extended set is shown in Table 5 
and Figure 13. One point that needs to be specially 
stated is that the visualization tool Visdom we use has 
an automatic zoom function when drawing, which 
automatically hides the blank area of auto-hide the 
chart. Therefore, the different training sets in Figure 13 
correspond to the vertical axis starting position and 
scaling of the chart. It's the same. The accuracy of the 
training set is lower than 95% and the loss curve of the 
test set shows a downward trend, indicating that there 
is no over-fitting phenomenon after expanding the data 
set; the accuracy of the test set is gradually reduced in 
the later period. The small description model tends to 
converge, and it can be seen that the model after the 
extended training set is higher than the extended 
training set. From Table 5, it can be seen that the 
training time after the expansion of the training set is 4 
hours and 53 minutes, and the best accuracy rate is 
increased by 2.31254% compared with the expansion. 
TABLE V.  RESULT AFTER DATA AUGMENTATION 
Train sample 
number 
test sample 
number 
Best test 
accuracy 
time 
73257 26032 90.01229 1h24min 
604388 26032 92.32483 6h17min 
 
 
Figure 13. Training of the model after adding data augmentation 
 
 
 
 
International Journal of Advanced Network, Monitoring and Controls          Volume 04, No.03, 2019 
52 
 
   
   
   
Figure 14. Figure 1 Test result 
IV. CONCLUSION 
The convolutional neural network applied in the 
SVHN dataset to improve the classic LeNet-5 network 
is: (1) modify the input layer to accept three-channel 
images; (2) switch to the more commonly used 
maximum pooling and Activation function, loss 
function; (3) introduction of gradient descent 
optimization algorithm RMSprop; (4) use L2 
regularization. The seven-layer convolutional neural 
network implemented in this paper achieves direct 
processing of color pictures without complicated 
preprocessing, which improves the versatility of the 
model, speeds up the training and effectively avoids 
over-fitting. In the end, both the training speed and the 
prediction accuracy are better than the domestic Ma 
Miao and others based on the experimental results of 
the improved LeNet-5. After expanding the dataset, I 
tried to run a maximum of 170 epoch, and there was no 
obvious improvement in the test accuracy. Therefore, 
the future improvement direction should still be based 
on the principle. We can consider deepening the 
network level to obtain more abundant features. 
 
REFERENCES 
 
[1] Mori S, Suen C Y, Yamamoto K. Historical review of OCR research 
and development. Proceedings of the IEEE, 1992, 80(7): 1029-1058. 
[2] Plamondon R, Srihari S N. Online and off-line handwriting 
recognition: a comprehensive survey. IEEE Transactions on pattern 
analysis and machine intelligence, 2000, 22(1): 63-84. 
[3] De Campos T E, Babu B R, Varma M. Character recognition in 
natural images. VISAPP (2), 2009, 7. 
[4] Yamaguchi T, Nakano Y, Maruyama M, et al. Digit classification on 
signboards for telephone number recognitio. IEEE, 2003: 359. 
[5] ZHANG Shuai, SU Shi-tao. Doorplate Identification for Mobile 
Robot in Hallway Based on Morphological. Modern Electronics 
Technique, 2011, 34(14): 7-9+12. 
[6] MA Li-ling. An algorithm based on ODLLTSA and SVM classier for 
door plate number recognition. Journal of Central South University 
(Science and Technology), 2011, 42: 789. 
[7] ZHOU Cheng-wei. Recognition of Numbers in Natural Scene with 
Convolutional Neural Network.Computer Technology and 
Development, 2017. 
 


Street View House Number Identification Based on Deep Learning

Mishra, Devesh

Cyber Security Cookbook for Practitioners

Not a member yet