Exeley Inc.

Xing, Ma

Fan, Yu

Haige, Yu

Yanxi, Wei

Wenhui, Yang

2018 International Conference on Sensor Network and Computer Engineering (ICSNCE 2018) 
19 
 
3D Target Recognition Based on Decision Layer Fusion 
Ma xing
a*
, Yu Fan
b
, Yu Haige, Wei Yanxi, Yang Wenhui 
School of computer science and engineering  
Xi’an Techonolgical University 
Xi’an ,710021, Shaanxi  
e-mail: 
a*
512066020@qq.com; 
b
yffshun@163.com
 
 
 
Abstract—Target recognition has always been a hot research 
topic in computer image and pattern recognition. This paper 
proposes a target recognition method based on decision layer 
fusion. ModelNet[1]—The 3D CAD model library，which is 
used to be identified. Features are extracted from the model's 
point cloud data and multi-view images. The image is 
identified using the AlexNet[2] network ,The point cloud is 
identified by the VoxNet[3] network. The fusion algorithm is 
used in the decision layer to complete the fusion of features. 
The results show that the proposed method improves the 
accuracy of object recognition. 
Keywords-Target Recognition; Convolutional Neural 
Network; Decision Fusion 
 
I. INTRODUCTION 
At present, the methods for identifying objects are 
mainly divided into two categories. The first is to identify 
the images generated by the objects, and the second is to 
identify the point clouds generated by the objects. 
In terms of image recognition, the current deep learning 
method has a high recognition rate. For instance, Xie etal. [4] 
adopt the multi-view depth image representation and propose 
multi-view deepextreme learning machine (MVD-ELM) to 
achieve fast and quality projective feature learning for 3D 
shapes. Zhu et al.[5] alsoproject 3D shapes into 2D space 
and use autoencoder for featurelearning on 2D images. In 
2012, the ImageNet contest champion model--Alexnet 
became a classical convolutional neural network image 
classification model. 
For the identification of point clouds, domestic and 
foreign scholars have done a lot of research. Some scholars 
use the method of manually extracting features to classify 
and identify. R.B.Rusu[6]and others used the relationship 
between the normal vectors of a region as a feature to 
classify objects for recognition. Yasir Salihet al. [7] used 
VFH as a feature and used a support vector machine as a 
classifier to classify and recognize point clouds. Manually 
extracting features requires very professional knowledge and 
rich experience. Convolutional neural networks can 
automatically extract features and classify them, and they are 
invariant to displacements, scaling, and other forms of rigid 
body changes. Some experts and scholars have used 
convolutional neural networks to classify and recognize 
point cloud images, of which the VoxNet network has the 
highest recognition rate. 
The above method achieves higher accuracy in target 
classification recognition. I have seen that image recognition 
is affected by factors such as lighting and viewing angles. 
The accuracy of point cloud recognition is lower than that of 
image recognition. Therefore, this paper integrates the image 
and the point cloud at the decision-making level to improve 
the accuracy of object recognition. 
II. NETWORK STRUCTURE 
A. Convolutional neural network 
Traditional shallow learning methods such as support 
vector machines require manually extracting image features 
and then sending the features into the classifier for training. 
This leads to a problem that the manually extracted feature is 
not necessarily the best description of the current image. 
Even if the selected feature is very suitable for the current 
image, when the external conditions of the object such as the 
angle, size, and illumination of the image change, the 
manually selected feature cannot adapt well to this change, 
and it is necessary to artificially adjust the selected feature 
according to the situation. Different from the traditional 
shallow learning method, the input of the convolutional 
neural network is the entire image. It continuously adjusts 
the parameters of the network through a learning algorithm 
and adaptively extracts the most significant features of the 
current image, which avoids manual intervention and saves a 
large amount of manpower, with the continuous updating of 
the input pictures, the essential features of the current picture 
are extracted with the times, ensuring the accuracy and 
efficiency of the recognition. As a special architecture that is 
particularly suitable for classifying images, compared with 
conventional shallow machine learning methods such as 
support vector machines, convolutional neural networks can 
be much smaller than normal when faced with large-scale 
high-resolution image classification problems. The method 
learns more picture information in the training time, and the 
classification accuracy is higher than the conventional 
method, which is due to its unique network architecture. 
The convolutional neural network consists of three parts: the 
convolutional layer, the down-sampling layer and the fully 
connected layer. The down-sampling is usually after the 
convolutional layer, alternating with the convolutional layer, 
and finally connected to the fully connected layer. 
Convolutional neural networks use local connections, weight 
sharing, and spatial or temporal correlation down-sampling 
to obtain good translation, scaling, and distortion invariance, 
making the extracted features more distinguishable. CNNs 
2018 International Conference on Sensor Network and Computer Engineering (ICSNCE 2018) 
20 
 
training includes forward propagation. The two processes of 
forward propagation and reverse propagation are the process 
of the input signal output from the input layer, through 
several hidden layers, and the output layer. The reverse 
propagation is the process of back propagation of the error 
signal from the output layer to the input layer. It mainly uses 
errors. The back propagation (Error Back Propagation, EBP) 
algorithm and gradient descent adjust the weights at each 
level of the network and is similar to the training process of 
an ordinary neural network. 
B. AlexNet network structure 
The standard convolutional neural network(CNN) is a 
special multilayer feed forward neural network. It has a deep 
network structure and is generally composed of an input 
layer, a convolutional layer, a down-sampling layer, a fully 
connected layer, and an output layer. The convolutional layer, 
the down-sampling layer, and the fully connected layer are 
hidden layers. AlexNet uses two GPU services. The model is 
divided into eight layers, five convolutional layers and three 
fully connected layers. Each convolutional layer contains the 
excitation function RELU and local response normalization 
(LRN) processing, and then after down-sampling (POOL). 
The network structure designed in this paper is shown in 
Figure 1. 
 
Figure 1. Multi-view image network structure diagram 
The input layer are images, there are 5 convolutional 
layers. The number of feature maps are 55, 27, 31, 13,and 6 
features respectively. The convolution kernel sizes 
respectively are 11, 5, 3, 3, and 3,Below the first two 
convolutional layers there is a largest pooled layer with an 
LRN layer for localized normalization. The largest pooled 
layer is to take the maximum value of the feature points in 
the domain. After processing through the convolution layer, 
many features are obtained. The amount of direct calculation 
is very large, and the increase of features is particularly 
prone to overfitting. Therefore, the network constructed in 
this paper is Each time a convolution process is performed, a 
Max-Pooling layer is added. The Dropout layer has a discard 
rate of 0.5. Dropout temporarily discards some networks in 
the training process with a certain probability, and each 
mini-batch discards different networks. It can reduce the 
amount of calculation, and more importantly it can prevent 
over-fitting. The number of neurons in two fully connected 
layers is 256 and 10, respectively. The last output layer is the 
Softmax[6]layer, which does not directly output the 
classification of the identified image, but rather the 
probability that the output image belongs to each 
classification. 
C. VoxNet Network Structure 
Figure 2 shows the network structure of VoxNet. The 
input layer accepts data as a form. There are a total of two 
convolutional layers, and the number of feature maps is 32, 
using the sum of the convolution kernels. The Dropout layer 
discard rates are 0.2 and 0.3, respectively, to prevent 
overfitting while reducing the amount of computation. The 
largest pooled layer, the filter used. Finally, there is a fully 
connected layer with 128 neurons and a Dropout layer with a 
discard rate of 0.4. The seventh layer is the output layer and 
the number of neurons is 10. 
Input Layer Cov3d Layer Dropout Layer
Cov3d Layer Fully Layer Dropout Layer
Max Pool 3d Layer Dropout Layer Output Layer
 
Figure 2. VoxNet network structure diagram 
D. Network convergence structure 
Multi sensor information fusion is to extract and integrate 
the same target image with a multi-source information 
channel for further processing. Information fusion can be 
divided into three layers: the fusion based on data layer, the 
fusion based on the feature layer, and the fusion based on the 
decision layer. The level of fusion was from low to high. So 
this paper uses the method of decision layer fusion. 
The feature fusion of the decision layer is usually the 
fusion of the prediction results of multiple classifiers. We 
extract various features from feature extraction algorithm. 
We assume that all kinds of features are independent of each 
other, and these features can separately predict the result of 
recognition separately. On this basis, we send data into their 
respective classifiers, get the prediction results of each 
classifier, then combine all the classifier's prediction to get 
the final recognition results, and complete the fusion of 
multiple features in the decision level. 
As shown in Figure 3, point cloud is extracted from 
VoxNet using point cloud feature, and AlexNet is used to 
extract image features. Softmax regression model is used to 
complete recognition and classification respectively. The 
fusion algorithm is used in the decision layer to complete the 
fusion of features. 
2018 International Conference on Sensor Network and Computer Engineering (ICSNCE 2018) 
21 
 
 
Figure 3. Fusion model 
III. . EXPERIMENTS 
A. Experiment environement 
The environment used in this experiment was 
TensorFlow-GPU 0.12.1 open source software library, 
Windows 7 operating system, and Nvidia GTX 950 graphics 
card. The data used in the experiment was ModelNet of 
Princeton University. ModelNet is a large-scale 3D CAD 
model database, similar to ImageNet in 2D images. 
B. Experiment Datasets 
This article uses the ModelNet40 dataset, where the point 
cloud dataset is from the dataset in PointNet[12], and the 2D 
image is from Multi-view. On the topside of Figure 4 is the 
point cloud image, at the bottom is a two-dimensional image. 
 
 
Figure 4. Point cloud image and two-dimensional image 
C. Linear Combination Coefficient Selection 
Static linear combination of the results of the AlexNet 
network and VoxNet network prediction results in the final 
prediction. Before linear combination, we need to determine 
the coefficients of each classifier's prediction result, which 
controls the relative importance of the results predicted by 
each classifier. It is very important to choose a suitable 
coefficient. The appropriate coefficient can fully exert the 
advantages of each classifier and make a joint decision. The 
final recognition accuracy rate will be better than the 
recognition accuracy of a single classifier. The use of 
inappropriate coefficients will result in the classification 
accuracy of the final joint decision even lower than that of a 
single classifier. 
There will be a Softmax classifier at the last level of the 
AlexNet and VoxNet network. For each input sample, the 
output of the Softmax classifier is a probability 
vector , that the sample may belong to. 
Where  represents the probability that the sample belongs 
to class n, and n represents the number of classes of all 
samples. With , the sum of the 
probabilities that a sample belongs to all classes is 1. In the 
object recognition of a single classifier, we will choose the 
probability. 
The class label corresponding to the largest element in 
the vector  is the class corresponding to the sample. In this 
chapter, for each test sample, the obtained recognition rate 
results are 、 , and the coefficients of each 
classifier are α and β respectively. Then we can complete the 
fusion of all base classifiers according to Eq. 1. 
  
After the fusion is complete, we get a k-dimensional 
vector,  where n is the number of 
categories. We take the largest element among them as the 
final sample tag. 
D. Recognition results 
In order to test the accuracy of this experimental method, 
the method of this paper is compared with the recognition 
accuracy of VoxNet and AlexNet. The experimental results 
are shown in Table 1. 
TABLE I. ACCURACY OF DIFFERENT METHODS 
recognition methods accuracy rate/% 
AlexNet 85 
VoxNet 83 
 
In this paper, the coefficients α and β before AlexNet and 
VoxNet are set to different values, the method with higher 
accuracy is used to set larger coefficients, the method with 
lower accuracy is set with smaller coefficients, and the 
comparison between different combinations of coefficients is 
2018 International Conference on Sensor Network and Computer Engineering (ICSNCE 2018) 
22 
 
accurate for the network. influences. The experimental 
results are shown in Table 2. 
TABLE II. EFFECTS OF DIFFERENT COMBINATIONS OF COEFFICIENTS ON 
NETWORK ACCURACY 
weight（ ） accuracy/% 
（0.5，0.5） 79.8 
（0.6，0.4） 81.2 
（0.7，0.3） 91 
（0.8，0.2） 87.5 
（0.9，0.1） 86.3 
 
From Table 2, the network has the highest recognition 
rate when α and β are set to 0.7 and 0.3. Compare this 
method with the recognition accuracy of VoxNet and 
AlexNet. Experimental results show that this method has the 
highest recognition rate. 
IV. CONCLUSION 
This paper presents a decision-level three-dimensional 
target fusion algorithm, we using different convolutional 
neural network frameworks to extract point cloud features 
and visual features of three-dimensional objects, respectively, 
and finally achieved effective fusion. Experimental results 
show that feature fusion in the decision-making layer is also 
an effective feature fusion method. This method improves 
the accuracy of object recognition. In the process of 
integration, a method with a higher recognition accuracy rate 
sets a larger coefficient, and a method with a low recognition 
accuracy rate sets a smaller coefficient, and the final 
accuracy rate is the highest. 
REFERENCES 
[1] Z. Wu, S. Song, A. Khosla, F. Yu, L. Zhang, X. Tang and J. Xiao. 3D 
ShapeNets: A Deep Representation for Volumetric Shapes. 
CVPR2015. 
[2] Krizhevsky A, Sutskever I, Hinton G E. ImageNet classification with 
deep convolutional neural networks[C]// International Conference on 
Neural Information Processing Systems. Curran Associates Inc. 
2012:1097-1105. 
[3] D. Maturana and S. Scherer. VoxNet: A 3D Convolutional Neural 
Network for Real-Time Object Recognition. IROS2015. 
[4] H. Su, S. Maji, E. Kalogerakis, E. Learned-Miller. Multi-view 
Convolutional Neural Networks for 3D Shape Recognition. 
ICCV2015. 
[5] Charles R. Qi, Hao Su, Kaichun Mo, and Leonidas J. Guibas. 
PointNet: Deep Learning on Point Sets for 3D Classification and 
Segmentation. CVPR 2017. 
[6] Z. Xie, K. Xu, W. Shan, L. Liu, Y. Xiong, H. Huang, Projective 
feature learning for 3D shapes with multiview depth images, in: 
Computer Graphics Forum, vol.34, Wiley Online Library, 2015, 
pp.1–11. 
[7] Z. Zhu, X. Wang, S. Bai, C. Yao, X. Bai, Deep learning 
representation using autoencoder for 3D shape retrieval, in: 
Proceedings of the International Conference on Security, Pattern 
Analysis, and Cybernetics(SPAC), IEEE, 2014, pp.279–284. 
[8] Holz D, Holzer S, Rusu R B, et al. Real-Time Plane Segmentation 
Using RGB-D Cameras[C]// Robot Soccer World Cup XV. 
Springer-Verlag, 2012:306-317. 
[9] Salih Y, Malik A S. Comparison of stochastic filtering methods for 
3D tracking[J]. Pattern Recognition, 2011, 44(10–11):2711-2737. 
[10] C. R. Qi, H. Su, M. Nießner, A. Dai, M. Yan, and L. 
Guibas.Volumetric and multi-view cnns for object classification on 
3d data. In Proc. Computer Vision and Pattern Recognition (CVPR), 
IEEE, 2016. 
[11] http://www.cnblogs.com/graphics/archive/2010/08/05/1793393.html 
The Princeton ModelNet. http://modelnet.cs. 
[12] The Princeton ModelNet. http://modelnet.cs.  
[13] R. B. Girshick, J. Donahue, T. Darrell, and J. Malik. Rich feature 
hierarchies for accurate object detection and semantic segmentation. 
In Proc. CVPR, 2014. 
 


3D Target Recognition Based on Decision Layer Fusion

Wang, Baoli

Wang, Junli

Zhipeng, Zhao

2018 International Conference on Sensor Network and Computer Engineering (ICSNCE 2018) 
77 
 
Design of Pellet Recycle Scraper System in Sand-Blasting Chamber 
Wang Baoli  
Bombardier Sifang (Qingdao)  
Transportation Ltd 
Shandong 266111, China 
Wang Junli  
Bombardier Sifang (Qingdao)  
Transportation Ltd 
Shandong 266111,China 
Zhao Zhipeng 
Bombardier Sifang (Qingdao)  
Transportation Ltd 
Shandong 266111, China 
e-mail: 760907040@qq.com 
 
 
 
 
 
 
 
 
Abstract—In the process of industrial sand-blasting, it is 
necessary to recycle the pellet. In order to solve the existing 
problems such as high cost, vulnerability, maintenance 
difficulty and low recycle efficiency in pneumatic recycling 
system and mechanical recycling system of pellet, we designed 
a trapezoidal scraper recycle system of pellet which powered 
by the cylinder and controlled by PLC to realize the pellet 
collection. The core component of the system is the scraper 
mechanical structure which mainly composed of scraper, 
scraper cylinder, baffle, scraper frame, bearing. This design 
has the advantages such as simple structure, resistant to 
breakdown, easy dismantlement. By installing pneumatic valve 
on the cylinder, the sand scraping speed can be adjusted. The 
trapezoidal scraper pellet recycling system efficiency is much 
higher after the contrast experiments between trapezoidal 
scraper and mechanical screw conveyor of the original 
equipment. 
Keywords-Pellet; Cylinder; PLC; Scraper; Trapezoid  
 
I. INTRODUCTION 
In order to improve the adhesion of the surface paint, it is 
necessary to sand-blast on the workpiece. During the 
sand-blasting process, pellet need to be recycled for the 
purpose of cost reduction [1-2]. At present, pneumatic 
recycling system [3] and mechanical recycling system [4-5] 
are widely used in our country for pellet recycle. However, 
problems such as high cost, vulnerability and maintenance 
difficulty still exist. Therefore, it is overwhelming important 
to design a recycle system for pellet of low cost, resistance to 
breakdown and easy maintenance. 
II. PELLET RECYCLE SYSTEM 
Scraper-typed pellet recycle system [6] designed by this 
design is mainly composed of 2 rod, 3 scraper, 5 cylinder, 6 
pellet-collect container and so on. As shown in Figure 1, 1 is 
trench, 4 is air rod and 7 is pellet material. The air rod is 
fixedly connected with the pull rod. Numerous scrapers with 
the same structure are fixedly connected with the pull rod. 
The scraper has the function of one-way reversing. When the 
air rod is retracted, the scraper will move the pellet material 
to the material collecting container. When the air rod is 
stretched out, the scraper will flip without pushing pellet, so 
that pellet will achieve a cycle of directional movement. By 
installing pneumatic valve on the cylinder, air rod speed can 
be adjusted. 
 
1
7
2 3 4 5
6
 
Figure 1. Scraper pellet recycle system 
 
III. MECHANICAL SYSTEM DESIGN 
A. Scraper structure 
The core component of the material recycle system is the 
scraper. For the scraper system designed in this design as 
shown in Fig. 2, Fig. 3, Fig. 4, Fig. 5, 1 is trench, 7 is the 
pellet material. The physical map is shown in Figure 3-4. 
Scraper system is mainly composed of 2 bearing mounting 
plates, 3 scraper frame, 4 vertical bearing, 5 lateral bearing, 6 
scraper, 8 baffle, 9 bayonet and 10 scraper cylinder as shown 
in Figure 3-1 and Figure 3-2. Scraper and scraper cylinder 
are in dynamic connection, with functions such as one-way 
flip and reverse scraping. The basic scraper structure is 
trapezoidal. The role of bayonet is to constrain the freedom 
for scraper up and down. The surface of the scraper frame is 
fixedly connected with the scraper cylinder and the baffle. 
Lateral bearing and vertical bearing are cross-typed with 
functions of guiding and undertaking weight respectively as 
shown in Figure 3-3. Lateral bearing and vertical bearing are 
mounted on the bearing mounting plate. Bearing mounting 
plate is fixedly installed on the inner wall of the trench. One 
end of the scraper frame is fixedly connected with the air rod. 
The pneumatic control valves and proximity switch are 
mounted onto the air cylinder. 
2018 International Conference on Sensor Network and Computer Engineering (ICSNCE 2018) 
78 
 
1
A
A
2
3
4
5
6
7
10
9
8
I
 
Figure 2. Scraper system 
 
10
6
9
8
7
3
 
Figure 3. A-A directional view 
 
3
4
2
5
8
 
Figure 4. I partial view 
 
 
Figure 5. Physical map of trapezoidal scraper 
B. Scraper’s working theory 
Once the scraper-typed pellet recycle is started, the 
scraper frame will move the trapezoid scraper forward and 
backward by scraper cylinder. When the scraper moves 
forward, the scraper will push the pellet forward due to the 
pulling force of the air cylinder and the blocking effect of the 
baffle up on the scraper. When the scraper moves backward, 
the baffle won’t have blocking effect on the scraper, the 
scraper rotates around the scraper cylinder under the reaction 
force of the pellet on the ground, it won’t push the pellet 
material. With such a reciprocating rhyme, directional 
movement of pellet material can be completed as well as 
pellet will be collected. Pneumatic valve can adjust the air 
rod running speed according to the current situation, hence 
adjust the scraping pellet speed by scraper. 
C. Scraper performance analysis 
 The scraper major components are cylinder, scraper, 
scraper frame, bearing and so on. It has advantages 
such as simple structure, low cost, easy 
dismantlement.  
 This design uses trapezoidal scraper and make full 
use of the trench surface to supervise the height of 
pellet piled in case it affects scraper flip. Its 
influence is most obvious when the width of trench 
is small.  
 The bearings are distant from the ground, which can 
prevent the accumulation of pellet from damaging 
the bearings to reduce the maintenance times.  
 The pneumatic speed control valve is installed onto 
the cylinder so the scraper speed can be reduced 
when there are fewer pellet, which can reduce 
scraper loss. 
IV. ELECTRICAL SYSTEM DESIGN  
According to requirements of the system, this design uses 
Siemens S7-200 PLC to control the scraper. This product is 
cheap, powerful with great practicality. PLC I/O allocation 
table is shown in Table 1. Its external wiring diagram is 
shown in Fig 6. Proximate to the switch, PLC detects the 
position of the air rod and controls direction change of the 
cylinder through solenoid valve. When the air rod fails to 
change direction, the PLC will stop and give alarms [7-8]. 
TABLE I. PLC I/O ALLOCATION  
Input Output 
Input component Address 
Output 
component 
Address 
SB1, start I0 Y1, cylinder out Q0 
SB2, stop I1 
Y2, cylinder 
contraction 
Q1 
SQ1, cylinder 
stretched out 
until the position 
of test 
I2 
BJ1, warning 
light 
Q2 
SQ2, cylinder 
contracted until 
the position of 
test 
I3   
2018 International Conference on Sensor Network and Computer Engineering (ICSNCE 2018) 
79 
 
SB1
SB2
SQ1
SQ2
Y1
Y2
BJ1
COM
I0
I1
I2
I3
Q0
Q1
Q2
DC24V
+24V
0V
 
Figure 6. Schema of electrical control 
 
V. EFFICIENCY COMPARISON BETWEEN TRAPEZOIDAL 
SCRAPER AND MECHANICAL PELLET CECYCLE SYSTEM 
In this design, trapezoidal scraper and the original 
mechanical screw conveyor pellet recycling efficiencies are 
compared. The same amount of pellet are placed in two 
ditches. With both systems under the best operating 
efficiency, statistics of recycle time will be collected 
according to 5 kg, 10 kg and 15 kg pellet. The result is 
shown in Table 2. 
TABLE II. PELLET RECOVERY TIME 
Recycled 
pellet quality 
(kg) 
Pellet recovery time(hour) 
Screw 
conveyor 
Trapezoidal 
scraper 
5 0.51 0.30 
10 1.11 0.62 
15 1.56 0.89 
Data show that under the same condition, the time 
required to complete the 5 kg, 10 kg, 15 kg recycle by 
trapezoidal scraper is significantly less than that of the 
mechanical screw conveyor thus prove that the trapezoidal 
scraper pellet recycle system efficiency is much higher. 
VI. CONCLUSION 
Scraped pellet recycle system is proposed in this design. 
It uses Siemens S7-200 PLC as the controller and it's core 
component is the trapezoidal scraper whose advantages are 
simple structure, resistance to breakdown and easy 
dismantlement. With comparison of experiments, it shows 
high efficiency in recycling pellet. Hence it is worthwhile of 
popularization. 
 
REFERENCES 
[1] Ye Yangxiang, Pan zhaoji. Application Manual for Coating 
Technology [M]. China Machine Press, 1997. 
[2] Yao Shoushan, Li Geyang, Hu wenbing. Surface Science and 
Technology [M]. China Machine Press, 2005. 
[3] Wang Xuefang. Analysis of Pneumatic Sand-Blasting System 
[J].Electric Power Locomotives and Urban Railway Vehicles, 2003, 
26 (2): 30-32. 
[4] Wang Yongli. Application Examples by Sand-Blasting Machine in 
the Coating Line [J]. Modern Coatings, 2013, 16 (3): 27-33. 
[5] Liu Fugui. Research on Screw Conveyor Machine of Pellet [J]. 
Rolling Stock Workers, 2001 (11): 9-11. 
[6] Luan Xianyu, Liu Hongnan. Design of Blasting Chamber Scraper 
Recycle System Based on PLC [J]. Mechanical Engineers, 2014 (6): 
158-159 
[7] Liu Huabo. Siemens S7-200 PLC programming and application case 
selection [M]. China Machine Press, 2009. 
[8] Liao Changchu. S7-200 PLC programming and application [M]. 
China Machine Press, 2014 
 
 
.
 


Design of Pellet Recycle Scraper System in Sand-Blasting Chamber

Ju, Yang

Bailin, Liu

Zhao, Zhixiang

2018 International Conference on Sensor Network and Computer Engineering (ICSNCE 2018) 
97 
 
A Collaborative Filtering Recommendation Algorithm with Improved Similarity 
Calculation 
 
Yang Ju
a
, Liu Bailin
b
 and Zhao Zhixiang
 
School of Computer Science and Engineering, Xi'an Technological University 
 Xi'an, 710021, Shaanxi 
State and Provincial Joint Engineering Lab. of Advanced Network and Monitoring Control  
Xi'an, 710021, China 
e-mail: 
a
yangju90@163.com, 
b
498194312@qq.com  
 
  
Abstract—In order to improve the accuracy of the proposed 
algorithm in collaborative filtering recommendation system, an 
Improved Pearson collaborative filtering (IP-CF) algorithm is 
proposed in this paper. The algorithm uses the user portrait, 
item characteristics and data of user behavior to compute the 
baseline predictors model. Instead of the traditional 
algorithm's similarity calculation, the prediction model is used 
to improve the accuracy of the recommendation algorithm. 
Experimental results on Moivelens dataset show that the IP-CF 
algorithm significantly improves the accuracy of the 
recommended results, and the RMSE and MAE evaluation 
results are better than the traditional algorithms. 
Keyword-Recommendation Algorithm; Collaborative 
Filtering; Similarity Calculation; Baseline Predictors Model 
I. INTRODUCTION 
With the rapid development of computer technology and 
network technology, the number of network information 
services and applications is growing rapidly. China Internet 
Network Information Center reported statistics, as of June 
2016, the size of China's Internet users reached 710 million, 
a total of 21.32 million new netizens in half a year
 [1]
. With 
the increase in the number of people on the Internet, Internet 
information has also seen explosive growth. How to find 
interesting and effective information in this vast data is a 
very difficult thing. In order to solve this problem, academia 
and industry put forward personalized recommendation 
system
[2]
. According to the user's personal information and 
historical habits, it can discover the potential interest of the 
user and recommend the resources of interest to the user 
actively. 
Personalized recommendation system is a special form of 
information filtering system
[3]
. The recommendation system 
can be divided into the following categories: collaborative 
filtering recommendation system, content-based 
recommendation system and hybrid recommendation system. 
Because of its wide applicability, strong interpretability and 
good stability, the collaborative filtering recommendation 
system based on neighborhood model is widely used in 
various fields. Therefore, this paper focuses on the 
collaborative filtering recommendation system based on 
neighborhood model. 
The accuracy of recommendation results in the 
recommendation system is the main index to measure the 
recommendation effect. Sarwar et al.
[4]
 proposed an item-
based collaborative filtering recommendation algorithm that 
looked into cosine-based similarity to compute the similarity 
between products. This method provides dramatically better 
performance than traditional recommendation algorithm, 
while at the same time providing better accuracy. Chen and 
Cheng
[5]
 use the rating data to compute the similarity 
between users, and use the ranking data as the weight of 
similarity calculation. Yang and Gu
[6]
 propose to use user 
behavior information to construct the user's interest points 
and use the interest points to compute the similarity between 
users. Experiments show that these methods are better than 
the classic collaborative filtering algorithm. However, these 
methods only consider the user-item behavioral data, and 
neglect the user portrait and item features, which causes 
deviations in the accuracy of the recommendation result. 
This paper improves similarity calculation method in 
collaborative filtering recommendation algorithm based on 
neighborhood model, and uses the user portrait, item 
characteristics and user-item behavior data to compute 
similarity. We experimentally evaluate our results and 
compare them to the classic collaborative filtering algorithm. 
Experiments suggest that the improved similarity calculation 
method can improve the accuracy of the recommended 
results. 
II. COLLABORATIVE FILTERING RECOMMENDATION 
ALGORITHM BASED ON NEIGHBORHOOD MODEL 
Collaborative filtering recommendation algorithm is to 
select the same custom hobby user groups, use other people's 
experience to meet their own needs, in order to achieve the 
purpose of reducing overhead. Typically, the workflow of a 
collaborative filtering system is: 
(1) Compute the similarity between users. 
(2) Determine the neighbor set. Find the k users whose 
user interest is the most similar through the similarity size, 
and set these users as the user sets. 
(3) According to the user sets prediction rating. The 
system recommends items that the users have rated highly 
but not yet being rated by this user. 
The most important thing in collaborative filtering 
algorithm is similarity calculation. For the calculation of 
2018 International Conference on Sensor Network and Computer Engineering (ICSNCE 2018) 
98 
 
similarity, the researchers put forward a variety of similarity 
calculation methods. 
Cosine-based similarity: For user u  and user v , ( )N u  
denotes the set of positive feedback items for user u , and 
( )N v  denotes the set of positive feedback items for user v . 
And similarity between items u  and v , denoted by 
( , )sim u v  is given by: 
)()(
)()(
),(
vNuN
vNuN
vusim

 
Pearson correlation coefficient[7-8]: In this case, similarity 
is computing based on the vector of the rating. Among them, 
u
r  in the formula has two forms in the traditional 
recommendations. One is the average rating of user u , and 
the other is the average rating of item i  scored by all users. 
The Pearson correlation is given by: 
2 2
( )( )
(u,v)
( ) ( )
uv
uv uv
ui u vi v
i I
ui u vi v
i I i I
r r r r
sim
r r r r

 
 

 

 

The similarity between users can be calculated by the 
above formula, and the similarity ranking of each user with 
other users can be obtained to obtain the nearest neighbor 
user set. After getting the user set, the next step is interest 
prediction computation. We can denote the prediction 
( , )p u i  as: 



)(),(
),(),(
iNKuSv
virvusimiup 
III. IMPROVED SIMILARITY MEASURES 
This paper considers the user rating data from the overall 
situation, introduces the characteristics of personal habits, 
item quality and category to improve the similarity 
computation formula. Thus the approximated correlation 
coefficient is given by: 







nmnm
nm
Pv
nini
Pv
mimi
Pi
ninimimi
nm
brbr
brbr
S
,,
,
22
,
)()(
))((

Here, ,m nS  denotes the similarity between user m and n. 
mi
r  is the raw rating of the user m  for the item i . ,m nP  
represent the user m, n common rating set. 
mi
b  is a baseline 
predictors for rating 
mi
r . The baseline predictors model is as 
follows: 



iGg
gimmi cbbb  
  denotes the intercept of the baseline predictors model. 
The parameters 
m
b  and 
i
b  indicate the deviations of user m  
and item i , respectively, from the average. The last term of 
the formula denotes the preference of the user on the item 
category. 
i
G  represents the set of categories to which the 
item belongs. Here we use an example to illustrate the 
baseline predictors model. Create a baseline predictors model 
for the rating of movie i  for user m . Assume that the mean 
rating of all the movie scores is 3.5. m  is a critical user, who 
tends to rate 0.3 stars lower than the average. i  is a movie 
with a relatively high standard, so its rating is 0.5 stars higher 
than the average rating. In addition, the movie i  belong to 
, ,
x y z
c c c , with a bias relative to the average of -0.05, 0.08, 
0.12. Therefore, the prediction for the movie i  rating by 
user m  is 3.5-0.3+0.5-0.15+0.18+0.12=3.85. For this 
formula, the purpose is to find 
m
b , 
i
b  and gc . This paper 
solves the problem by solving the least-squares problem[9-11]. 
The cost function formula is as follows: 



iGg
giumimi cbbre         
)()(min
222
),(
2   
  

Ii Ii Gg
gi
Um
m
im
mi
i
cbbe 
   
  
In the above (6) formula, r
mi
 is the true value of user for 
item. In formula (7),   represents the set of user ratings for 
items, U  represents user set, and I  represents the set of 
item. The solution process is to obtain the best fitting 
m
b , 
i
b  
and gc  by minimizing the first term 
2
( , )
( )
mi
m i
e

  in 
equation (7). The second item is the L1 regular, that is added 
to prevent overfitting. The size of   indicates the degree of 
intervention to fit, and the larger the general   is, the 
smoother the fitting curve is. 
Because the proposed model of this paper is different 
from the traditional one, the data matrix cannot be directly 
applied to the training of the model. So the matrix of the 
training data is restructured, adding personal habits, item 
quality and category bias. For example, when a movie i was 
released, it was called a masterpiece of elements such as 
comics, entertainment, suspense, etc. These classified data 
were useful for the model but could not be used. Through the 
transformation of the data format, useful information is used, 
and the information is vectorized according to the 
classification categories. Each row of data through the 
transformation training matrix can be expressed 
as: 1 1 1{( , ), ... , ... , ... , | }m n j uiu i u u i i c c r c G . 
2018 International Conference on Sensor Network and Computer Engineering (ICSNCE 2018) 
99 
 
TABLE I.  TRAINING DATA MATRIX 
),( iu  1u  ... mu  1i  ... ni  1c  ... jc  uir  
(1,1) 1 ... 0 1 ... 0 1 ... 0 4 
... 
... 
... 
... 
... 
... 
... 
... 
... 
... 
... 
(1,n) 1 ... 0 0 ... 1 0 ... 0 2 
... 
... 
... 
... 
... 
... 
... 
... 
... 
... 
... 
(m,n) 0 ... 1 0 ... 1 1 ... 1 5 
 
TABLE II.  MOIVELENS DATASET INFORMATION 
Version users movies size Sparsity 
ML 943 1682 100K 93.695% 
 
The training data matrix is shown in Table . 
The complete algorithm steps are described as follows: 
a) Compute the baseline predictors model. According to 
the formula (5) combined with the rating matrix, personal 
habits, item quality and category, the least square method is 
used to solve for 
mi
b , 
m
b , 
i
b , and 
i
g
g G
c

 . The solution 
formulas are (6) and (7). 
b) Compute similarity. Using formula (4), compute the 
similarity between each two users. 
c) Getting a set of nearest neighbors. According to the 
similarity computed in the steps (b), we sort the nearest 
neighbors of user m  that need to be predicted, and determine 
the relevant user set 
k
m
S  according to sorting order and k. 
d) Rating prediction. According to formula (3), system 
recommends items that the users have rated highly but not 
yet being rated by this user. 
IV. EXPERIMENTAL EVALUATION 
A. Data Set And Evaluation Metrics 
In order to verify the actual recommendation effect of the 
proposed algorithm (Improved Pearson Similarity 
Collaborative Filtering, IP-CF) in this paper, the MoiveLens 
film data set was used for verification. This data set consists 
of: ①100,000 ratings (1-5) from 943 users on 1682 movies. 
②User data and item data have simple feature portraits. ③
Users with less complete personal portraits and fewer 
comments in the data have been cleaned. The dataset 
information is shown in Table . 
Select the root mean square error (RMSE) and mean 
absolute error (MAE) to evaluate the accuracy of the 
recommendation algorithm on the rating data[12-13]. The 
smaller the value, the higher the accuracy of the prediction. 
For a user u  and item i  in the test set, uir  is the actual 
rating, 
uirˆ  is the predicted rating, and T is the total number 
 
Figure 1.   Comparison of precision between IPCF algorithm and 
traditional algorithm 
of items that need to be predicted. The RMSE and MAE 
formulas are as follows: 
 


Tiu
uiui rr
T
RMSE
,
2
ˆ
1




Tiu
uiui rr
T
MAE
,
ˆ
1

B. Experimental Results 
1)  Comparison of accuracy of recommendation 
algorithms 
The traditional cosine similar algorithm (Cosine-CF) and 
Pearson similar algorithm (Pearson-CF) were compared with 
the proposed algorithm (IP-CF). We tested them on our data 
sets by computing RMSE and MAE. The size k of similar 
user set is from 5 to 180. Figure 1 shows the experimental 
results. 
It can be observed from the results that the RMSE and 
MAE values of the improved similarity algorithm proposed 
in this paper decrease with the increase of the neighborhood. 
When the number of near-neighbor sets reaches a certain 
amount, it tends to a fixed value. The traditional 
collaborative filtering algorithm (Pearson similarity and 
cosine similarity) needs to find the optimal result, if the 
number is too large, it will affect the accuracy of the 
recommendation result. Overall, the RMSE and MAE of the 
rating prediction are 0.82% and 1.16% lower than the  
 
 
Figure 2.  Comparison between the baseline predictors model and the 
traditional model 
2018 International Conference on Sensor Network and Computer Engineering (ICSNCE 2018) 
100 
 
traditional algorithm respectively. The IP-CF algorithm has 
better accuracy. 
2) Comparison of accuracy of baseline predictors 
The following experiments verify the effectiveness of the 
baseline predictors model. The experiment compares the user 
mean model (BU), basic baseline predictors model (BP), and 
improved baseline predictors model (UBP). Experimental 
results show that the improved baseline predictors model 
significantly improves the accuracy of the baseline 
prediction. 
V. CONCLUSION 
In this paper, a collaborative filtering recommendation 
algorithm based on improved similarity computation is 
proposed, which takes into account user portrait, item 
characteristics and user behavior data in recommendation 
process. Experiments have shown that user portraits and item 
features played an important role in improving the accuracy 
of recommendations, and which are an important basis for 
analyzing potential needs. Secondly, we found that in the 
Top-N recommendation, the number of neighbors and the 
evaluation index are not a positive or negative relationship, 
and the size of the neighbor will affect the accuracy of the 
recommendation. Our further work will research the 
relationship between the number of neighbors and the 
effectiveness of recommendations, especially how to choose 
the best neighbor value to improve the accuracy of 
recommendations.  
 
REFERENCES 
[1] Statistical Report on Internet Development in China[R]. China 
Internet Network Information Center, 2017. 
[2] Borkar V, Carey MJ, Li C. Inside Big Data management:ogres, 
onions, or parfaits[C]. Proceedings of the 15th International 
Conference on Extending Database Technology. ACM, 2012:3-14. 
[3] WANG Guoxia, LIU Heping. Survey of personalized 
recommendation system[J]. Computer Engineering and Applications, 
2012, 48(7), 66-76. 
[4] Sarwar B, Karypis G, Konstan J, et al. Item based collaborative 
filtering recommendation algorithms[C]. Proc 10th Int’l WWW Conf, 
Hong Kong, 2001:1-5. 
[5] Chen YL , Cheng LC. A novel collaborative filtering approach for 
recommending ranked items[C]. Expert Systems with Applications, 
2008, 34(4): 2396 -2405. 
[6] Yang MH, Gu ZM. Personalized recommendation based on partial 
similarity of interests[C]. Advanced Data Mining and Applications 
Proceedings, 2006, 4093:509-516. 
[7] FU He-gang, WANG Zhu-wei. Improvement of Item-Based 
Collaborative Filtering Algorithm[J]. Journal of Chongqing 
University of Technology(Natural Science), 2010(9):71-72． 
[8] GUO Lei, MA Jun. Incorporating Item Relations for Social 
Recommendation[J]. Chinese Journal of Computers, 2014, 37(1):220-
227. 
[9] X. Luo, Xia, Y. and Zhu, Q. Incremental collaborative filtering 
recommender based on regularized matrix factorization[J]. 
Knowledge-Based Systems, 2012, 27, pp.271-280. 
[10] SUN Chen, XI Hongsheng, GAO Rong. A Recommendation-Support 
Model Using Neighborhood-Based Linear Least Squares Fitting[J]. 
Journal of Xi'an Jiaotong University, 2015, 49(6), 78-83. 
[11] C. Rana, and Jain, S.K. A study of the dynamic features of 
recommender systems[J]. Artificial Intelligence Review, 2015, 43(1), 
pp.141-153. 
[12] Item-network-based collaborative filtering: A personalized 
recommendation method based on a user’s item network[J]. Taehyun 
Ha, Sangwon Lee. Information Processing and Management. 2017(5). 
[13] Gai Li, Zhiqiang Zhang, et al. One-Class Collaborative Filtering 
Based on Rating Prediction and Ranking Prediction[J]. Knowledge-
Based Systems. 2017. 
 


A Collaborative Filtering Recommendation Algorithm with Improved Similarity Calculation

Rzepnikowska, Weronika

Kochański, Andrzej

A role for the GDAP1 gene in the molecular pathogenesis of Charcot-Marie-Tooth disease

Morosanu, Cezar Octavian

Jurca, Razvan Lucian

Simonca, Larisa

Ilies, Roxana Flavia

Moldovan, Remus

Florian, Ioan Stefan

Filip, Gabriela Adriana

Experimental cerebral hemispherectomy in rodent models. A systematic review of current literature

REMENÁR, MATEJ

KAMLÁROVÁ, ANNA

HARICHOVÁ, JANA

ZÁMOCKÝ, MARCEL

FERIANC, PETER

The Heavy-Metal Resistance Determinant of Newly Isolated Bacterium from a Nickel-Contaminated Soil in Southwest Slovakia

DETERMINATION OF MECHANICAL PROPERTIES OF METHYL METHACRYLATE ADHESIVE (MMA)

BRACHACZEK, Wacław

INFLUENCE OF SELECTED ADMIXTURES ON THE MICROSTRUCTURE OF RENOVATION PLASTER MORTARS

Aldhafiri, Sarah

Mahmoud, Huda

Al-Sarawi, Mohammed

Ismail, Wael A.

S1 
 
Supplementary Material 
Aldhafiri S. et al., Natural Attenuation Potential of Polychlorinated Biphenyl-Polluted Marine Sediments, 
Polish Journal of Microbiology, 2018, Vol. 67, No 1 
 
 
Table S1 
PCB congener concentrations (ng/g dry weight) in sediments collected from seven 
stations form Shuwaikh harbor. The hatched cells represent levels below the experimental 
detection limit (DL). 
   
 
 
Sample 
ID 
Sample 
ID 
Sample 
ID 
Sample 
ID 
Sample 
ID 
Sample 
ID 
Sample 
ID 
 
  W1 W2 W3 W4 W5 W6 W7  
PCB 
congener 
Cl  
atoms 
4 
detected 
5 
detected 
9 
detected 
2 
detected 
35 
detected 
27 
detected 
3 
detected 
TOTAL 
PCB 
28 & 31 3 < DL < DL < DL < DL 14.55 < DL < DL 14.55 
33 3     2.88   2.88 
44 4     14.64   14.64 
49 4   1.72  18.86 11.14  31.73 
52 4 21.76    53.66 14.26  89.68 
70 4   1.93  21.86 4.64  28.42 
74 4     5.01   5.01 
82 5     4.70 3.77  8.47 
87 5  1.58 1.55  34.76 18.51  56.40 
95 5     45.35 14.35  59.70 
99 5  1.69 1.53  42.95 13.64  59.81 
101 5     78.56 27.47  106.03 
105 5   0.84  19.30 20.25  40.39 
110 5 3.44 2.95 2.89 1.79 76.38 31.94 0.68 120.06 
118 5 2.54 2.69 2.56 1.79 65.54 30.69 0.72 106.51 
128 6     13.50 15.24  28.74 
132 6     21.18 22.89  44.08 
138 6     83.61 81.78  165.39 
149 6  2.80 2.48  2.69 26.43 0.65 35.06 
151 6   1.01  13.91 7.18  22.10 
153 6     74.28 29.22  103.50 
156 6     8.62 9.72  18.34 
169 6        0 
170 7     15.11 9.90  25.01 
171 7     4.37 3.04  7.41 
177 7     7.94 3.92  11.86 
180 7     33.72 11.37  45.09 
183 7     7.41 3.36  10.78 
187 7     14.55 4.73  19.28 
S2 
 
191 7     1.35   1.35 
194 8     4.50 4.08  8.58 
195 8     5.83   5.83 
201 8     2.63   2.63 
205 8     4.13 3.43  7.56 
206 9        0 
208 9        0 
209 10 49.44    20.38 41.41  111.23 
TOTAL  77.17 11.70 16.51 3.58 838.72 468.38 2.05 1418.10 
S3 
 
Table S2 
Surrogate (
13
C-labeled) PCB recovery data (%). 
 Sample 
ID 
Sample 
ID 
Sample 
ID 
Sample 
ID 
Sample 
ID 
Sample 
ID 
Sample 
ID 
13
C-
PCB 
W1 W2 W3 W4 W5 W6 W7 
28 88 58 73 67 77 77 97 
52 70 42 60 59 75 78 86 
101 99 55 73 73 93 109 100 
138 80 44 59 61 74 88 79 
153 133 66 85 88 107 113 113 
180 115 69 90 93 110 128 125 
209 91 56 80 85 82 109 110 
 
S4 
 
Uncultured bacteria (DGGE gel bands recovered from the sediment samples): 
 DEFINITION  Uncultured bacterium DGGE gel band F17 16S ribosomal RNA 
gene, partial sequence. ACCESSION   KU663378 
 DEFINITION  Uncultured bacterium DGGE gel band F18 16S ribosomal RNA 
 gene, partial sequence. ACCESSION   KU663379 
 DEFINITION  Uncultured bacterium  DGGE gel band F19 16S ribosomal RNA 
gene, partial sequence. ACCESSION   KU663380 
 DEFINITION  Uncultured bacterium DGGE gel band F21 16S ribosomal RNA 
gene, partial sequence. ACCESSION   KU663381 
 DEFINITION  Uncultured bacterium DGGE gel band F22 16S ribosomal RNA 
gene, partial sequence. ACCESSION   KU663382 
 DEFINITION  Uncultured bacterium  DGGE gel band R23 16S ribosomal RNA 
gene, partial sequence. ACCESSION   KU663383 
 
 DEFINITION  Uncultured bacterium DGGE gel band R_23 16S ribosomal RNA 
gene, partial sequence. ACCESSION   KU663384 
 DEFINITION  Uncultured bacterium DGGE gel band F241 16S ribosomal RNA 
gene, partial sequence. ACCESSION   KU663385 
 DEFINITION  Uncultured bacterium  DGGE gel band R_24 16S ribosomal RNA 
gene, partial sequence. ACCESSION   KU663386 
 DEFINITION  Uncultured bacterium DGGE gel band F27 16S ribosomal RNA 
gene, partial sequence. ACCESSION   KU663387 
 DEFINITION  Uncultured bacterium DGGE gel band F28 16S ribosomal RNA 
gene, partial sequence. ACCESSION   KU663388 
 DEFINITION  Uncultured bacterium  DGGE gel band R_28 16S ribosomal RNA 
gene, partial sequence. ACCESSION   KU663389 
 DEFINITION  Uncultured bacterium DGGE gel band F30 16S ribosomal RNA 
gene, partial sequence. ACCESSION   KU663390 
 DEFINITION  Uncultured bacterium DGGE gel band R31 16S ribosomal 
RNAgene, partial sequence. ACCESSION   KU663391 
 DEFINITION  Uncultured bacterium DGGE gel band S 16S ribosomal RNA 
gene, partial sequence. ACCESSION   KU663392 
 
 
 
 
 
Sequences of DGGE bands recovered from the biphenyl enrichment Culture: 
S5 
 
DEFINITION  Paenibacillus sp. W21A 16S ribosomal RNA gene, partial sequence. 
ACCESSION   KU682428 
DEFINITION  Paenibacillus sp. W21B 16S ribosomal RNA gene, partial sequence. 
ACCESSION   KU682429 
DEFINITION  Paenibacillus sp. W24 16S ribosomal RNA gene, partial sequence. 
ACCESSION   KU682430 


Natural Attenuation Potential of Polychlorinated Biphenyl-Polluted Marine Sediments

Zhongsheng, Wang

Sen, Yang

Hairui, Dong

Quadrotor Formation Inversion Control Method Based on Unit Quaternion

Not a member yet