Computer and Modernization

Application of Improved ORB Algorithm in Image Matching

FAN Xin-nan, GU Ya-fei, NI Jian-jun

2019, 0(02): 1. doi:10.3969/j.issn.1006-2475.2019.02.00

Asbtract ( 343 )

PDF (6536KB) ( 234 )

References | Related Articles | Metrics

In the field of computer vision, image matching is a core issue. In order to improve the accuracy of image feature matching and enhance its anti-interference ability, in view of the insufficiency of ORB (oriented FAST and rotated BRIEF), an improved method based on ORB is proposed. This algorithm sets self-adaptive threshold in the process of detecting feature points. Furthermore, the outer points which do not conform to the geometric characteristics of images are removed after the coarse matching, then the affine invariant constraint is chosen to pick out the exact matching points. The experimental results show that the proposed method has greatly improved the matching quality and has low computing cost. In addition, the proposed method has strong robustness for different blur and exposure degrees.

Somatosensory Interaction Method Based on Deep Learning

TANG Hui, WANG Qing, CHEN Hong，GUO Hao

2019, 0(02): 7. doi:10.3969/j.issn.1006-2475.2019.02.002

Asbtract ( 215 )

PDF (3081KB) ( 115 )

References | Related Articles | Metrics

With Microsoft’s announcement of a permanent discontinuation of Kinect products in October 2017, there is an urgent need for a Kinect replacement in the field of somatosensory interaction. This article uses a normal monocular camera to read the video stream in real time. The Faster-RCNN network is used to detect the position of the human body and frame the human body. The non-maximum suppression algorithm is improved, and a linear weighting function is introduced to reduce the detection frame score of the IOU greater than the threshold instead of becoming zero. Secondly, according to the obtained detection frame, the CPM network is detected by the key points of the human body, and the coordinate position of the whole body skeleton point is outputted, and Center Loss is introduced to increase the cohesiveness and inter-class difference of the intra-class features of the key points. Finally, according to the template matching method, a control instruction of the somatosensory interaction is generated according to the recognition result. The method of this paper reduces the dependence on professional equipment, simplifies the complexity of somatosensory interaction, and has important value for promoting the popularity of somatosensory and expanding the scope of human-computer interaction.

An Image Style Transfer Algorithm Based on Multi-dimension Histograms Matching

LI Yu-ni, WANG Peng, XIAO Jian-li

2019, 0(02): 15. doi:10.3969/j.issn.1006-2475.2019.02.003

Asbtract ( 254 )

PDF (1396KB) ( 123 )

References | Related Articles | Metrics

Based on the classical image style transfer algorithm using histogram matching, a new image style transfer algorithm has been proposed, which uses multi-dimension histograms matching. Firstly, different dimension histograms are generated. Then, the histogram matching is performed between the original and target images. As a result, the images after style transfer using different dimension histograms are obtained. Finally, the image having the same style with the target image is produced by fusing the images obtained in the histogram matching step. Experimental results show that the proposed algorithm can ensure that the original and target images obtain the best similarity in global scale. Meanwhile, the proposed algorithm can also make the generated image save the most details from the original image. Namely, the generated image has the similarity between the original and target images as much as possible and the local details from the original image are saved as many as possible simultaneously.

A Part Image Mosaic Method Based on Sparse Feature Points

ZHANG Qin1,2, JIA Yuan1, WANG Yao-bin1

2019, 0(02): 19. doi:10.3969/j.issn.1006-2475.2019.02.004

Asbtract ( 182 )

PDF (6423KB) ( 101 )

References | Related Articles | Metrics

Aiming at the poor results that many image mosaic methods based on the feature points are used to parts image with weak texture, few features and many similar regions, an improved method is proposed. Firstly, the feature points are extracted by the FAST method, and the candidate points sets for matching are generated from them. Secondly, the gray feature values are sampled by template region, and the points matching results are obtained via setting the rotation angle and scaling search regions with the structural similarity index measurement (SSIM) method. Finally, the rotation, scaling, translation parameters are worked out by the points matching results. The last results of the parameters are gained through 3σ principle removing abnormal values. The experiment results show that the new method could get good rotation, scaling, translation parameters and mosaic image when angle search region is［-45°,+45°］ and scaling search region is［0.5,1.5］.

An Improved Image Dehazing Algorithm Based on Dark Channel Prior

BAO Bin1, LI Ya-gang2

2019, 0(02): 27. doi:10.3969/j.issn.1006-2475.2019.02.005

Asbtract ( 192 )

PDF (3129KB) ( 111 )

References | Related Articles | Metrics

Aiming at the problem that the dark channel prior image dehazing algorithm produces color distortion and dim image in offwhite scenery or bright sky, an improved method based on dark channel prior is proposed. By correcting the computing problem of transmission rate which causes color distortion, the visual effect of the image can be improved. In addition, an enhanced free-haze image is obtained by reducing the high brightness value of the three color channels and using the mean value method. The experimental results show that the proposed method largely eliminates color distortion in the bright areas of the dehazed image and has better color degrees.

Optimized Image Classification Method by Double Entropy Fast Extraction ROI

ZHAO Xiao-lei1, XU Xi-bin2

2019, 0(02): 31. doi:10.3969/j.issn.1006-2475.2019.02.006

Asbtract ( 173 )

PDF (2139KB) ( 134 )

References | Related Articles | Metrics

A multi-features optimized image classification method based on region of interest (ROI) extracting method using color entropy extreme value and color entropy mutual information is proposed. Firstly, the most relevant region is determined by the color entropy extreme value, then the continuous ROI region is determined by using entropy mutual information to grow sub-region quickly. The Dense-SIFT characteristic description is extracted based on the ROI region, and a visual dictionary is generated by K-means method. In order to use the spatial local information, the pyramid matching method is adopted. Finally, the characteristics are input into SVM for classification. In the Caltech101 and Caltech256 databases, 8 data sets are selected for experiment. The average classification accuracy is improved by 6.86% obtained by using ROI extraction algorithm and the convergence rate is improved by nearly half. After adding the color entropy and the color third moments, the classification accuracy is further increased by 2.36%, it is 9.22% higher than before improvement totally.

Application of Adaptive Weight CV Model in Ultrasonic Phased Array Image Segmentation

LIU Yong-luo1, WANG Wen-qiang2, MA Li-wu2

2019, 0(02): 37. doi:10.3969/j.issn.1006-2475.2019.02.007

Asbtract ( 161 )

PDF (1493KB) ( 120 )

References | Related Articles | Metrics

Image segmentation is the key link in the 3D reconstruction of ultrasonic phased array NDT, and the accuracy and efficiency of the segmentation are the important guarantee for the accuracy and real time of the 3D reconstruction of the image. The ultrasonic phased array NDT image contains a lot of noise and intensity in homogeneity. Traditional CV model uses fixed iterative step to segment the ultrasonic phased array NDT image, which leads to low segmentation efficiency and low precision. This paper makes use of the characteristics of the over segmentation of the traditional watershed algorithm to make a watershed transformation to the image. The number of pixels and intensity information in each region are counted, and a weight matrix is obtained. The matrix is introduced into the CV model, and the adaptive weight CV model is obtained. In the iterative process of level set function, the weight matrix can adaptively adjust the iteration step size according to the image information. The experiment shows that, in the segmentation of the ultrasonic phased array image, compared with the CV model and the LBF model, the weighted CV model proposed in this paper has higher efficiency and segmentation accuracy.

Classification and Identification of Magnetic Resonance Images of #br# Duchenne Muscular Dystrophy with Convolutional Neural Network

ZHANG Ming-huan1, CHEN Ying1, SHEN Ying2, CHENG Ai-lan2, LIU Xiao-qing2

2019, 0(02): 43. doi:10.3969/j.issn.1006-2475.2019.02.008

Asbtract ( 162 )

PDF (1248KB) ( 119 )

References | Related Articles | Metrics

Duchenne muscular dystrophy (DMD) is a fatal skeletal muscle hereditary disease. The conventional treatment is invasive, which incurs great sufferings. Therefore, this paper explores a non-invasive detection method on the basis of magnetic resonance images (MRI) of the patients. 485 experimental MRIs are obtained with the guidance of senior physicians of neuromuscular department. These images are divided into two control groups: the patient group and the healthy group; each group includes two weighted images, T1 and T2. A 10-hidden layer depth convolutional neural network (CNN) is designed and used to directly read T1 and T2, and classify them. The results show: firstly, by increasing the numbers of network parameters and iterative optimizations, the accuracies of image recognition have reached 99.2% and 98.9% respectively; secondly, both T1 and T2 can be used to well distinguish between patient and healthy groups; thirdly, in comparison with KNN, LR, DT and SVM algorithms, the accuracy of classification with the CNN algorithm is best. In particular, the CNN algorithm improves the recognition accuracy of T2 images, and greatly explores the utilization value of T2 images. Therefore, using CNN for DMD image classification and recognition, because of its high accuracy, lossless image information and other characteristics, it is expected to provide an objective and effective auxiliary diagnosis means for clinical; this is a new exploration of application of artificial intelligence in the field of DMD non-invasive detection.

Human Pose Recognition Based on CNN

ZHOU Yi-kai, WANG Yu, ZHAO Yong-fei, YUAN Yan

2019, 0(02): 49. doi:10.3969/j.issn.1006-2475.2019.02.009

Asbtract ( 408 )

PDF (1748KB) ( 208 )

References | Related Articles | Metrics

Human posture recognition is one of the important research topics in human-computer interaction. With the development of machine learning and neural networks, the research methods and results tend to be diversified, and the application value of gesture recognition is becoming more and more extensive. This paper constructs a convolutional neural network model, which has 11 layers. It convolves and pools five kinds of human poses in the sampled data set, and finally enters the fully connected layer for classification, thus completing the training and identification of the data set. The results show that compared with the machine learning method, the recognition performance of the model is more excellent, and the complex feature extraction mode design is eliminated, so that the network itself extracts features to identify and classify better.

Longitudinal Shredded Paper Stitching Method Based on Edge Matching

LIU Wei, WANG Jun-min, LIU Ke-qin

2019, 0(02): 55. doi:10.3969/j.issn.1006-2475.2019.02.010

Asbtract ( 161 )

PDF (1424KB) ( 129 )

References | Related Articles | Metrics

This article elaborates on a method of recovery for torn slivers of rules. The algorithm first transforms the image into the corresponding pixel matrix, and the matrix is processed by two vales. Next, according to the continuity of text strokes, the matching matrix of edge information is calculated, and then the similarity matrix between fragments is calculated according to the continuity between texts. Finally, combining the information of matching degree matrix and similarity matrix, when matching a certain piece, the corresponding next one is highly consistent with the matching degree matrix as well as the similarity matrix, so as to achieve the splicing and restoration of broken paper. The verification of experimental shredding stitching shows that the effect is good, the method is accurate and convenient. It is feasible.

A Random Maximum Likelihood Algorithm Based on Limited PSO Initial Space

CAI Li-ping, TIAN Hui, CHEN Hai-hua, HU Jia-liang

2019, 0(02): 60. doi:10.3969/j.issn.1006-2475.2019.02.011

Asbtract ( 162 )

PDF (847KB) ( 127 )

References | Related Articles | Metrics

Aiming at the problem of large computational complexity due to multidimensional nonlinear optimization in the DOA estimation of the random maximum likelihood algorithm, a SML algorithm is proposed to limit the search space of particle swarm optimization algorithm. This algorithm overcomes a defect that when the ESPRIT algorithm defines the initial space of PSO, the signal cannot be directly processed by ESPRIT algorithm when the array structure is a non-uniform linear array and the signal is a coherent signal, and we need to adopt a set of preprocessing techniques, which increases the complexity of algorithm calculation. The key point of the proposed algorithm is to use the hypothesis technique to determine the initialization point instead of the solution of the ESPRIT algorithm, the initial solution space of the PSO algorithm is determined by combining the CRRAM. This method eliminates the need for preprocessing techniques, and the algorithm that defines the PSO initialization space greatly reduces the computational complexity of the SML algorithm. Experimental results show that the proposed algorithm provides fairly good initial values for both coherent and incoherent cases. Finally, the proposed algorithm is compared with many existing algorithms, and the validity and accuracy of the proposed algorithm are verified.

Vehicle Transit Time Prediction Based on Multi-model Fusion

LIU Yin-ping, MA Shao-hui

2019, 0(02): 66. doi:10.3969/j.issn.1006-2475.2019.02.012

Asbtract ( 118 )

PDF (1376KB) ( 95 )

References | Related Articles | Metrics

The time that a vehicle passes through a certain road network is one of an important indicator to measure traffic congestion degree. In order to improve the prediction accuracy of vehicle transit time, it is necessary to consider not only the influence of data acquisition accuracy but also the choice of model. This paper proposes a method of the multi-model fusion to predict vehicle transit time, and it is found that the multi-model fusion has higher prediction accuracy. Vehicle traffic monitoring data of three intersections of a highway is empirical data. Comparing the multi-model fusion algorithm with a single model, it shows the application potential of multi-model fusion algorithm in the field of traffic congestion management.

Trajectory Prediction Based on Gaussian Mixture-Bayesian Model

ZHU Kun

2019, 0(02): 72. doi:10.3969/j.issn.1006-2475.2019.02.013

Asbtract ( 346 )

PDF (3752KB) ( 170 )

References | Related Articles | Metrics

Nowadays, real-time, accurate and reliable track prediction of moving objects plays a very important role in traffic management system, military mechanized battlefield and safe driving system， which has been applied more and more widely in the market, namely intelligent prediction. Intelligent prediction can provide accurate location-based services, and it can also recommend optimal routes to car owners based on pre-judgment, which has become a hot spot of research on mobile object database. Aiming at the shortcomings of the existing methods, a Gaussian mixture-Bayesian trajectory prediction model is proposed. The experimental results show that the GM-BM model can predict the most likely trajectory by adjusting the weight of the neutron model of the mixed model under the normal traffic flow of road section. After calculation, the prediction accuracy is improved by at least 10.00% compared with the single model under the same parameter setting.

Solving Cube Robot System Based on Two-stage Bidirectional Search

ZHANG Jin-long, ZOU Yu-long, YANG Bin, YAO Can-jie, ZHENG Yao-zong

2019, 0(02): 82. doi:10.3969/j.issn.1006-2475.2019.02.014

Asbtract ( 240 )

PDF (2546KB) ( 217 )

References | Related Articles | Metrics

In recent years, the robots of the Rubiks cube have emerged in an endless stream. The current Rubiks cube robots have defects such as instability, high cost and large size. For this, a new type of Rubiks cube robot is developed. Rubiks cube design uses a wide range of knowledge, including mechanical design, system control, visual inspection, algorithms and so on. In order to make it completely intelligent, the third-order cube is arbitrarily restored in the shortest time. The overall design structure consists of a camera and a two-finger robot. The camera captures the image and transmits it to the host computer. The STM32F103C8T6 minimum system controls the rotation of the robot arm fingers. In the case of low cost, the Rubiks cube is restored faster and more accurately than human calculations.

Sentiment Analysis Based on Bi-LSTM with Multimedia Information

DING Yan1,2, BAO Yan1,2, HU Xiao1,2

2019, 0(02): 88. doi:10.3969/j.issn.1006-2475.2019.02.015

Asbtract ( 135 )

PDF (708KB) ( 109 )

References | Related Articles | Metrics

With the rapid development of information technology, the construction of intelligent government is in full swing in China. In order to serve society better, sentiment analysis is to be important in the future. However, due to the variety of the media data, such as the content of tweet, the title of topics, the reply and the limitation of the content, we not only need to analyze the content of tweets, but require to pay attention to some media information, such as reply and topics. While few work study media information to model context, this paper proposes a new model CBi-LSTM (Content Bi-LSTM) to study sentiment classification of multimedia information, such as reply, topic and so on. Experiments show that the fusion of text information and multimedia information can improve sentiment classification remarkably.

Commodity Attributes Extracting in Chinese Shopping Reviews Based on Bi-LSTM and CRF

ZHANG Shi-lin

2019, 0(02): 93. doi:10.3969/j.issn.1006-2475.2019.02.016

Asbtract ( 275 )

PDF (728KB) ( 119 )

References | Related Articles | Metrics

With the improvement of the evaluation system of e-commerce system, the content of online shopping reviews plays a very important role in guiding consumers shopping. However, consumers cant find attributes and evaluations about attributes directly from a lot of reviews. Compared with constructing knowledge base and traditional machine learning methods, we need to summarize complex features and rules manually to extract attributes and attribute evaluations. This paper applies the method of Bi-directional Long Short-Term Memory (Bi-LSTM), Conditional Random Fields (CRF) and POS features to realize automatic extraction of commodity attributes and attributes evaluations in the reviews. This avoids summarizing the rules and has more domain universality. Through testing camera, menswear and child safety seat, the three commodity areas have obtained the macro precision of 86.74% and the macro recall of 85.89%.

Virtual Flip Book System Based on Infrared Induction and Multimedia Technology

SHEN Yue-jie, ZHAO Zheng-xu, ZUO Zong-cheng

2019, 0(02): 98. doi:10.3969/j.issn.1006-2475.2019.02.017

Asbtract ( 142 )

PDF (2614KB) ( 109 )

References | Related Articles | Metrics

As a new propaganda method of popular science, the virtual book flipping system not only displays large amount of information, but also shows novel forms. It is easier to attract users to stop and watch, so as to achieve the goal of popularizing science. By comparing the existing virtual flip book systems and analyzing the main applications of them, a virtual flip book system using infrared induction technology and computer multimedia technology is proposed. The system first uses the Flash software to make content into an e-book with the effect of turning the book, and then sends the content to the book screen through the projection device. Two infrared cameras on the side of the book screen are used to capture the visitors’ wave from left to right or from right to left. Then the wave motion is converted into interactive instruction to drive Flash e-book to do book flipping effect. The system is suitable for a variety of users to use, in line with reading habits, and no special hardware requirements, low production cost, suitable for the general public to promote.

Big Data Log Collection System Based on Self-controlled Container Platform

WU Xin-quan, YANG Jun

2019, 0(02): 102. doi:10.3969/j.issn.1006-2475.2019.02.018

Asbtract ( 184 )

PDF (801KB) ( 147 )

References | Related Articles | Metrics

With the application of technologies such as cloud computing, virtualization, container cloud, and domestically-controlled and controllable requirements, more and more services will be deployed on domestic servers and self-controlled container cloud. In the process of running the service, we need to obtain log data for a series of monitoring, statistics, analysis and forecasting work. However, due to the particularity of the domestic server and the particularity of self-controlled container cloud, the traditional log collection methods (including real-time acquisition and centralized acquisition) are not well suited for self-controlled container cloud. In order to improve the quality and performance of the service and ensure a certain comprehensiveness of the log data, an adaptive data acquisition algorithm is proposed, which can self-adjust the log data collection and transmission according to the load of the server. When the server load is low, the number of data collection and transmission is increased, even in real-time collection; when the server load is high, the collection and transmission of log data is reduced, thereby reducing the pressure on the server load, and improving the quality and performance of the service itself. Finally, theoretical and experimental analysis proves that the data acquisition algorithm can effectively alleviate the pressure of domestic servers while ensuring a certain comprehensiveness of data.

Network Confrontation Deduction System Based on Offensive and Defensive Action Chain

YIN Fa, AI Zhong-liang

2019, 0(02): 107. doi:10.3969/j.issn.1006-2475.2019.02.019

Asbtract ( 269 )

PDF (4625KB) ( 131 )

References | Related Articles | Metrics

Network confrontation deduction is an important part of task planning and evaluation decision in network confrontation. In view of the characteristics of wide dimension, multi-objective and short burst of network confrontation, the traditional deduction system can not complete the requirements of network confrontation deduction. In order to solve this problem, this paper designs a network confrontation deduction system based on the offensive and defensive action chain. The system consists of database, visualization platform, attack and defense chain structure, deduction control engine, model construction, application layer and other subsystems. The modules of each subsystem cooperate with each other to drive the overall network against the deduction system in accordance with the deduction plot, and form a chain action plan or structure with offensive and defensive game. This paper calls it the offensive and defensive action chain. The proposal of the offensive and defensive action chain enables the deductive system to better perform the scheduling and decision making during the deduction process, and also provides a reference research direction for the network attack and defense confrontation.

Optimization and Test of LB-AGR Routing Protocol #br# in Underwater Acoustic Network Test-bed

LI Chong1,2, DU Xiu-juan1,2, WANG Li-juan1,2

2019, 0(02): 112. doi:10.3969/j.issn.1006-2475.2019.02.020

Asbtract ( 175 )

PDF (3303KB) ( 93 )

References | Related Articles | Metrics

Aiming at the current situation that the research on underwater acoustic network are mostly based on simulation, the LB-AGR protocol is implemented, tested, analyzed and optimized on the real time monitoring test-bed of Qinghai lake. Further we analyze the problems arose in the procedure of experiment: counting to infinity of node’s level and status deadlock of Sink node. Design optimization is carried out by setting timer and improving the implement scheme of LB-AGR protocol, and the optimized protocol is retested. Test results show that the optimized LB-AGR scheme not only solves the problems of counting to infinity of node’s level and status deadlock of Sink node, but also delivers data more quickly and stably.

A New Transmission Line Foreign Object Detection Network Structure: TLFOD Net

SHEN Mao-dong1, PEI Jian1, FU Xin-yang1, ZHANG Jun-ling1, GONG Fan-kui1, LIU Xia2

2019, 0(02): 118. doi:10.3969/j.issn.1006-2475.2019.02.021

Asbtract ( 273 )

PDF (4564KB) ( 128 )

References | Related Articles | Metrics

Floating foreign bodies suspended on high voltage transmission lines may cause great harm to power transmission, but the existing object detection methods can not effectively identify irregular objects. This paper proposes a new network structure for foreign body detection: TLFOD Net (Transmission Line Foreign Object Detection Net). According to the characteristics of foreign bodies, this paper designs the TLFOD Net network structure, which mainly includes feature extraction network, region proposal network and classified regression network, optimizes suitable candidate boxes and proposes end-to-end joint training mode to improve the performance of TLFOD Net. Through image reversal technology, the number of training sets is increased. The experimental results show that TLFOD Net improves the detection speed and accuracy significantly compared with the existing networks.

An Improved HK Social Network Modeling Method

CHEN Jing-yi, XU Ming-hai, YANG Xi, DU Fan

2019, 0(02): 123. doi:10.3969/j.issn.1006-2475.2019.02.022

Asbtract ( 202 )

PDF (605KB) ( 99 )

References | Related Articles | Metrics

With the deepening of research on complex networks, social network modeling has become one of research hotspots. Based on Holme and Kim (HK) model, we put forward an improved HK social network model. We not only consider “preference connections” and “triangular structures”, but also consider two network growth patterns named “internal evolution” and “external extension” when adding new nodes in the network. A creative way of network dynamic evolution including nodes saturation and links refreshing is put forward based on traditional one-way growth network. The simulation results show that the improved model has power-law degree distribution, larger clustering coefficient and smaller average shortest path which satisfies small-world effects and scale-free properties at the same time. The whole social network model is spiraling in the process of link building and blocking and reproduces characteristics of real social network better.

Table of Content