Computer and Modernization

Fast Paper Edge Detection Method Based on HED Network

ZHAO Qi-wen, XU Kun, XU Yuan

2021, 0(05): 1-5.

Asbtract ( 483 )

PDF (2564KB) ( 185 )

References | Related Articles | Metrics

The Holistically-nested Edge Detection （HED） network is one of the deep learning network models with better edge detection performance at present. However, when the HED is used for edge detection of paper, the detection speed is slow and cannot meet the real-time requirements. On the premise of ensuring the detection accuracy, this paper proposes a fast paper edge detection method based on HED network. This article uses the lightweight network MobileNetV2 as the HED backbone network, and removes the last two bottleneck modules of the MobileNetV2 network and the convolutional layer with a large number of output channels to further accelerate the detection speed. In addition, the pooling layer in the network is removed, and a 5×5 convolutional layer with a step length of 1 is added to improve the detection accuracy. A paper data set MPDS containing a variety of situations is produced, the method proposed in this paper is trained and tested on MPDS. The experimental results show that the proposed model increases the ODS and OIS indicators to 0.867 and 0.876, respectively. The detection speed is 42.68 FPS. The method proposed in this paper can quickly and accurately detect the edge of the paper and meet the requirements of the desktop enhancement system for paper detection.

Cloth Simulation Filtering Algorithm Based on Elevation Normalization

CHEN Xi-liang, WANG Xue , BI Xiao-wei

2021, 0(05): 6-12.

Asbtract ( 343 )

PDF (4346KB) ( 139 )

References | Related Articles | Metrics

The point cloud filtering process is a very important part of LiDAR data processing, that is, to separate ground points and non-ground points in point cloud data so as to lay the foundation for subsequent data processing. Based on the traditional progressive mathematical morphology filtering and cloth simulation filtering methods， this paper considers that the effect of progressive morphological filtering on ground point separation is acceptable, that is, it can basically retain all ground points. However, due to the weak adaptability of terrain, the height difference threshold is also unstable with the change of terrain slope, some non-ground points are easy to be regarded as ground points, and cloth simulation filtering has the advantage of high efficiency of algorithm operation, and the filtering effect of cloth simulation filtering in flat terrain is better than that in areas with large terrain undulations. Based on the results of progressive morphological filtering, the coarse DEM raster data of the target area is established, and then the elevation value of the target point cloud is normalized to eliminate the influence of terrain changes on the cloth simulation filtering. Finally, the experimental results of using three sets of standard data samples on the official website of ISPRS show that the type I error is reduced compared to the result of progressive morphological filtering, the type II error is reduced compared to the result of cloth simulation filtering, and the total error is also reduced, so a better filtering effect is achieved.

Modeling Method of 3D Printing Full Contact Insole Based on Triply Periodic Minimal Surface

JIANG Chao-qun, TONG Jing, GUO Ming-deng, LU Rong-jie, CHEN Zheng-ming

2021, 0(05): 13-19.

Asbtract ( 491 )

PDF (4484KB) ( 144 )

References | Related Articles | Metrics

The full-contact insole can reduce the peak plantar pressure to improve and prevent neurotic ulcer symptoms in diabetic foot. The traditional full-contact insole design method is complex to operate. In this paper, a novel 3D printing modeling method for full-contact insole is proposed based on triply periodic minimal surface （TPMS）. Through three steps of data collection, full-contact model construction and model porousness, a full-contact insole based on TPMS structure is constructed and produced using 3D printing technology. First, the user foot model and the scan insole model are collected. Then the lower boundary of the pre-designed insole model is approximated to the scan insole model by Laplace transform algorithm, and the full-contact insole model is constructed by adjusting the upper surface of the pre-designed insole model to the bottom of foot. Finally, the Marching Cubes mesh reconstruction method is applied to reconstruct the full-contact insole into a TPMS structure based mesh model. The experiment verifies that the method proposed in this paper can design a full-contact insole with the ability to reduce the peak pressure of the sole.

A Hand Gesture Segmentation Method Based on Style Transfer

CHEN Ming-yao, XU Kun, LI Xiao-xuan

2021, 0(05): 20-25.

Asbtract ( 192 )

PDF (3554KB) ( 116 )

References | Related Articles | Metrics

Hand gesture segmentation based on fully convolutional networks excessively dependents on the accurate per-pixel annotations of training data. At the same time, the features lack enough context information, which often leads to misclassification with intra-class inconsistency. In order to solve the above issues, a hand gesture segmentation method based on style transfer is proposed. Firstly, the first five layers of hand gesture segmentation network in HGR-Net are selected as the backbone network, and the context information enhancement layer is added to each layer of backbone network. In the context information enhancement layer, global average pool operation and channel attention mechanism are adopted to enhance the weight of the discrimination feature and ensure the continuity of context information in features, so as to solve the intra-class inconsistency. Secondly, in order to improve the generalization ability of the hand gesture segmentation module proposed by this paper, and address the cross-domain samples segmentation problem, a domain adaptive method based on style transfer is proposed. The pre-trained VGG model is used to transfer the source domain testing sample, so as to make the source domain testing sample have both its content and the style of the target domain training sample. Testing on the OUHANDS dataset, the mIoU and MPA values of the proposed method are 0.9143 and 0.9363 respectively, and they are 3.2 and 1.8 percentage points higher than those of HGR-Net. Testing on the self-collection dataset with the style transfer method, the mIoU and MPA values are respectively 19 and 23 percentage points higher than without this method. The domain adaptive method based on style transfer provides a new idea for cross-domain segmentation of unlabeled samples.

RGB-D Salient Object Detection Based on Depth Image Gain

WEI Ji-peng, QIN Guo-feng

2021, 0(05): 26-30.

Asbtract ( 153 )

PDF (2079KB) ( 114 )

References | Related Articles | Metrics

Depth information has been proved to be practical information for salient object detection, but it is still a matter worth exploring that how can the depth information and RGB information complement each other better so as to achieve higher performance. To this end, this paper proposes a RGB-D salient object detection method based on depth image gain. A gain subnet is added to the double-branch network structure, and the gain of the depth image for saliency detection is obtained by the method of saliency map difference, which is used as a pseudo GT for the gain subnet pre-training. The three-branch network separately obtains RGB features, depth features, and depth gain information, and finally fuses the features of the three branches to obtain the final salient object detection result, the gain information provides the fusion basis for the two branch feature fusion. The experimental results of salient object detection based on depth image gain show that the salient object foreground object obtained by this method is more prominent, and it has better performance on multiple experimental datasets.

A Deep Learning Laser Point Cloud Data Classification Method Using PCA

HUANG Wu-chao, HAN Ling, HUANG Bo-xue, YANG Zhao-hui

2021, 0(05): 31-37.

Asbtract ( 211 )

PDF (2574KB) ( 97 )

References | Related Articles | Metrics

In order to improve the classification accuracy of airborne LiDAR data and avoid the time-consuming point cloud multi-feature extraction, the article extracts the relative elevation feature of the point cloud data based on the point cloud denoising, and proposes a network model based on the combination of PCA data dimension reduction and Point-Net. The acquired relative elevation features and original features are input into the network after dimensionality reduction, and the global features extracted by the Point-Net network model are used for point cloud classification, and the label after each point classification is returned. The classification results are visualized according to the coordinate information and label of the point cloud, and the
classification of airborne LiDAR point cloud data is realized. Finally, the accuracy analysis of the classification results is carried out. The classification experiment shows that the point cloud classification results obtained by this method are better.

Application of Improved Adaptive Hybrid Ant Colony Algorithm in MRCPSP

YAN Sen, WANG Yu-mei

2021, 0(05): 38-43.

Asbtract ( 195 )

PDF (892KB) ( 95 )

References | Related Articles | Metrics

Based on the background of Multi-mode Resource Constrained Project Scheduling Problem (MRCPSP), an adaptive hybrid ant colony algorithm is proposed to solve the balance problem between the convergence speed of ant colony algorithm and restraining the possibility of falling into the local optimum. The range of parameters can be adaptively adjusted synchronously with the operation of the algorithm and ants can obtain random parameter values in the parameter value range, forming a hybrid ant colony. The algorithm introduces the pioneer scout ant, reward mechanism of elite ant colony with ranking factor and upper and lower pheromone limits to optimize the pheromone update strategy. At the same time, the time uncertainty problem in MRCPSP can be solved by using this algorithm based on fuzzy theory. Finally, the simulation results show that, compared with other heuristic project scheduling optimization algorithms, this algorithm can improve the global search ability and the quality of solution, and better solve the MRCPSP problem, so it has higher practical application value.

Variable-length Motif Mining of Time Series Based on Matrix Profile

ZHU Xu, ZHU Xiao-xiao, WANG Ji-min

2021, 0(05): 44-50.

Asbtract ( 337 )

PDF (813KB) ( 100 )

References | Related Articles | Metrics

Existing variable-length motif discovery algorithms have the problems that its speed is slow, its scalability is poor, and its results include meaningless phantoms such as too short, too long, or ordinary matching. A time series variable-length motif mining algorithm based on Matrix Profile is proposed. The algorithm uses the STOMP algorithm as a subroutine, and uses the lower bound distance combined with the incremental calculation to accelerate the process of extracting candidate motifs. The length similarity condition and the equivalent class method of motif group are used to remove the meaningless motifs that are too short, too long, or trival matched. Experiments on the dataset UCR show that proposed algorithm can effectively filter the meaningless motifs when the variable-length motifs are found, and has high efficiency and accuracy.

Flood Forecasting of Small and Medium Rivers Based on Integrated Learning

WANG Ji-min, JI Chang-zheng, LI Jia-huan, CAO Ying

2021, 0(05): 51-58.

Asbtract ( 198 )

PDF (2013KB) ( 95 )

References | Related Articles | Metrics

In order to solve the problems that the traditional data-driven flood forecasting method has large prediction error and the subnetworks in the traditional ensemble learning forecasting method can’t interact with each other, on the basis of single model predisction, the heterogeneous BP, CNN, LSTM neural networks are selected to establish a neural network integrated flood forecasting model based on negative correlation learning, and the overall error-variance decomposition and bifurcation decomposition of the model are carried out by explicitly adding regularization term, which makes the subnetworks in the integrated neural network incompletely independent, so as to ensure the diversity of the ensemble model and improve the prediction accuracy of the final model. The experiment in Tunxi basin of Anhui Province shows that the model based on negative correlation learning can effectively forecast the flood process, and the prediction accuracy is higher than the traditional single model.

Feature Weighted CLSVSM

NIU Feng-gao, YAN Tao

2021, 0(05): 59-65.

Asbtract ( 127 )

PDF (803KB) ( 107 )

References | Related Articles | Metrics

The rational and effective representation of document information using spatial vectors has a larger impact on text clustering and retrieval results. The Co-occurrence Latent Semantic Vector Space Model (CLSVSM) deeply excavates the co-occurrence latent semantic information between document feature words and improves the performance of document clustering. Based on CLSVSM, this paper first introduces word frequency information, then, the introduced word frequency is used as a weight to assign the co-occurrence strength in CLSVSM, and finally constructs feature weighted CLSVSM. The clustering effect of feature weighted CLSVSM on Chinese data is as follows: compared with CLSVSM and Word2vec text models, the F value is increased respectively by nearly 2.4% and 5.2%; compared with 90%CLSVSM_K and Word2vec text models， the entropy value is reduced respectively by nearly 3.1% and 9.0%; compared with the word frequency CLSVSM and TF-IDF models, the clustering effect is improved. The clustering effect of feature weighted CLSVSM on English data is similar to that of other models. The stability of feature weighted CLSVSM needs to be improved, which is limited by the completeness of keyword frequency information expression.

Penalized Matrix Decomposition Based on CLSVSM and Its Application in Text Topic Clustering

NIU Feng-gao, FENG Shi-jia, HUANG Chen

2021, 0(05): 66-72.

Asbtract ( 200 )

PDF (1068KB) ( 94 )

References | Related Articles | Metrics

Reasonable representation of text information plays an important role in text topic clustering and retrieval. Aiming at the problem of high dimension of text representation model, penalized matrix decomposition （PMD） is studied based on the co-occurrence potential semantic vector space model （CLSVSM）, and the vector is sparsely constrained by PMD to extract core features, so as to realize the reconstruction of original data. Through co-occurrence analysis theory and PMD method, the semantic information between features is deeply mined and the semantic kernel function （PMD_K） is constructed. The methods proposed in this paper are applied to text topic clustering, the experimental results show that the clustering effect of PMD and PMD_K is obviously better than that of other methods. Taking the F value as an example, the F value of PMD_K method is 21.9% higher than that of the previous 95%CLSVSM_K method. Combining PMD with text representation model not only improves the efficiency and accuracy of text topic clustering, but also avoids the complex computation of high-dimensional matrix.

Network Public Opinion Monitoring Based on Improved Floyd Algorithm

LI Jin-ze, WU Wen-hao, LI Kai-hang

2021, 0(05): 73-77.

Asbtract ( 192 )

PDF (980KB) ( 103 )

References | Related Articles | Metrics

With the rapid development of computer information technology in China, the network public opinion system has received more and more attention in the field of Internet, but there are still many problems in the management of network public opinion in China, the most concentrated problem is the imperfect response mechanism. In view of this, this study first describes the connotation of Floyd algorithm, describes the operation steps of Floyd algorithm, proposes the control strategy based on Floyd improved algorithm, analyzes the application and model of Floyd improved algorithm in network public opinion, and focuses on three factors affecting network public opinion monitoring technology. Finally, the relevant data published by ***.com and are used as public opinion hot data to verify and compare the proposed model algorithm. The results show that the improved algorithm based on Floyd is significantly better than other algorithms in various indicators, but when the experimental array reaches a certain upper limit, the reuse rate limit still appears. It is hoped that this study can provide some help for our government to strengthen the analysis of Internet public opinion, to achieve the ability to respond to social emergencies, and to improve the public opinion management ability of government agencies in the Internet environment.

Digital Resource Allocation Based on Cloud Computing and Deep Belief Network

WU Jun-cai, LIU Xue-fang

2021, 0(05): 78-82.

Asbtract ( 139 )

PDF (814KB) ( 97 )

References | Related Articles | Metrics

Aiming at the problems of poor rationality and low efficiency in the allocation of digital resources in university libraries, this paper proposes a method of library digital resources allocation based on cloud computing and deep belief network. Firstly, the cloud computing technology is used to construct the model of library digital resource allocation. Secondly, the deep belief network is used to train the characteristics of library digital resource nodes, and reasonable resource allocation is carried out. Finally, the experimental analysis and comparison of library digital resource allocation with other methods are carried out. The analysis results show that the proposed method can not only improve the quality and efficiency of library digital resource allocation, but also reduce the time of library digital resource allocation.

A Low Energy Wireless Routing Algorithm for Cluster Head Selection by Residual Energy Filtering

ZHANG Yan-hu, YAN Li-juan

2021, 0(05): 83-87.

Asbtract ( 164 )

PDF (808KB) ( 89 )

References | Related Articles | Metrics

Based on LEACH protocol, an improved self-organizing routing algorithm for wireless sensor network is proposed. The algorithm is improved in the cluster head generation of the original LEACH protocol. In the process of cluster head generation, the algorithm compares the current node residual energy with the average residual energy of all wireless sensor network nodes, prevents the nodes whose residual energy is less than the average residual energy of the whole network from being selected as cluster heads, so as to further optimize the balance of energy consumption of nodes in the whole network and effectively delay the dead time of nodes. In the cluster head election stage, the random selection of LEACH is replaced by purposeful screening so as to reduce the energy consumption of wireless sensor network and extend the network life cycle. The test is carried out by MATLAB simulation software. The experimental results show that the improved algorithm can improve the life cycle of wireless network, balance the energy consumption of wireless network, increase network throughput, and effectively delay the dead time of wireless network nodes.

A Method for Mobile Community Detection Based on Multi-dimensional Informational Fusion

SHU Peng, DU Qing-wei

2021, 0(05): 88-92.

Asbtract ( 140 )

PDF (832KB) ( 110 )

References | Related Articles | Metrics

To address the issue that the traditional community detection algorithm is difficult to apply to large-scale complex heterogeneous mobile networks, a mobile network model is constructed using mobile network usage detail record （UDR） and users’ social relationship data, and a method for mobile community detection based on multi-dimensional informational fusion is proposed, called BNMF-NF. Firstly, the paper comprehensively considers the user’s social relationship and spatiotemporal behavior, and gives the user’s social similarity, spatiotemporal distribution similarity and topic preference similarity. Then, the weighted network fusion method is used to fuse multi-dimensional similarity relations to construct a user similarity network. Finally, the community structure of the mobile network is detected by the use of the bounded non-negative matrix factorization. Experimental results on Foursquare and telecom data sets show that the method can effectively detect the community structure in the mobile network.

Application System Identification Method Oriented to Unbalanced Datasets

DONG Yan-hui, XIAO Jun-bi, ZHANG Hong-xia, YANG Yong-jin, JI Zhi-bin

2021, 0(05): 93-97.

Asbtract ( 142 )

PDF (1030KB) ( 102 )

References | Related Articles | Metrics

Aiming at the problem that traditional flow-based analysis methods cannot achieve effective identification of application systems in the oilfield local area network environment, this paper designs an application system identification framework for imbalanced data sets, WEBCLA, which uses the improved SMOTE algorithm based on Gini gain (GSMOTE) combined with the XGBoost classification algorithm to effectively identify web-based application systems. Specifically, the GSMOTE algorithm proposed by this paper over-samples the minority classes to effectively alleviate the problem of imbalance in recognition samples, and combines the XGBoost classification algorithm to identify the application system. Through experiments on real data sets, the results show that the method proposed in this paper has a significant improvement in recall rate compared with the traditional method, which is about 112.8% higher than the ordinary integrated method, and about 10.8% higher than the method without sampling processing. It can effectively solve the application system identification problem in the oil field LAN.

Modeling and Simulation of Infectious Diseases in Heterogeneous Population with Repeated Infection Based on SI1SI2R Model

SUN Lu, XUE Xiao-fei, CHENG Niu-niu

2021, 0(05): 98-104.

Asbtract ( 168 )

PDF (2933KB) ( 102 )

References | Related Articles | Metrics

Because the infectious diseases with the possibility of repeated infection are not hindered by antibodies or vaccines, their spread ranges are more extensive than ordinary infectious diseases, which seriously endangers human physical and mental health and affects social harmony and stability. It is necessary to explore the transmission law of infectious diseases, so as to formulate more targeted coping strategies. Based on the classical infectious disease model, SI1SI2R model of considering the characteristics of repeated infection of infectious diseases is constructed, and the interaction and transmission rules between heterogeneous individuals are defined. The transmission law of infectious diseases with repetitive infection characteristics among heterogeneous groups is explored by using ABM simulation modeling method. The simulation results show that the infection range of this kind of infectious diseases increases with the increase of secondary infection rate, and the composition of heterogeneous transmission groups not only affects the proportion of infected people, but also affects the time to reach the peak. Therefore, it is not only necessary to reduce the possibility of repeated infection of such diseases, but also necessary to put forward more efficient management schemes according to the heterogeneity of transmission groups.

A Detection Approach of Physical Host Status Anomalousness Based on Linear Regression and Least Squares

XU Sheng-chao, SONG Juan, PAN Huan

2021, 0(05): 105-111.

Asbtract ( 136 )

PDF (1503KB) ( 84 )

References | Related Articles | Metrics

A detection approach of physical host status anomalousness based on linear regression and least squares called EPADA （Efficient Physical host status Anomalousness Detection Approach） is proposed. EPADA can predict the CPU utilization for a period of time in the future based on the history of usage in each host. It is used in the live migration process to predict over-loaded and under-loaded hosts. When a host becomes over-loaded, some virtual machines migrate to other hosts to reduce SLA violation. When a host becomes under-loaded, the host switches to the sleep mode for reducing power consumption. EPADA is implemented and simulated by CloudSim. Simulation results show the good performance of EPADA.

Research Hotspot Analysis of AGV Path Planning Based on CiteSpace

LIU Yu-fei, ZHANG Xu-mei, LIANG Xiao-lei

2021, 0(05): 112-119.

Asbtract ( 169 )

PDF (1691KB) ( 117 )

References | Related Articles | Metrics

AGV path planning plays an important role in improving the efficiency of material handling. With the development of technology and the gradual expansion of the application scope of AGV, the problem of path planning has attracted the attention of scholars. This paper analyzes Chinese and foreign literatures based on CNKI and WOS by using the method of literature statistical analysis and visualization tool CiteSpace software. First, the research hotspots are obtained by keyword frequency and centralized sorting. The Chinese literature research hotspots include path planning algorithm, AGV scheduling, time window and laser navigation. The foreign literature research hotspots include flexible manufacturing system, material handling system, layout, AGV system and algorithm. Secondly, the hotspots are reviewed one by one. Then the Chinese and foreign literatures are compared and the similarities and differences of the Chinese and foreign literatures are analysed. Finally, the research of AGV path planning is prospected, and the future research will focus on the following aspects: combinatorial optimization of algorithms, targeted modeling of application scenarios and collaborative research on related issues.

Identification of Platen Switch State Based on Transfer Learning Strategy

CHEN Xiang, ZOU Qing-nian, XIE Shao-yu, CHEN Cui-qiong

2021, 0(05): 120-126.

Asbtract ( 198 )

PDF (2142KB) ( 108 )

References | Related Articles | Metrics

In order to realize the automatic inspection of the platen state in substation and improve the reliability and security of substation operation, a platen switch state identification algorithm based on transfer learning strategy is proposed. Firstly, the network parameters trained by Inception-V3 in dataset ImageNet are used to obtain the pre-trained model. Secondly, the trained bottleneck layer feature parameters are extracted to the target network as the feature extractor of the target platen switch image dataset. Then, the support vector machine algorithm based on particle swarm optimization is constructed to complete the platen switch state recognition. By comparing with the experimental results of commonly used deep learning network in learning efficiency and learning accuracy, the effectiveness and superiority of the proposed algorithm are verified. It also shows that transfer learning combined with convolution neural network can solve the problem of small samples in power equipment inspection and improve the accuracy and efficiency of platen switch state recognition.

Table of Content