Loading...
  • 中国计算机学会会刊
  • 中国科技核心期刊
  • 中文核心期刊

Current Issue

    • A survey of fast convolution algorithms
      LI Chuang, LIU Zong-lin, LIU Sheng, LI Yong, XU Xue-gang, XIA Yi-min
      2021, 43(10): 1711-1719. doi:
      Abstract ( 411 )   PDF (605KB) ( 478 )     
      Convolutional neural network is one of the most widely applied directions of deep learning algorithms. At present, the application of convolutional neural network is not only in the field of science and technology, but also in medical, military and other fields, and has played a huge role in related fields. Convolution is the most core part of convolutional neural network, and the computation amount of convolution accounts for more than 70% of the time of the whole network. Therefore, it is very important to study the acceleration of convolution operation. Firstly, the convolution algorithms in recent years are introduced, and their complexity is analyzed. The advantages and disadvantages of these algorithms are summarized. Finally, the possible breakthroughs in theoretical research and application are discussed and prospected.

      A strategy-proof auction mechanism for resource allocation in edge computing systems
      CHI Lai-xin, YANG Xu-tao, XIE Ning, ZHANG Xue-jie
      2021, 43(10): 1720-1729. doi:
      Abstract ( 168 )   PDF (626KB) ( 204 )     
      The use of auction mechanisms to allocate computing resources is a major challenging problem in the current field of edge computing. However, most of the current researches are based on a single type of resources and unable to meet the strategy-proof requirement. Aiming at this issue, this paper designs a strategy-proof auction mechanism for resource allocation in edge computing systems. The auction mechanism can allocate multiple resources in the form of virtual machines. It can improve the resource utilization and the social welfare effectively by taking into account the density of resource requirements, deployment constraints, and resource capacity simultaneously. It calculates the critical price as the payment price by dichotomy so that it can increase the calculation speed and meets the strategy-  proof requirement. Experimental results show that the mechanism can improve the resource utilization and the social welfare significantly and calculate the result in a short time. 
      Multiple-kernel clustering based on compressed subspace alignment
      OU Qi-yuan, ZHU En
      2021, 43(10): 1730-1735. doi:
      Abstract ( 194 )   PDF (469KB) ( 180 )     
      In recent years, multiple-kernel clustering (MKC) has achieved remarkable progress in fusing information from multi-source to boost the performance of clustering. However, denoting n as the sample number, the O(n2)  memory consumption and the  O(n3) computational consumption limit the practicality of these methods. In this paper, we redesign the formulation of subspace segmentation-based MKC, thereby reducing its memory and computational complexity to O(n)  and  O(n2), respectively. In the proposed algorithm, maned Compressed Subspace Alignment based Multiple Kernel Clustering(CSA-MKC), we sample only a part of the data to reconstruct the whole dataset. Specifically, in our design, a consensus sampling matrix is learned simultaneously with the information fusion process, so as to make the generated anchor point set more suitable for data reconstruction across different views. Consequently, the discriminative capability of the reconstruction matrix is improved, and the performance of clustering is enhanced. Moreover, since our algorithm is straightforward for parallelization, through the acceleration of GPU, our algorithm can achieve superior performance against the compared state-of-the-art methods on six datasets with square time cost. 



      A parallel processing approach for video big data based on Spark Streaming framework
      ZHANG Yuan-ming, YU Jia-rui, LU Jia-wei, GAO Fei, XIAO Gang
      2021, 43(10): 1736-1743. doi:
      Abstract ( 186 )   PDF (1059KB) ( 202 )     
      Video devices are widely used in public areas, smart transportation, industrial production and many other fields. Video data has typical characteristics of huge volume, fast speed, sparse value and completely unstructured data. To achieve higher processing performance for video big data, a parallel processing approach based on Spark Streaming is proposed. A parallel processing framework based on Spark Streaming is designed. Especially, parallel strategies for inter-frame independent algorithms and inter-frame correlation algorithms are given in detail. The former strategy maps the de-redundant video frames to different nodes with data parallelism, and the latter maps the operators of the algorithm to different nodes based on the dependency relationship. The parallel processing approach is evaluated with real video big data. A parallel detection algorithm for elevator passenger number and a parallel detection algorithm for elevator door anomalies are designed. When the number of nodes increases to 16, the speedup of the elevator passenger number detection algorithm is 615%, and the speedup of elevator door anomaly detection is 253%.

      A four-wing memristive chaotic system based on hyperbolic sine function and its FPGA implementation
      BAI Yu-long, PAN Xing-yu, DUAN Ji-kai, YANG Yang
      2021, 43(10): 1744-1749. doi:
      Abstract ( 153 )   PDF (1059KB) ( 142 )     
      Using the hyperbolic sine function memristor as the positive feedback term, a four- dimensional chaos model with four wings is designed. Firstly, the system is numerically solved using Runge-Kutta algorithm, and the stability of the system is analyzed. It is found that the system has only one equilibrium point as saddle point. Moreover, the dynamic analysis of the system is carried out, and the Lyapunov exponent and bifurcation diagram with the system parameters are drawn. Besides, the Lyapunov dimension of the system is calculated and the corresponding system is obtained. With the change of the motion states when the parameters change, it is found that there are many motion forms of the system, such as periods and chaos. Finally, a chaotic circuit system is designed by FPGA, and the results observed by oscilloscope are basically consistent with the numerical results, which lays a foundation for the application of memristor chaotic system in communication.

      A user location prediction model based on GRU network in crowd sensing environment
      ZHANG An-ran, LIAO Yi-wei, ZHAO Guo-sheng, WANG Jian
      2021, 43(10): 1750-1757. doi:
      Abstract ( 146 )   PDF (835KB) ( 144 )     
      In the case of the sparse distribution of users in the perception area, predicting the user's location in advance is the key to improve the task completion rate of the crowd sensing system. This paper presents a user location prediction model based on the gated recurrent unit. Firstly, a model of the crowd sensing system is constructed, and the application of the participatory sensing based on location is realized. Secondly, the data set of the user's location is normalized, and by combining the multidimensional characteristics of the user's historical location data, a gated recurrent unit structure is constructed. Finally, the actual trajectory data set in the vehicles networks is used to train the model, and the Adam algorithm is used to optimize the performance parameters of the user position prediction model based on the gated recurrent unit. The simulation results show that, compared with the RNN model and the LSTM model, the prediction mean square error of the proposed model is reduced by 22% and 18% respectively, and has the advantage of strong implementability in processing sequence data.

      An improved differential privacy parameter setting and data optimization algorithm
      HU Yu-gu, GE Li-na,
      2021, 43(10): 1758-1765. doi:
      Abstract ( 236 )   PDF (748KB) ( 147 )     
      Data perturbation based on differential privacy is a hotspot of privacy protection technology. In order to realize differential privacy protection for sensitive data and improve the usability of data as much as possible, reasonable Settings of privacy parameters and optimization of noised data are the key technologies. The privacy parameter setting algorithm RBPPA and the optimization algorithm DPSRUKF are proposed in this paper. RBPPA constructs a fine-grained privacy parameter setting scheme based on the reputation of data visitors and contributors, and is associated with data privacy degree and access rights value. DPSRUKF uses Square-Root Unscented Kalman Filter to process noisy data, which improves the usability of differential private data. Experimental results show that this algorithm can realize fine-grained setting of privacy parameters and improve the accuracy of noisy data. It not only provides data security for sensitive data for applications, but also provides high usability of data for data visitors.

      An improved 3DDV-Hop localization algorithm based on 3D coordinate correction
      LUO Shi-zhang, ZHANG Jing, WANG Jian-min
      2021, 43(10): 1766-1772. doi:
      Abstract ( 105 )   PDF (1213KB) ( 140 )     
      Aiming at the problem that the traditional 3DDV-Hop algorithm has a large positioning error for unknown nodes, various improved positioning algorithms based on hop count and hop distance calculation methods have been proposed, but the improved methods of hop count and hop distance calculation in various algorithms need to be optimized and the coordinates of unknown nodes have not been refined twice. Therefore, an improved 3DDV-Hop positioning algorithm based on three-dimensional coordinate correction is proposed. The algorithm sets three kinds of communication radii and hop weights for nodes to reduce the calculation error of hop count and hop distance, and construct a cross area of cube to refine the coordinates of unknown nodes twice. The comparative analysis of experimental results shows that the improved 3DDV-Hop localization algorithm based on 3D coordinate correction can significantly reduce the average localization error of unknown nodes. 

      Repair analysis of concurrent event process model based on Petri net
      YANG Hui-hui, FANG Xian-wen, SHAO Chi-feng
      2021, 43(10): 1773-1780. doi:
      Abstract ( 118 )   PDF (671KB) ( 146 )     
      At present, the process model can be mined from a large number of event logs to replay most of the logs. However, a few logs that deviate from the process model are also valid, and model repair is a good way to make the event log fit more with the process model. This paper proposes a repair analysis method of concurrent event process model based on Petri net. Firstly, the optimal alignment between the event log and the process model is found to filter out the concurrent events for repair. Secondly, the filtered concurrent events are reconstructed by using the proposed reconstruction sub-process repair method. Finally, the algorithm is embedded into the original model to realize the model repair, and the reasonable effectiveness of the method is illustrated through a concrete example. The repaired model can completely replay the given event log, and it can avoid the occurrence of redundant behavior caused by the loop, while also preserving the use value of the original model to the maximum extent.

      Design and implementation of a change event driven microservice composition platform
      WANG Xin, LIU Xiao-yan, ZHANG Kai-qi, WANG Xing, YAN Xin
      2021, 43(10): 1781-1788. doi:
      Abstract ( 140 )   PDF (660KB) ( 154 )     
      The field of microservice composition is still in its immature stage and needs a more complete solution. Therefore, a microservice composition platform based on change events is designed. Compared with the existing solutions, the platform supports fine-grained data access control at the domain-specific language level and further enriches the language expression. The Spring Cloud Netflix ecosystem is introduced to solve the problem that the microservices in the platform cannot be deployed dynamically and improve the robustness. For the events that trigger the composition of the microservices, the change detection of Web resources is added. That it, the content and structure of the XML document are compared, and the existing comparison method at the content level of the XML document is optimized, and the similarity measure of content and structure is comprehensively considered. The experimental results show that the microservice composition platform is more complete and rich, with emphasis on privacy protection, and stronger fault tolerance, and improves the algorithm time efficiency.



      Unsupervised image style transfer based on generating adversarial network
      LAN Tian, XIN Yue-lan, YIN Xiao-fang, LIU Wei-ming, JIANG Xing-yu
      2021, 43(10): 1789-1795. doi:
      Abstract ( 221 )   PDF (826KB) ( 164 )     
      Unsupervised transfer of image style is a very important and challenging problem in the field of computer vision. Unsupervised image style migration is intended to map images of a given class to similar images of other classes. In general, pairwise matching data sets are difficult to obtain, which greatly limits the transformation model of image style migration. Therefore, in order to avoid this limitation, this paper improves the existing unsupervised image style transfer method and adopts an improved cycle consistency adversarial network to conduct unsupervised image style transfer. Firstly, in order to improve the training speed of the network and avoid the phenomenon of gradient disappearing, this paper introduces Densenet network into the traditional cycle consistent network generator. In terms of improving the performance of generators, the generator network introduces the attention mechanism to output better images. In order to reduce the structural risk of the network, spectral normalization is used in each convolutional layer of the network. In order to verify the effectiveness of the proposed method, experiments are carried out on the datasets of monet2photo, vangogh2photo, and facades, 
      the experimental results show that the average of Inception score and FID distance evaluation index are improved.

      A SDI video image segmentation system based on Zynq
      WANG Wei-chen, TU Hai-yang, WANG Wei-ming, ZHAO Xiao-bo
      2021, 43(10): 1796-1802. doi:
      Abstract ( 130 )   PDF (993KB) ( 163 )     
      To make up for the shortcomings of the traditional video image splitter such as weak anti-interference ability, low frame rate, and complex design, Xilinx Zynq XC7Z035 FPGA heterogeneous platform is selected and integrated with SDI technology. The high-definition digital serial decoding chip TW6874 is used to synchronously collect 4 digital video images, and output BT.1120 data to FPGA, in order to realize the separate display of 4 channels of video. In order to meet the resolution and frame rate requirements of video images, pixel resampling of video image data is first performed. Secondly, AXI4-Stream Data FIFO is used for line input buffering, which is flexible in processing data and easy to expand, which provides a basis for further integration of algorithms. AXI4-Stream Data FIFO generates s_axi_s2mm_tlast signal for each line of 960 data and handshake with AXI DMA, buffers the data in DDR3 SDRAM, and buffers the next buffer address after 540 lines. AXI DMA has 3 video images per channel buffer, thus completing the three-buffer design, to ensure that the video image is not torn. Finally, the cached data is output to the SMPTE SDI IP core for display. The experimental results show that the system realizes the 4-channel SDI video image segmentation, the system resource utilization rate is low, the video image frame rate is high, the layer is obvious, and no tearing and distortion occurs.

      An improved Mask RCNN algorithm based on adaptive-threshold non-maximum suppression
      WANG Mei, LI Dong-xu, CHEN Lin-lin, FAN Si-meng, XU Chuan-hai, YANG Er-long
      2021, 43(10): 1803-1809. doi:
      Abstract ( 171 )   PDF (1079KB) ( 188 )     
      Target detection algorithms under big data often have the problem of missed target detection and repeated detection. To solve this problem, Mask RCNNAT-NMS algorithm based on AT-NMS is proposed. Firstly, a deformable convolution module is added on the basis of ResNet to enhance the ability to extract multi-layer convolution features of the target. Secondly, 
      the AT-NMS algorithm is used to extract the in-depth information of the candidate target area in the RPN (Regional Candidate Network) stage. Thirdly, the positioning of the target is more accurate through two quantitative processing of ROI Align. Finally, three branches are used to achieve target instance segmentation, target classification and target border regression. The experimental results on the PASCAL-VOC2012 and Indoor CVPR_09 data sets show that, compared with the mask RCNN algorithm, the Mask RCNNAT-NMS algorithm reduces the repeated detection rate and the target missed detection rate, and improves the recognition accuracy. It can be seen that Mask RCNNAT-NMS algorithm can alleviate the problem of target missing and repeated detection caused by fixed threshold, and improve the detection accuracy on this basis.


      Image quality evaluation based on parallel small CNN#br#
      #br#
      CAO Yu-dong, CAI Xi-biao
      2021, 43(10): 1810-1816. doi:
      Abstract ( 127 )   PDF (844KB) ( 135 )     
      Objective evaluation of image quality is widely used in many image processing tasks. A non-reference image quality evaluation algorithm is proposed based on small parallel-mode convolutional neural networks under deep learning technology. Convolution operation and parallel multi-scale input could learn not only rich feature, but also subtle feature. Firstly, the Gaussian image pyramid is used to obtain different scale distorted images as the input of 4 small-scale single-layer convolutional neural networks. After convolution and pooling, 4 feature vectors are output, and the learned feature vectors are merged and then mapped into image quality prediction scores through fully connected regression. Para- meters are optimized through two serial stages to improve the accuracy of the model. Experimental test- ing results show that the designed small network model is effective, and the proposed algorithm has higher performance than the current comparative algorithms and has good stability and strong generalization ability.

      A nonlinear time series prediction algorithm based on combination model
      YU Qiong, TIAN Xian
      2021, 43(10): 1817-1825. doi:
      Abstract ( 143 )   PDF (1883KB) ( 146 )     
      In order to solve the problem of low construction efficiency and low accuracy of the nonlinear time series prediction model in complex systems, a Hurst-EMD prediction algorithm based on combination model is proposed. This algorithm uses EMD algorithm to decompose the nonlinear time series into individual IMFs representing the characteristics of the original series, and then introduces Hurst exponent to integrate the similar IMF into new components. Finally, LS-SVR and ARIMA models are used for combinational prediction. In this algorithm, the process of sequence classification and integration is designed, the number of calculation is optimized, and an efficient and accurate prediction model is constructed. In order to verify the validity of the model, the public data set of Shanghai stock index and real traffic flow data are used for testing. The experimental results show that the improved HURST-EMD combination model has better prediction accuracy while improving the prediction efficiency.

      A collaborative filtering recommendation algorithm based on tag extension
      CHEN Hai-long, YAN Wu-yue, SUN Hai-jiao, CHENG Miao
      2021, 43(10): 1826-1832. doi:
      Abstract ( 118 )   PDF (545KB) ( 122 )     
      Most recommendation algorithms that use the relationship between tags and users and items have to face the problem of sparse tags caused by different individual users. Different users will have different tags for the items. Aiming at the problem of sparse user-tag and item-tag matrix due to the randomness of user labeling, a collaborative filtering recommendation algorithm based on tag extension is proposed. The label similarity based on the label is calculated according to the user's labeling behavior, and the label similarity based on the label semantics is calculated according to the semantics of the label marked by the user. The similarity of tags is evaluated in terms of user behavior and label semantics, and the tag similarity is used to expand each item-tag to reduce the sparseness of the matrix generated by the association relationship between items and tags. Experimental results show that running the algorithm on the dataset MovieLens improves the accuracy.


      A tobacco storage moldy prediction method based on one-dimensional convolutional neural network
      ZHAI Nai-qi, YUN Li-jun, YE Zhi-xia, WANG Yi-bo, LI Ya-zhao
      2021, 43(10): 1833-1837. doi:
      Abstract ( 180 )   PDF (491KB) ( 168 )     
      Aiming at the problem of mildew during the storage of tobacco leaves, the traditional prevention and control measures are not effective, and the existing tobacco leaf mildew prediction model has low accuracy, which cannot effectively reduce the occurrence of tobacco leaf mildew. In order to improve the accuracy of tobacco leaf mildew state prediction, a method based on one-dimensional convolution deep neural network (1D-CNN) is proposed. Based on the collection of terminal sensor data, it is stan- dardized and processed to obtain the model's training features. A 1D-CNN is trained to predict the mildew state of tobacco leaves, and the network structure is optimized. The experimental results show that the proposed method has higher prediction accuracy than other traditional models. Finally, an intelligent monitoring system for tobacco leaf storage mildew is designed and implemented to realize the real-time prediction function of tobacco leaf mildew, and good results are achieved.

      A self-adaptive clustering algorithm without neighborhood parameter k and cluster number c
      ZHANG Bo-kai, YANG De-gang, FENG Ji,
      2021, 43(10): 1838-1847. doi:
      Abstract ( 124 )   PDF (891KB) ( 127 )     
      Traditional clustering methods often cannot avoid the selection of neighborhood parameters and the number of clusters. The optimal selection of these parameters in different shapes of data is hard to choose, and this choice is depending on prior knowledge. Aiming at the above parameter selection problem, this paper proposes a natural neighbors based border peeling clustering algorithm (NaN-BP), which can obtain satisfactory clustering results without setting the neighborhood parameters and the number of clusters. The core idea of the algorithm is to adaptively iterate to a logarithmic stable state and obtain neighborhood information according to the distribution characteristics of the data set, then mark and strip the boundary points according to the neighborhood information, and finally gather the core points as the center of the data cluster. Extensive comparative experiments is conducted on data sets of different scales and distributions, and satisfactory experimental results verify the adaptability and effectiveness of the algorithm.

      Military named entity recognition based on self-attention and Lattice-LSTM
      LI Hong-fei, LIU Pan-yu, WEI Yong
      2021, 43(10): 1848-1855. doi:
      Abstract ( 189 )   PDF (632KB) ( 211 )     
      The identification of military named entities can provide automatic auxiliary support for intelligence analysis, command and decision-making, and is the key technical means to improve the intelligence of command information system. Because of the differences in Chinese and English language characteristics, Chinese entity recognition must first part the text, and word-breaks will lead to the accumulation of errors in the recognition of named entities. In addition, the identification of named entities in a piece of text may be related only to local information, and each word contributes differently to other entities, and too much redundant information can only negatively affect the identification of named entities. In response to the above problems, we propose a network model of Lattice-Long Memory Neural Network (LSTM) combined with self-attention mechanisms. The Lattice LSTM structure enables the identification of proper nouns in sentences and integrates potential word information into character-based LSTM-CRF models. Self-attention structures can capture syntactic or semantic features between words in the same sentence. Model experiments were conducted on a small sample set that we labeled ourselves, and the results show that our model achieves the desired effect.


      An imbalanced data classification algorithm based on DPC clustering resampling combined with ELM
      DONG Hong-cheng, WEN Zhi-yun, WAN Yu-hui, YAN Fei-yang,
      2021, 43(10): 1856-1863. doi:
      Abstract ( 156 )   PDF (736KB) ( 139 )     
      The combination of sampling technology and ELM classification algorithm can improve the classification accuracy of a small number of samples, but most existing sampling methods that combine ELM do not take into account the imbalance of the sample and the distribution within the sample. The sampling technique is too single, resulting in low efficiency of the classification model and low recognition rate of a small number of samples. In order to solve this problem, this paper proposes an imba- lanced data classification algorithm based on DPC clustering resampling combined with ELM. First, a mixed sampling model is constructed to balance the data set in two cases according to the degree of imbalance of the data set. Secondly, the DPC clustering algorithm is used to analyze and deal with the majority and minority classes on this model respectively. It can solves the problem of intra-class imbalance and noise in the data, so that the two types of samples are relatively balanced. Finally, the obtained ba- lanced data sets are classified using the ELM classification algorithm. Compared with the same type of classification algorithm, the two classification performance indexes F-Measure and G-mean of the proposed algorithm are significantly improved on the experimental data set.