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

    • Compressed page walk cache
      JIA Chao-yang, ZHANG Dun-bo, WANG Qiong, SHEN Li
      2020, 42(09): 1521-1528. doi:
      Abstract ( 220 )   PDF (606KB) ( 186 )     
      General-Purpose Graphics Processing Units (GPGPUs) his been widely used in modern high performance computing systems. The Single-Instruction Multi-Thread (SIMT) execution model of GPGPUs results in a lower page hit rate, and requires Page Walk Cache (PWC) to reduce the actual number of page table accesses for irregular applications. There is a lot of redundant information in the traditional PWC and the capacity is limited, so the actual effect is not good. We analyze the information redundancy in the traditional PWC and propose a new structure: Compressed PWC. Compressed PWC completely eliminates redundant information and compresses the space while keeping the same search overhead unchanged, so that PWC can record more page table access history, thereby effectively reducing the number of page table accesses during the address translation. Experimental results indicate that, compared with the traditional PWC of the same capacity, compressed PWC can significantly improve the efficiency of virtual-to-physical address translation.


      An automatic model splitting strategy generation method for model parallel training
      WANG Li, GUO Zhen-hua, CAO Fang, GAO Kai, ZHAO Ya-qian, ZHAO Kun
      2020, 42(09): 1529-1537. doi:
      Abstract ( 276 )   PDF (804KB) ( 136 )     
      With the increase of the training data scale and the increasing complexity of the model, the training cost of the deep neural network is getting higher and higher, which requires higher computational power for the computing platform. In recent years, AI accelerators (such as FPGA, TPU, AI chip, etc.) based on heterogeneous distributed training have emerged endlessly, providing the hardware foundation for the parallelization of deep neural network. In order to make full use of all kinds of hardware resources, the researchers need to set a variety of different work force and hardware architecture AI accelerator computing platforms for neural network model training. Therefore, in the model paralle- lism training, how to efficient use all sorts of AI accelerator computing resources and realize the training mission in a variety of load balancing on the accelerator is the hot issue researchers concern about. This paper proposes a method that can automatically generate the model splitting strategy based on static network model, and map the model splitting strategy to model training, so as to realize the task assignment of network layers on different AI accelerators. The model allocation strategy automatically generated based on this method can efficiently utilize all computing resources on a single computing platform and ensure the load balancing of model training tasks among various devices. Compared with the current manual splitting strategy, it has higher timeliness, saves the generation time of the splitting strategy by more than 100 times, and reduces the uncertainty caused by human factors.



      An energy trace preprocessing method for SoC
      CAI Xiao-min, LI Ren-fa, LI Shao-qing, SHEN Gao, KUANG Shi-jie
      2020, 42(09): 1538-1543. doi:
      Abstract ( 181 )   PDF (812KB) ( 119 )     
      Differential power attack is one of the most effective methods to reveal the key of cryptographic devices. Therefore, an energy trace classification method based on energy model is proposed to locate the significant points that have strong correlation with the key. Two energy models (hamming distance and hamming weight) are used, and a few sampling points near the significant point are intercepted to form the significant interval. The SoC chip with 128-bit AES algorithm under different keys is attacked. The actual test results show that the length of the optimal characteristic interval obtained by hamming distance energy model is short. Finally, the characteristic point location method based on hamming distance energy model has low sample calculation and reduces a large amount of samples.



      Real-time scheduling and schedulability analysis for mobile system
      CHEN Cong, HONG Zhong, CHEN Yang-yang, ZHANG Shi, JIANG Jian-min, ​
      2020, 42(09): 1544-1555. doi:
      Abstract ( 122 )   PDF (772KB) ( 103 )     
      In order to ensure the safety of a complex real-time mobile system, the behavior of the mobile system needs to be modeled and analyzed by formal method. It is an important goal to judge whether the moving objects collide with each other during normal operation so as to verify the safety of a real-time mobile system. For achieving this goal, it is necessary to isolate and analyze each moving object within the system, which needs the supports from scheduling theory. However, traditional scheduling theory is based on coarse-grained task scheduling and cannot be directly used in fine-grained isolation analysis. To solve this problem, a mobile system can be modeled through Time Dependency Structure (TDS), which is a more fine-grained formal model based on events. The real-time scheduling method is defined on the basis of TDS, and the safety of the system can be judged by investigating the nature of real-time scheduling and analyzing the schedulability of isolation.
      A RLWE-based two-factor three-party  authentication key exchange protocol
      SHEN Yan-mei, LI Ya-ping, WANG Yan, WANG Hui, HUANG Li-juan
      2020, 42(09): 1556-1562. doi:
      Abstract ( 131 )   PDF (1539KB) ( 117 )     
      In order to enable the Diffie-Hellman-style key exchange protocol on the lattice to achieve authentication and is suitable for large-scale communication in the client-server-client mode, a two-factor three-party authentication key exchange protocol based on Ring Learning With Error (RLWE) is proposed. The protocol uses passwords and biometrics as long-term keys for the client, enabling the server to explicitly authenticate the client. Firstly, the advantages of the difficult problem of error learning on the ring (short key and cipher text size and high operating efficiency) are used to construct the cryptosystem. Secondly, the server passes ring elements through password and biometric hash values, and combines D-type error coordination. The mechanism enables the communicating party to obtain a random and even session key. The final analysis shows that the protocol is suitable for large-scale communication, improves the communication volume, has higher security attributes, and can resist the password impersonation attacks of users.

      Application security risk assessment  of state grid edge computing
      GUO Hao, HE Xiao-yun, SUN Xue-jie, CHEN Hong-song, LIU Zhou-bin, XIE Jing
      2020, 42(09): 1563-1571. doi:
      Abstract ( 180 )   PDF (952KB) ( 117 )     
      According to a series of the national cyber security level protection and risk assessment standards and the characteristics of electric power information systems, a risk assessment model for application security of state grid edge computing is proposed. Then, the vulnerability scanning tools AWVS and AppScan are used to target security vulnerability evaluation and risk assessment experiments on the open source web application target software BWAPP that integrates the latest security vulnerabilities. Finally, the fuzzy analytic hierarchy method is used to comprehensively evaluate the security of Web application security. Based on the test results of the application security, the security assessment data are compiled to realize the verification of the application security risk assessment of the state grid edge computing.


      Quality compensation analysis of the watermarking  method based on singular matrix robustness
      WANG Wen-bing, LIU Sheng-li
      2020, 42(09): 1572-1577. doi:
      Abstract ( 108 )   PDF (602KB) ( 97 )     
      The quality compensation for the watermarking method based on the singular matrix robustness can improve the invisibility of watermarks to some extent. To optimize the compensation parameters, the paper analyzes the embedding principle of the watermarking method, finds out the relationship between pixel modifications and singular matrix changes, and studies the variation rules of pixels before and after quality compensation, thereby obtaining the optimal compensation parameters. Based on the optimal parameters, the paper further studies the relationship between quality compensation and robustness, and determines these factors that affect the compensation effect. The final experiments verify the quality compensation effect, and the comparison with similar methods proves that the proposed quality compensation method can improve the performance of watermarking.



      A high spatial temporal fusion method based on deep learning and super resolution reconstruction
      ZHANG Yong-mei, HUA Rui-min, MA Jian-zhe, HU Lei
      2020, 42(09): 1578-1586. doi:
      Abstract ( 199 )   PDF (1253KB) ( 179 )     
      Aiming at the "space-time conflict" of remote sensing images, a high spatial-temporal fusion algorithm based on improved STARFM is proposed. SRCNN is used for the super-resolution reconstruction of low-resolution images. Due to the large difference in resolution between the two groups of fusion images, the network training is difficult. Firstly, both of the two groups are sampled to an intermediate resolution, and low-resolution images are reconstructed by SRCNN with high-resolution images as their prior knowledge. Secondly, the obtained intermediate resolution images are resampled, and then they are reconstructed by SRCNN with original high-resolution images as their prior knowledge. The resulting reconstructed images have higher PSNR and SSIM than the images resampled by interpolation, alleviating the systematic error caused by the sensor difference. The STARFM fusion method uses expert knowledge to extract artificial features in selecting "Spectrally Similar Neighbor Pixels" and computer their weights. Based on the basic concept of STARFM, an automatic feature extraction method using SRCNN as the basic framework is realized. The experimental results show that this method has lower MSE value than the original STARFM, which further improves the quality of spatial-temporal fusion.


      An improved DSST real-time target tracking algorithm based on TLD framework
      HUANG Hao-miao, ZHANG Jiang, ZHANG Jing, BAO Jun-rong
      2020, 42(09): 1587-1598. doi:
      Abstract ( 187 )   PDF (1429KB) ( 143 )     
      In view of the image blurring caused by fast target movement, it is difficult for the DSST algorithm to distinguish between the target and the background information. The filter is cyclically shifted during the training phase to collect dense samples, which easily results in boundary effect and leads to the tracking drift problem. Therefore, this paper proposes an improved DSST real-time target trac- king algorithm (TLD-DSST) that incorporates the TLD framework. The algorithm improves the position filter of the DSST algorithm, adds the weight coefficient matrix through the spatial regularization me- thod to reduce the response of the non-target area, and performs rough positioning of the target under fast motion. At the same time, a naive Bayesian classifier is introduced to improve the TLD detector, in order to improve the detector's ability to distinguish between the target and the background information. Moreover, the optimal similarity matching is performed on the position of the DSST target response and the target area obtained by the TLD detector, so as to get the precise positioning result. The TLD detector positive and negative sample online update mechanism is used to continuously optimize the robustness of the algorithm. Experimental results show that the TLD-DSST algorithm has high accuracy and success rate for target tracking in complex scenarios such as fast motion.

      Apple detection and grading based on color and fruit-diameter
      FAN Ze-ze, LIU Qian, CHAI Jie-wei, YANG Xiao-feng, LI Hai-fang
      2020, 42(09): 1599-1607. doi:
      Abstract ( 184 )   PDF (3005KB) ( 198 )     
      Apple is one of the main producing fruits and the main economic crops in many areas. Detecting and grading apples through the image of apple trees under natural environment is helpful to promote the modernization of fruit industry. Combining deep learning with traditional methods, this paper proposes a fruit detection and grading method combining color and apple diameter. In order to improve the detection rate of unobvious targets and the precision of bounding boxes when illumination or fruit coloration is uneven, the convolutional neural network is used to construct an apple detection model and detect apple on feature maps of two scales. b*, (1.8b* -L*) color components of the image in bounding boxes in CIELAB color space are extracted, the image is binarized, and the target contour is accurately extracted to correct the bounding boxes. Experimental results show that the precision is 91.60% and the F1-score value is 87.62%. According to the image and actual size mapping method, the apple diameter is calculated to achieve the apple grading. Experimental results show that the grading accuracy is 90%.

      A surface defect recognition algorithm based on improved SSD model
      LI Lan, XI Shu-shu, ZHANG Cai-bao, MA Hong-yang
      2020, 42(09): 1608-1615. doi:
      Abstract ( 208 )   PDF (829KB) ( 188 )     
      Surface defect of workpiece is an important factor that affects the performance of mechanical equipment. Fast and efficient detection is the focus of current research. In order to solve the problem of workpiece surface defect detection, a detection method based on SSD model is proposed. By proposing DH-Mobilenet network to replace VGG16 network in SSD structure, this method simplifies the detection model and reduces the computation. At the same time, the inverse residual block is used to predict the position, and the dilated convolution is used to replace the down sampling operation to avoid information loss. Scanning electron microscope is used to obtain the surface image of workpiece, and the workpiece surface defect data set is established and expanded. Finally, three kinds of high frequency defects, namely fragment, peeling off and pear ditch, are trained and tested, and the results are compared with the original models of YOLO, Faster R-CNN and SSD. The test results show that this method can detect the surface defects of the workpiece more accurately and quickly, which provides a new idea for the defect detection in the actual industrial scene.

      A double row license plate segmentation method based on adaptive threshold projection
      MA Yong-jie, CHEN Zhen-yu, MA Yun-ting
      2020, 42(09): 1616-1624. doi:
      Abstract ( 159 )   PDF (1031KB) ( 112 )     
      Aiming at the problem that there are few algorithms and low correct segmentation rate of double row license plate number in China, an adaptive projection method for double row license plate segmentation is proposed. Firstly, the image of license plate in HSV (Hue Saturation Value) space is binarized. Secondly, the double line license plate is segmented into a single line license plate by the adaptive projection method. Finally, the characters on the license plate number are segmented. Experiments show that this method can effectively segment the double line license plate with white characters on blue background, red characters on white background, black characters on green background, and black characters on yellow background, and the correct segmentation rate of the double line electric bicycle brand is higher. At the same time, this method overcomes the shortcomings of the deep learning method such as long training time and a large number of data sets. The proposal t is a simple and efficient segmentation method.



      Complexity analysis and finite time synchronization of fractional Rucklidge system
      NA Ge-si, ZHAO Hai-yan
      2020, 42(09): 1625-1631. doi:
      Abstract ( 158 )   PDF (740KB) ( 100 )     
      Nonlinear system is a complex subject in the field of finance, and its complexity change and control methods have always been the key and difficult research points for experts and scholars. Earlier studies of nonlinear systems were based on integer-order nonlinear systems. With the deepening of the research on fractional-order systems, people have gradually realized that the complexity of fractional-order systems is generally higher than that of integer-order systems. Analyzing and studying its complex characteristics is helpful to define system parameters and increase the complexity of the system. This paper takes a fractional Rucklidge system as an example to study its dynamic and complex characteristics. For the fractional-order nonlinear system, a synchronous controller is designed, which can be controlled and simulated synchronously. Simulation results show that the controller can achieve synchronization in a short time.




      Event extraction in political diplomacy based on similar semantics and dependency syntax
      CUI Ying
      2020, 42(09): 1632-1639. doi:
      Abstract ( 123 )   PDF (973KB) ( 97 )     
      Based on the research of news data in the field of political diplomacy, aiming at the pro- blems of extraction difficulty, low recall rate and accuracy rate of event extraction based on traditional pattern matching, and low accuracy rate of event extraction based on deep learning method in specific field, this paper proposes an event extraction method in the field of political diplomacy based on similar semantics and dependency syntax. The proposed method extends the event triggers by similarity calculation of semantic description, which lays a foundation for accurate recognition of event types. Furthermore, based on patterns, text dependency syntactic parsing is used to identify and extract event elements in political and diplomatic fields, so as to achieve a structured description of events. The accuracy of the extraction results is obviously better than that of the end-to-end event extraction model based on the deep neural network, and it can be used for reference and implementation in other specific fields. Finally, the main difficulties and application prospects of event extraction are discussed and summarized.



      Complex network community detection algorithm based on deep encoder 
      ZHANG Shi-jin, ZHANG Sheng, TIAN Ji-biao, WU Zhi-qiang, DAI Wei-kai
      2020, 42(09): 1640-1648. doi:
      Abstract ( 155 )   PDF (1592KB) ( 124 )     
      Complex network is a typical representation of complex systems. Community structure is one of the most important structural characteristics of complex network. Aiming at the problem that the current community detection algorithms have low community detection accuracy and is not suitable for large-scale networks, a Deep Auto-encoder and EForest (DA-EF) algorithm and an influence diffusion similarity index are proposed. The DA-EF algorithm combines a multi-layer auto-encoder with a EForest to form a two-level cascade model, transforms the similarity matrix into low dimension and higher order feature matrices through dimensionality reduction and characterization learning, and finally uses K-means to obtain community detection results. The cascade structure greatly reduces the time complexity of the algorithm while maintaining the same depth of the algorithm. The simulation results show that, compared with similar algorithms such as K-means, Spectral and CoDDA, the proposed algorithm has the best NMI and modularity Q values, and the lowest running time of clustering on synthetic datasets and real datasets. It has the advantages of high accuracy and high efficiency. In the performance experiment of the algorithm, the rationality and effectiveness of the cascade structure, the depth of the auto-encoder, and the similarity index of the algorithm are verified.


      Short-term hydrothermal scheduling based  on hybrid chaotic krill herd algorithm
      XIAO Xiong, GAO Miao, CHEN Gong-gui
      2020, 42(09): 1649-1660. doi:
      Abstract ( 150 )   PDF (1585KB) ( 115 )     
      Short-Term Hydrothermal Scheduling (STHS) is a nonlinear, multi-constrained and time-varying optimization problem. When the valve point effect is considered, the problem becomes non- convex and more complicated. In order to improve the search ability of the Krill Herd Algorithm (KHA) in the STHS problem, the hybrid chaotic map is introduced to improve the global convergence speed of KHA. In order to avoid premature convergence of the algorithm, by recording the number of times that the fuel cost values of the best individual in each generation remain unchanged and making the decision that a positional mutation in the non-positionally dominant individual within its feasible domain, a hybrid chaotic krill herd algorithm (HCKHA) is proposed. HCKHA and KHA, CKHA were applied to the standard STHS test systems such as "four hydro and three thermal plants" and "four hydro and ten thermal plants", independently. The simulation results show that HCKHA has better optimization ability, fuel cost values and transmission loss values than KHA, CKHA and the optimization methods in other related literatures.
      A sequence recommendation algorithm based on knowledge graph embedding and multiple neural networks
      SHEN Dong-dong, WANG Hai-tao, JIANG Ying, CHEN Xing
      2020, 42(09): 1661-1669. doi:
      Abstract ( 226 )   PDF (728KB) ( 183 )     
      Recurrent neural networks play an important role in sequence recommendation. However, in recommendation, the user's behavior sequences are far more complicated than the sentences in natural language processing or images in computer vision. A single recurrent neural network structure is difficult to fully mine user preferences, so this paper proposes a new sequence recommendation algorithm that takes into account both the time information and content information of the sequence. This algorithm is mainly divided into two parts: improved item embedding and sequence preference learning. Firstly, an item embedding method that incorporates a knowledge graph is used to generate high-quality item vectors. Secondly, a sequence modeling method combining convolutional neural networks with long-term and short-term memory neural networks is proposed. Furthermore, an attention-based framework is proposed, which dynamically combines user's points of interest. This algorithm is compared with the traditional methods and the existing advanced methods of the same type on the public data set MovieLens10M. The experimental results show that the average reciprocal ranking MRR@N of the recommended evaluation index and the recall rate Recall@N is improved significantly, which verifies the effectiveness of the proposed algorithm.

      Biomedical event extraction based on  deep contextual word representation and self-attention
      WEI You, LIU Mao-fu, HU Hui-jun,
      2020, 42(09): 1670-1679. doi:
      Abstract ( 182 )   PDF (669KB) ( 121 )     
      Biomedical event extraction is one of the most significant and challenging tasks in biome- dical text information extraction, which has attracted more attentions in recent years. The two most important subtasks in biomedical event extraction are trigger recognition and argument detection. Most of the preceding methods consider trigger recognition as a classification task but ignore the sentence-level tag information. Therefore, a sequence labeling model based on bidirectional long short-term memory (Bi-LSTM) and conditional random field (CRF) is constructed for trigger recognition, which separately uses the static pre-trained word embedding combined with character-level word representation and the dynamic contextual word representation based on the pre-trained language model as model inputs. Meanwhile, for the event argument detection task, a self-attention based multi-classification model is proposed to make full use of the entity and entity type features. The F1-scores of trigger recognition and overall event extraction are 81.65% and 60.04% respectively, and the experimental results show that the proposed method is effective for biomedical event extraction.

      Proximal policy optimization and adversarial learning based dialog generation
      CAI Yue, YOU Jin-guo, DING Jia-man
      2020, 42(09): 1680-1689. doi:
      Abstract ( 152 )   PDF (661KB) ( 103 )     
      Dialogue generation is the key research direction of natural language processing. Generative adversarial nets (GAN) have recently been well applied in the field of dialog generation. In order to further improve the quality of dialogue generation and solve the low efficiency of model training caused by the discriminative model return reward low utilization rate in the GAN training process, this paper proposes a dialogue generation algorithm (PPO_GAN) based on proximal policy optimization (PPO). 
      The algorithm, via GAN, generates a dialogue through the generation model, and distinguishes between generated dialogue and real dialogue through the discriminant model. The GAN is trained by proximal policy optimization method, which can deal with the situation that the back propagation of GAN cannot be differentiated when the dialogue is generated. While ensuring the monotonic non-reduction training of the generated model, the rewards obtained by the discriminant model can be reused by limiting the gra- dient of the generated model iteration. The experimental results show that, compared with dialog gene- ration algorithm such as the maxinum likelihood estimation  and Adver-REGS, the PPO_GAN algorithm improves the efficiency of dialogue training and the quality of dialog generation.

      UAV path planning based on  improved particle swarm optimization
      WANG Yi-hu, WANG Si-ming
      2020, 42(09): 1690-1696. doi:
      Abstract ( 348 )   PDF (642KB) ( 243 )     
      Aiming at the problem that the traditional Particle Swarm Optimization (PSO) algorithm is easy to fall into the local optimum when it solves the UAV path planning problem, the chemotactic operation and migration operation of the Bacteria Foraging Algorithm (BFA) are introduced in the PSO algorithm to improve its optimization ability. Firstly, based on the UAV (Unmanned Aerial Vehicle) flight environment, a three-dimensional elevation environment model is established, and the fitness function is established by using the path length cost, the obstacle risk cost and the elevation cost. Se- condly, based on the analysis of the principles and characteristics of particle swarm algorithm and bacterial foraging algorithm, the improvement methods and specific procedures of the algorithm are given. Finally, the MATLAB simulation verification shows that the hybrid algorithm effectively improves the defects of the particle swarm optimization algorithm. Compared with the traditional PSO algorithm, the optimization accuracy and stability of the hybrid algorithm are significantly improved in UAV path planning.