Current Issue

• 2022, No. 01　Published:25 January 2022 Last issue    Next issue
  Columns in this issue:

• Implementation and optimization of sparse matrix vector multiplication based on RISC-V vector instruction

  GU Yue, ZHAO Yin-liang

  2022, 44(01): 1-8. doi:

  Abstract ( 442 )   PDF (714KB) ( 264 )   　　

  Open source instruction set architecture RISC-V has the advantages of high performance, modularization, simplicity, easy extension, etc., and is widely used in the Internet of Things, cloud computing and other fields. The V module of its vector expansion part supports matrix numerical calculation well. As an important part of matrix numerical calculation, sparse matrix vector multiplication (SpMV) has profound research significance and value. Using the vector configurability and addressing characteristics of RISC-V instruction set, vector multiplication of sparse matrix based on CSR, ELLPACK and HYB compressed format is vectorized respectively. Meanwhile, considering that the sparse matrix is extremely sparse and the number of non-zero elements in each row fluctuates greatly, the HYB storage format is improved by compressing the storage of row vectors with low density of non-zero elements and adjusting the HYB segmentation threshold, which significantly improves the computational efficiency and storage efficiency.

  Research and implementation of dynamic migration of peripheral resources in an embedded dual operating systems architecture

  CUI Zhen-li, LUO Yu

  2022, 44(01): 9-15. doi:

  Abstract ( 103 )   PDF (659KB) ( 105 )   　　

  With the increasingly performance improvement and function expansion of embedded devices, the single operating system architecture is becoming difficult to meet the increased requirement of more and more complex applications. Therefore, a dual operating system architecture is proposed by the academic and industrial community. However, how to reasonably configure limited peripheral resources is an important problem remained to be solved by the dual operating system architecture. Based on the binding principle of peripheral interrupts and CPU cores in Huawei HiSilicon Hi3559AV100 SOC dual operating system architecture, a dynamic migration solution of peripheral resources during system operation is proposed, and experiments were carried out on an embedded device equipped with the SOC. The results show that the proposed solution has good feasibility and reliability during the operation.
  

  
  

  Task allocation of crowdsourcing for maximizing satisfaction with preference matching

  GUO Jia-yu, FU Xiao-dong , YUE Kun, LIU Li, FENG Yong, LIU Li-jun

  2022, 44(01): 16-26. doi:

  Abstract ( 129 )   PDF (897KB) ( 131 )   　　

  Task allocation mechanism plays an important role on crowdsourcing task quality. However, the existing allocation methods do not consider the bilateral user preferences of crowdsourcing, the accuracy of the allocation results remains to be improved, and there are many crowdsourcing users who are not satisfied with the current allocated crowdsourcing tasks, causing the phenomenon of low quality of crowdsourcing task accomplishment. Therefore a crowdsourcing task allocation method based on preference matching is proposed. Firstly, this method compute satisfaction and construct satisfaction matrix by using the preference of crowdsourcing tasks and workers. Secondly, by referring to the idea of stable matching and taking into account the bilateral preferences of crowdsourcing subjects, the method makes crowdsourcing subjects as satisfied as possible with the current allocation objects so as to guarantee the stability of the allocation results. Then, by using the idea of stable matching, the crowdsourcing task allocation problem is modeled as an optimization problem to find the maximum satisfaction of tasks under stable matching rules. Finally, greedy algorithm is used to solve the problem and a crowdsourcing task allocation scheme is obtained. The rationality and effectiveness of the method are verified by experiments, which shows that the method improves the accuracy of the allocation and effectively reduces the number of invalid allocations, thus improving the quality of crowdsourcing task allocation. 
  

  
  

  Analog circuit fault diagnosis based on representation learning

  TAN En-min, WANG Chen

  2022, 44(01): 27-35. doi:

  Abstract ( 76 )   PDF (992KB) ( 97 )   　　

  Aiming at the problems of multiple features, high noise and long fault diagnosis time in analog circuit fault diagnosis, an analog circuit fault diagnosis model based on H-DELM is proposed. The architecture unit of the model is a deep extreme learning machine (DELM-AE) with double random hidden layers. Two random hidden layers are used to encode features and one output layer is used to decode features. The H-DELM model is constructed by stacking DELM-AE in hierarchical structure. Because DELM-AE can represent the features and the output is the same as the original input information, H-DELM can copy the original input data as much as possible, and then learn more expressive and compact features. Finally, the verification is carried out by two circuits: quadruple operational amplifier double-order high-pass filter and two-stage four-op-amp biquad lowpass filter. The experimental results show that the model is feasible in analog circuit fault diagnosis. Compared with other model, it is proved that the proposed model has high robustness, the classification speed is about 1s, and the accuracy rate of fault classification can reach 100%.
  
  
  
  

  
  

  Research progress in resistive switching mechanism and materials of memristor

  DENG Ya-feng, WEI Zi-jian, WANG Dong

  2022, 44(01): 36-47. doi:

  Abstract ( 587 )   PDF (1245KB) ( 661 )   　　

  Memristor can integrate information storage and logic operation into one electronic device, which will break the traditional von Neumann computer architecture, and its application prospect is immeasurable. Firstly, the development course and basic concept of memristor are introduced. Secondly, the resistive  switching mechanism of memristor and the choice of its material are summarized. The known resistive  switching mechanism of memristor can be mainly divided into three categories: anion-dominant resistive switching mechanism, cation-dominant resistive switching mechanism, and pure electron-dominant resistive switching mechanism. At the same time, the characteristics of different types of materials in memristor application are described in detail. Then the application of memristor in Boolean calculation and neuromorphic system is discussed. Finally, the future development direction of memristor and the possible problems in its practical application are expected.
  

  
  

  Review on four-layer load balancing technology in data center network

  LI Li , WANG Shuo, HUANG Tao, LIU Yun-jie,

  2022, 44(01): 48-59. doi:

  Abstract ( 225 )   PDF (1035KB) ( 152 )   　　

  Aiming at the problem of data centers difficulty in adapting to traffic growth for horizontal expansion and ensuring connection consistency, this paper explains the important role of layer-4 load balancing technology in coping with high concurrent access and improving resource utilization, and sorts out four-layer load balancing modules domestic and international. The advantages and disadvantages of load balancers deployed in different ways are summarized, and the application and latest development of network programmable forwarding technology in the field of layer-4 load balancing are analyzed. Finally, The development of load balancing technology under the new network situation is further prospected and future research directions are discussed.
  

  
  

  An exact thumbnail-preserving image encryption scheme

  HOU Xing-wang, ZHAO Ruo-yu, ZHANG Yu-shu

  2022, 44(01): 60-67. doi:

  Abstract ( 215 )   PDF (711KB) ( 154 )   　　

  With the popularity of mobile phones, computers, and tablets, photos are more easily available in daily life, and people are used to storing a large number of photos in the cloud. While enjoy- ing the convenience brought by cloud storage, users are also threatened by privacy leaks. Although scholars have devised many image encryption schemes to prevent privacy leakage, they often ignore the usability of images. Recently, Tajik et al. proposed an ideal thumbnail preservation encryption scheme to balance image privacy and usability well, but this scheme only used two pixels as a group during the process of encryption and thus leads to the low efficiency. Therefore, an image encryption scheme using a segmentation method is proposed. This scheme uses three pixels as a group to do encryption, so as to keep the thumbnail of the ciphertext image consistent with the thumbnail of the plaintext image. The proposed scheme has higher efficiency than the Tajiks. Experiments show that this scheme can make the ciphertext image accurately maintain the same thumbnail as the plaintext image, which balances privacy and usability.
  
  
  

  Design and simulation of a free-space laser-communicating network topology algorithm

  ZHOU Yang, DONG Yu-hui, YAO Xu, LIU Qiang, Sun Yan-tao, ZHANG Liang

  2022, 44(01): 68-74. doi:

  Abstract ( 88 )   PDF (868KB) ( 121 )   　　

  his paper studies the topology controlling problem for mobile nodes in free-space laser-communicating network. A heuristic topology algorithm based on set partitioning is proposed for laser-communicating Ad-hoc network in specific scenes, where the connectivity of all nodes and the stability of all links can be achieved according to the position and attitude of the nodes and the constraint conditions of building laser links. An almost optimal fully connected stable topology can be generated by heuristic iterations, thus enabling the network layer communication for global nodes. The network diameter can be optimized by the algorithm, and the topology can be repaired by set merging when link disconnection occurs. The simulation result indicates that the topology stability and the network throughput is optimized, and the topology can be successfully repaired. Besides, the execution time in different scales conforms to the time complexity of Freud algorithm, i.e. O(n3).
  
  

  An improved unconstrained optimization 3D-DV-Hop localization

  ZHANG Jing, LI Yu,

  2022, 44(01): 75-83. doi:

  Abstract ( 73 )   PDF (684KB) ( 113 )   　　

  Aiming at the disadvantages of the traditional 3D DV-Hop positioning algorithm with large positioning errors and heavy calculation of machine learning and bionic algorithms, an improved unconstrained optimization 3D DV-Hop positioning algorithm is proposed. In terms of hop count, a two- communication radius strategy is adopted to calculate the minimum hop value. In terms of hop distance, a square cost function method is proposed to optimize the anchor node hop distance value, and its weighted hop distance value is used as the unknown node hop distance value. Finally, according to the unconstrained solution idea of the constraint problem, the weighted error is minimized and then solved. Through simulation comparison with traditional algorithms and various improved algorithms under three conditions, it is verified that the optimization algorithm can significantly reduce the positioning error under the condition of low calculation amount.
  

  
  

  Image caption generation based on residual dense hierarchical information

  WANG Xi, ZHANG Kai, LI Jun-hui, KONG Fang

  2022, 44(01): 84-91. doi:

  Abstract ( 85 )   PDF (746KB) ( 99 )   　　

  The current mainstream method of image caption generation is based on deep neural networks, especially the self-attention mechanism model. However, the traditional deep neural network layers are stacked linearly, which makes the information captured by the low-level network not be able to be reflected in the high-level network and not fully utilized. Therefore, this paper proposes a method based on dense residual network to obtain hierarchical semantic information to generate high-quality image captions. First of all, in order to make full use of the layer information of the network and extract the local features of each layer in the deep network, this paper proposes Layer RDense (Layer Residual Dense), which carries out dense residual connections between layers. Secondly, SubRDense (Sublayer Residual Dense) is proposed. It uses a dense residual network in the sub-layers of each layer of the network at the Decoder side, in order to better integrate image features and image description information. The experimental results based on the MSCOCO 2014 dataset show that the proposed LayerRDense and SubRDense networks can further improve the performance of image caption generation.

  Research and application of an improved wavelet soft threshold function in image denoising

  XU Jing-xiu, ZHANG Qing

  2022, 44(01): 92-101. doi:

  Abstract ( 146 )   PDF (1559KB) ( 278 )   　　

  For the conventional wavelet soft threshold denoising method, the wavelet coefficients of the image before and after processing are different, resulting in serious image distortion after de- noising. In order to further improve the denoising effect and improve the ability of denoising and detail preserving, the threshold selection method and the threshold function are improved. The improved method determines the threshold value through the length of each sub-band of wavelet transform to realize the adaptive and accurate quantification of the threshold value. The improved soft threshold function uses the hyperbolic tangent function to replace the symbol function, and gradually compresses the wavelet coefficients within the absolute value range of the threshold by nonlinear function, so that the improved threshold function has better continuity and stronger stability. The simulation results show that the peak signal-to-noise ratio (PSNR) and structure similarity of the improved wavelet threshold denoising method are improved by 48% and 80.6% respectively. It is concluded that, compared with the conventional wavelet threshold denoising method, the new improved wavelet soft threshold denoising method can effectively suppress the noise while retaining the details of the original image, and the image qua- lity is significantly improved.
  
  

  Correcting distorted document images on smartphones

  ZHOU Li, FENG Bai-ming, GUAN Yu, FANG Ge

  2022, 44(01): 102-109. doi:

  Abstract ( 555 )   PDF (1089KB) ( 150 )   　　

  Distorted document images usually appear in smartphone images, which causes a lot of inconvenience for users. These images affect text recognition, post-processing, etc. Existing distorted document image correction methods have some limitations such as single correction type and unsatisfactory correction effect. To solve the above problems, an distorted document image correction method based on minimizing re-projection is proposed. Firstly, the connected domain of text lines is obtained by detecting text field contour and merging text field contour. Secondly, PCA (Principal Component 
  Ana- lysis) is adopted to generate key points of text in the connected domain of the row. Finally, resampled parameters are obtained by minimizing the distance between the key points and their projection points. The distorted pages are re-projected to minimize the document image correction. After correction, the recognition rate is effectively improved. Compared with the existing methods, a better recognition effect is achieved. Moreover, ablation experiments are used to verify the improvement effect of text field merging and minimization re-projection on the recognition performance respectively.
  

  
  

  A semi supervised image annotation method based on LDA and convolutional neural network

  WANG Bao-cheng, LIU Li-jun, HUANG Qing-song,

  2022, 44(01): 110-117. doi:

  Abstract ( 80 )   PDF (635KB) ( 97 )   　　

  With the continuous emergence of intelligent devices, the number of pictures increases rapidly. However, many images are not fully utilized because they are not labeled. In order to solve this problem, a semi supervised image annotation method based on LDA and convolutional neural network is proposed. Firstly, all text information in the image training set is put into LDA to generate text tagging words. Secondly, the convolutional neural network is used to obtain the high-level visual features of the image, and the convolutional neural network is optimized by adding attention mechanism and modifying loss function. Thirdly, the label words generated by LDA are combined with the high-level visual features of the obtained image, and the semi supervised learning is used to complete the model training. Finally, the correlation between the tagging words and the prediction results using the final model are combined to complete the final tagging of the image. Comparative experiments on the IAPR TC-12 image data set show that the proposed labeling method is more accurate.
  
  

  An image saliency area style transfer method combining visual attention mechanism

  WANG Yang, YU Zhen-xin, LU Jia,

  2022, 44(01): 118-123. doi:

  Abstract ( 102 )   PDF (790KB) ( 157 )   　　

  Performing style transfer on part of an image usually results in style overflow and insigni- ficant effects after stylization of smaller areas. Aiming at this problem, a style transfer algorithm for image saliency regions is proposed. Firstly, according to the characteristics of the human visual attention mechanism, the saliency regions in the training image data set are labeled, and the fast semantic segmentation model is used for training to obtain a binary mask image containing the saliency regions of the image. Then, by simplifying the network layer structure of the fast neural style transfer model, and adopting the instance regularization layer in the generating network part, a more realistic overall style transfer result is obtained. Finally, the binary mask image obtained by semantic segmentation is combined with the overall style transfer image, and the final result image is output. A comparative experiment was carried out on the Cityscapes dataset and the Microsoft COCO 2017 dataset. The results show that the local target area in the image is stylized uniformly and delicately, and can be well integrated with the background area. While a more realistic style transfer effect is achieved, operating efficiency is more dominant.
  

  
  

  Multi-focus image fusion with improved sparse representation and integrated energy sum

  ZHANG Gui-cang, WANG Jing, SU Jin-feng

  2022, 44(01): 124-131. doi:

  Abstract ( 72 )   PDF (905KB) ( 87 )   　　

  In order to solve the problem of limited retention of detail information in the multi-focus image fusion algorithm, a multi-focus image fusion algorithm with improved sparse representation and integrated energy sum is proposed. Firstly,the non-subsampled shearlet transform is used on the source image to obtain low-frequency and high-frequency coefficient matrix. Secondly, the image block is extracted from the low-frequency coefficient matrix through the sliding window technique,a joint local adaptive dictionary is constructed, and the sparse representation coefficients are calculated using the orthogonal matching tracking algorithm. Then, the sparse after fusion is obtained using the variance energy weighting rule coefficients, and the fused low-frequency coefficient matrix is obtained through the reverse sliding window technique. Thirdly, for the high-frequency coefficients, the integration rule of the integrated energy sum is proposed to obtain the fused high-frequency coefficient matrix. Finally, the fusion image is obtained by inverse transformation. The experimental results show that the algorithm can retain more detailed information and has certain advantages in visual quality and objective evaluation.
  

  
  

  
  

  Application of DenseNet in voiceprint recognition

  ZHANG Yu-jie, ZHANG Zan

  2022, 44(01): 132-137. doi:

  Abstract ( 125 )   PDF (562KB) ( 146 )   　　

  In order to improve the recognition performance of voiceprint recognition technology, DenseNet is applied to the spectrogram to realize voiceprint recognition. DenseNet is optimized from two aspects: improving the computing efficiency of the network and enhancing the characterization ability of voiceprint features. Depth separable convolution is used to reduce the amount of network parameters, and center loss function item is increased to improve the characterization ability of voiceprint features. The training results show that, through the depth separable convolution, the network parameters are reduced by 25.5%, and the model size is reduced by 24.6%; The simulation results show that the increase of the center loss item makes the clustering effect of the voiceprint feature more obvious and improves the characterization ability of voiceprint features. Therefore, the improved DenseNet can achieve good recognition results in the field of spectrogram and voiceprint recognition.
  
  
  

  A survey on quality evaluation of machine generated texts

  QIN Ying

  2022, 44(01): 138-148. doi:

  Abstract ( 138 )   PDF (644KB) ( 130 )   　　

  The quality evaluation of machine generated texts largely affects the research of Natural Language Generation (NLG), and has become a bottleneck restricting the development of the field. This paper reviews on the quality evaluation of various NLG tasks in a broad sense including machine translation, automatic summarization, dialogue, image captioning and machine writing with thorough summarization. The paper introduces the features, pros and cons of human evaluation and automatic metrics respectively as well as some open evaluation resources. This review analyzes the different perspective and applications of various evaluation tasks.  The comparative analysis of different evaluation methods can provide reference for method fusion and exploration of key issues. Overall, the quality evaluation of machine-generated language is still limited to the superficial comparison of linguistic forms, and there are many challenges in deeper evaluation at the level of semantic and coherence or cohesion. Based on the analysis of difficulties and current developments, the paper proposes the research tendencies of quality evaluation of generated texts.
  

  
  

  A sentiment unit representation method based on layer hierarchy

  ZHANG Bao-hua, LI En-lin, ZHANG Hua-ping, SHANG Jian-yun

  2022, 44(01): 149-158. doi:

  Abstract ( 102 )   PDF (599KB) ( 95 )   　　

  Sentiment word is the basic unit in the task of sentiment analysis, so sentiment lexicon plays an important role in sentiment analysis. Currently, the sentiment lexicon building methods only use word formation and semantic information, but ignore the context. Based on this, for some words with unknown semantics, it is difficult for traditional semantic methods to obtain the semantic weight, and for some words that have new usage due to context changes, it is difficult to calculate their true weight using semantic methods. To rectify the problem, a sentiment analysis hierarchy from the word to chapter is proposed. Each layer has a representation method and sentiment value calculation formula corresponding to the upper layer, which subdivides the analysis unit from sentence dimensions into word dimensions. Based on this, this paper proposes an automatic construction method for sentiment lexicon based on the character and the context of sentiment word. This method can calculate the weight of sentiment word by using the public sentiment lexicon, the word formation of sentiment word, and the contextual sentiment tendency of sentiment word, and the obtained result is more accurate. Experiments on the real dataset of social networks show that the sentiment unit constructed in this paper has a 3% improvement in accuracy compared with the previous methods. At the same time, the sentiment unit can be used directly in sentiment analysis tasks and the accuracy of sentiment analysis has a 9% improvement in rule-based sentiment analysis experiments and a 3% improvement in deep learning methods.
  

  
  

  Recognition  and division of aircraft flight action based on MRF model

  YAN Ting-long, LI Ying, WANG Feng-qin

  2022, 44(01): 159-164. doi:

  Abstract ( 101 )   PDF (510KB) ( 119 )   　　

  Military aircraft flight action have strong randomness and ambiguity. In order to realize the recognition and division of military aircraft flight action, a Markov Random Field (MRF) based recognition and division algorithm is proposed. The flight data segment is divided and clustered to realize the recognition and division of flight actions. Simulation experiments show that, compared with traditional flight action  recognition algorithms, the flight action recognition algorithm based on the MRF model has a higher recognition rate.
  
  
  
  

  Overview on sentiment analysis of microblog

  WANG Chun-dong, ZHANG Hui, MO Xiu-liang, YANG Wen-jun

  2022, 44(01): 165-175. doi:

  Abstract ( 305 )   PDF (785KB) ( 245 )   　　

  With the rapid growth of the number of microblog users, some emotions and opinions carried in microblog have a growing impact on the society, especially some negative emotions related to the personal safety of the public, which may affect the stability of the society. Therefore, it is of great significance to analyze the sentiment of microblog. The content of microblog sentiment analysis includes the acquisition of microblog corpus, the preprocessing of microblog corpus and the methods of sentiment analysis. The commonly used sentiment analysis methods include the method based on emotion dictionary, the method based on machine learning, and the method based on depth learning. With the widespread use of attention mechanism in NLP field, many researchers began to integrate attention mechanism into deep learning model for sentiment analysis, which greatly improves the accuracy of sentiment analysis. The BERT model proposed by Google is also based on attention mechanism essentially, which has made a breakthrough in the field of sentiment analysis.