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 hardware acceleration method for the normalized
      product correlation algorithm and its FPGA implementation
      LI Hong-jun, GUO Yang, JIA Run
      2019, 41(11): 1905-1910. doi:
      Abstract ( 155 )   PDF (944KB) ( 255 )      Review attachment
      With the trend that cruise missiles aim at high efficient attack and intelligent enhancement, only using DSP software to process images cannot satisfy the higher real-time requirements for terrain matching and target recognition. Therefore, according to the advantage of FPGA implementation on computation accelerating, this paper proposes a simplification method for the normalized product correlation algorithm formula, which has the characteristics of high precision and high speed, and designs the product correlation’s hardware acceleration unit and multi-channel parallel computing architecture. FPGA implementation verifies that the proposal can satisfy the real-time requirements for terrain matching and target recognition of the new generation of cruise missiles.
       
      Fault diagnosis of analog circuits based
      on deep extreme learning machine
      YAN Xue-long,MA Run-ping
      2019, 41(11): 1911-1918. doi:
      Abstract ( 215 )   PDF (1098KB) ( 199 )      Review attachment
      Aiming at the problems of feature extraction and long model training time in analog circuit fault diagnosis, an analog circuit fault diagnosis algorithm based on deep extreme learning machine is adopted. The algorithm introduces the idea of auto encoder in deep learning into the extreme learning machine to construct a depth network, and forms more abstract advanced features from the underlying fault features. It can learn data features autonomously, avoiding additional feature extraction and selection. Finally, two analog circuits of Sallen-Key and four operational amplifiers with double quadratic high-pass filtering are taken as examples to carry out the simulation research. The experimental results verify the feasibility of fault diagnosis in analog circuits, and also show that the model has fast learning speed, good generalization ability and strong diagnosis ability. The classification accuracy of fault diagnosis can reach 100%, and the diagnosis time is about 0.3 s.
       
      OpenCL computing platform based
      on domestic software and hardware
      AN Ting-yu1,GUO Bao-bao2
      2019, 41(11): 1919-1923. doi:
      Abstract ( 167 )   PDF (396KB) ( 188 )      Review attachment
      With the development of intelligent computing and big data applications, the demand for acceleration components such as GPU is increasing. Based on the requirements for accelerated computing by running display and control applications on domestic fundamental software and hardware platforms, this paper studies the portability and implementation of OpenCL computing platforms, and preliminarily explores the GPU computing on domestic software and hardware platforms. The studied computing platforms  include Mesa, ROCm, Pocl and Beignet. Finally, the idea and solution of how to port and adapt ROCm to domestic platforms are given.

       
      Optimized FPGA memory allocation for image processing
      CHEN Kai-feng,LIANG Jian-ru
      2019, 41(11): 1924-1929. doi:
      Abstract ( 130 )   PDF (649KB) ( 152 )     
      Field Programmable Gate Array (FPGA) has broad prospects in computer vision applications. However, limited memory resources of FPGA are difficult to meet the performance, size and power requirements of the application scenarios. To solve this problem, this paper studies the resource allocation of on-chip memory, designs a partition balancing algorithm to minimize resource usage and power consumption, and implements it on the platform. The experimental results show that, compared with the commercial FPGA's advanced synthesis tools, the proposed algorithm improves the utilization rate by 60% and reduces the dynamic power consumption by up to 70%. In the experiment of the advanced algorithm MeanShift tracking, the experimental data shows that the partition balancing algorithm can reduce the total power consumption by up to 30% without affecting the key performance.
      Key words:
      A networking scheme for highly
      dynamic flying ad hoc networks
      HONG Jie1,2,ZHANG De-hai1
      2019, 41(11): 1930-1938. doi:
      Abstract ( 249 )   PDF (926KB) ( 267 )     
      The high-speed movement of nodes and the rapid change of network topology severely affect the performance of highly dynamic flying ad hoc networks. A networking scheme suitable for highly dynamic mobile ad hoc networks is proposed. The scheme focuses on QoS requirements and the highly dynamic characteristics of the network topology,reasonably configures different levels of network protocols, and adapts to highly dynamic application scenarios. The planar network is taken as an example to evaluate the performance of the FANET network in a highly dynamic environment. The experimental results verify that this networking scheme has certain feasibility and can meet the requirements of highly dynamic flying ad hoc networks. In addition, this paper proposes the optimal node number to maintain good network performance under the individual mobile mode and a counter measure strategy for adapting each network layer to highly dynamic topology changes.


       
      VC Chain:An alliance audio-video
      copyright blockchain system
      CHEN Zi-hao,LI Qiang,GAN Jun,ZHANG Chao,LI Zu-rui
      2019, 41(11): 1939-1948. doi:
      Abstract ( 246 )   PDF (1107KB) ( 327 )      Review attachment
      With the rapid development of the audio-video industry, the copyright protection of audio-video productions has attracted more and more attentions. The centralized management of the traditional digital copyright system leads to the disadvantages of easy disclosure and tampering of copyright data, which becomes a difficult problem of audio-video copyright protection. How to use technologies to determine the copyright in the case of infringement and copyright disputes becomes the key problem. To solve this problem, an alliance audio-video copyright blockchain system (VC Chain) based on the improved PBFT consensus algorithm is proposed. The system prevents data leakage or malicious tampering, and provides a more reliable process of digital copyright confirmation and tracing by cryptography technology, which improves the security of audio-video digital copyright. The experimental results show that the new consensus algorithm optimizes the system in terms of consensus speed and data communication times. Finally, some problems of digital copyright management are pointed out and their corresponding solutions are given.
       

       
      A mixed ring loading problem with penalty cost
      GUAN Li,FENG Yan-xue
      2019, 41(11): 1949-1953. doi:
      Abstract ( 108 )   PDF (374KB) ( 127 )     
      In the past more than 20 years, the ring loading problem has been studied widely. The ring loading problem with penalty cost is the generalized form of the ring loading problem. There are some research results on undirected rings and directed rings. This paper proposes a mixed ring loading problem with penalty cost. Given a mixed ring C and a set of requests, each request  rj has a demand  dj and a penalty pj. When the request  rj is accepted, its traffic can be transmitted along the ring clockwise and counter clockwise. When the request   rj is rejected, the penalty   pj is generated to minimize the sum of the maximum load among all links on the ring and the total penalty cost of the rejected requests. When the demand can be split, a  2-approximation algorithm is given by using the LP-rounding technique. Further, a  1.58-approximation algorithm is obtained by using the random rounding technique. Similarly, when the demand cannot be split, a  3-approximation algorithm and a (1.58+ε)-approximation algorithm are given, where ε>0  is a constant.
       
      A graphical password scheme based on
      multi-stroke graphics and numbers
      GU Yan-bo,LI Jing-wen,HUO Jin-ping,XI Xiao-hui
      2019, 41(11): 1954-1960. doi:
      Abstract ( 97 )   PDF (608KB) ( 137 )     
      Android Unlock Pattern (AUP) is currently the most widely used graphical password unlocking scheme on portable devices like mobile phones and pads. However, the available password in practice occupies only a fraction of the graphical password space, and the uneven distribution of the password caused by the user’s habitual choices makes the security of AUP far lower than that in theory and makes the password much easier to be cracked by attackers. This paper proposes a graphical password scheme based on multi-stroke graphics and numbers (MSDGP),which recommends a password of the corresponding difficulty level according to the user’s selection. This scheme is similar with AUP. Due to the difference of the numbers and their locations, the graphical password space is greatly enlarged. Besides, the scheme can solve the uneven distribution caused by the user’s habitual choices as the password is recommended by the system, which effectively avoids brute cracks and dictionary attacks and owns a higher security.
       
      A texture synthesis coverless information
      hiding method based on LBP
      WEI Wei-yi,WANG Yu,A Cheng-feng
      2019, 41(11): 1961-1967. doi:
      Abstract ( 151 )   PDF (939KB) ( 193 )     
      In order to improve the embedded capacity and anti-interference capability of the coverless steganography algorithm, this paper proposes a texture synthesis information hiding scheme based on LBP texture analysis. Firstly, the original small-size texture image is selected and divided into uniform pixel blocks, the LBP value of each pixel in the image blocks is  calculated, and the LBP value with the largest LBP distribution is taken as the representive information of the image block. Secondly, when hiding the secret information, the pseudo-random sequence is generated with a specified key to determine the position of the texture candidate block placed on the white paper, then the candidate block is selected according to the value of the secret information and placed on the designated position on the white paper, and the remaining blank areas are filled by the texture synthesis method. Inversely, when extracting secret information, the position of the steganography image block is obtained according to the pseudo-random sequence generated by the key, then the LBP value of each image block with the largest distribution is calculated to obtain the secret information. Experimental results show that the steganography image generated by this method has good visual effect and further improves the embedded capacity and anti-interference ability.
       
      A new virtual reality software
      system development method
      ZHOU Zhe-hong1,2,3,XUE Jin-yun1,3,HUANG Jie-wen1,2,3
      2019, 41(11): 1968-1975. doi:
      Abstract ( 123 )   PDF (656KB) ( 188 )     
      Virtual reality technology is a comprehensive technology involving computer graphics, multimedia technology, human-computer interaction and artificial intelligence. It has a wide range of applications in education, medical, entertainment, military and many other fields. All of these technologies are ultimately implemented by computer software. This makes the software of the virtual reality system very large and complex. The programming techniques of traditional software development methods focus on processing text data, which obviously cannot meet the needs of developing virtual reality software. Based on the successful PAR method and PAR platform of the author's R&D team, according to the characteristics of the virtual reality system, this paper explores a new virtual reality software system development method, and further expands and improves the existing multimedia processing technology in the PAR platform, the formal modeling technology and the automatic generation system of C# and other high-level language programs.

       
      A modeling and scheduling algorithm
      for multi-platform avionic resources
      SHI Wen-jie,ZHANG Yu-qi,LI Kui
      2019, 41(11): 1976-1984. doi:
      Abstract ( 127 )   PDF (1618KB) ( 186 )     
      With the development of avionics systems, it is particularly important to rely on the network to build multi-platform avionics resources including different aircraft avionic resources. By comprehensively utilizing the resources of different flight platforms, the advantages of different platform resources can be utilized and cooperation is used to enhance the task handling ability. Resource management and task scheduling of multi-platform avionics systems are their core functions. In order to simulate and verify the resource management function of multi-platform avionics systems and further study the resource scheduling method, the paper studies the modeling method and the scheduling algorithm of multi-platform avionic resources. The paper reasonably schedules hardware resources on multi-platform avionics systems to increase the task acceptance rate. Firstly, multi-platform avionics resources are modeled by multi-level hierarchical topology, and the requirements of multi-platform avionics tasks are analyzed. Secondly, based on the SST (Sliding-Scheduled Tenant) adaptive scheduling algorithm, factors such as sensors and priorities are added to achieve higher acceptance rate. The algorithm completes a series of task requests in the demand allocation process of avionic resources. Finally, the algorithm is implemented in CloudSim simulation  environment. The experimental results are analyzed comprehensively from different scenarios, which show that the proposed algorithm has a higher request acceptance rate than the original algorithm.
      Pixel-level skin segmentation and face color grading
      WU Cong-zhong1,HOU Guo-song1,DING Zheng-long2,XU Liang-feng1,ZHAN Shu1
      2019, 41(11): 1985-1990. doi:
      Abstract ( 129 )   PDF (731KB) ( 206 )      Review attachment
      Skin is the largest organ in the human body, and skin color is more convenient and stable than other biological properties of the human body. Therefore, it is very meaningful to design an effective skin color grading system. In this paper, the skin color grading system is divided into two parts: skin segmentation and skin color grading. For the skin segmentation,a multi-scale feature fusion network is built under the framework of the generative adversarial network. Compared with the traditional semantic segmentation networks, the proposed segmentation model makes full use of the information of each layer's feature map. In the face color grading experiment, the SVM classifier and the BP neural network are trained with 1 000 images in the normalized rgb, HSV, and Lab color spaces. 128 skin images are used as test sets, and the correct rate is between 73% and 76%. Then,the color features are combined with the LBP texture features of the skin region to do the learning. The correct rate of the SVM classifier is 85%, and the correct rate of the BP neural network is 91%.
       
      A product image retrieval method
      based on SHN model
      HE Zhou-yu1,FENG Xu-peng2,LIU Li-jun1,HUANG Qing-song1,3
      2019, 41(11): 1991-1999. doi:
      Abstract ( 115 )   PDF (930KB) ( 188 )     
      In recent years, with the rapid development of the e-commerce industry, how to quickly and accurately find the required goods through image information in the huge product library has important application value.Aiming at the characteristics such as the large scale of product image data, the large difference of data between classes, the large difference between the scales of photographed products and the loss of detailed information in compressed images, an Spatial pyramid pooling -Hash-Net (SHN)model combining spatial pyramid pooling strategy and hash learning is proposed as the feature extraction part of the product image retrieval method. In order to improve the robustness of the model to image deformation, the spatial pyramid pooling strategy is adopted to achieve multi-scale feature fusion. In order to make the learned hash code have better independence, the quantization error loss and additional weights are used to constrain the hash code.The method preserves the original image information and solves the negative effects caused by image scale changes, and it can realize fast product image retrieval through hash coding.The experimental results show that the mAP value of this method reaches 91.986 3%, and the time for completing a search is 0.034 856 s. The image retrieval performance is better than the current mainstream methods.
       
      Unsupervised learning for small
      objects detection in retinal images
       
      SUN Yi-fei,WU Ji-gang,ZHANG Xin-peng
      2019, 41(11): 2000-2006. doi:
      Abstract ( 286 )   PDF (651KB) ( 289 )      Review attachment
      Small object detection is an unresolved problem for image processing, especially for medical image processing. Microaneurysm (MA) is a kind of small objects in retinal images. It has small size, low local contrast, and more noise interference, so it is difficult to be detected.Traditional detection methods require manual extraction of features, making it difficult to accurately detect MA.The detection based on deep learning requires a large amount of complex preparatory work, and it is difficult to solve the imbalance problem between positive and negative samples, which is easy to cause over-fitting.Sparse autoencoder (SAE) is an unsupervised machine learning algorithm that efficiently extracts the features of samples in an environment with unbalanced sample data.Therefore, an unsupervised learning method based on SAE is proposed to detect MA. The weights and offsets of SAE are updated by backpropagation to extract the features of the samples, and the extracted features are used to train softmax to achieve accurate detection of MA.In order to evaluate the performance of the method, three databases (Retinopathy Online Challenge, DIARETDB1 and E-ophtha-MA) are used to carry out experiments.Experimental results show that the method can accurately detect MA in retinal image and obtain higher accuracy and sensitivity. The accuracy rates are 98.5%, 87.2%, and 92.6% respectively, and the sensitivity are 99.9%, 99.8%, and 98.7% respectively.
       
      An image data augmentation algorithm
      based on convolutional neural networks
      JIANG Yun,ZHANG Hai,CHEN Li,TAO Sheng-xin
      2019, 41(11): 2007-2016. doi:
      Abstract ( 291 )   PDF (1223KB) ( 281 )      Review attachment
      Improving the generalization ability and reducing the over-fitting risk is the research focus of deep convolutional neural networks. Occlusion is one of the critical factors affecting the generalization ability of convolutional neural networks. It is usually hoped that the models after complex training can have a good generalization for occlusion images.In order to reduce the over-fitting risk and improve the robustness of the model to random occlusion image recognition, this paper proposes an activation feature processing algorithm. During the training process, the input image is occluded by processing the maximum activation feature map of a convolutional layer, then the occluded new image is used as a new input to the network to go on training the model. The experimental results show that the proposed algorithm can improve the classification performance of multiple convolutional neural network models on different datasets and the trained models have excellent robustness to the identification of random occlusion images.
       
      Knowledge measure of hesitant
      intuitionistic fuzzy set and its applications
      ZHANG Rong-rong,LI Yong-ming
      2019, 41(11): 2017-2026. doi:
      Abstract ( 147 )   PDF (480KB) ( 170 )     
      The hesitant intuitionistic fuzzy set integrates the advantages of the intuitionistic fuzzy set and the hesitant fuzzy set, and can deal with the uncertainty problems with inconsistent preferences of decision makers perfectly. By considering the hesitancy and fuzziness of information provided by decision makers, the axiomatic definition of the knowledge measure on the hesitant intuitionistic fuzzy set is proposed. Besides, a class of parametric knowledge measure model that comply the axiom is developed, which can not only effectively depict the information in the hesitant intuitionistic fuzzy set, but also indicate the attitude characteristics of decision makers. Next, based on the discussion of the parameters in the knowledge measure, an amount of knowledge measures representing different attitude characteristics of decision makers are obtained. Furthermore, it is verified that the knowledge measure is proportional to the change of the attitude coefficient. Finally, a multiple attribute group decision making method based on the constructed knowledge measure of the hesitant intuitionistic fuzzy set is proposed. Meanwhile, a practical application, i.e., air conditioning installation company selection of an Internet company, demonstrates the effectiveness and practicability of the developed measure.

       

       

       
      A user churn prediction method
       based on multi-model fusion
      YE Cheng,ZHENG Hong,CHENG Yun-hui
      2019, 41(11): 2027-2032. doi:
      Abstract ( 149 )   PDF (804KB) ( 170 )     
      Accurate user churn prediction ability facilitates improving user retention rate, increasing user count and increasing profitability. Most of the existing user churn prediction models are single model or simple integration of multiple models, and the advantages of multi-model integration are not fully utilized.This paper draws on the idea of Bootstrap Sampling in random forests, proposes an improved Stacking ensemble method, and applies the method to the real data set to predict the user churn. Through the experimental comparison on the validation set, the proposed method is better than the classical Stacking ensemble method with the same structure in the terms of the F1-score, recall rate and prediction accuracy of user churn. When the appropriate structure is adopted, the performance can surpass the optimal performance on the base classifier.
       
      A hybrid recommendation model based on incremental
      collaborative filtering and latent semantic analysis
      LIU Hui1,2,3,WAN Cheng-feng1,2,WU Xiao-hao1,2
      2019, 41(11): 2033-2039. doi:
      Abstract ( 122 )   PDF (587KB) ( 142 )     
      Traditional news recommendation systems regularly update recommendation models, which cannot adjust recommendation lists dynamically according to the change of user preferences. In order to solve this problem, we propose a hybrid recommendation model (IULSACF). It includes two key parts: an item-based incremental update collaborative filtering algorithm and a key word frequency based latent sementic analysis algorithm. The hybrid recommendation model dynamically adjusts the recommendation list by incrementally updating the similarity list of items in the item-based incremental update collaborative filtering module, and combines the latent semantic analysis algorithm to ensure the relevance of recommended articles. Experimental results show that the proposed IULSACF algorithm is superior to traditional recommendation methods in all evaluation indexes.
       
      A collaborative filtering recommendation algorithm
      based on bipartite graph partitioning co-clustering
      HUANG Le-le1,MA Hui-fang1,2,LI Ning3,YU Li1
      2019, 41(11): 2040-2047. doi:
      Abstract ( 144 )   PDF (678KB) ( 255 )     
      To accurately and actively provide users with  potentially interested information or services is the main task of a recommender system. Collaborative filtering is one of the most widely used recommendation algorithms, whereas it is suffering the issue of data sparsity that severely degrades recommendation quality. To address this issue, we propose a collaborative filtering recommendation algorithm based on bipartite graph partitioning co-clustering, called BPCF. Firstly, users and items are constructed into a bipartite graph for co-clustering, which is then mapped to the low-dimensional feature space. Then, the proposed algorithm computes the two types of improved similarities (cluster preference similarity and rating similarity) according to the clustering results and combines them. Based on the combined similarity, the user-based approach and item-based approach are adopted, respectively, to predict for an unknown target rating and these prediction results are fused. Experimental results show that the proposed method outperforms the state-of-the-art co-clustering collaborative filtering recommendation algorithms.
       
      A hybrid recommendation algorithm combining
      trust degree and project association
      YANG Feng-rui1,2,3,WU Xiao-hao 1,2,WAN Cheng-feng1,2
      2019, 41(11): 2048-2054. doi:
      Abstract ( 86 )   PDF (602KB) ( 153 )     

      Recommendation is an important means of dealing with information overload in the era of big data. Traditional recommendation algorithms have relatively low accuracy and reliability. Aiming at the cold start issue of new users and projects, we propose a hybrid recommendation algorithm   based on probability matrix decomposition (HR-TP). Firstly, the user's invisible trust relationship is mined from the perspective of the user’s rating. Then the label context is used to measure the relationship between items according to user characteristics. The relationship matrix is fused with the probability matrix model to make recommendation. Experiments show that the proposed method achieves better results in recommendation accuracy compared with conventional methods.