Table of Content

2021-09-10, Volume 41 Issue 9

Article List

Full Abstract

Previous Issue    Next Issue

Artificial intelligence

Semi-supervised classification algorithm based on weight diversity

MAO Mingze, CAO Ruihao, YAN Chungang

2021, 41(9):  2473-2480.  DOI: 10.11772/j.issn.1001-9081.2020111872

Abstract ( )   PDF (1236KB) ( )  

References | Related Articles | Metrics

In real life, many data samples of systems can be easily obtained, but only a small part of accurate laabels can be obtained. In order to obtain a better classification learning model, a semi-supervised classification algorithm based on weight diversity was proposed by introducing semi-supervised learning and improving Unlabeled Data to Enhance Ensemble Diversity (UDEED), namely UDEED⁺. In UDEED⁺, based on the prediction disagreement of unlabeled data by base learners, the loss of weight diversity was proposed. The disagreement between base learners was represented by the cosine similarity of the weights of base learners. The diversity of model was fully expanded from different perspectives of loss function, and the unlabeled data were used to encourage the diversity representation of ensemble learners in the process of model training, so as to improve the performance and generalization of the classification learning model. Comparative experiments were conducted on 8 UCI public datasets with semi-supervised algorithms of UDEED algorithm, Safe Semi-Supervised Support Vector Machine (S4VM) and Semi-Supervised Weak-Label (SSWL). Compared with UDEED, UDEED⁺ has the accuracy and F1 score improved by 1.4 percentage points and 1.1 percentage points respectively; compared with S4VM, UDEED⁺ has the accuracy and F1 score improved by 1.3 percentage points and 3.1 percentage points respectively; compared with UDEED, UDEED⁺ has the accuracy and F1 score improved by 0.7 percentage points and 1.5 percentage points respectively. Experimental results illustrate that the increase of weight diversity can improve the classification performance of the model, verifying its positive effect on the improvement of the classification performance of UDEED⁺.

Extreme learning machine optimization based on hidden layer output matrix

SUN Haoyi, WANG Chuanmei, DING Yiming

2021, 41(9):  2481-2488.  DOI: 10.11772/j.issn.1001-9081.2020111791

Abstract ( )   PDF (1706KB) ( )  

References | Related Articles | Metrics

Aiming at the problem of the error existed from the hidden layer to the output layer of Extreme Learning Machine(ELM), it was found the analysis revealed that the error came from the process of solving the Moore-Penrose generalized inverse matrix H^† of the hidden layer output matrix H,that revaled the matrix H^† H was deviated from the identity matrix. The appropriate output matrix H was able to be selected according to the degree of deviation to obtain a smaller training error. According to the definitions of the generalized inverse matrix and auxiliary matrix,the target matrix H^† H and the error index L21-norm were firstly determined. Then,the experimental analysis showed that the L21-norm of H^† H was linearly related to the ELM error. Finally,Gaussian filtering was introduced to reduce the noise of the target matrix,which effectively reduced the L21-norm of the target matrix and the ELM error,achieving the purpose of optimizing the ELM algorithm.

Syntax-enhanced semantic parsing with syntax-aware representation

XIE Defeng, JI Jianmin

2021, 41(9):  2489-2495.  DOI: 10.11772/j.issn.1001-9081.2020111863

Abstract ( )   PDF (1042KB) ( )  

References | Related Articles | Metrics

Syntactic information, which is syntactic structure relations or dependency relations between words of a complete sentence, is an important and effective reference in Natural Language Processing (NLP). The task of semantic parsing is to directly transform natural language sentences into semantically complete and computer-executable languages. In previous semantic parsing studies, there are few efforts on improving the efficiency of end-to-end semantic parsing by using syntactic information from input sources. To further improve the accuracy and efficiency of the end-to-end semantic parsing model, a semantic parsing method was proposed to utilize the source-side dependency relation information of syntax to improve the model efficiency. As the basic idea of the method, an end-to-end dependency relation parser was pre-trained firstly. Then, the middle representation of the parser was used as syntax-aware representation, which was spliced with the original word embedding representation to generate a new input embedding representation, and this obtained input embedding representation was used in the end-to-end semantic parsing model. Finally, the model fusion was carried out by the transductive fusion learning. In the experiments, the proposed model was compared with the baseline model Transformer and the related works in the past decade. Experimental results show that, on ATIS, GEO and JOBS datasets, the semantic parsing model integrating dependency syntax-aware representation and transductive fusion learning achieves the best accuracy of 89.1%, 90.7%, and 91.4% respectively, which exceeds the performance of the Transformer. It verifies the effectiveness of introducing the dependency relation information of syntax.

Chinese description of image content based on fusion of image feature attention and adaptive attention

ZHAO Hong, KONG Dongyi

2021, 41(9):  2496-2503.  DOI: 10.11772/j.issn.1001-9081.2020111829

Abstract ( )   PDF (1520KB) ( )  

References | Related Articles | Metrics

Aiming at the problem that the existing Chinese description models of image content based on attention mechanism cannot focus on the key content without weakening or missing attention information, a Chinese description model of image content based on fusion of image feature attention and adaptive attention was proposed. An encode-decode structure was used in this model. Firstly, the image features were extracted in the encoder network, and the attention information of all feature regions of the image was extracted by the image feature attention. Then, the decoder network was used to decode the image features with attention weights to generate hidden information, so as to ensure that the attention information was not weakened or missed. Finally, the visual sentry module of self-adaptive attention was used to focus on the key content in the image features again, so that the main content of the image was able to be extracted more accurately. Several evaluation indices including BLEU, METEOR, ROUGEL and CIDEr were used to verify the models, the proposed model was compared with the image description models based on self-adaptive attention or image feature attention only, and the proposed model had the evaluation value of CIDEr improved by 10.1% and 7.8% respectively. Meanwhile, compared with the baseline model Neural Image Caption (NIC) and the Bottom-Up and Top-Down (BUTD) attention based image description model, the proposed model had the evaluation index value of CIDEr increased by 10.9% and 12.1% respectively. Experimental results show that the image understanding ability of the proposed model is effectively improved, and the score of each evaluation index of the model is better than those of the comparison models.

Multi-layer encoding and decoding model for image captioning based on attention mechanism

LI Kangkang, ZHANG Jing

2021, 41(9):  2504-2509.  DOI: 10.11772/j.issn.1001-9081.2020111838

Abstract ( )   PDF (1112KB) ( )  

References | Related Articles | Metrics

The task of image captioning is an important branch of image understanding. It requires not only the ability to correctly recognize the image content, but also the ability to generate grammatically and semantically correct sentences. The traditional encoder-decoder based model cannot make full use of image features and has only a single decoding method. In response to these problems, a multi-layer encoding and decoding model for image captioning based on attention mechanism named MLED was proposed. Firstly, Faster Region-based Convolutional Neural Network (Faster R-CNN) was used to extract image features. Then, Transformer was employed to extract three kinds of high-level features of the image. At the same time, the pyramid fusion method was used to effectively fuse the features. Finally, three Long Short-Term Memory (LSTM) Networks were constructed to decode the features of different layers hierarchically. In the decoding part, the soft attention mechanism was used to enable the model to pay attention to the important information required at the current step. The proposed model was tested on MSCOCO dataset and evaluated by BLEU, METEOR, ROUGE-L and CIDEr. Experimental results show that on the indicators BLEU-4, METEOR and CIDEr, the model is increased by 2.5 percentage points, 2.6 percentage points and 8.8 percentage points compared to the Recall what you see (Recall) model respectively, and is improved by 1.2 percentage points, 0.5 percentage points and 3.5 percentage points compared to the Hierarchical Attention-based Fusion (HAF) model respectively. The visualization of the generated description sentences show that the sentence generated by the proposed model can accurately reflect the image content.

Entity association query system based on enterprise knowledge graph construction

YU Dunhui, WAN Peng, WANG She

2021, 41(9):  2510-2516.  DOI: 10.11772/j.issn.1001-9081.2020111768

Abstract ( )   PDF (2446KB) ( )  

References | Related Articles | Metrics

Concerning the problem of low semantic relevance between nodes and low query efficiency in the current knowledge graph query, an entity-related query method was proposed,and then a knowledge gragh based enterprise query system was designed and implemented base on it. In this method, a four-layer filtering model was adopted. And firstly, the common paths of the target node were found through path search, so that the query nodes with a low degree of relevance were filtered out, and the filtering set was obtained. Then, the relevance degrees of the filtering set's attributes and relationships were calculated in the middle two layers, after that, the graph set filtering was performed based on the dynamic threshold. Finally, the entity relevance and relationship relevance scores was integrated and sorted to obtain the final query result. Experimental results on real enterprise data show that compared with traditional graph query algorithms such as Ness and NeMa, the proposed method reduces the query time by an average of 28.5%, and at the same time increases the filtering performance by an average of 29.6%, verifying that the algorithm can efficiently complete the task of query and display entities associated with the target.

Joint extraction method of entities and relations based on subject attention

LIU Yaxuan, ZHONG Yong

2021, 41(9):  2517-2522.  DOI: 10.11772/j.issn.1001-9081.2020111842

Abstract ( )   PDF (977KB) ( )  

References | Related Articles | Metrics

Extracting entities and relations is crucial for building large-scale knowledge graph and different knowledge extraction tasks. Based on the pre-trained language model, an entity-oriented joint extraction method combining subject attention was proposed. In this method, the key information of the subject was extracted by using Convolutional Neural Network (CNN) and the dependency relationship between the subject and the object was captured by the attention mechanism. Followed by the above, a Joint extraction model based on Subject Attention (JSA) was built. In experiments on public dataset New York Times corpus (NYT) and the dataset of artificial intelligence built by distant supervision, the F1 score of the proposed model was improved by 1.8 and 8.9 percentage points respectively compared with Cascade binary tagging framework for Relational triple extraction (CasRel).

Deep unsupervised discrete cross-modal hashing based on knowledge distillation

ZHANG Cheng, WAN Yuan, QIANG Haopeng

2021, 41(9):  2523-2531.  DOI: 10.11772/j.issn.1001-9081.2020111785

Abstract ( )   PDF (1705KB) ( )  

References | Related Articles | Metrics

Cross-modal hashing has attracted much attention due to its low storage cost and high retrieval efficiency. Most of the existing cross-modal hashing methods require the inter-instance association information provided by additional manual labels. However, the deep features learned by pre-trained deep unsupervised cross-modal hashing methods can also provide similar information. In addition, the discrete constraints are relaxed in the learning process of Hash codes, resulting in a large quantization loss. To solve the above two issues, a Deep Unsupervised Discrete Cross-modal Hashing (DUDCH) method based on knowledge distillation was proposed. Firstly, combined with the idea of knowledge transfer in knowledge distillation, the latent association information of the pre-trained unsupervised teacher model was used to reconstruct the symmetric similarity matrix, so as to replace the manual labels to help the supervised student method training. Secondly, the Discrete Cyclic Coordinate descent (DCC) was adopted to update the discrete Hash codes iteratively, thereby reducing the quantization loss between the real-value Hash codes learned by neural network and the discrete Hash codes. Finally, with the end-to-end neural network adopted as teacher model and the asymmetric neural network constructed as student model, the time complexity of the combination model was reduced. Experimental results on two commonly used benchmark datasets MIRFLICKR-25K and NUS-WIDE show that compared with Deep Joint-Semantics Reconstructing Hashing (DJSRH), the proposed method has the mean Average Precision (mAP) in image-to-text/text-to-image tasks increased by 2.83 percentage points/0.70 percentage points and 6.53 percentage points/3.95 percentage points averagely and respectively, proving its effectiveness in large-scale cross-modal retrieval.

Automatic patent price evaluation based on recurrent neural network

LIU Zichen, LI Xiaojuan, WEI Wei

2021, 41(9):  2532-2538.  DOI: 10.11772/j.issn.1001-9081.2020111887

Abstract ( )   PDF (1027KB) ( )  

References | Related Articles | Metrics

Patent price evaluation is an important part of intellectual property right transactions. When evaluating patent prices, the impact of the market, law, and technical dimensions on patent prices was not considered effectively by the existing methods. And the market factor of patent plays an important role in the evaluation of patent prices. Aiming at the above problem, an automatic patent price evaluation method based on recurrent neural network was proposed. In this method, based on the market approach, various other factors were considered comprehensively, and the Gated Recurrent Unit (GRU) neural network method was used to realize the automatic evaluation of patent prices. Example tests show that, with the qualitative evaluation results of experts as the benchmark, the average relative accuracy of the proposed method is 0.85. And this average relative accuracy of the proposed method is increased by 3.66%, 4.94% and 2.41% of the average relative accuracies of Analytic Hierarchy Process (AHP), rough set theory method and Back Propagation (BP) neural network method respectively.

Credit scoring model based on enhanced multi-dimensional and multi-grained cascade forest

BIAN Lingzhi, WANG Zhijie

2021, 41(9):  2539-2544.  DOI: 10.11772/j.issn.1001-9081.2020111796

Abstract ( )   PDF (1204KB) ( )  

References | Related Articles | Metrics

Credit risk is one of the main financial risks which commercial banks are faced with, while traditional credit scoring methods cannot effectively make use of the existing feature learning methods, resulting in low prediction accuracy. To solve this problem, an enhanced multi-dimensional and multi-grained cascade forest method was proposed to build credit scoring model, with the use of the idea of residual learning, the multi-dimensional and multi-grained cascade residual Forest (grcForest) model was built, which greatly increased the extracted features. Besides, the multi-dimensional multi-grained scanning was used to extract features of the raw data as many as possible, which improved the efficiency of feature extraction. The proposed model was compared with the existing statistical and machine learning methods on four credit scoring datasets, and evaluated by Area Under Curve (AUC) and accuracy. The AUC of the proposed model was 1.13% and 1.44% higher then that of the Light Gradient Boosting Machine (LightGBM) and the eXtreme Gradient Boosting (XGBoost). Experimental results show that the proposed model performs best in the prediction.

Trajectory prediction model of social network users based on self-supervised learning

DAI Yurou, YANG Qing, ZHANG Fengli, ZHOU Fan

2021, 41(9):  2545-2551.  DOI: 10.11772/j.issn.1001-9081.2020111859

Abstract ( )   PDF (1050KB) ( )  

References | Related Articles | Metrics

Aiming at the existing problems in user trajectory data modeling such as the sparsity of check-in points, long-term dependencies and complex moving patterns, a social network user trajectory prediction model based on self-supervised learning, called SeNext, was proposed to model and train the user trajectory to predict the next Point Of Interest (POI) of the user. First, data augmentation was utilized to expand the training trajectory samples, which solved the problem of the deficiency of model generalization capability caused by insufficient data and too few footprints of some users. Second, Recurrent Neural Network (RNN), Convolutional Neural Network (CNN) and attention mechanism were adopted into the modeling of current and historical trajectories respectively, so as to extract effective representations from high-dimensional sparse data to match the most similar moving patterns of users in the past. Finally, SeNext learned the implicit representations in the latent space by combining self-supervised learning and introducing contrastive loss Noise Contrastive Estimation (InfoNCE) to predict the next POI of the user. Experimental results show that compared to the state-of-the-artVariational Attention based Next (VANext)model, SeNext improves the prediction accuracy about 11% on Top@1.

Micro-expression recognition algorithm based on convolutional block attention module and dual path networks

NIU Ruihua, YANG Jun, XING Lanxin, WU Renbiao

2021, 41(9):  2552-2559.  DOI: 10.11772/j.issn.1001-9081.2020111743

Abstract ( )   PDF (1663KB) ( )  

References | Related Articles | Metrics

Micro-expression is a facial movement that humans make when they are trying to hide their true emotions. It has the typical characteristics of short duration and small amplitude. Concerning the problems of the difficulty in recognition and the unsatisfactory recognition effect of micro-expression, a micro-expression recognition algorithm based on Convolutional Block Attention Module (CBAM) and Dual Path Networks (DPN), namely CBAM-DPN, was proposed. Firstly, data fusion of typical micro-expression datasets was performed. Then, the change values of pixels in the sequence frames were analyzed to determine the position of the apex frame, after that, image enhancement was performed to the apex frame. Finally, based on the CBAM-DPN network, the features of the enhanced micro-expression apex frame was effectively extracted, and a classifier was constructed to recognize the micro-expression. The Unweighted F1-score (UF1) and Unweighted Average Recall (UAR) of the model after optimization can reach 0.720 3 and 0.729 3 respectively, which are improved by 0.048 9 and 0.037 9 respectively compared with those of the DPN model, and are improved by 0.068 3 and 0.078 7 respectively compared with those of the CapsuleNet model. Experimental results show that the CBAM-DPN algorithm combined with the advantages of CBAM and DPN can enhance the information extraction ability of small features, and effectively improve the performance of micro-expression recognition.

Improved ant colony optimization algorithm for path planning based on turning angle constraint

LI Kairong, LIU Shuang, HU Qianqian, TANG Yiyuan

2021, 41(9):  2560-2568.  DOI: 10.11772/j.issn.1001-9081.2020111713

Abstract ( )   PDF (1445KB) ( )  

References | Related Articles | Metrics

Concerning the problems that basic Ant Colony Optimization (ACO) is easy to fall into the local optimum, and has too long path and excessive turning angles during path search, an improved ACO algorithm based on turning angle constraint was proposed. Firstly, the initial pheromone concentration of the area between the starting point and the target point was enhanced to avoid the initial blind search. Then, the A^* algorithm's evaluation function and the turning angle constraint factor were added to the heuristic function. In this way, the node with the shortest path length and least number of turns was able to be selected at the next step. Finally, the distribution principle of wolf pack algorithm was introduced in the pheromone updating part to enhance the influence of high-quality population. At the same time, the Max and Min Ant System (MMAS) algorithm was used to limit the pheromone concentration to avoid the algorithm being trapped into the local optimum. Matlab simulation showed that compared with the traditional ACO, the improved algorithm was able to shorten the planned path length by 13.7%, reduce the number of turns by 64.3% and decrease the accumulated turning angle by 76.7%. Experimental results show that the improved ACO algorithm can effectively solve the global path planning problem and avoid the excessive energy loss of mobile robots.

Data science and technology

Impact and enhancement of similarity features on link prediction

CAI Biao, LI Ruicen, WU Yuanyuan

2021, 41(9):  2569-2577.  DOI: 10.11772/j.issn.1001-9081.2020111744

Abstract ( )   PDF (4634KB) ( )  

References | Related Articles | Metrics

Link prediction focuses on the design of prediction algorithms that can describe a given network mechanism more accurately to achieve the prediction result with higher accuracy. Based on an analysis of the existing research achievements, it is found that the similarity characteristics of a network has a great impact on the link prediction method used. In networks with low tag similarity between nodes, increasing the tag similarity is able to improve the prediction accuracy; in networks with high tag similarity between nodes, more attention should be paid to the contribution of structural information to link prediction to improve the prediction accuracy. Then, a tag-weighted similarity algorithm was proposed by weighting the tags, which was able to improve the accuracy of link prediction in networks with low similarity. Meanwhile, in networks with relatively high similarity, the structural information of the network was introduced into the node similarity calculation, and the accuracy of link prediction was improved through the preferential attachment mechanism. Experimental results on four real networks show that the proposed algorithm achieves the highest accuracy compared to the comparison algorithms Cosine Similarity between Tag Systems (CSTS), Preferential Attachment (PA), etc. According to the network similarity characteristics, using the proposed corresponding algorithm for link prediction can obtain more accurate prediction results.

Differential privacy high-dimensional data publishing method via clustering analysis

CHEN Hengheng, NI Zhiwei, ZHU Xuhui, JIN Yuanyuan, CHEN Qian

2021, 41(9):  2578-2585.  DOI: 10.11772/j.issn.1001-9081.2020111786

Abstract ( )   PDF (1281KB) ( )  

References | Related Articles | Metrics

Aiming at the problem that the existing differential privacy high-dimensional data publishing methods are difficult to take into account both the complex attribute correlation between data and computational cost, a differential privacy high-dimensional data publishing method based on clustering analysis technology, namely PrivBC, was proposed. Firstly, the attribute clustering method was designed based on the K-means++, the maximum information coefficient was introduced to quantify the correlation between the attributes, and the data attributes with high correlation were clustered. Secondly, for each data subset obtained by the clustering, the correlation matrix was calculated to reduce the candidate space of attribute pairs, and the Bayesian network satisfying differential privacy was constructed. Finally, each attribute was sampled according to the Bayesian networks, and a new private dataset was synthesized for publishing. Compared with PrivBayes method, PrivBC method had the misclassification rate and running time reduced by 12.6% and 30.2% averagely and respectively. Experimental results show that the proposed method can significantly improve the computational efficiency with ensuring the data availability, and provides a new idea for the private publishing of high-dimensional big data.

Parallel decompression algorithm for high-speed train monitoring data

WANG Zhoukai, ZHANG Jiong, MA Weigang, WANG Huaijun

2021, 41(9):  2586-2593.  DOI: 10.11772/j.issn.1001-9081.2020111173

Abstract ( )   PDF (1272KB) ( )  

References | Related Articles | Metrics

The real-time monitoring data generated by high-speed trains during running are usually processed by variable-length coding compression technology, which is convenient for transmission and storage. However, this method will complicate the internal structure of the compressed data, so that the corresponding data decompression process must follow the composition order of the compressed data, which is inefficient. In order to improve the decompression efficiency of high-speed train monitoring data, a parallel decompression algorithm for high-speed train monitoring data was proposed with the help of the speculation technology. Firstly, the structural characteristics of high-speed train monitoring data were studied, and the internal dependence that affects data division was analyzed. Secondly, the speculation technology was used to clean up internal dependence, and then, the data were divided into different parts tentatively. Thirdly, the division results were decompressed in a distributed computing environment in parallel. Finally, the parallel decompression results were combined together. Through this way, the decompression efficiency of high-speed train monitoring data was improved. Experimental results showed that on the computing cluster composed of 7 computing nodes, compared with the serial algorithm, the speedup of the proposed speculative parallel algorithm was about 3, showing a good performance of this algorithm. It can be seen that this algorithm can improve the monitoring data decompression efficiency significantly.

Cyber security

Reflective cross-site scripting vulnerability detection based on fuzzing test

NI Ping, CHEN Wei

2021, 41(9):  2594-2601.  DOI: 10.11772/j.issn.1001-9081.2020111770

Abstract ( )   PDF (1152KB) ( )  

References | Related Articles | Metrics

In view of the low efficiency, high false negative rate and high false positive rate of Cross-Site Scripting (XSS) vulnerability detection technology in current World Wide Web (WWW) applications, a reflective XSS vulnerability detection system based on fuzzing test was proposed. First, the Web crawler technology was used to crawl the Web page links with specified depth in the whole website and analyze them, so as to extract the potential user injection points. Secondly, a fuzzing test case was constructed according to the grammatical form of the attack payload, and an initial weights was set for each element, according to the injected probe vector, the output point type was obtained to select the corresponding attack grammatical form for constructing potential attack payload, and it was mutated to form a mutated attack payload as the request parameter. Thirdly, the website response was analyzed and the weights of the elements were adjusted to generate a more efficient attack payload. Finally, this proposed system was compared horizontally with OWASP Zed Attack Proxy (ZAP) and Wapiti systems. Experimental results show that the number of potential user injection points found by the proposed system is increased by more than 12.5%, the false positive rate of the system is dropped to 0.37%, and the false negative rate of the system is lower than 2.23%. At the same time, this system reduces the number of requests and saves the detection time.

Intrusion detection model based on semi-supervised learning and three-way decision

ZHANG Shipeng, LI Yongzhong, DU Xiangtong

2021, 41(9):  2602-2608.  DOI: 10.11772/j.issn.1001-9081.2020111883

Abstract ( )   PDF (936KB) ( )  

References | Related Articles | Metrics

Aiming at the situation that the existing intrusion detection models perform poorly on unknown attacks and have extremely limited labeled data, an intrusion detection model named SSL-3WD based on Semi-Supervised Learning (SSL) and Three-Way Decision (3WD) was proposed. In SSL-3WD model, the excellent performance of 3WD in the case of insufficient information was used to meet the assumption of sufficient redundancy of data information in SSL. Firstly, the 3WD theory was used to classify network behavior data, then some appropriate "pseudo-labeled" samples were selected according to the classification results to form a new training set to expand the original dataset. Finally, the classification process was repeated to obtain all the classifications of network behavior data. On the NSL-KDD dataset, the detection rate of the proposed model was 97.7%, which was 5.8 percentage points higher than that of the adaptive integrated learning intrusion detection model Multi-Tree, which has the highest detection rate in the comparison methods. On the UNSW-NB15 dataset, the accuracy of the proposed model reached 94.7% and the detection rate reached 96.3%, which were increased by 3.5 percentage points and 6.2 percentage points respectively compared with those of the best performing one in the comparison methods, the intrusion detection model based on Stack Nonsymmetric Deep Autoencoder (SNDAE). The experimental results show that the proposed SSL-3WD model improves the accuracy and detection rate of network behavior detection.

Detection method of domains generated by dictionary-based domain generation algorithm

ZHANG Yongbin, CHANG Wenxin, SUN Lianshan, ZHANG Hang

2021, 41(9):  2609-2614.  DOI: 10.11772/j.issn.1001-9081.2020111837

Abstract ( )   PDF (893KB) ( )  

References | Related Articles | Metrics

The composition of domain names generated by the dictionary-based Domain Generation Algorithm (DGA) is very similar to that of benign domain names and it is difficult to effectively detect them with the existing technology. To solve this problem, a detection model was proposed, namely CL (Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) network). The model includes three parts:character embedding layer, feature extraction layer and fully connected layer. Firstly, the characters of the input domain name were encoded by the character embedding layer. Then, the features of the domain name were extracted by connecting CNN and LSTM in serial way through the feature extraction layer. The n-grams features of the domain name were extracted by CNN and the extracted result were sent to LSTM to learn the context features between n-grams. Meanwhile, different combinations of CNNs and LSTMs were used to learn the features of n-grams with different lengths. Finally, the dictionary-based DGA domain names were classified and predicted by the fully connected layer according to the extracted features. Experimental results show that when the CNNs select the convolution kernel sizes of 3 and 4, the proposed model achives the best performance. In the four dictionary-based DGA family experiments, the accuracy of the CL model is improved by 2.20% compared with that of the CNN model. And with the increase of the number of sample families, the CL network model has a better stability.

Energy data access control method based on blockchain

GE Jihong, SHEN Tao

2021, 41(9):  2615-2622.  DOI: 10.11772/j.issn.1001-9081.2020111844

Abstract ( )   PDF (1175KB) ( )  

References | Related Articles | Metrics

In order to solve the problems of energy data tampering, leakage and data ownership disputes in the process of data sharing between enterprises and departments of energy internet, combined with the characteristics of blockchain-traceability and hard to be tampered with, an energy data access control method based on blockchain multi-chain architecture was proposed, which can protect user privacy and realize cross-enterprise and cross-department access control of energy data at the same time. In this method, the combination of supervision chain and multi-data-chain was used to protect the privacy of data and improve the scalability. The method of storing data on the chain and storing original data under the chain alleviated the storage pressure of the blockchain.By using the outsourcing supported multi-authority attribute-based encryption technology, the fine-grained access control of energy data was realized. Experimental simulation results show that in the proposed method, the blockchain network has availability, and outsourcing supported multi-authority attribute-based encryption technology has advantages in functionality and computing cost. Therefore, the proposed method can achieve fine-grained access control of energy data while protecting user privacy.

Audio encryption algorithm in fractional domain based on cascaded chaotic system

XU Liyun, YAN Tao, QIAN Yuhua

2021, 41(9):  2623-2630.  DOI: 10.11772/j.issn.1001-9081.2020122044

Abstract ( )   PDF (2308KB) ( )  

References | Related Articles | Metrics

In order to ensure the security of audio signals in communication transmission, a fractional domain audio encryption algorithm based on cascaded chaotic system was proposed. Firstly, the audio signal was grouped. Secondly, the chaotic system was used to obtain the orders of fractional Fourier transform, and the order corresponding to each group data changed dynamically. Thirdly, the sampling fractional Fourier discrete transform with less computational complexity was used to obtain the fractional domain spectrum data of each group. Finally, the cascaded chaotic system was used to perform data encryption to the fractional domain of each group in turn, so as to realize the overall encryption of the audio signals. Experimental results show that the proposed algorithm is extremely sensitive to the key, and has the waveform and fractional domain spectrum of obtained encrypted signal more uniformly distributed and less correlated compared with those of the original signal. At the same time, compared with the frequency domain encryption and fixed-order fractional domain encryption methods, the proposed algorithm can effectively increase the key space while reducing the computational complexity. It can be seen that the proposed algorithm can satisfy the real-time and secure transmission requirements of audio signals effectively.

Advanced computing

Dynamic mapping method for heterogeneous multi-core system under thermal safety constraint

AN Xin, YANG Haijiao, LI Jianhua, REN Fuji

2021, 41(9):  2631-2638.  DOI: 10.11772/j.issn.1001-9081.2020111870

Abstract ( )   PDF (1107KB) ( )  

References | Related Articles | Metrics

The heterogeneous multi-core platform provides flexibility for system design by integrating different types of processing cores, so that applications can dynamically select different types of processing cores according to their requirements and realize efficient operation of applications. With the development of semiconductor technology, the number of integrated cores on a single chip has increased, making the modern multi-core processors have a higher power density, and this will cause the chip temperature to rise, which will eventually cause a certain negative impact on the system performance. To make the performance advantages of heterogeneous multi-core processing system fully utilized, a dynamic mapping method was proposed to maximize the performance of heterogeneous multi-core systems under the premise of satisfying temperature safe power. In this method, two heterogeneous indices of heterogeneous multi-core systems including core type and thermal susceptibility were considered to determine the mapping scheme:the first heterogeneous index is the core type. Different types of processing cores have different characteristics, so they are suitable for processing different applications. The second heterogeneous index is thermal susceptibility. Different processing core positions on the chip have different thermal susceptibility. The processing cores closer to the center receive more heat transfer from other processing cores, so that they have higher temperature. For the above, a neural network performance predictor was created to match threads to processing core types, and the Thermal Safe Power (TSP) model was used to map the matched threads to specific locations on the chip. Experimental results show that the proposed method achieves about 53% increase of the average number of instructions executed by the program in each clock cycle-Instruction Per Cycle (IPC) under the premise of ensuring thermal safety constraints compared with the common Round Robin Scheduler (RRS).

Computation offloading policy for machine learning in mobile edge computing environments

GUO Mian, ZHANG Jinyou

2021, 41(9):  2639-2645.  DOI: 10.11772/j.issn.1001-9081.2020111734

Abstract ( )   PDF (1127KB) ( )  

References | Related Articles | Metrics

Concerning the challenges of the diversity of data sources, non-independent and identical distribution of data and the heterogeneity of both computing capabilities and energy consumption of edge devices in Internet of Things (IoT), a computation offloading policy in Mobile Edge Computing (MEC) network that deploys both centralized learning and federated learning was proposed. Firstly, a system model of computation offloading related to both centralized learning and federated learning was built, considering the network transmission delay, computation delay and energy consumption of centralized learning and federated learning models. Then, with the system delay minimization as optimization object, considering the constraints of energy consumption and the training times based on machine learning accuracy, a computation offloading optimization model for machine learning was constructed. After that, the game for this computation offloading was formulated and analyzed. Based on the analysis results, an Energy-Constrained Delay-Greedy (ECDG) algorithm was proposed, which found the optimal solutions for the model via a two-stage policy of greedy decision and energy-constrained decision updating. Compared to the centralized-greedy and Federated Learning with Client Selection (FedCS) algorithms, ECDG algorithm has the lowest average learning delay, which is 1/10 of that in the centralized-greedy algorithm, and 1/5 of that in the FedCS algorithm. The experimental results show that, ECDG algorithms can automatically select the optimal machine learning models by computation offloading so that it can efficiently reduce the average machine learning delay, improve the energy efficiency of edge devices and satisfy the Quality of Service (QoS) requirements of IoT applications.

Relay computation and dynamic diversion of computing-intensive large flow data

LIAO Jia, CHEN Yang, BAO Qiulan, LIAO Xuehua, ZHU Zhousen

2021, 41(9):  2646-2651.  DOI: 10.11772/j.issn.1001-9081.2020111725

Abstract ( )   PDF (1199KB) ( )  

References | Related Articles | Metrics

In view of the problems such as the slow computation of large flow data, the high computation pressure on the server, a set of relay computation and dynamic diversion model of computing-intensive large flow data was proposed. Firstly, in the distributed environment, the in-memory data storage technology was used to determine the computation amounts and complexity levels of the computation tasks. At the same time, the nodes were sorted by the node resource capacity, and the tasks were dynamically allocated to different nodes for parallel computing. Meanwhile, the computation tasks were decomposed by a relay processing mode, so as to guarantee the performance and accuracy requirements of high flow complex computing tasks. Through analysis and comparison, it can be seen that the running time of multiple nodes is shorter than that of the single node, and the computation speed of multiple nodes is faster than that of the single node when dealing with data volume of more than 10 000 levels. At the same time, when the model is applied in practice, it can be seen that the model can not only reduce the running time in high concurrency scenarios but also save more computing resources.

Loop-level speculative parallelism analysis of kernel program in TACLeBench

MENG Huiling, WANG Yaobin, LI Ling, YANG Yang, WANG Xinyi, LIU Zhiqin

2021, 41(9):  2652-2657.  DOI: 10.11772/j.issn.1001-9081.2020111792

Abstract ( )   PDF (1190KB) ( )  

References | Related Articles | Metrics

Thread-Level Speculation (TLS) technology can tap the parallel execution potential of programs and improve the utilization of multi-core resources. However, the current TACLeBench kernel benchmarks are not effectively analyzed in TLS parallelization. In response to this problem, the loop-level speculative execution analysis scheme and analysis tool were designed. With 7 representative TACLeBench kernel benchmarks selected, firstly, the initialization analysis was performed to the programs, the program hot fragments were selected to insert the loop identifier. Then, the cross-compilation was performed to these fragments, the program speculative thread and the memory address related data were recorded, and the maximun potential of the loop-level parallelism was analyzed. Finally, the program runtime characteristics (thread granularity, parallelizable coverage, dependency characteristics) and the impacts of the source code on the speedup ratio were comprehensively discussed. Experimental results show that:1) this type of programs is suitable for TLS acceleration, compared with serial execution results, under the loop structure speculative execution, the speedup ratios for most programs are above 2, and the highest speedup ratio in them can reach 20.79; 2) by using TLS to accelerate the TACLeBench kernel programs, most applications can effectively make use of 4-core to 16-core computing resources.

Improved feature selection and classification algorithm for gene expression programming based on layer distance

ZHAN Hang, HE Lang, HUANG Zhangcan, LI Huafeng, ZHANG Qiang, TAN Qing

2021, 41(9):  2658-2667.  DOI: 10.11772/j.issn.1001-9081.2020111801

Abstract ( )   PDF (1220KB) ( )  

References | Related Articles | Metrics

Concerning the problem that the interpretable mapping relationship between data features and data categories do not be revealed by general feature selection algorithms. on the basis of Gene Expression Programming (GEP),by introducing the initialization methods, mutation strategies and fitness evaluation methods,an improved Feature Selection classification algorithm based on Layer Distance for GEP(FSLDGEP) was proposed. Firstly,the selection probability was defined to initialize the individuals in the population directionally, so as to increase the number of effective individuals in the population. Secondly, the layer neighborhood of the individual was proposed, so that each individual in the population would mutate based on its layer neighborhood, and the blind and unguided problem in the process of mutation was solved。Finally, the dimension reduction rate and classification accuracy were combined as the fitness value of the individual, which changed the population evolutionary mode of single optimization goal and balanced the relationship between the above two. The 5-fold and 10-fold verifications were performed on 7 datasets, the functional mapping relationship between data features and their categories was given by the proposed algorithm, and the obtained mapping function was used for data classification. Compared with Feature Selection based on Forest Optimization Algorithm (FSFOA), feature evaluation and selection based on Neighborhood Soft Margin (NSM), Feature Selection based on Neighborhood Effective Information Ratio (FS-NEIR)and other comparison algorithms, the proposed algorithm has obtained the best results of the dimension reduction rate on Hepatitis, Wisconsin Prognostic Breast Cancer (WPBC), Sonar and Wisconsin Diagnostic Breast Cancer (WDBC) datasets, and has the best average classification accuracy on Hepatitis, Ionosphere, Musk1, WPBC, Heart-Statlog and WDBC datasets. Experimental results shows that the feasibility, effectiveness and superiority of the proposed algorithm in feature selection and classification are verified.

Improved butterfly optimization algorithm based on cosine similarity

CHEN Jun, HE Qing

2021, 41(9):  2668-2677.  DOI: 10.11772/j.issn.1001-9081.2020111776

Abstract ( )   PDF (1469KB) ( )  

References | Related Articles | Metrics

Aiming at the problems that Butterfly Optimization Algorithm (BOA) tends to fall into local optimum and has poor convergence, a Multi-Strategy Improved BOA (MSBOA) was proposed. Firstly, the cosine similarity position adjustment strategy was introduced to the algorithm, rotation transformation operator and scaling transformation operator were used to update the positions, so as to effectively maintain the population diversity of the algorithm. Secondly, dynamic switching probability was introduced to balance the transformation between the local phase and the global phase of the algorithm. Finally, a hybrid inertia weight strategy was added to accelerate convergence. Solving 16 benchmark test functions, as well as the Wilcoxon rank-sum test and CEC2014 test functions were to verify, the effectiveness and robustness of the proposed algorithm. Experimental results show that compared with BOA, some BOAs with different improvement strategies and some swarm intelligence algorithms, MSBOA has significant improvement in convergence accuracy and convergence speed.

Network and communications

Local differential privacy protection mechanism for mobile crowd sensing with edge computing

LI Zhuo, SONG Zihui, SHEN Xin, CHEN Xin

2021, 41(9):  2678-2686.  DOI: 10.11772/j.issn.1001-9081.2020111787

Abstract ( )   PDF (1255KB) ( )  

References | Related Articles | Metrics

Aiming at the problem of the difficulty in privacy protection and the cost increase caused by privacy protection in the user data submission stage in Mobile Crowd Sensing (MCS), CS-MVP algorithm for joint privacy protection and CS-MAP algorithm for independent privacy protection of the attributes of user submitted data were designed based on the principle of Local Differential Privacy (LDP). Firstly, the user submitted privacy model and the task data availability model were constructed on the basis of the attribute relationships. And CS-MVP algorithm and CS-MAP algorithm were used to solve the availability maximization problem under the privacy constraint. At the same time, in the edge computing supported MCS scenarios, the three-layer architecture for MCS under privacy protection of the user submitted data was constructed. Theoretical analysis proves the optimality of the two algorithms under the data attribute joint privacy constraint and data attribute independent privacy constraint respectively. Experimental results show that under the same privacy budget and amount of data, compared with LoPub and PrivKV, the accuracy of user submitted data recovered to correct sensor data based on CS-MVP algorithm and CS-MAP algorithm is improved by 26.94%, 84.34% and 66.24%, 144.14% respectively.

Deep learning-based joint channel estimation and equalization algorithm for C-V2X communications

CHEN Chengrui, SUN Ning, HE Shibiao, LIAO Yong

2021, 41(9):  2687-2693.  DOI: 10.11772/j.issn.1001-9081.2020111779

Abstract ( )   PDF (1086KB) ( )  

References | Related Articles | Metrics

In order to effectively improve the Bit Error Rate (BER) performance of communication system without significantly increasing the computational complexity, a deep learning based joint channel estimation and equalization algorithm named V-EstEqNet was proposed for Cellular-Vehicle to Everything (C-V2X) communication system by using the powerful ability of deep learning in data processing. Different from the traditional algorithms, in which channel estimation and equalization in the communication system reciever were carried out in two stages respectively, V-EstEqNet considered them jointly, and used the deep learning network to directly correct and restore the received data, so that the channel equalization was completed without explicit channel estimation. Specifically, a large number of received data were used to train the network offline, so that the channel characteristics superimposed on the received data were learned by the network, and then these characteristics were utilized to recover the original transmitted data. Simulation results show that the proposed algorithm can track channel characteristics more effectively in different speed scenarios. At the same time, compared with the traditional channel estimation algorithms (Least Squares (LS) and Linear Minimum Mean Square Error (LMMSE)) combining with the traditional channel equalization algorithms (Zero Forcing (ZF) equalization algorithm and Minimum Mean Square Error (MMSE) equalization algorithm), the proposed algorithm has a maximum BER gain of 6 dB in low-speed environment and 9 dB in high-speed environment.

Phase shift model design for 6G reconfigurable intelligent surface

WANG Dan, LIANG Jiamin, LIU Jinzhi, ZHANG Youshou

2021, 41(9):  2694-2698.  DOI: 10.11772/j.issn.1001-9081.2020111836

Abstract ( )   PDF (808KB) ( )  

References | Related Articles | Metrics

In order to solve the problem of high energy consumption of relay communication and high difficulty in the construction of 5G base stations, the research on Reconfigurable Intelligent Surface (RIS) technology was introduced in 6G mobile communication. Aiming at the problem of characteristic loss and instability of the truncated Hadamard matrix and Discrete Fourier Transform (DFT) matrix when constructing intelligent surfaces, a new RIS phase shift model design scheme of constructing unitary matrix based on Hankel matrix and Toeplitz matrix was proposed. The characteristics of the unitary matrix were used to minimize the channel error and improve the reliability of the communication channel. The simulation results show that compared with that of non-RIS-assisted communication, the user receiving rate of RIS-assisted communication can obtain a gain of 1 (bit·s^-1)/Hz when the number of RIS units is 15. With the increase of the number of RIS units, the gain of the user receiving rate will be more and more significant. When the number of RIS units is 4, compared to the method of using DFT matrix to construct intelligent reflecting surfaces, the methods of using the two obtained unitary matrices to construct reflecting surfaces have higher reliability, and can obtain the performance gain of about 0.5 dB.

Time-varying channel estimation method based on sliding window filtering and polynomial fitting

JING Xinghong, SUN Guodong, HE Shibiao, LIAO Yong

2021, 41(9):  2699-2704.  DOI: 10.11772/j.issn.1001-9081.2020122035

Abstract ( )   PDF (912KB) ( )  

References | Related Articles | Metrics

The Long Term Evolution based Vehicle to Everything (LTE-V2X) standard follows the LTE standard's frame format and uses a block-type pilot assisted Single-Carrier Frequency-Division Multiple Access (SC-FDMA) system for channel estimation. However, due to the time-varying characteristics of the V2X channel, large technical challenges are brought to the channel estimation at the receiver. Therefore, a time-varying channel estimation method based on sliding window filtering and polynomial fitting was designed. Aiming at the noise problem at pilot symbols, based on Least Squares (LS) method, an adaptive-length sliding window filtering was adopted for noise reduction, so as to ensure the channel estimation accuracy of pilot symbols. Furthermore, according to the size of the Doppler frequency shift of data symbols, an adaptive-order polynomial fitting method was designed to track the channel changes at data symbols. The simulation results show that the proposed method has a good denoising effect based on LS method. In the case of low-speed movement, the estimation accuracy of the proposed method is between those of LS method and Linear Minimum Mean Square Error (LMMSE) method. In the case of high-speed movement, the proposed method can fit the time-varying channel characteristics better, and its performance exceeds that of the channel estimation method of LMMSE method combined with linear interpolation. The above results show that the proposed method has better adaptability than the comparison methods and is suitable for LTE-V2X communication scenarios with different channel noises and terminal moving speeds.

Multimedia computing and computer simulation

Object tracking algorithm of fully-convolutional Siamese networks with rotation and scale estimation

JI Zhangjian, REN Xingwang

2021, 41(9):  2705-2711.  DOI: 10.11772/j.issn.1001-9081.2020111805

Abstract ( )   PDF (2425KB) ( )  

References | Related Articles | Metrics

In the object tracking task, Fully-Convolutional Siamese networks (SiamFC) tracking method has problems of tracking errors or inaccurate tracking results caused by the rotation and scale variation of objects. Therefore, a SiamFC tracking algorithm with rotation and scale estimation was proposed, which consists of location module and rotation-scale estimation module. Firstly, in the location module, the tracking position was obtained by using SiamFC algorithm, and this position was adjusted by combining the rotation and scale information. Then, in the rotation-scale estimation module, as the image rotation and scale variation were converted into translational motions in log-polar coordinate system, the object search area was transformed from Cartesian coordinate system to log-polar coordinate system, so that the scale and rotation angle of the object were estimated by using correlation filtering technology. Finally, an object tracking model which can simultaneously estimate object position, rotation angle and scale variation was obtained. In the comparison experiments, the proposed algorithm had the success rate and accuracy of 57.7% and 81.4% averagely on Visual Tracker Benchmark 2015 (OTB2015) dataset, and had the success rate and accuracy of 51.8% and 53.3% averagely on Planar Object Tracking in the wild (POT) dataset with object rotation and scale variation. Compared with the success rate and accuracy of SiamFC algorithm, those of the proposed algorithm were increased by 13.5 percentage points and 13.4 percentage points averagely. Experimental results verify that the proposed algorithm can effectively solve the tracking challenges caused by object rotation and scale variation.

General object detection framework based on improved Faster R-CNN

MA Jialiang, CHEN Bin, SUN Xiaofei

2021, 41(9):  2712-2719.  DOI: 10.11772/j.issn.1001-9081.2020111852

Abstract ( )   PDF (2181KB) ( )  

References | Related Articles | Metrics

Aiming at the problem that current detectors based on deep learning cannot effectively detect objects with irregular shapes or large differences between length and width, based on the traditional Faster Region-based Convolutional Neural Network (Faster R-CNN) algorithm, an improved two-stage object detection framework named Accurate R-CNN was proposed. First of all, a novel Intersection over Union (IoU) metric-Effective Intersection over Union (EIoU) was proposed to reduce the proportion of redundant bounding boxes in the training data by using the centrality weight. Then, a context related Feature Reassignment Module (FRM) was proposed to re-encode the features by the remote dependency and local context information of objects, so as to make up for the loss of shape information in the pooling process. Experimental results show that on the Microsoft Common Objects in COntext (MS COCO) dataset, for the bounding box detection task, when using Residual Networks (ResNets) with two different depths of 50 and 101 as the backbone networks, Accurate R-CNN has the Average Precision (AP) improvements of 1.7 percentage points and 1.1 percentage points respectively compared to the baseline model Faster R-CNN, which are significantly than those of the detectors based on mask with the same backbone networks. After adding mask branch, for the instance segmentation task, when ResNets with two different depths are used as the backbone networks, the mask Average Precisions of Accurate R-CNN are increased by 1.2 percentage points and 1.1 percentage points respectively compared with Mask Region-based Convolutional Neural Network (Mask R-CNN). The research results illustrate that compared to the baseline model, Accurate R-CNN achieves better performance on different datasets and different tasks.

Indoor scene recognition method combined with object detection

XU Jianglang, LI Linyan, WAN Xinjun, HU Fuyuan

2021, 41(9):  2720-2725.  DOI: 10.11772/j.issn.1001-9081.2020111815

Abstract ( )   PDF (1357KB) ( )  

References | Related Articles | Metrics

In the method of combining Object detection Network (ObjectNet) and scene recognition network, the object features extracted by the ObjectNet and the scene features extracted by the scene network are inconsistent in dimensionality and property, and there is redundant information in the object features that affects the scene judgment, resulting in low recognition accuracy of scenes. To solve this problem, an improved indoor scene recognition method combined with object detection was proposed. First, the Class Conversion Matrix (CCM) was introduced into the ObjectNet to convert the object features output by ObjectNet, so that the dimension of the object features was consistent with that of the scene features, as a result, the information loss caused by inconsistency of the feature dimensions was reduced. Then, the Context Gating (CG) mechanism was used to suppress the redundant information in the features, reducing the weight of irrelevant information, and increasing the contribution of object features in scene recognition. The recognition accuracy of the proposed method on MIT Indoor67 dataset reaches 90.28%, which is 0.77 percentage points higher than that of Spatial-layout-maintained Object Semantics Features (SOSF) method; and the recognition accuracy of the proposed method on SUN397 dataset is 81.15%, which is 1.49 percentage points higher than that of Hierarchy of Alternating Specialists (HoAS) method. Experimental results show that the proposed method improves the accuracy of indoor scene recognition.

Remote sensing scene classification based on bidirectional gated scale feature fusion

SONG Zhongshan, LIANG Jiarui, ZHENG Lu, LIU Zhenyu, TIE Jun

2021, 41(9):  2726-2735.  DOI: 10.11772/j.issn.1001-9081.2020111778

Abstract ( )   PDF (3143KB) ( )  

References | Related Articles | Metrics

There are large differences in shape, texture and color of images in remote sensing image datasets, and the classification accuracy of remote sensing scenes is low due to the scale differences cased by different shooting heights and angles. Therefore, a Feature Aggregation Compensation Convolution Neural Network (FAC-CNN) was proposed, which used active rotation aggregation to fuse features of different scales and improved the complementarity between bottom features and top features through bidirectional gated method. In the network, the image pyramid was used to generate images of different scales and input them into the branch network to extract multi-scale features, and the active rotation aggregation method was proposed to fuse features of different scales, so that the fused features have directional information, which improved the generalization ability of the model to different scale inputs and different rotation inputs, and improved the classification accuracy of the model. On NorthWestern Polytechnical University REmote Sensing Image Scene Classification (NWPU-RESISC) dataset, the accuracy of FAC-CNN was increased by 2.05 percentage points and 2.69 percentage points respectively compared to those of Attention Recurrent Convolutional Network based on VGGNet (ARCNet-VGGNet) and Gated Bidirectional Network (GBNet); and on Aerial Image Dataset (AID), the accuracy of FAC-CNN was increased by 3.24 percentage points and 0.86 percentage points respectively compared to those of the two comparison networks. Experimental results show that FAC-CNN can effectively solve the problems in remote sensing image datasets and improve the accuracy of remote sensing scene classification.

3D point cloud face recognition based on deep learning

GAO Gong, YANG Hongyu, LIU Hong

2021, 41(9):  2736-2740.  DOI: 10.11772/j.issn.1001-9081.2020111826

Abstract ( )   PDF (1375KB) ( )  

References | Related Articles | Metrics

In order to enhance the robustness of the 3D point cloud face recognition system for multiple expressions and multiple poses, a deep learning-based point cloud feature extraction network was proposed, namely ResPoint. The modules such as grouping, sampling and local feature extraction (ResConv) were used in the ResPoint network, and skip connection was used in ResConv module, so that the proposed network had good recognition results for sparse point cloud. Firstly, the nose tip point was located by the geometric feature points of the face, and the face area was cut with this point as the center. The obtained area had noisy points and holes, so Gaussian filtering and 3D cubic interpolation were performed to it. Secondly, the ResPoint network was used to extract features of the preprocessed point cloud data. Finally, the features were combined in the fully connected layer to realize the classification of 3D faces. In the experiments on CASIA 3D face database, the recognition accuracy of the ResPoint network is increased by 5.06% compared with that of the Relation-Shape Convolutional Neural Network (RS-CNN). Experimental results show that the ResPoint network increases the depth of the network while using different convolution kernels to extract features, so that the ResPoint network has better feature extraction capability.

Facial landmark detection based on ResNeXt with asymmetric convolution and squeeze excitation

WANG Hebing, ZHANG Chunmei

2021, 41(9):  2741-2747.  DOI: 10.11772/j.issn.1001-9081.2020111847

Abstract ( )   PDF (2305KB) ( )  

References | Related Articles | Metrics

Cascaded Deep Convolutional Neural Network (DCNN) algorithm is the first model that uses Convolutional Neural Network (CNN) in facial landmark detection and the use of CNN improves the accuracy significantly. This strategy needs to perform regression processing to the data between the adjacent stages repeatedly, resulting in complex algorithm procedure. Therefore, a facial landmark detection algorithm based on Asymmetric Convolution-Squeeze Excitation-Next Residual Network (AC-SE-ResNeXt) was proposed with only single-stage regression to simplify the procedure and solve the non-real-time problem of data preprocessing between adjacent stages. In order to keep the accuracy, the Asymmetric Convolution (AC) module and the Squeeze-and-Excitation (SE) module were added to Next Residual Network (ResNeXt) block to construct the AC-SE-ResNeXt network model. At the same time, in order to fit faces in complex environments such as different illuminations, postures and expressions better, the AC-SE-ResNeXt network model was deepened to 101 layers. The trained model was tested on datasets BioID and LFPW respectively. The overall mean error rate of the model for the five-point facial landmark detection on BioID dataset was 1.99%, and the overall mean error rate of the model for the five-point facial landmark detection on LFPW dataset was 2.3%. Experimental results show that with the simplified algorithm procedure and end to end processing, the improved algorithm can keep the accuracy as cascaded DCNN algorithm, while has the robustness significantly increased.

Frontier and comprehensive applications

Consensus of time-varying multi-agent systems based on event-triggered impulsive control

CHAI Jie, GUO Liuxiao, SHEN Wanqiang, CHEN Jing

2021, 41(9):  2748-2753.  DOI: 10.11772/j.issn.1001-9081.2020111843

Abstract ( )   PDF (903KB) ( )  

References | Related Articles | Metrics

For the consensus problem of time-varying multi-agent systems under time-varying topology connection, an event-triggered impulsive control protocol was proposed. In this protocol, for each agent, the controller would be updated only when the related state error exceeded a threshold, and the control inputs would be carried out only at the event triggering instants, and continuous communication between agents was avoided. This protocol would greatly reduce the cost of communication and control for network consensus. The sufficient conditions for the multi-agent systems with time-varying characteristics to achieve consensus under event-triggered impulsive control were analyzed based on the algebraic graph theory, Lyapunov stability and impulsive differential equation. At the same time, it was proved theoretically that there was no Zeno behavior in the event-triggered time sequences. Finally, the effectiveness of the obtained theoretical conclusion was verified through several numerical simulations.

Co-evolutionary simulation regarding emergency logistics in major public health risk governance

GONG Ying, HE Yanting, CAO Cejun

2021, 41(9):  2754-2760.  DOI: 10.11772/j.issn.1001-9081.2020111728

Abstract ( )   PDF (1077KB) ( )  

References | Related Articles | Metrics

To enhance the efficiency of emergency logistics during the process of major public health risk governance, an efficient emergency logistics collaboration mechanism was designed based on the analysis of the behavioral characteristics of government and logistics enterprise. An evolutionary game model between local government and logistics enterprise was established to investigate the evolutional laws and paths of local government's supervision and logistics enterprise's collaboration. Then, the feasibility and effectiveness of the proposed model were verified based on the numerical simulation. The results indicate that the emergency logistics collaboration mechanism in major public health risk governance significantly depends on local government's supervision compared with the collaboration mechanism of commercial logistics and it makes the collaboration level of logistics enterprise fluctuate between 0.25 and 0.9 repeatedly. After the establishment of a dynamic reward and punishment mechanism for local government, the obtained collaboration level of logistics enterprise stabilizes at 0.46 when the number of games reaches 30. It can be seen that this dynamic reward and punishment mechanism improves the stability of emergency logistics collaboration mechanism significantly.

Real-time fall detection method based on threshold and extremely randomized tree

LIU Xiaoguang, JIN Shaokang, WEI Zihui, LIANG Tie, WANG Hongrui, LIU Xiuling

2021, 41(9):  2761-2766.  DOI: 10.11772/j.issn.1001-9081.2020111816

Abstract ( )   PDF (1152KB) ( )  

References | Related Articles | Metrics

Aiming at the problem that wearable device-based fall detection cannot have good accuracy real-timely, a real-time fall detection method based on the fusion of threshold and extremely randomized tree was proposed. In this method, the wearable devices only needed to calculate the threshold value and did not need to ensure the accuracy of fall detection, which reduced the amount of calculation; at the same time, the host computer used the extremely randomized tree algorithm to ensure the accuracy of fall detection. Most of the daily actions were filtered by the wearable devices through the threshold method, so as to reduce the amount of action data detected by the host computer. In this way, the proposed method had high accuracy of fall detection in real time. In addition, in order to reduce the false positive rate of fall detection, the attitude angle sensor and the pressure sensor were integrated into the wearable devices, and the feedback mechanism was added to the host computer. When the detection result was false positive, the wrong detected sample was added to the non-fall dataset for retraining through the host computer. Through this kind of continuous learning, the model would generate an alarm model suitable for the individual. And this feedback mechanism provided a new idea for reducing the false positive rate of fall detection. Experimental results show that in 1 259 test samples, the proposed method has an average accuracy of 99.7% and the lowest false positive rate of 0.08%.

Degree centrality based method for cognitive feature selection

ZHANG Xiaofei, YANG Yang, HUANG Jiajin, ZHONG Ning

2021, 41(9):  2767-2772.  DOI: 10.11772/j.issn.1001-9081.2020111794

Abstract ( )   PDF (2920KB) ( )  

References | Related Articles | Metrics

To address the uncertainty of cognitive feature selection in brain atlas, a Degree Centrality based Cognitive Feature Selection Method (DC-CFSM) was proposed. First, the Functional Brain Network (FBN) of the subjects in the cognitive experiment tasks was constructed based on the brain atlas, and the Degree Centrality (DC) of each Region Of Interest (ROI) of the FBN was calculated. Next, the difference significances of the subjects' same cortical ROI under different cognitive states during executing cognitive task were statistically compared and ranked. Finally, the Human Brain Cognitive Architecture-Area Under Curve (HBCA-AUC) values were calculated for the ranked regions of interest, and the performances of several cognitive feature selection methods were evaluated. In the experiments on functional Magnetic Resonance Imaging (fMRI) data of mental arithmetic cognitive tasks, the values of HBCA-AUC obtained by DC-CFSM on the Task Positive System (TPS), Task Negative System (TNS), and Task Support System (TSS) of the human brain cognitive architecture were 0.669 2, 0.304 0 and 0.468 5 respectively. Compared with Extremely randomized Trees (Extra Trees), Adaptive Boosting (AdaBoost), random forest, and eXtreme Gradient Boosting (XGB), the recognition rate for TPS of DC-CFSM was increased by 22.17%, 13.90%, 24.32% and 37.19% respectively, while its misrecognition rate for TNS was reduced by 20.46%, 29.70%, 44.96% and 33.39% respectively. DC-CFSM can better reflect the categories and functions of the human brain cognitive system in the selection of cognitive features of brain atlas.

De novo peptide sequencing by tandem mass spectrometry based on graph convolutional neural network

MOU Changning, WANG Haipeng, ZHOU Piyu, HOU Xinhang

2021, 41(9):  2773-2779.  DOI: 10.11772/j.issn.1001-9081.2020111875

Abstract ( )   PDF (11373KB) ( )  

References | Related Articles | Metrics

In proteomics, de novo sequencing is one of the most important methods for peptide sequencing by tandem mass spectrometry. It has the advantage of being independent on any protein databases and plays a key role in the determination of protein sequences of unknown species, monoclonal antibodies sequencing and other fields. However, due to its complexity, the accuracy of de novo sequencing is much lower than that of the database search methods, therefore the wide application of de novo sequencing is limited. Focused on the issue of low accuracy of de novo sequencing, denovo-GCN, a de novo sequencing method based on Graph Convolutional neural Network (GCN) was proposed. In this method, the relationships between peaks in mass spectrometry were expressed by using graph structure, and the peak features were extracted from each corresponding peptide cleavage site. Then the amino acid type at the current cleavage site was predicted by GCN, and finally a complete sequence was formed step by step. Three significant parameters affecting the model were experimentally determined, including the GCN model layer number, the combination of ion types and the number of spectral peaks used for sequencing, and datasets of a wide variety of species were used for experimental comparison. Experimental results show that, the peptide-level recall of denovo-GCN is 4.0 percentage points to 21.1 percentage points higher than those of the graph theory-based methods Novor and pNovo, and is 2.1 percentage points to 10.7 percentage points higher than that of DeepNovo, which adopts Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) network.

Cattle eye image feature extraction method based on improved DenseNet

ZHENG Zhiqiang, HU Xin, WENG Zhi, WANG Yuhe, CHENG Xi

2021, 41(9):  2780-2784.  DOI: 10.11772/j.issn.1001-9081.2020101533

Abstract ( )   PDF (1024KB) ( )  

References | Related Articles | Metrics

To address the problem of low recognition accuracy caused by vanishing gradient and overfitting in the cattle eye image feature extraction process, an improved DenseNet based cattle eye image feature extraction method was proposed. Firstly, the Scaled exponential Linear Unit (SeLU) activation function was used to prevent the vanishing gradient of the network. Secondly, the feature blocks of cattle eye images were randomly discarded by DropBlock, so as to prevent overfitting and strengthen the generalization ability of the network. Finally, the improved dense layers were superimposed to form an improved Dense convolutional Network (DenseNet). Feature information extraction recognition experiments were conducted on the self-built cattle eyes image dataset. Experimental results show that the recognition accuracy, precision and recall of the improved DenseNet are 97.47%, 98.11% and 97.90% respectively, and compared to the network without improvement, the above recognition accuracy rate, precision rate, recall rate are improved by 2.52 percentage points, 3.32 percentage points, 2.94 percentage points respectively. It can be seen that the improved network has higher precision and robustness.