Computer and Modernization

Application of Hybrid CTC/Attention Model in Mandarin Recognition

XU Hong-kui, ZHANG Zi-feng, LU Jiang-kun, ZHOU Jun-jie, HU Wen-ye, JIANG Tong-tong

2022, 0(08): 1-6.

Asbtract ( 358 )

PDF (1151KB) ( 103 )

References | Related Articles | Metrics

The end-to-end speech recognition model based on Connectionist Temporal Classification （CTC） has the advantages of simple structure and automatic alignment, but the recognition accuracy needs to be further improved. This paper introduces the attention mechanism to form a hybrid CTC/Attention end-to-end model. This method adopts the multi-task learning approach, combining the alignment advantage of CTC with the context modeling advantage of attention mechanism. The experimental results show that when the 80-dimensional FBank feature and the 3-dimensional pitch feature are selected as the acoustic features, and the VGG-Bidirectional long short-time memory network is selected as the encoder for Chinese Mandarin recognition, the character error rate of this hybrid model is reduced by about 6.1% compared with the end-to-end model based on CTC, after the external language model is connected, the character error rate is further reduced by 0.3%. Compared with the traditional baseline model, the character error rate also decreased significantly.

Chinese Word Segmentation Model Based on Attention-BIGRU-CRF

ZHOU Hui, XU Ming-hai, XU Xiao-dong

2022, 0(08): 7-12.

Asbtract ( 298 )

PDF (1062KB) ( 95 )

References | Related Articles | Metrics

Natural language processing is an important branch of the development of artificial intelligence, and Chinese word segmentation is the first step of natural language processing. Improving the efficiency of Chinese word segmentation can improve the accuracy of the results of natural language processing. Therefore, an Attention-BIGRU-CRF-CRF model is proposed in this paper. Firstly, the Chinese text is transformed into vector form through word vector conversion, and then the BIGRU is used for serialization learning. Then, the attention mechanism is introduced to calculate the correlation between the input and output of BIGRU to obtain more accurate vector values, Finally, the vector value is spliced with the vector value serialized by BIGRU as the input of CRF layer, and the label prediction result is obtained. The simulation results show that the F1 values of Attention-BIGRU-CRF model in the corpus of people’s daily 2014 and MSRA are 97.34% and 98.25% respectively, and the word segmentation rate of processed text is 248.1 KB/s. Therefore, the model integrating attention mechanism and BIGRU-CRF network can not only improve the accuracy of word segmentation, but also improve the time efficiency of word segmentation.

Improved Binary Harmony Search Algorithm for Solving Multidimensional Knapsack Problem

LIU Ya-wen, JIANG Yan, PAN Da-zhi

2022, 0(08): 13-19.

Asbtract ( 245 )

PDF (1602KB) ( 85 )

References | Related Articles | Metrics

Harmony search （HS） is a meta-heuristic method that has been applied widely to continuous optimization problems. The Multidimensional Knapsack Problem（MKP）is a kind of typical combinatorial optimization problems. In order to solve this problem, an Improved Binary Harmony Search（IBHS）algorithm was proposed. The proposed algorithm generated a binary population through a Bernoulli random process, introduced a dynamic adaptive parameter p in the candidate harmony generation operator, and coordinated the global search and local search of the algorithm through the adaptive adjustment of the algorithm parameter p,and proposed an effective method to measure the multidimensional weighted value density of commodities for binary individual correction and optimization; the introduction of an elite local search mechanism for collaborative optimization was improved the convergence speed of IBHS. By solving 10 sets of typical multidimensional knapsack examples of different scales and comparing with Greedy Binary Lion Swarm Optimization （GBLSO）algorithm, Modified Binary Differential Evolution（MBDE）algorithm and Binary Modified Harmony （BMHS）algorithm, the experimental results show that the proposed algorithm has fast convergence efficiency, high optimization accuracy and good robustness when solving MKP.

Anti-collision Shortest Path Planning Based on Improved Dijkstra Algorithm

HUANG Yi-hu, YU Ya-nan

2022, 0(08): 20-24.

Asbtract ( 399 )

PDF (781KB) ( 132 )

References | Related Articles | Metrics

Multiple UAVs may face the contradiction of track conflict when performing operational tasks. Therefore, an improved Dijkstra algorithm is proposed to realize the function of multiple UAVs to find the shortest and non-conflicting route. In the process of searching and traversing each track node by the classical Dijkstra algorithm, the variable length backtracking array of precursor nodes of each node is introduced to record all precursor nodes contained in each node, and all feasible shortest length routes from the starting point to the target point of each task are found. Then the time window conflict judgment model is introduced to separate the non-conflicting routes from all feasible routes of each task. Once all routes conflict, the conflict node in one of the shortest routes is treated as a temporary obstacle point and the shortest route that does not conflict with other tasks is re found by changing the backtracking array. Matlab software is used to design and write programs to verify the algorithm. The experiments show that the improved algorithm can plan all the shortest and non-conflicting routes contained in each task when multiple UAVs perform operational tasks and the planning efficiency of task set has been significantly improved.

An Improved KNN Medical Classification Algorithm Based on FLANN

GUO Kai, AI Ju-mei

2022, 0(08): 25-29.

Asbtract ( 226 )

PDF (1025KB) ( 78 )

References | Related Articles | Metrics

In this paper, by studying the application and analysis of KNN （k-nearest neighbor） algorithm in the field of disease prediction, two shortcomings of KNN are summarized, and the F_KNN （cyclic search nearest neighbor） algorithm is proposed: 1） for faults of KNN large amount of calculation and low efficiency, this paper uses the FLANN （quick nearest neighbor search） to loop search the nearest point of sample under test, record the number of nearest neighbor points as nearest neighbor ideas set, calculate using the sample subset to replace the complete treatment, can reduce the amount of calculation, greatly improve the efficiency of the KNN algorithm; 2） In view of the shortcoming of KNN that it is difficult to classify high-dimensional data sets, AHP （analytic hierarchy process） is adopted in this paper to study the correlation of characteristic attributes of samples, and appropriate parameters are used to assign weights, which improves the accuracy of KNN algorithm. In this paper, a set of cerebral apoplexy data sets are used to test the optimized algorithm, and the experimental results show that the accuracy of F_KNN is 96.2%. Compared with the traditional KNN, it improves the classification performance and greatly improves the efficiency of the algorithm. When dealing with high dimensional and large data sets, F_KNN algorithm has obvious advantages and has a good application prospect.

Adaptive Division of Sound Velocity Profile Based on Gradient Difference

MA Qian, DUAN Yi, XU Dong,

2022, 0(08): 30-35.

Asbtract ( 197 )

PDF (1169KB) ( 81 )

References | Related Articles | Metrics

Sound velocity profile represents the functional relationship between sound velocity and depth, which is often used in underwater acoustic location. The vertical axis represents the depth of the sea and the horizontal axis represents the speed of sound. In view of the contradiction between the positioning accuracy and the amount of calculation in the high-precision underwater acoustic positioning system, an adaptive layered algorithm of sound velocity profile is proposed. According to the characteristics of sound velocity changing in a limited range, the algorithm searches the sound velocity profile layer by layer. The gradient difference between sound velocity layers is used to represent the change of sound velocity, and the nodes of sound velocity change are found out. The original characteristics of sound velocity profile are retained to simplify the sound velocity profile. The simulation results show that according to the threshold value, the original sound velocity profile can be reduced by 70% after adaptive layered processing. Compared with D-P algorithm, smaller positioning error can be obtained when selecting the appropriate threshold. It has good engineering application value.

Link Prediction in Directed Network Combining Node Centrality and Degree-related Clustering

CHEN Guang-fu, LIAN Yan-ping,

2022, 0(08): 36-42.

Asbtract ( 212 )

PDF (1293KB) ( 74 )

References | Related Articles | Metrics

Most existing link prediction methods of directed networks only focus on link direction information and reciprocal link information, but ignore the contribution of node importance and degree correlation clustering, which leads to the decrease of prediction accuracy. To solve these problems, a directed network link prediction algorithm based on node centrality and degree correlation clustering is proposed. Firstly, it uses node centrality to calculate the number of neighbors of any node to measure the influence of node. Secondly, the node degree correlation clustering coefficient method is extended to the directed network to evaluate the clustering ability of nodes, and it is combined with the network coordination coefficient to obtain the high clustering ability of nodes. Finally, a parametric directed network link prediction index is proposed by integrating the above two kinds of information. On the six real world network compared with recent representative methods, the AUPR and AUC of the proposed algorithm are improved 33% and 1.6% respectively.

Association Analysis of Image Genetic Data Based on Group Sparse Joint Leraning

ZHAO Ying-li, ZHU Xu

2022, 0(08): 43-49.

Asbtract ( 181 )

PDF (2274KB) ( 75 )

References | Related Articles | Metrics

The development of image genetics has greatly promoted the research of mental diseases. It mainly analyzes and mines multimodal data to find out the disease-related pathogenesis. However, the data usually show the characteristics of group correlation or multiple feature correlation. It is difficult to find the relevant disease mechanism by traditional methods, which is prone to the problem of too sparse. To solve the above problems, this paper introduces the regularization term l1,2 norm which can achieve intra-group sparsity and inter-group smoothing, and jointly punishes canonical correlation analysis with the l2,1 norm which can achieve inter-group sparsity and intra-group smoothing. By optimizing the correlation between data, the feature selection of two-modal data sets with related group features and intra-group features is realized. The results of simulation experiments show that this method can not only accurately estimate the correlation coefficient between the two groups of data, but also select the relevant inter-group and intra-group features. On the real schizophrenia data set, this method can find more susceptibility genes and risk brain regions related to schizophrenia.

Cross-modal Retrieval Based on Context Fusion and Multi-similarity Learning

ZENG Yi-bin, GE Hong

2022, 0(08): 50-56.

Asbtract ( 255 )

PDF (932KB) ( 79 )

References | Related Articles | Metrics

Most of cross-modal retrieval methods do not fully exploit the interaction between heterogeneous data. To solve the problem, a novel method called Context Fusion and Multi-Similarity Learning （CFMSL） is proposed. To exploit the interactions between different modal data, the context fusion is adapted to aggregate different modal information. The generative module is used to generate discriminative representations by optimizing the pair similarity loss in the common subspace, which maximizes the intra-class similarity and minimizes the inter-class similarity for cross-modal alignment. Moreover, the re-ranking strategy based on single modality and fused multi-modality is proposed during evaluation phase, appropriately adjusting the final retrieval results to improve the performance. The experiments demonstrate that our proposed method achieves competitive results in cross-modal retrieval tasks on several widely-used image-text datasets, such as Pascal Sentences, Wikipedia, and NUS-WIDE-10K.

Point of Interest Recommendation Combined with Dynamic Multiple Types of Information

FENG Shen, YU Yue-cheng, ZHANG Zong-hai

2022, 0(08): 57-64.

Asbtract ( 184 )

PDF (1079KB) ( 81 )

References | Related Articles | Metrics

For the current mainstream point-of-interest recommendation algorithm, on the one hand, the user’s historical check-in data needs to be used, and on the other hand, the user’s long-term and short-term preferences need to be considered at the same time. However, existing methods tend to ignore the user preference information implied in user reviews and ignore the differences in the dependence of different users on long-term and short-term preferences. In view of the above limitations, a method of POI recommendation combined with dynamic multiple types of information （DMGCR） is proposed. Firstly, the attention mechanism is used to capture the user’s attention to different POI, so as to quantitatively describe the user’s long-term preference for POI. Secondly, the review information is combined with location and category information, and Bi-directional Long-Short Term Memory is used to learn the semantic features implicit in the review text. In this way, the user’s short-term preferences can be accurately portrayed on the basis of capturing the user’s emotional tendency toward POI. Finally, a comprehensive prediction function of user preferences integratedynamic multiple types of information is designed. Then the quantitative calculation of the recommendation probability of the next POI can be realized. Experimental results on multiple data sets verified the effectiveness and superiority of this method in recommendation performance.

Combined Estimation Method of Fuel Conesumption Based on Cluster Analysis

LI Shu, ZHANG Wei-ye, WANG Kun, DUAN Zhao-bin

2022, 0(08): 65-69.

Asbtract ( 172 )

PDF (1865KB) ( 70 )

References | Related Articles | Metrics

Aiming at the problem of single discontinuous missing data and continuous missing data in the carbon emission report, the estimation error of using a single method is large, a combined estimation method based on cluster analysis is proposed. The method firstly uses the K-medoids clustering algorithm to classify the data into single discontinuous missing data and continuous missing data, and then uses the Naive Bayes （NB） method to estimate the single discontinuous data, uses Dynamic Time Warping （DTW） method to estimate the continuous missing data, and finally evaluates the estimation results at 1%, 2%, and 3% root mean square error. The simulation results show that the NB-DTW combination method based on cluster analysis can effectively reduce the estimation error, which is 9.3%, 12.1% and 12.96% lower than the NB method at 1%, 2% and 3% root mean square error, respectively, and reduced by 35.46%, 43.62% and 55.04% respectively than DTW method.

Car-following Model Considering Multi-preceding Vehicles’ Information Feedback and Backward Looking Effect

HUI Fei, XI Hui, ZHANG Kai-wang, WEI Si

2022, 0(08): 70-77.

Asbtract ( 241 )

PDF (1795KB) ( 79 )

References | Related Articles | Metrics

In order to improve the stability of the car-following model in the Internet of vehicles environment, based on the classic OVCM model, an extended car-following model is deduced with the consideration of the influence of comprehensive information, including the effect of backward looking, the speed difference of multiple preceding vehicles and the optimal speed memory of multiple preceding vehicles on the stability of traffic flow. The neutral stability judgment conditions is obtained by linear stability analysis and the numerical simulation experiments and analysis are carried out. The experimental results show that, under the same initial disturbance conditions, the proposed model has a larger traffic flow stable area and a smaller speed fluctuation range than the OV, FVD, and OVCM models. Especially when the number of vehicles ahead k, the sensitivity coefficient of backward looking λi and the sensitivity coefficient of memory effect γi are k=3,λi=［0.2,0.15,0.1］,γi=［0.1,0.08,0.06］, the average speed fluctuation rate of the vehicle is less than 0.1%. Consequently, the proposed model can effectively reduce the impact of disturbances and enhance the steady-state maintenance of traffic flow.

An Improved Grey Wolf Algorithm for Flexible Job Shop Scheduling Problem

TIAN Yun-na, TIAN Yuan, LIU Xue, ZHAO Yan-lin

2022, 0(08): 78-85.

Asbtract ( 252 )

PDF (1216KB) ( 71 )

References | Related Articles | Metrics

Flexible job shop scheduling problem is a typical scheduling problem in the field of intelligent manufacturing. It is one of the most key links in manufacturing process planning and management. An effective solution method is of great practical significance to improve production efficiency. Based on the classical grey wolf algorithm, an improved grey wolf algorithm is proposed to solve the flexible job shop scheduling problem with the goal of optimizing the makespan. Firstly, the algorithm adopts the weight based coding form to discretize the continuous coding in the classical wolf swarm algorithm. Secondly, the random walk strategy is added in the iterative optimization process to enhance the local search ability, then the tail elimination strategy is added in the population updating process to avoid local optimization, increase the population diversity and reasonably expand the search range of the algorithm. The simulation results on the standard example show that the improved wolf swarm algorithm has obvious improvement in the optimization ability than the classical grey wolf algorithm. Compared with other intelligent optimization algorithms, the algorithm proposed in this paper has better optimization performance in each example.

Mobile Edge Computing Task Offloading Model and Algorithm Based on Energy Consumption and Delay Optimization

ZHAN Jun-wei, ZHUANG Yi

2022, 0(08): 86-93.

Asbtract ( 361 )

PDF (1302KB) ( 94 )

References | Related Articles | Metrics

With the rise of mobile edge computing, how to handle the offloading of edge computing tasks has become one of the hot research issues. For the multi-task-multi-edge server scenario, this paper first proposes a mobile edge computing task offloading model based on energy and delay optimization. This model takes into account the remaining power of the device, and uses the delay and energy consumption weighting factors to calculate the total cost of edge devices.And it has the advantages of prolonging equipment use time, reducing task offloading delay and energy consumption. Then we propose a mobile edge computing task offloading algorithm based on an improved genetic algorithm, which converts the problem of solving the optimal offloading decision into a problem of solving the population optimal solution. Comparative simulation experiment results show that the task offloading model and algorithm proposed in this paper can effectively solve the task offloading problem. The improved task offloading algorithm has a more accurate solution, can avoid the local optimal solution, and is helpful to find the best task offloading decision.

Data Acquisition Architecture Based on Domestic TCM Chip Encryption and Edge Cloud Collaborative Architecture

LIU Qiang, LI Qiao, BAO Xiao

2022, 0(08): 94-98.

Asbtract ( 211 )

PDF (1041KB) ( 85 )

References | Related Articles | Metrics

Before the audit of internal units, the audit department needs to collect and sort out all kinds of materials as the audit basis, including Internet public data and confidential data within the unit. When collecting confidential data inside the object, it is usually necessary to transmit data through encrypted media or auditors to work directly on the site. This paper proposes an acquisition architecture based on domestic TCM chip encryption and edge cloud collaboration. It can simultaneously collect the Internet public data and the confidential data inside the object, and encrypt data through the TCM chip solidified in the hardware to ensure that only the designated machine can read the confidential materials. The audit data edge cloud collaborative collection architecture based on domestic TCM chip encryption can realize the cross network transmission of classified data. According to the feedback of actual useres, taking the transmission of classified data using encrypted media as the comparison scenario, after adopting the collaborative collection architecture, it can quickly enter the audit work locally, and the data acquisition efficiency can be improved by 30% on average. The working time of non local audit is reduced by 3 days~4 days to effectively improve the audit efficiency.

Insider Threat Detection Based on Hybrid N-Gram and XGBoost Theory

SUN Dan, RAO Lan-xiang, SHI Wei-li, MENG Sha-sha, HU Shao-wen, HU Bi-wei, YING Song

2022, 0(08): 99-105.

Asbtract ( 245 )

PDF (1019KB) ( 80 )

References | Related Articles | Metrics

With the establishment and improvement of the network security mechanism of the government、 the enterprises and the Institutions, the threshold for attacking the target system from the outside is getting higher and higher. So the insider threats are gradually increasing. The internal threats are different from external threats. The attackers are mainly from internal users, so it makes the attacks more concealed and harder to be detected. The paper first analyzes user behaviors in the public SEA data set, then proposes an insider threat detection based on hybrid N-Gram and XGBoost theory, using the big data and machine learning methods. Three feature extraction methods: bag-of-words, N-Gram, and vocabulary are used for experimental comparison and N value experimental screening. The internal threat detection method based on the hybrid N-Gram model and XGBoost algorithm has a better detection effect than one-dimensional data and two-dimensional data. The effect of combining the different features of the four-dimensional data on the feature subset is better. The specificity reaches 0.23, the sensitivity reaches 27.65, the accuracy reaches 0.94, and the F1 value reaches 0.97. Comparing the 4 evaluation indicators of specificity, sensitivity, accuracy, and F1 value, the feature extraction method based on hybrid N-gram is more effective in detection than traditional bag-of-words and vocabulary feature extraction methods. This detection method not only improves the discrimination of internal threat detection signatures, but also improves the accuracy of feature extraction and calculation performance.

Multi-focus Image Fusion Based on Quadtree Decomposition and Adaptive Focus Measure

WANG Ji-wei, QU Huai-jing, WEI Ya-nan, XIE Ming, XU Jia, ZHANG Zhi-sheng, ZHANG Han-yuan

2022, 0(08): 106-113.

Asbtract ( 223 )

PDF (8545KB) ( 74 )

References | Related Articles | Metrics

In order to overcome the problem that the block-based fusion method is sensitive to the size of the block and the artifacts in the fused image, a new multi-focus image fusion method based on quadtree decomposition and adaptive focus measurement is proposed. Firstly, according to a new adaptive focus measure based on the sum of modified Laplacian （SML） and guided filtering, it is used to obtain the focus map of the source image. Then, using a new quadtree decomposition strategy and combining the obtained focus map, the source image is further decomposed into tree blocks of optimal size, the focus area is detected from the tree blocks, and a decision map is formed. Finally, the consistency of the decision graph is optimized and verified, and a fully focused image is reconstructed. Through experiments on public multi-focus image data sets, visual quality and objective indicators are compared with 9 advanced fusion methods. The experimental results show that the fusion method proposed in this paper has achieved better performance.

Medical Glass Bottle Mouth Defect Detection Method Based on Convolutional Autoencoder

REN Qiu-lin, REN De-jun, LI Xin, YAN Zong-yi, CAO Lin-jie, TANG Hong

2022, 0(08): 114-120.

Asbtract ( 275 )

PDF (2962KB) ( 89 )

References | Related Articles | Metrics

To solve the problem of a small number and variety of defect samples in defect detection, which makes it difficult to provide enough data for training of supervised deep learning models, this paper uses a large amount of easily accessible positive sample data （without defects） in industrial production to built an auto-encoder model of convolutional network with Encoder-Decoder structure, and embeds the convolutional attention module combining spatial and channel attention in the encoder to enhance the network feature extraction ability. We added the contextual information module in the encoding stage to obtain a larger perceptual field and reduce the computational requirements. Meanwhile, multi-scale structural similarity MS-SSIM and L1 loss were combined to improve the quality of reconstructed images, and peak signal-to-noise ratio （PSNR） was used to measure reconstruction error and discriminate anomalies. The experimental results show that the proposed medical glass bottle defect detection method can accurately detect the defect data and segment the defect region with 99.45% accuracy, 97.63% recall rate, 0.55% miss detection rate, and 2.93% false detection rate. The method can accurately detect the glass bottle defect and locate the defect area, and the image reconstruction time is short, only about 10.37 ms, which can achieve accurate and efficient automated product quality inspection.

Multistage-transformer Large-factor Network: Reference-based Super-resolution

CHEN Tong, ZHOU Deng-wen

2022, 0(08): 121-126.

Asbtract ( 386 )

PDF (2750KB) ( 67 )

References | Related Articles | Metrics

Image super-resolution （SR） refers to reconstructing the corresponding high-resolution copy from a low-resolution image. Aiming at solving the problem of inaccurate reconstruction of SR in the cases of super-large magnification （8×, 16×）, a multi-level transformer super-magnification reconstruction network is proposed （MTLF）. MTLF performs multi-level stacking of multiple transformers to process features of different scales, and uses the attention weights, which are obtained by the transformer and then improved by the modified attention module, to synthesize finer textures. In the end, the features of all magnifications fuse into a super-large-scale SR image. Experiments resenlts show that MTLF is superior to the state-of-the-art methods （including single-image super-resolution and Ref-based super-resolution methods） in terms of peak signal-to-noise ratio and visual effects. In particular, MTLF achieves fairly good results in the ultimate magnification （32×） scenario.

Table of Content