Loading...

Table of Content

    10 May 2022, Volume 42 Issue 5
    Artificial intelligence
    Text classification model combining word annotations
    Xianfeng YANG, Jiahe ZHAO, Ziqiang LI
    2022, 42(5):  1317-1323.  DOI: 10.11772/j.issn.1001-9081.2021030489
    Asbtract ( )   HTML ( )   PDF (1662KB) ( )  
    Figures and Tables | References | Related Articles | Metrics

    The traditional text feature representation method cannot fully solve the polysemy problem of word. In order to solve the problem, a new text classification model combining word annotations was proposed. Firstly, by using the existing Chinese dictionary, the dictionary annotations of the text selected by the word context were obtained, and the Bidirectional Encoder Representations from Transformers (BERT) encoding was performed on them to generate the annotated sentence vectors. Then, the annotated sentence vectors were integrated with the word embedding vectors as the input layer to enrich the characteristic information of the input text. Finally, the Bidirectional Gated Recurrent Unit (BiGRU) was used to learn the characteristic information of the input text, and the attention mechanism was introduced to highlight the key feature vectors. Experimental results of text classification on public THUCNews dataset and Sina weibo sentiment classification dataset show that, the text classification models combining BERT word annotations have significantly improved performance compared to the text classification models without combining word annotations, the proposed BERT word annotation _BiGRU_Attention model has the highest precision and recall in all the experimental models for text classification, and has the F1-Score of reflecting the overall performance up to 98.16% and 96.52% respectively.

    Short text classification method by fusing corpus features and graph attention network
    Shigang YANG, Yongguo LIU
    2022, 42(5):  1324-1329.  DOI: 10.11772/j.issn.1001-9081.2021030508
    Asbtract ( )   HTML ( )   PDF (1121KB) ( )  
    Figures and Tables | References | Related Articles | Metrics

    Short text classification is an important research problem of Natural Language Processing (NLP), and is widely used in news classification, sentiment analysis, comment analysis and other fields. Aiming at the problem of data sparsity in short text classification, by introducing node and edge weight features of corpora, based on Graph ATtention network (GAT), a new graph attention network named Node-Edge GAT (NE-GAT) by fusing node and edge weight features was proposed. Firstly, a heterogeneous graph was constructed for each corpus, Gravity Model (GM) was used to evaluate the importance of word nodes, and edge weights were obtained through Point Mutual Information (PMI) between nodes. Secondly, a text-level graph was constructed for each sentence, node importance and edge weights were integrated into the update process of nodes. Experimental results show that, the average accuracy of the proposed model on the test sets reaches 75.48%, which is better than those of the models such as Text Graph Convolution Network (Text-GCN), Text-Level-Graph Neural Network (TL-GNN) and Text classification method for INductive word representations via Graph neural networks (Text-ING). Compared with original GAT, the proposed model has the average accuracy improved by 2.32 percentage points, which verifies the effectiveness of the proposed model.

    Emotional map of emergency based on sentiment analysis and influence evaluation
    Liqing QIU, Fushuai QU
    2022, 42(5):  1330-1338.  DOI: 10.11772/j.issn.1001-9081.2021040654
    Asbtract ( )   HTML ( )   PDF (3347KB) ( )  
    Figures and Tables | References | Related Articles | Metrics

    Aiming the spread of negative network public opinions in emergencies, a research method of emotional map of emergency based on sentiment analysis and influence evaluation was proposed. In the proposed method, a sentiment analysis model based on multi-head self-attention mechanism and Bi-directional Long Short-Term Memory network (Bi-LSTM) was proposed to evaluate website users’ emotional tendencies. Meanwhile, a point influence evaluation algorithm combining weighted degree and K-shell value was proposed to measure users’ influences. Based on the above models, the emotional map of emergency was constructed, which effectively improved the accuracy and scientificity of the emotional map. Taking “7.7 Anshun Bus Falling into Lake Incident” as an example, the life cycle of an emergency was divided into four stages such as outbreak stage, spread stage, maturity stage and decline stage, which were used to separately generate the emotional maps for visualization analysis. Experimental results show that, the F1-score of the proposed sentiment analysis model on the hotel review dataset is 9.92 percentage points and 2.5 percentage points higher than that of Recurrent Neural Networks for Text Classification (Text-RNN) model in positive and negative aspects respectively. On the Karate network, the discrimination and accuracy of the proposed influence evaluation algorithm are 46.89 percentage points and 29.05 percentage points higher than those of the K-shell algorithm respectively. By building the emotional map based on social networks, relevant department can find the opinion leaders and their tendencies, thereby grasping the development trend of online public opinion, and reducing the influence of negative emotions on society.

    News recommendation method with knowledge graph and differential privacy
    Li’e WANG, Xiaocong LI, Hongyi LIU
    2022, 42(5):  1339-1346.  DOI: 10.11772/j.issn.1001-9081.2021030527
    Asbtract ( )   HTML ( )   PDF (1421KB) ( )  
    Figures and Tables | References | Related Articles | Metrics

    The existing recommendation method with knowledge graph and privacy protection cannot effectively balance the noise of Differential Privacy (DP) and the performance of recommender system. In order to solve the problem, a News Recommendation method with Knowledge Graph and Privacy protection (KGPNRec) was proposed. Firstly, the multi-channel Knowledge-aware Convolutional Neural Network (KCNN) model was adopted to merge the multi-dimensional feature vectors of news title, entities and entity contexts of knowledge graph to improve the accuracy of recommendation. Secondly, based on the attention mechanism, the noise with different magnitudes was added in the feature vectors according to different sensitivities to reduce the impact of noise on data analysis. Then, the uniform Laplace noise was added to weighted user feature vectors to ensure the security of user data. Finally, the experimental analysis was conducted on real news datasets. Experimental results show that, compared with the baseline methods such as Privacy-Preserving Multi-Task recommendation Framework (PPMTF) and recommendation method based on Deep Knowledge-aware Network (DKN), the proposed KGPNRec can protect user privacy and ensure the prediction performance of method. For example, on the Bing News dataset, the Area Under Curve (AUC) value, accuracy and F1-score of the proposed method are improved by 0.019, 0.034 and 0.034 respectively compared with those of PPMTF.

    Long short-term session-based recommendation algorithm combining paired coding scheme and two-dimensional conventional neural network
    Xueqin CHEN, Tao TAO, Zhongwang ZHANG, Yilei WANG
    2022, 42(5):  1347-1354.  DOI: 10.11772/j.issn.1001-9081.2021030467
    Asbtract ( )   HTML ( )   PDF (1011KB) ( )  
    Figures and Tables | References | Related Articles | Metrics

    The session-based recommendation algorithm based on Recurrent Neural Network (RNN) can effectively model the long-term dependency in the session, and can combine the attention mechanism to describe the main purpose of the user in the session. However, it cannot bypass the items that are not related to the user’s main purpose in the process of session modeling, and is susceptible to their influence to reduce the recommendation accuracy. In order to solve problems, a new paired coding scheme was designed, which transformed the original input sequence embedding vector into a three-dimensional tensor representation, so that non-adjacent behaviors were also able to be linked. The tensor was processed by a two-dimensional Conventional Neural Network (CNN) to capture the relationship between non-adjacent items, and a Neural Attentive Recommendation Machine introducing two-dimensional COnvolutional neural network for Session-based recommendation (COS-NARM) model was proposed. The proposed model was able to effectively skip items that were not related to the user’s main purpose in the sequence. Experimental results show that the recall and Mean Reciprocal Rank (MRR) of the COS-NARM model on multiple real datasets such as DIGINETICA are improved to varying degrees, and they are better than those of all baseline models such as NARM and GRU-4Rec+. On the basis of the above research, Euclidean distance was introduced into the COS-NARM model, and the OCOS-NARM model was proposed. Euclidean distance was used to directly calculate the similarity between interests at different times to reduce the parameters of model and reduce the complexity of model. Experimental results show that the introduction of Euclidean distance further improves the recommendation effect of the OCOS-NARM model on multiple real datasets such as DIGINETICA, and makes the training time of the OCOS-NARM model shortened by 14.84% compared with that of the COS-NARM model, effectively improving the training speed of model.

    Feature selection algorithm based on neighborhood rough set and monarch butterfly optimization
    Lin SUN, Jing ZHAO, Jiucheng XU, Xinya WANG
    2022, 42(5):  1355-1366.  DOI: 10.11772/j.issn.1001-9081.2021030497
    Asbtract ( )   HTML ( )   PDF (1375KB) ( )  
    Figures and Tables | References | Related Articles | Metrics

    The classical Monarch Butterfly Optimization (MBO) algorithm cannot handle continuous data well, and the rough set model cannot sufficiently process large-scale, high-dimensional and complex data. To address these problems, a new feature selection algorithm based on Neighborhood Rough Set (NRS) and MBO was proposed. Firstly, local disturbance, group division strategy and MBO algorithm were combined, and a transmission mechanism was constructed to form a Binary MBO (BMBO) algorithm. Secondly, the mutation operator was introduced to enhance the exploration ability of this algorithm, and a BMBO based on Mutation operator (BMBOM) algorithm was proposed. Then, a fitness function was developed based on the neighborhood dependence degree in NRS, and the fitness values of the initialized feature subsets were evaluated and sorted. Finally, the BMBOM algorithm was used to search the optimal feature subset through continuous iterations, and a meta-heuristic feature selection algorithm was designed. The optimization performance of the BMBOM algorithm was evaluated on benchmark functions, and the classification performance of the proposed feature selection algorithm was evaluated on UCI datasets. Experimental results show that, the proposed BMBOM algorithm is significantly better than MBO and Particle Swarm Optimization (PSO) algorithms in terms of the optimal value, worst value, average value and standard deviation on five benchmark functions. Compared with the optimized feature selection algorithms based on rough set, the feature selection algorithms combining rough set and optimization algorithms, the feature selection algorithms combining NRS and optimization algorithms, the feature selection algorithms based on binary grey wolf optimization, the proposed feature selection algorithm performs well in the three indicators of classification accuracy, the number of selected features and fitness value on UCI datasets, and can select the optimal feature subset with few features and high classification accuracy.

    Human learning optimization algorithm based on learning psychology
    Han MENG, Liang MA, Yong LIU
    2022, 42(5):  1367-1374.  DOI: 10.11772/j.issn.1001-9081.2021030505
    Asbtract ( )   HTML ( )   PDF (1244KB) ( )  
    Figures and Tables | References | Related Articles | Metrics

    Aiming at the problems of low optimization accuracy and slow convergence of Simple Human Learning Optimization (SHLO) algorithm, a new Human Learning Optimization algorithm based on Learning Psychology (LPHLO) was proposed. Firstly, based on Team-Based Learning (TBL) theory in learning psychology, the TBL operator was introduced, so that on the basis of individual experience and social experience, team experience was added to control individual learning state to avoid the premature convergence of algorithm. Then, the memory coding theory was combined to propose the dynamic parameter adjustment strategy, thereby effectively integrating the individual information, social information and team information. And the abilities of the algorithm to explore locally and develop globally were better balanced. Two examples of knapsack problem of typical combinatorial optimization problems, 0-1 knapsack problem and multi-constraint knapsack problem, were selected for simulation experiments. Experimental results show that, compared with the algorithms such as SHLO algorithm, Genetic Algorithm (GA) and Binary Particle Swarm Optimization (BPSO) algorithm, the proposed LPHLO has more advantages in optimization accuracy and convergence speed, and has a better ability to solve the practical problems.

    Multi-label classification algorithm based on non-negative matrix factorization and sparse representation
    Yongchun BAO, Jianchen ZHANG, Shouxin DU, Junjun ZHANG
    2022, 42(5):  1375-1382.  DOI: 10.11772/j.issn.1001-9081.2021050706
    Asbtract ( )   HTML ( )   PDF (773KB) ( )  
    Figures and Tables | References | Related Articles | Metrics

    Traditional multi-label classification algorithms are based on binary label prediction. However, the binary labels can only indicate whether the data has relevant categories, so that they contain less semantic information and cannot fully represent the label semantic information. In order to fully mine the semantic information of label space, a new Multi-Label classification algorithm based on Non-negative matrix factorization and Sparse representation (MLNS) was proposed. In the proposed algorithm, the non-negative matrix factorization and sparse representation technologies were combined to transform the binary labels of data into the real labels, thereby enriching the label semantic information and improving the classification effect. Firstly, the label latent semantic space was obtained by the non-negative matrix factorization of label space, and the label latent semantic space was combined with the original feature space to form a new feature space. Then, the global similarity relation between samples was obtained by the sparse coding of the obtained feature space. Finally, the binary label vectors were reconstructed by using the obtained similarity relation to realize the transformation between binary labels and real labels. The proposed algorithm was compared with the algorithms such as Multi-Label classification Based on Gravitational Model (MLBGM), Multi-Label Manifold Learning (ML2), multi-Label learning with label-specific FeaTures (LIFT) and Multi-Label classification based on the Random Walk graph and the K-Nearest Neighbor algorithm (MLRWKNN) on 5 standard multi-label datasets and 5 evaluation metrics. Experimental results show that, the proposed MLNS is better than the compared multi-label classification algorithms in multi-label classification, the proposed MLNS ranks top1 in 50% cases, top 2 in 76% cases and top 3 in all cases.

    Multi-label image classification method based on global and local label relationship
    Wei REN, Hexiang BAI
    2022, 42(5):  1383-1390.  DOI: 10.11772/j.issn.1001-9081.2021071240
    Asbtract ( )   HTML ( )   PDF (4088KB) ( )  
    Figures and Tables | References | Related Articles | Metrics

    Considering the difficulty of modeling the interaction between labels and solidification of global label relationship in multi-label image classification tasks, a new Multiple-Label image classification method based on Global and Local Label Relationship (ML-GLLR) was proposed by combining self-attention mechanism and Knowledge Distillation (KD) method. Firstly, Convolutional Neural Network (CNN), semantic module and Dual Layer Self-Attention (DLSA) module were used by the Local Label Relationship (LLR) model to model local label relationship. Then, the KD method was used to make LLR learn global label relationship. The experimental results on the public datasets of MicroSoft Common Objects in COntext (MSCOCO) 2014 and PASCAL VOC challenge 2007 (VOC2007) show that, LLR improves the mean Average Precision (mAP) by 0.8 percentage points and 0.6 percentage points compared with Multiple Label classification based on Graph Convolutional Network (ML-GCN) respectively, and the proposed ML-GLLR increases the mAP by 0.2 percentage points and 1.3 percentage points compared with LLR. Experimental results show that, the proposed ML-GLLR can not only model the interaction between labels, but also avoid the problem of global label relationship solidification.

    Cross-domain person re-identification method based on attention mechanism with learning intra-domain variance
    Daili CHEN, Guoliang XU
    2022, 42(5):  1391-1397.  DOI: 10.11772/j.issn.1001-9081.2021030459
    Asbtract ( )   HTML ( )   PDF (2210KB) ( )  
    Figures and Tables | References | Related Articles | Metrics

    To solve severe performance degradation problem of person re-identification task during cross-domain migration, a new cross-domain person re-identification method based on attention mechanism with learning intra-domain variance was proposed. Firstly, ResNet50 was used as the backbone network and some modifications were made to it, so that it was more suitable for person re-identification task. And Instance-Batch Normalization Network (IBN-Net) was introduced to improve the generalization ability of model. At the same time, for the purpose of learning more discriminative features, a region attention branch was added to the backbone network. For the training of source domain, it was treated as a classification task. Cross-entropy loss was utilized for supervised learning of source domain, and triplet loss was introduced to mine the details of source domain samples and improve the classification performance of source domain. For the training of target domain, intra-domain variance was considered to adapt the difference in data distribution between the source domain and the target domain. In the test phase, the output of ResNet50 pool-5 layer was used as image features, and Euclidean distance between query image and candidate image was calculated to measure the similarity of them. In the experiments on two large-scale public datasets of Market-1501 and DukeMTMC-reID, the Rank-1 accuracy of the proposed method is 80.1% and 67.7% respectively, and its mean Average Precision (mAP) is 49.5% and 44.2% respectively. Experimental results show that, the proposed method has better performance in improving generalization ability of model.

    Lightweight human pose estimation method based on non-local high-resolution network
    Qixiang SUN, Ning HE, Jingzun ZHANG, Chen HONG
    2022, 42(5):  1398-1406.  DOI: 10.11772/j.issn.1001-9081.2021030512
    Asbtract ( )   HTML ( )   PDF (3303KB) ( )  
    Figures and Tables | References | Related Articles | Metrics

    Human pose estimation is one of the basic tasks in computer vision, which can be applied to the fields such as action recognition, games, and animation production. The current designs of deep network model mostly use deepening the network to obtain better performance. As a result, the demand for computing resources is beyond the computing power of embedded devices and mobile devices, and the requirements of actual applications can not be met. In order to solve the problems, a new lightweight network model integrating Ghost module structure was proposed, that is, the Ghost module was used to replace the basic module in the original high-resolution network, thereby reducing the number of network parameters. In addition, a non-local high-resolution network was designed, that is, the non-local network module was fused in the 1/32 resolution stage of the network, so that the network has the ability to obtain global features, thereby improving the accuracy of human pose estimation, and reducing the network parameters while ensuring the accuracy of model. Experiments were carried out on the human pose estimation datasets such as Max Planck Institut Informatik (MPII) and the Common Objects in COntext (COCO).Experimental results indicate that, compared with the original high-resolution network, the proposed network model has the accuracy of human pose estimation increased by 1.8 percentage points with the number of network parameters reduced by 40%.

    Proposal-based aggregation network for single object tracking in 3D point cloud
    Yi ZHUANG, Haitao ZHAO
    2022, 42(5):  1407-1416.  DOI: 10.11772/j.issn.1001-9081.2021030533
    Asbtract ( )   HTML ( )   PDF (3836KB) ( )  
    Figures and Tables | References | Related Articles | Metrics

    Compared with 2D RGB-based images, 3D point clouds retain the real and rich geometric information of objects in space to deal with vision challenge with scale variation in the single object tracking problem. However, the precision of 3D object tracking is affected by the loss of information brought by the sparsity of point cloud data and the deformation caused by the object position changing. To solve the above two problems, a proposal-based aggregation network composed of three modules was proposed in an end-to-end learning pattern. In this network, the 3D bounding box was determined by locating object center in the best proposal to realize the single object tracking in 3D point cloud. Firstly, the point cloud data of both templates and search areas was transferred into bird’s-eye view pseudo images. In the first module, the feature information was enriched through spatial and cross-channel attention mechanisms. Then, in the second module, the best proposal was given by the anchor-based deep cross-correlation Siamese region proposal subnetwork. Finally, in the third module, the object features were extracted through region of interest pooling operation by the best proposal at first, and then, the object and template features were aggregated, the sparse modulated deformable convolution layer was used to deal with the problems of point cloud sparsity and deformation, and the final 3D bounding box was determined. Experimental results of the comparison between the proposed method and the state-of-the-art 3D point cloud single object tracking methods on KITTI dataset show that: in comprehensive experiment of car, the proposed method has improved 1.7 percentage points on success rate and 0.2 percentage points on precision in real scenes; in multi-category extensive experiment of car, van, cyclist and pedestrian, the proposed method has improved the average success rate by 0.8 percentage points, and the average precision by 2.8 percentage points, indicating that the proposed method can solve the single object tracking problem in 3D point cloud and make the 3D object tracking results more accurate.

    Pruning of YOLOv4 based on rank information in industrial scenes
    Xiao QIN, Miao CHENG, Shaobing ZHANG, Lian HE, Xiangwen SHI, Pinxue WANG, Shang ZENG
    2022, 42(5):  1417-1423.  DOI: 10.11772/j.issn.1001-9081.2021030448
    Asbtract ( )   HTML ( )   PDF (2320KB) ( )  
    Figures and Tables | References | Related Articles | Metrics

    In the Radio Frequency IDentification (RFID) real-time defect detection task in industrial scenes, the deep learning target detection algorithms such as You Only Look Once (YOLO) are often adopted in order to ensure the detection precision and speed. However, these algorithms are still difficult to meet the speed requirement of industrial detection, and the corresponding network models cannot be deployed on resource-constrained devices. To solve these problems, the YOLO model must be pruned and compressed. A new network pruning method of the weighted fusion of feature information richness and feature information diversity based on rank information was proposed. Firstly, the unpruned model was loaded and reasoned, and the rank information of the corresponding feature maps of the filters was obtained in forward propagation to measure the feature information richness. Secondly, according to the different pruning rates, the rank information was clustered or the similarity of the rank information was calculated to measure the feature information diversity. Finally, the importance degrees of the corresponding filters were obtained after the weighted fusion and were sorted, and the filters with low importance were cut off. Experimental results show that, for YOLOv4, when the pruning rate is 28.87% and the weight of feature information richness is 0.75, the proposed method has the mean Average Precision (mAP) improved by 2.6%-8.9% compared with the method that uses rank information of the feature maps alone, and the model pruned by the proposed method even has the mAP increased by 0.4% and the model parameters reduced by 35.0% compared with the unpruned model, indicating that the proposed method is conducive to the model deployment.

    Lung nodule classification algorithm based on neural network architecture search
    Xinlin XIE, Yi XIAO, Xinying XU
    2022, 42(5):  1424-1430.  DOI: 10.11772/j.issn.1001-9081.2021050813
    Asbtract ( )   HTML ( )   PDF (1632KB) ( )  
    Figures and Tables | References | Related Articles | Metrics

    Lung nodule classification is an important task in the diagnosis of early-stage lung cancer. Although the lung nodule classification methods based on deep learning can achieve good classification accuracy, they have the problems of complex model and poor interpretability. Therefore, a lung nodule classification algorithm based on neural network architecture search was proposed. Firstly, the attention residual convolution cell was regarded as the basic unit of the search space, and the partial order pruning method was used as the search strategy to construct the neural network architecture for searching 3D classification network, thereby achieving the balance between network performance and search speed. Then, the multi-scale channels and spatial attention modules were constructed in the network to improve the interpretability of feature description and categorical inference. Finally, the stacking method was used to merge the searched network architectures with multiple models to obtain accurate prediction results of classification of benign and malignant lung nodules. Compared with the state-of-the-art lung nodule classification methods, the proposed algorithm has better classification performance and faster convergence on the widely-used lung nodule classification dataset LIDC-IDRI. Moreover, the proposed algorithm has the specificity and precision reached 95.37% and 93.42% respectively, showing it can achieve accurate classification of benign and malignant lung nodules.

    Remote sensing image scene classification based on effective channel attention
    Zhen QU, Kunting LI, Zhixi FENG
    2022, 42(5):  1431-1439.  DOI: 10.11772/j.issn.1001-9081.2021030464
    Asbtract ( )   HTML ( )   PDF (2678KB) ( )  
    Figures and Tables | References | Related Articles | Metrics

    The methods based on artificially designed features cannot extract high-level information from remote sensing images and previously used Convolutional Neural Network (CNN) such as VGGNet and ResNet cannot focus on distinguishable classification features in remote sensing images. In order to solve the problems, a novel method called ECA-ResNeXt-8-SVM was proposed based on Effective Channel Attention (ECA) mechanism for remote sensing image scene classification. In order to build an effective model, a deep feature extraction network called ECA-ResNeXt-8 embedded with the ECA module was designed, and the end-to-end learning was used to make network lay emphasis on channels with distinguishable classification features. At the same time, Support Vector Machine (SVM) was utilized to replace the fully connected layer as the classifier of the extracted deep features, which helped to improve the classification accuracy and generalization ability of model. On the experimental dataset UC Merced Land-Use, the classification accuracy of the proposed model reaches 95.81%, which is increased by 6% and 18% compared to SE-ResNeXt50 and ResNeXt50 networks respectively. When the classification accuracy is 75%, the proposed model has the training time reduced by 82% and 81% compared to the two above networks respectively. Experimental results show that the proposed model can reduce the convergence time of model effectively and improve the classification accuracy for remote sensing image scene.

    Low contrast filament sizing defect detection method of non-woven fabric based on deep feature fusion
    Yongshuai LU, Yingjie TANG, Xinran MA
    2022, 42(5):  1440-1446.  DOI: 10.11772/j.issn.1001-9081.2021050834
    Asbtract ( )   HTML ( )   PDF (2419KB) ( )  
    Figures and Tables | References | Related Articles | Metrics

    In order to solve the problem of poor detection effect of traditional image processing methods for the low contrast filament sizing defects in non-woven fabric production process, a low contrast filament sizing defect detection method of non-woven fabric based on Convolutional Neural Network (CNN) was proposed. Firstly, the collected non-woven fabric images were preprocessed to construct a defect dataset of filament sizing. Then, an improved convolutional neural network and a multi-scale feature sampling fusion module were used to construct an encoder to extract the semantic information of low contrast filament sizing defects, and a skip connection was used in the decoder to achieve multi-scale feature fusion for optimizing the upsampling module. Finally, the low contrast defect detection of filament sizing was realized by training the network model on the constructed dataset. Experimental results show that, the proposed method can effectively locate and detect the low contrast filament sizing defects on non-woven fabric. The Mean Intersection over Union (MIoU) and category Mean Pixel Accuracy (MPA) of the proposed method can reach 77.32% and 86.17% respectively, and the average detection time of single sample of the proposed method is 50 ms, which can meet the requirements of industrial production.

    Data science and technology
    Cross-layer data sharing based multi-task model
    Ying CHEN, Jiong YU, Jiaying CHEN, Xusheng DU
    2022, 42(5):  1447-1454.  DOI: 10.11772/j.issn.1001-9081.2021030516
    Asbtract ( )   HTML ( )   PDF (1841KB) ( )  
    Figures and Tables | References | Related Articles | Metrics

    To address the issues of negative transfer and difficulty of information sharing between loosely correlated tasks in multi-task learning model, a cross-layer data sharing based multi-task model was proposed. The proposed model pays attention to fine-grained knowledge sharing, and is able to retain the memory ability of shallow layer shared experts and generalization ability of deep layer specific task experts. Firstly, multi-layer shared experts were unified to obtain public knowledge among complicatedly correlated tasks. Then, the shared information was transferred to specific task experts at different layers for sharing partial public knowledge between the upper and lower layers. Finally, the data sample based gated network was used to select the needed information for different tasks autonomously, thereby alleviating the harmful effects of sample dependence to the model. Compared with the Multi-gate Mixture-Of-Experts (MMOE) model, the proposed model improved the F1-score of two tasks by 7.87 percentage points and 1.19 percentage points respectively on UCI census-income dataset. The proposed model also decreased the Mean Square Error (MSE) value of regression task to 0.004 7 and increased the Area Under Curve (AUC) value of classification task to 0.642 on MovieLens dataset. Experimental results demonstrate that the proposed model is suitable to improve the influence of negative transfer and can learn public information among complicated related tasks more efficiently.

    Imbalanced data classification algorithm based on ball cluster partitioning and undersampling with density peak optimization
    Xuewen LIU, Jikui WANG, Zhengguo YANG, Qiang LI, Jihai YI, Bing LI, Feiping NIE
    2022, 42(5):  1455-1463.  DOI: 10.11772/j.issn.1001-9081.2021050736
    Asbtract ( )   HTML ( )   PDF (1551KB) ( )  
    Figures and Tables | References | Related Articles | Metrics

    It is an effective hybrid strategy for imbalanced data classification of integrating cost-sensitivity and resampling methods into the ensemble algorithms. Concerning the problem that the misclassification cost calculation and undersampling process less consider the intra-class and inter-class distributions of samples in the existing hybrid methods, an imbalanced data classification algorithm based on ball cluster partitioning and undersampling with density peak optimization was proposed, named Boosting algorithm based on Ball Cluster Partitioning and UnderSampling with Density Peak optimization (DPBCPUSBoost). Firstly, the density peak information was used to define the sampling weights of majority samples, and the majority ball cluster with “neighbor cluster” was divided into “area misclassified easily” and “area misclassified hardly”, then the sampling weight of samples in “area misclassified easily” was increased. Secondly, the majority samples were undersampled based on the sampling weights in the first iteration, then the majority samples were undersampled based on the sample distribution weight in every iteration. And the weak classifier was trained on the temporary training set combining the undersampled majority samples with all minority samples. Finally, the density peak information of samples was combined with the categorical distribution of samples to define the different misclassification costs for all samples, and the weights of samples with higher misclassification cost were increased by the cost adjustment function. Experimental results on 10 KEEL datasets indicate that, the number of datasets with the highest performance achieved by DPBCPUSBoost is more than that of the imbalanced data classification algorithms such as Adaptive Boosting (AdaBoost), Cost-sensitive AdaBoost (AdaCost), Random UnderSampling Boosting (RUSBoost) and UnderSampling and Cost-sensitive Boosting (USCBoost), in terms of evaluation metrics such as Accuracy, F1-Score, Geometric Mean (G-mean) and Area Under Curve (AUC) of Receiver Operating Characteristic (ROC). Experimental results verify that the definition of sample misclassification cost and sampling weight of the proposed DPBCPUSBoost is effective.

    Density peak clustering algorithm based on adaptive nearest neighbor parameters
    Huanhuan ZHOU, Bochuan ZHENG, Zheng ZHANG, Qi ZHANG
    2022, 42(5):  1464-1471.  DOI: 10.11772/j.issn.1001-9081.2021050753
    Asbtract ( )   HTML ( )   PDF (5873KB) ( )  
    Figures and Tables | References | Related Articles | Metrics

    Aiming at the problem that the nearest neighbor parameters need to be set manually in density peak clustering algorithm based on shared nearest neighbor, a density peak clustering algorithm based on adaptive nearest neighbor parameters was proposed. Firstly, the proposed nearest neighbor parameter search algorithm was used to automatically obtain the nearest neighbor parameters. Then, the clustering centers were selected through the decision diagram. Finally, according to the proposed allocation strategy of representative points, all sample points were clustered through allocating the representative points and the non-representative points sequentially. The clustering results of the proposed algorithm was compared with those of the six algorithms such as Shared-Nearest-Neighbor-based Clustering by fast search and find of Density Peaks (SNN?DPC), Clustering by fast search and find of Density Peaks (DPC), Affinity Propagation (AP), Ordering Points To Identify the Clustering Structure (OPTICS), Density-Based Spatial Clustering of Applications with Noise (DBSCAN), and K-means on the synthetic datasets and UCI datasets. Experimental results show that, the proposed algorithm is better than the other six algorithms on the evaluation indicators such as Adjusted Mutual Information (AMI), Adjusted Rand Index (ARI) and Fowlkes and Mallows Index (FMI). The proposed algorithm can automatically obtain the effective nearest neighbor parameters, and can better allocate the sample points in the edge region of the cluster.

    Clustering algorithm based on local gravity and distance
    Jie DU, Yan MA, Hui HUANG
    2022, 42(5):  1472-1479.  DOI: 10.11772/j.issn.1001-9081.2021030515
    Asbtract ( )   HTML ( )   PDF (3200KB) ( )  
    Figures and Tables | References | Related Articles | Metrics

    The Density Peak Clustering (DPC) algorithm cannot accurately select the cluster centers for the datasets with various density and complex shape. The Clustering by Local Gravitation (LGC) algorithm has many parameters which need manual adjustment. To address these issues, a new Clustering algorithm based on Local Gravity and Distance (LGDC) was proposed. Firstly, the local gravity model was used to calculate the ConcEntration (CE) of data points, and the distance between each point and the point with higher CE value was determined according to CE. Then, the data points with high CE and high distance were selected as cluster centers. Finally, the remaining data points were allocated based on the idea that the CE of internal points of the cluster was much higher than that of the boundary points. At the same time, the balanced k nearest neighbor was used to adjust the parameters automatically. Experimental results show that, LGDC achieves better clustering effect on four synthetic datasets. Compared with algorithms such as DPC and LGC, LGDC has the index of Adjustable Rand Index (ARI) improved by 0.144 7 on average on the real datasets such as Wine, SCADI and Soybean.

    Cyber security
    Reversible data hiding algorithm in encrypted domain based on secret image sharing
    Zexi WANG, Minqing ZHANG, Yan KE, Yongjun KONG
    2022, 42(5):  1480-1489.  DOI: 10.11772/j.issn.1001-9081.2021050823
    Asbtract ( )   HTML ( )   PDF (4022KB) ( )  
    Figures and Tables | References | Related Articles | Metrics

    The current reversible data hiding algorithms in encrypted domain have the problems that the ciphertext images carrying secret have poor fault tolerance and disaster resistance after embedding secret data, once attacked or damaged, the original image cannot be reconstructed and the secret data cannot be extracted. In order to solve the problems, a new reversible data hiding algorithm in encrypted domain based on secret image sharing was proposed, and its application scenarios in cloud environment were analyzed. Firstly, the encrypted image was divided into n different ciphertext images carrying secret with the same size. Secondly, in the process of segmentation, the random quantities in Lagrange interpolation polynomial were taken as redundant information, and the mapping relationship between secret data and each polynomial coefficient was established. Finally, the reversible embedding of the secret data was realized by modifying the built-in parameters of the encryption process. When k ciphertext images carrying secret were collected, the original image was able to be fully recovered and the secret data was able to be extracted. Experimental results show that, the proposed algorithm has the advantages of low computational complexity, large embedding capacity and complete reversibility. In the (3,4) threshold scheme, the maximum embedding rate of the proposed algorithm is 4 bit per pixel (bpp), and in the (4,4) threshold scheme, the maximum embedding rate of the proposed algorithm is 6 bpp. The proposed algorithm gives full play to the disaster recovery characteristic of secret sharing scheme. Without reducing the security of secret sharing, the proposed algorithm enhances the fault tolerance and disaster resistance of ciphertext images carrying secret, improves the embedding capacity of algorithm and the disaster recovery ability in the application scenario of cloud environment, and ensures the security of carrier image and secret data.

    Android malware family classification method based on code image integration
    Mo LI, Tianliang LU, Ziheng XIE
    2022, 42(5):  1490-1499.  DOI: 10.11772/j.issn.1001-9081.2021030486
    Asbtract ( )   HTML ( )   PDF (3025KB) ( )  
    Figures and Tables | References | Related Articles | Metrics

    Code visualization technology is rapidly popularized in the field of Android malware research once it was proposed. Aiming at the problem of insufficient representation ability of code image converted from single DEX (classes.dex) file, a new Android malware family classification method based on code image integration was proposed. Firstly, the DEX, XML (androidManifest.xml) and decompiled JAR (classes.jar) files in the Android application package were converted to three gray-scale images, and the Bilinear interpolation algorithm was used for the scaling of gray images in different sizes. Then, the three gray-scale images were integrated into a three-dimensional Red-Green-Blue (RGB) image for training and classification. In terms of classification model, the Soft Threshold (ST) Block+ResNeSt(STResNeSt) was proposed by combining the soft threshold denoising block with Split-Attention based ResNeSt. The proposed model has the strong anti-noise ability and is able to pay more attention to the important features of code image. To handle the long-tail distribution of data in the training process, Class Balance Loss (CB Loss) was introduced after data augmentation, which provided a feasible solution to the over-fitting caused by the imbalance of samples. On the Drebin dataset, the accuracy of integrated code image is 2.93 percentage points higher than that of DEX gray-scale image, the accuracy of STResNeSt is improved by 1.1 percentage points compared with the Residual Neural Network (ResNet), the scheme of data augmentation combined with CB Loss improves the F1 score by up to 2.4 percentage points. Experimental results show that, the average classification accuracy of the proposed method reaches 98.97%, which can effectively classify the Android malware family.

    Improved multi-primary-node consensus mechanism based on practical Byzantine fault tolerance
    Xiuli REN, Lei ZHANG
    2022, 42(5):  1500-1507.  DOI: 10.11772/j.issn.1001-9081.2021050772
    Asbtract ( )   HTML ( )   PDF (1899KB) ( )  
    Figures and Tables | References | Related Articles | Metrics

    The high communication complexity of Practical Byzantine Fault Tolerance (PBFT) consensus protocol will lead to low consensus efficiency, the failure or the existing of Byzantine behavior of the single primary node will lead to the stop of consensus process. In order to solve these problems, an Improved Multi-primary-node Practical Byzantine Fault Tolerance (IMPBFT) consensus mechanism was proposed. Firstly, the number of effective consensus rounds of nodes was calculated by the number of consensus rounds of nodes, the number of consensus rounds with Byzantine behavior and the priority values assigned to the nodes, and several primary nodes were selected according to the size of effective consensus rounds. Then, the original consensus mechanism was improved to make all nodes use the improved consensus mechanism for consensus. Finally, pipeline was introduced to implement the concurrent execution of IMPBFT consensus. In the pipeline operation, multi-stage messages of different rounds’ consensus were signed together, and no fixed cycle was used to control the pipeline. Theoretical research and experimental results show that, the multi-primary-node structure of IMPBFT is more secure and stable than the consensus structure of single primary node. Compared with PBFT and Credit-Delegated Byzantine Fault Tolerance (CDBFT) consensus with square level traffic, the proposed IMPBFT reduces the traffic to linear level. The IMPBFT has better performance than PBFT and CDBFT in terms of transaction throughput, scalability and transaction delay. The IMPBFT using the “multi-stage messages signed together with no fixed cycle” pipeline has improved the transaction throughput by 75.2% compared with the IMPBFT without pipeline.

    Advanced computing
    Resource load prediction model based on long-short time series feature fusion
    Yifei WANG, Lei YU, Fei TENG, Jiayu SONG, Yue YUAN
    2022, 42(5):  1508-1515.  DOI: 10.11772/j.issn.1001-9081.2021030393
    Asbtract ( )   HTML ( )   PDF (2857KB) ( )  
    Figures and Tables | References | Related Articles | Metrics

    Resource load prediction with high accuracy can provide a basis for real-time task scheduling, thus reducing energy consumption. However, most prediction models for time series of resource load make short-term or long-term prediction by extracting the long-time series dependence characteristics of time series and neglecting the short-time series dependence characteristics of time series. In order to make a better long-term prediction of resource load, a new edge computing resource load prediction model based on long-short time series feature fusion was proposed. Firstly, the Gram Angle Field (GAF) was used to transform time series into image format data, so as to extract features by Convolutional Neural Network (CNN). Then, the CNN was used to extract spatial features and short-term data features, the Long Short-Term Memory (LSTM) network was used to extract the long-term time series dependent features of time series. Finally, the extracted long-term and short-term time series dependent features were fused through dual-channel to realize long-term resource load prediction. Experimental results show that, the Mean Absolute Error (MAE), Root Mean Square Error (RMSE) and R-squared(R2) of the proposed model for CPU resource load prediction in Alibaba cloud clustering tracking dataset are 3.823, 5.274, and 0.815 8 respectively. Compared with the single-channel CNN and LSTM models, dual-channel CNN+LSTM and ConvLSTM+LSTM models, and resource load prediction models such as LSTM Encoder-Decoder (LSTM-ED) and XGBoost, the proposed model can provide higher prediction accuracy.

    Cloud computing task scheduling based on orthogonal adaptive whale optimization
    Jinquan ZHANG, Shouwei XU, Xincheng LI, Chongyang WANG, Jingzhi XU
    2022, 42(5):  1516-1523.  DOI: 10.11772/j.issn.1001-9081.2021050806
    Asbtract ( )   HTML ( )   PDF (1821KB) ( )  
    Figures and Tables | References | Related Articles | Metrics

    Aiming at the problems such as long task completion time, high task execution cost and unbalanced system load in task scheduling, a new cloud computing task scheduling method based on Orthogonal Adaptive Whale Optimization Algorithm (OAWOA) was proposed. Firstly, the Orthogonal Experimental Design (OED) was applied to the population initialization and global search stages to improve and maintain the population diversity, avoid the algorithm from falling into local convergence too early. Then, the adaptive exponential decline factor and bidirectional search mechanism were used to further strengthen the global search ability of the algorithm. Finally, the fitness function was optimized to enable the algorithm to achieve multi-objective optimization. Through the simulation experiments, the proposed algorithm was compared with Whale Optimization Algorithm (WOA), Particle Swarm Optimization (PSO) algorithm, Bat Algorithm (BA) and two other improved WOAs. Experimental results show that, when the task scale is 50 and 500, the proposed algorithm achieves better convergence effect, has the total time and total cost of the obtained system executing tasks lower than those of other algorithms, and has the load balancing degree only lower than that of BA. In conclusion, the proposed algorithm shows significant advantages in reducing the total time and cost of system executing tasks and improving the system load balancing.

    Parallel design and implementation of minimum mean square error detection algorithm based on array processor
    Shuai LIU, Lin JIANG, Yuancheng LI, Rui SHAN, Yulin ZHU, Xin WANG
    2022, 42(5):  1524-1530.  DOI: 10.11772/j.issn.1001-9081.2021030460
    Asbtract ( )   HTML ( )   PDF (1972KB) ( )  
    Figures and Tables | References | Related Articles | Metrics

    In massive Multiple-Input Multiple-Output (MIMO) systems, Minimum Mean Square Error (MMSE) detection algorithm has the problems of poor adaptability, high computational complexity and low efficiency on the reconfigurable array structure. Based on the reconfigurable array processor developed by the project team, a parallel mapping method based on MMSE algorithm was proposed. Firstly, a pipeline acceleration scheme which could be highly parallel in time and space was designed based on the relatively simple data dependency of Gram matrix calculation. Secondly, according to the relatively independent characteristic of Gram matrix calculation and matched filter calculation module in MMSE algorithm, a modular parallel mapping scheme was designed. Finally, the mapping scheme was implemented based on Xilinx Virtex-6 development board, and the statistics of its performance were performed. Experimental results show that, the proposed method achieves the acceleration ratio of 2.80, 4.04 and 5.57 in Quadrature Phase Shift Keying (QPSK) uplink with the MIMO scale of 128×4128×8 and 128×16, respectively, and the reconfigurable array processor reduces the resource consumption by 42.6% compared with the dedicated hardware in the 128×16 massive MIMO system.

    Real root isolation algorithm for exponential function polynomials
    Xinyu GE, Shiping CHEN, Zhong LIU
    2022, 42(5):  1531-1537.  DOI: 10.11772/j.issn.1001-9081.2021030440
    Asbtract ( )   HTML ( )   PDF (503KB) ( )  
    Figures and Tables | References | Related Articles | Metrics

    For addressing real root isolation problem of transcendental function polynomials, an interval isolation algorithm for exponential function polynomials named exRoot was proposed. In the algorithm, the real root isolation problem of non-polynomial real functions was transformed into sign determination problem of polynomial, then was solved. Firstly, the Taylor substitution method was used to construct the polynomial nested interval of the objective function. Then, the problem of finding the root of the exponential function was transformed into the problem of determining the positivity and negativity of the polynomial in the intervals. Finally, a comprehensive algorithm was given and applied to determine the reachability of rational eigenvalue linear system tentatively. The proposed algorithm was implemented in Maple efficiently and easily with readable output results. Different from HSOLVERand numerical calculation method fsolve, exRoot avoids discussing the existence of roots directly, and theoretically has termination and completeness. It can reach any precision and can avoid the systematic error brought by numerical solution when being applied into the optimization problem.

    Network and communications
    Multi-user computation offloading and resource optimization policy based on device-to-device communication
    Yu LI, Xiping HE, Lianggui TANG
    2022, 42(5):  1538-1546.  DOI: 10.11772/j.issn.1001-9081.2021030458
    Asbtract ( )   HTML ( )   PDF (2244KB) ( )  
    Figures and Tables | References | Related Articles | Metrics

    With the significant increase of computation-intensive and latency-intensive applications, Mobile-Edge Computing (MEC) was proposed to provide computing services for users at the network edge. In view of the limited computing resources of edge servers at the Base Stations (BSs) and the long latency of long-distance computation offloading of users at the network edge, a multi-user computation offloading and resource optimization policy based on Device-to-Device (D2D) communication was proposed. The D2D was integrated into MEC network to directly offload tasks to neighbor users for executing in D2D mode, which was able to further reduce offloading latency and energy consumption. Firstly, the joint optimization problem of multi-user computation offloading and multi-user computing resource allocation was modelled with the optimization objective of minimizing the total system computing cost including latency and energy consumption. Then, the solution of this problem was considered as a D2D pairing process, and the multi-user computation offloading and resource optimization policy algorithm was proposed based on stable matching. Finally, the optimization allocation policy of D2D offloading was solved iteratively. The characteristics such as stability, optimality and complexity of the proposed algorithm were analyzed by theoretical proof. Simulation results show that, the proposed algorithm can effectively reduce the total system computing cost by 10%-30% compared with the random matching algorithm, and the performance of the proposed algorithm is very close to the optimal exhaustive search algorithm, indicating that the proposed policy based on D2D offloading is helpful to improve latency and energy consumption performance.

    Resource control of infectious disease in multi-layer star coupling network
    Si ZHANG, Bishan ZHANG, Zhongjun MA
    2022, 42(5):  1547-1553.  DOI: 10.11772/j.issn.1001-9081.2021030491
    Asbtract ( )   HTML ( )   PDF (2054KB) ( )  
    Figures and Tables | References | Related Articles | Metrics

    Concerning that the existing infectious disease spread model do not consider the influence and mechanism of specific special network structure and resource factors on controlling infectious disease outbreak, a discrete dynamic propagation model was established by combining the two-layer star coupling network with the Susceptible-Infected-Susceptible (SIS) model of infectious disease. In this model, the structural characteristics and the concept of average degree of the star network were used to derive the discrete equations of the proportion of infected population in every layer with resources and various parameters. Theory analysis and simulation experimental results indicate that, the multi-layer star coupling infectious disease spread network has resource thresholds. When the node is a leaf node, the network has two resource thresholds. Increasing the number of resources to control the spread of infectious diseases is only effective between the two resource thresholds. At this time, the proportion of population infected with infectious diseases decreases with the increase of resources invested. When the node is a central node, the resource threshold in the network reduces from two to one with the increase of proportion of infected population in other layers. Additionally, the control effect of the coupling strength of the inter-layer central node and the inter-layer leaf node on the epidemic varies with the location of the nodes.

    Computer software technology
    Cross-project defect prediction method based on feature selection and TrAdaBoost
    Li LI, Kexin SHI, Zhenkang REN
    2022, 42(5):  1554-1562.  DOI: 10.11772/j.issn.1001-9081.2021050867
    Asbtract ( )   HTML ( )   PDF (2257KB) ( )  
    Figures and Tables | References | Related Articles | Metrics

    Cross-project software defect prediction can solve the problem of few training data in prediction projects. However, the source project and the target project usually have the large distribution difference, which reduces the prediction performance. In order to solve the problem, a new Cross-Project Defect Prediction method based on Feature Selection and TrAdaBoost (CPDP-FSTr) was proposed. Firstly, in the feature selection stage, Kernel Principal Component Analysis (KPCA) was used to delete redundant data in the source project. Then, according to the attribute feature distribution of the source project and the target project, the candidate source project data closest to the target project distribution were selected according to the distance. Finally, in the instance transfer stage, the TrAdaBoost method improved by the evaluation factor was used to find out the instances in the source project which were similar to the distribution of a few labeled instances in the target project, and establish a defect prediction model. Using F1 as the evaluation index, compared with the methods such as cross-project software defect prediction using Feature Clustering and TrAdaBoost (FeCTrA), Cross-project software defect prediction based on Multiple Kernel Ensemble Learning (CMKEL), the proposed CPDP-FSTr had the prediction performance improved by 5.84% and 105.42% respectively on AEEEM dataset, enhanced by 5.25% and 85.97% respectively on NASA dataset, and its two-process feature selection is better than the single feature selection process. Experimental results show that the proposed CPDP-FSTr can achieve better prediction performance when the source project feature selection proportion and the target project labeled instance proportion are 60% and 20% respectively.

    Multimedia computing and computer simulation
    Image super-resolution reconstruction network based on multi-channel attention mechanism
    Ye ZHANG, Rong LIU, Ming LIU, Ming CHEN
    2022, 42(5):  1563-1569.  DOI: 10.11772/j.issn.1001-9081.2021030498
    Asbtract ( )   HTML ( )   PDF (3016KB) ( )  
    Figures and Tables | References | Related Articles | Metrics

    The existing image super-resolution reconstruction methods are affected by texture distortion and details blurring of generated images. To address these problems, a new image super-resolution reconstruction network based on multi-channel attention mechanism was proposed. Firstly, in the texture extraction module of the proposed network, a multi-channel attention mechanism was designed to realize the cross-channel information interaction by combining one-dimensional convolution, thereby achieving the purpose of paying attention to important feature information. Then, in the texture recovery module of the proposed network, the dense residual blocks were introduced to recover part of high-frequency texture details as many as possible to improve the performance of model and generate high-quality reconstructed images. The proposed network is able to improve visual effects of reconstructed images effectively. Besides, the results on benchmark dataset CUFED5 show that the proposed network has achieved the 1.76 dB and 0.062 higher in Peak Signal-to-Noise Ratio (PSNR) and Structural SIMilarity (SSIM) compared with the classic Super-Resolution using Convolutional Neural Network (SRCNN) method. Experimental results show that the proposed network can increase the accuracy of texture migration, and effectively improve the quality of generated images.

    Image super-resolution reconstruction based on parallel convolution and residual network
    Huifeng WANG, Yan XU, Yiming WEI, Huizhen WANG
    2022, 42(5):  1570-1576.  DOI: 10.11772/j.issn.1001-9081.2021050742
    Asbtract ( )   HTML ( )   PDF (2730KB) ( )  
    Figures and Tables | References | Related Articles | Metrics

    The existing image super-resolution reconstruction algorithms can improve the overall visual effect of the image or promote the objective evaluation value of the reconstructed image, but have poor balanced improvement effect of image perception effect and objective evaluation value, and the reconstructed images lack high-frequency information, resulting in texture blur. Aiming at the above problems, an image super-resolution reconstruction algorithm based on parallel convolution and residual network was proposed. Firstly, taking the parallel structure as the overall framework, different convolution combinations were used on the parallel structure to enrich the feature information, and the jump connection was added to further enrich the feature information and fuse the output to extract more high-frequency information. Then, an adaptive residual network was introduced to supplement information and optimize network performance. Finally, perceptual loss was used to improve the overall quality of the restored image. Experimental results show that, compared with the algorithms such as Super-Resolution Convolutional Neural Network (SRCNN), Very Deep convolutional network for Super-Resolution (VDSR) and Super-Resolution Generative Adversarial Network (SRGAN), the proposed algorithm has better performance in image reconstruction and has clearer detail texture of the enlarged effect image. In the objective evaluation, the Peak Signal-To-Noise Ratio (PSNR) and Structural SIMilarity (SSIM) of the proposed algorithm in ×4 reconstruction are improved by 0.25 dB and 0.019 averagely and respectively compared with those of SRGAN.

    Single image de-raining algorithm based on semi-supervised learning
    Yongru QIU, Guangle YAO, Jie FENG, Haoyu CUI
    2022, 42(5):  1577-1582.  DOI: 10.11772/j.issn.1001-9081.2021030492
    Asbtract ( )   HTML ( )   PDF (3937KB) ( )  
    Figures and Tables | References | Related Articles | Metrics

    The images collected in rainy days usually have some phenomena that affect the image quality, such as the background object blocked by rain streaks and the image deformation, which have serious impact on the subsequent image analysis and application. Recently, numerous de-raining algorithms based on deep learning have been proposed and achieve good results. Most algorithms adopt supervised learning, that is, training the model on the synthetic rainy image dataset with paired labels due to the difficulty in acquiring clean background images without rain streaks from real-world rainy images. However, there are differences between synthetic and real-world rainy images on brightness, transparency, and shape of rain streaks. Thus, most de-raining algorithms based on supervised learning have poor generalization ability to real-world rainy images. Therefore, in order to improve the rain removal effect of de-raining models on the real-world rainy images, a single image de-raining algorithm based on semi-supervised learning was proposed. In the model training process of the proposed algorithm, the synthetic and real-world rainy images were added, and the difference of the first-order and second-order statistic information of feature vectors converted from the both input images were minimized to make the features of the both have same distribution. Meanwhile, in view of the complex and diverse characteristics of rain streaks, a multi-scale network was introduced to obtain richer image features and improve the performance of model. Experimental results show that, on the Rain100H dataset of synthetic rainy images, compared with Joint Deraining Network (JDNet), Synthetic-to-Real transfer learning (Syn2Real), the proposed algorithm improves the Peak Signal-to-Noise Ratio (PSNR) and Structural SIMilarity (SSIM) by at least 0.66 dB and 0.01 respectively. While removing rain streaks, the proposed algorithm can retain image details and color information to the greatest extent. At the same time, with the reduction of distribution discrepancy, the proposed algorithm achieves better performance on the real-world rainy images with strong generalization ability, compared with the de-raining algorithms such as JDNet and Syn2Real. The proposed algorithm is highly independent, can be applied to the existing de-raining algorithms based on supervised learning and significantly improve their de-raining effects.

    Power line inspection aerial image stitching based on Order-Aware network internal point screening network
    Lichuan HUI, Wanyu LI, Yilin CHEN
    2022, 42(5):  1583-1590.  DOI: 10.11772/j.issn.1001-9081.2021030493
    Asbtract ( )   HTML ( )   PDF (9373KB) ( )  
    Figures and Tables | References | Related Articles | Metrics

    The texture of power line inspection images with parallax variation is complex, the number of paired matching points obtained by traditional algorithms is less and the registration accuracy is low, which seriously affect the stitching effect of power line inspection unmanned aerial vehicle image. In order to solve the problems, a new image stitching method based on improved Order-Aware Network (OANet) was proposed. Firstly, the Accelerated KAZE (AKAZE) algorithm was adopted to match the power line inspection images to be stitched roughly. Secondly, the Squeeze-and-Excitation Networks (SENet) was added to the Order-Aware module in OANet, which helped to enhance the grasping ability of the network for both the local and global context information, and more accurate paired matching points were obtained. Then, the Mesh-based Photometric Alignment (MPA) algorithm was used to register the images to be stitched. Finally, the optimal suture line in the overlapping area was calculated by the content compressed sensing algorithm to complete image stitching. The number of correct matching points of the improved OANet network is about 10% higher than that of the original OANet network with time consumption increased by 10 ms on average. Compared with the registration stitching algorithms such as As-Projective-As-Possible (APAP) algorithm, Adaptive As-Natural-As-Possible (AANAP) algorithm and MPA algorithm, the proposed algorithm has the highest stitching quality with the root mean square error of the overlapping area of the images to be stitched is 0 and no distortion in the non-overlapping area. Experimental results show that, the proposed algorithm can stitch the aerial images of power line inspection quickly and stably.

    Image matching algorithm based on transmission tower area extraction
    Kegui GUO, Rui CAO, Neng WAN, Xiao WANG, Yue YIN, Xuming TANG, Junlin XIONG
    2022, 42(5):  1591-1597.  DOI: 10.11772/j.issn.1001-9081.2021050796
    Asbtract ( )   HTML ( )   PDF (3145KB) ( )  
    Figures and Tables | References | Related Articles | Metrics

    In order to solve the problem of low matching quality of the traditional feature extraction and matching algorithm in Unmanned Aerial Vehicle (UAV) visual localization, a new image matching algorithm based on transmission tower area extraction was proposed. Firstly, the image was divided into several overlapping grid areas, and the feature points were extracted by a two-layer pyramid structure for each area to ensure the uniform distribution of feature points. Then, the Line Segment Detector (LSD) algorithm was used to extract the lines in the images, the transmission tower support areas were extracted on the basis of special structure of transmission tower. Finally, the feature points in the transmission tower areas and the background areas were matched respectively in continuous images to further estimate the camera motion. In the rotation and translation estimation experiment, compared with the traditional Oriented Features from Accelerated Segment Test(FAST) and Rotated Binary Robust Independent Elementary Features (BRIEF) (ORB) feature extraction and matching algorithm, the proposed algorithm has the feature matching accuracy improved by 10.1 percentage points, the mean value of relative pose error reduced by 0.049. In the UAV inspection experiment, the relative error of the UAV trajectory estimation by using the proposed algorithm is 2.89%, which indicates that the proposed algorithm can achieve the robust and accurate estimation of the UAV’s pose during the real-time flying around the tower.

    Cattle body size measurement method based on Kinect v4
    Jianmin ZHAO, Cheng ZHAO, Haiguang XIA
    2022, 42(5):  1598-1606.  DOI: 10.11772/j.issn.1001-9081.2021030532
    Asbtract ( )   HTML ( )   PDF (3608KB) ( )  
    Figures and Tables | References | Related Articles | Metrics

    Aiming at the complexity of image background and difficulty of feature point extraction in cattle body size measurement based on machine vision, a new cattle body size measurement method based on Kinect v4 sensor was proposed. In this method, the color and depth images were collected, and the body size data were calculated by the body feature points extracted by the combination of algorithms such as object detection, Canny edge detection, and three-point arc curvature. Firstly, an image dataset of feature parts of cattle body size was created, and the deep learning You Only Look Once v5 (YOLOv5) target detection algorithm was used to detect feature part information of cattle body size in order to reduce the interference of other parts of cattle body and background on the extraction of body size measuring points. Secondly, with the help of Canny edge detection, contour extraction and other image processing algorithms in Open source Computer Vision (OpenCV) image processing library, the key contours with measuring points of cattle body size were obtained. Then, the algorithms such as polynomial fitting and three-point arc curvature were performed on the key contours to extract the measuring points of cattle body size in two-dimensional image. Finally, the depth information was used to convert the measuring point information in two-dimensional image to three-dimensional coordinate system, and the cattle body size measurement method was designed in three-dimensional coordinate system with the RANdom SAmple Consensus (RANSAC) algorithm. Through the comparison between the experimental measurement results with the sensor and the side of cattle body at different angles and manual measurement results in a complex environment, it can be seen that the average relative error of withers height is 0.76%, the average relative error of body oblique length is 1.68%, the average relative error of body straight length is 2.14 %, and the average relative error of hip height is 0.76% in cattle body measurement data. Experimental results show that the proposed method has high measurement accuracy in complex environment.

    Frontier and comprehensive applications
    Runoff forecast model based on graph attention network and dual-stage attention mechanism
    Hexuan HU, Huachao SUI, Qiang HU, Ye ZHANG, Zhenyun HU, Nengwu MA
    2022, 42(5):  1607-1615.  DOI: 10.11772/j.issn.1001-9081.2021050829
    Asbtract ( )   HTML ( )   PDF (2505KB) ( )  
    Figures and Tables | References | Related Articles | Metrics

    To improve the accuracy of watershed runoff volume prediction, and considering the lack of model transparency and physical interpretability of data-driven hydrological model, a new runoff forecast model named Graph Attention neTwork and Dual-stage Attention mechanism-based Long Short-Term Memory network (GAT-DALSTM) was proposed. Firstly, based on the hydrological data of watershed stations, graph neural network was introduced to extract the topology of watershed stations and generate the feature vectors. Secondly, according to the characteristics of hydrological time series data, a runoff forecast model based on dual-stage attention mechanism was established to predict the watershed runoff volume, and the reliability and transparency of the proposed model were verified by the model evaluation method based on attention coefficient heat map. On the Tunxi watershed dataset, the proposed model was compared with Graph Convolution Neural network (GCN) and Long Short-Term Memory network (LSTM) under each prediction step. Experimental results show that, the Nash-Sutcliffe efficiency coefficient of the proposed model is increased by 3.7% and 4.9% on average respectively, which verifies the accuracy of GAT-DALSTM runoff forecast model. By analyzing the heat map of attention coefficient from the perspectives of hydrology and application, the reliability and practicability of the proposed model were verified. The proposed model can provide technical support for improving the prediction accuracy and model transparency of watershed runoff volume.

    Prediction model of transaction pricing in internet freight transport platform based on combination of dual long short-term memory networks
    Youzhi LI, Zhihua HU, Chun CHEN, Peibei YANG, Yajing DONG
    2022, 42(5):  1616-1623.  DOI: 10.11772/j.issn.1001-9081.2021030504
    Asbtract ( )   HTML ( )   PDF (2220KB) ( )  
    Figures and Tables | References | Related Articles | Metrics

    Prediction results of transaction pricing of transport service orders in internet freight transport platform are the direct reflections of both platform operation strategy and carrier decision, and influences both platform benefits and the healthy development of carrier market significantly. Taking internet freight transport platform of SF Express network as an example, the data were preprocessed through missing value processing and categorical data conversion. Aiming at the prediction precision problem of transaction pricing in internet freight transport platform, a new prediction model of transaction pricing in internet freight transport platform based on combination of dual Long Short-Term Memory networks(LSTM) was designed, and the prediction results were analyzed by K-means clustering. Compared with the models such as LSTM, Support Vector Regression (SVR), Long Short-Term-Memory combined with Support Vector Regression (LSTM-SVR), and combination of grey GM(1,1) and Back Propagation (BP) (GM(1,1)-BP), the combination model of dual LSTM has the lowest Mean Absolute Error (MAE), Mean Square Error (MSE), Mean Absolute Percentage Error (MAPE) and highest R square (R2), which is 9.90, 402.54, 1.48 and 0.999 97 respectively. The evaluation results of predicted order transaction pricing by using K-means clustering analysis are consistent with the actual values. Experimental results indicate that, the proposed combination model of dual LSTM has effectiveness and precise prediction effect of transaction pricing in internet freight transport platform.

    Stock market volatility prediction method based on improved genetic algorithm and graph neural network
    Xiaohan LI, Huading JIA, Xue CHENG, Taiyong LI
    2022, 42(5):  1624-1633.  DOI: 10.11772/j.issn.1001-9081.2021030519
    Asbtract ( )   HTML ( )   PDF (1762KB) ( )  
    Figures and Tables | References | Related Articles | Metrics

    Aiming at the difficulty in selecting stock valuation features and the lack of time series relational dimension features during the prediction of stock market volatility by intelligent algorithms such as Support Vector Machine (SVM) and Long Short-Term Memory (LSTM) network, in order to accurately predict stock volatility and effectively prevent financial market risks, a new stock market volatility prediction method based on Improved Genetic Algorithm (IGA) and Graph Neural Network (GNN) named IGA-GNN was proposed. Firstly, the data of stock market trading index graph was constructed based on the time series relation between adjacent trading days. Secondly, the characteristics of evaluation indexes were used to improve Genetic Algorithm (GA) by optimizing crossover and mutation probabilities, thereby realizing the node feature selection. Then, the weight matrix of edge and node features of graph data was established. Finally, the GNN was used for the aggregation and classification of graph data nodes, and the stock market volatility prediction was realized. In the experiment stage, the studied number of total evaluation indexes of stock was 130, and 87 effective evaluation indexes were extracted from the above by IGA under GNN method, making the number of stock evaluation indexes reduced by 33.08%. The proposed IGA was applied to the intelligent algorithms for feature extraction. The obtained algorithms has the overall prediction accuracy improved by 7.38 percentage points compared with the intelligent algorithms without feature extraction. Compared with applying the traditional GA for feature extraction of the intelligent algorithms, applying the proposed IGA for feature extraction of the intelligent algorithms has the total training time shortened by 17.97%. Among them, the prediction accuracy of IGA-GNN method is the highest, which is 19.62 percentage points higher than that of GNN method without feature extraction. Compared with the GNN method applying the traditional GA for feature extraction, the IGA-GNN method has the training time shortened by 15.97% on average. Experimental results show that, the proposed method can effectively extract stock features and has good prediction effect.

    PID parameter tuning of brushed direct-current motor based on improved genetic algorithm
    Yanfei LIU, Zheng PENG, Yihui WANG, Zhong WANG
    2022, 42(5):  1634-1641.  DOI: 10.11772/j.issn.1001-9081.2021050745
    Asbtract ( )   HTML ( )   PDF (3093KB) ( )  
    Figures and Tables | References | Related Articles | Metrics

    Aiming at the complicated and time-consuming problems of brushed Direct-Current (DC) motor Proportion Integral Differential (PID) parameter tuning, a PID parameter tuning method based on improved Genetic Algorithm (GA) was proposed. Firstly, a fitness enhanced elimination through selection rule was proposed, which improved the selection process of traditional GA. Then, a gene infection crossover method was proposed to ensure the increase of the average fitness value in the evolution process. Finally, the unnecessary copy operation in traditional GA was deleted to improve the running speed of the algorithm. Modeling and simulation analysis were carried out through the motor transfer function. Experimental results show that, compared with conventional tuning methods, the proposed improved GA can significantly improve the PID parameter tuning effect. At the same time, compared with the traditional GA, the improved GA reduces the evolutionary generation number required to achieve the same evolutionary effect by 79%, and increases the running speed of the algorithm by 4.1%. The proposed improved GA improves GA from the two key operation steps of selection and crossover, and is applied to PID parameter tuning to make the rise time less, the stability time shorter, and the overshoot smaller.

    Industrial process control method based on local policy interaction exploration-based deep deterministic policy gradient
    Shaobin DENG, Jun ZHU, Xiaofeng ZHOU, Shuai LI, Shurui LIU
    2022, 42(5):  1642-1648.  DOI: 10.11772/j.issn.1001-9081.2021050716
    Asbtract ( )   HTML ( )   PDF (2120KB) ( )  
    Figures and Tables | References | Related Articles | Metrics

    In order to achieve the stable and precise control of industrial processes with non-linearity, hysteresis, and strong coupling, a new control method based on Local Policy Interaction Exploration-based Deep Deterministic Policy Gradient (LPIE-DDPG) was proposed for the continuous control of deep reinforcement learning. Firstly, the Deep Deterministic Policy Gradient (DDPG) algorithm was used as the control strategy to greatly reduce the phenomena of overshoot and oscillation in the control process. At the same time, the control strategy of original controller was used as the local strategy for searching, and interactive exploration was used as the rule for learning, thereby improving the learning efficiency and stability. Finally, a penicillin fermentation process simulation platform was built under the framework of Gym and the experiments were carried out. Simulation results show that, compared with DDPG, the proposed LPIE-DDPG improves the convergence efficiency by 27.3%; compared with Proportion-Integration-Differentiation (PID), the proposed LPIE-DDPG has fewer overshoot and oscillation phenomena on temperature control effect, and has the penicillin concentration increased by 3.8% in yield. In conclusion, the proposed method can effectively improve the training efficiency and improve the stability of industrial process control.

2025 Vol.45 No.3

Current Issue
Archive
Honorary Editor-in-Chief: ZHANG Jingzhong
Editor-in-Chief: XU Zongben
Associate Editor: SHEN Hengtao XIA Zhaohui
Domestic Post Distribution Code: 62-110
Foreign Distribution Code: M4616
Address:
No. 9, 4th Section of South Renmin Road, Chengdu 610041, China
Tel: 028-85224283-803
  028-85222239-803
Website: www.joca.cn
E-mail: bjb@joca.cn
WeChat
Join CCF