Computer and Modernization

Box Office Prediction Model Based on SA-EW-LSTM

LANG Kun, NIU Chunhui, LI Chenqiong, ZENG Suyu

2025, 0(05): 1-9. doi:10.3969/j.issn.1006-2475.2025.05.001

Asbtract ( 31 )

PDF (2701KB) ( 37 )

References | Related Articles | Metrics

Film box office is usually affected by many factors. However， the word-of-mouth， as a key factor， is often neglected by traditional prediction models. In order to improve the prediction accuracy， a novel box office prediction model based on the sentiment analysis， the entropy weight method and the LSTM neural network is presented in this paper. Firstly， the input index system is constructed by selecting eight influence factors， namely， word-of-mouth， box office of the previous day， box office of the same day last week， ticket price， service fee， screening rate， holiday or not， and search index. Secondly， the method of sentiment analysis is used to analyze the text of film review， and the sentiment index is used to quantify the word-of-mouth factor. Then， in order to quantify the impact of different factors on the box office， the entropy weight method is used to assign weights to different factors. Finally， the Long Short-Term Memory neural network is applied to predict the box office. Simulation results indicate that the prediction accuracy of the presented SA-EW-LSTM model reaches 94.9% and 94.8% on two different data sets， respectively， which is obviously superior to the other five classical models， and the effectiveness of the presented model is verified.

An Interactive Mining Approach for Spatial Co-location Patterns Incorporating#br# User Interest Preferences

BAO Xuguang1, 2 , CHEN Zhiwei1, LI Qiaochen1, 2, JIANG Chengcheng1

2025, 0(05): 10-20. doi:10.3969/j.issn.1006-2475.2025.05.002

Asbtract ( 25 )

PDF (3496KB) ( 28 )

References | Related Articles | Metrics

The mining of spatial co-location patterns is a crucial research focus within the field of spatial data mining. Its objective is to identify subsets of spatial features where instances frequently coexist in space. However， conventional approaches often rely solely on pattern frequency as a measure of user interest， disregarding subjective user preferences. This results in the extraction of numerous spatial co-location patterns that are uninteresting to users and may impact their subsequent decisions. Therefore， an interactive mining method of spatial co-location patterns combined with user’s interest preference is proposed. Firstly， a new similarity measure index for co-location patterns is introduced by combining Jaccard similarity and semantic similarity. Secondly， personalized clustering based on user feedback is utilized to extract user preferences. Subsequently， the Stochastic Degenerate Tree-Augmented Naive Bayes Integration Model （SDTANI） is proposed to establish a prediction model that integrates both user interest and preference. Finally， an interactive mining algorithm framework is employed to assist users in identifying interesting patterns. Experimental results using synthetic datasets with varying sizes and real datasets demonstrate that the proposed method outperforms other methods in terms of accuracy， particularly with respect to F1-score. Furthermore， it effectively mines spatial co-location patterns based on users’ interests.

Interference Suppression Algorithm for Millimeter Wave Radar Based on Independent Component Analysis

WANG Xing1, ZHONG Haili1, YU Yang2, LI Zhentao2, BAI Chuang1

2025, 0(05): 21-27. doi:10.3969/j.issn.1006-2475.2025.05.003

Asbtract ( 15 )

PDF (3825KB) ( 21 )

References | Related Articles | Metrics

To solve the problem of interference between millimeter-wave radars resulting in decreased signal-to-noise ratio and target detection failure， an interference suppression algorithm based on independent component analysis was proposed. This algorithm can accurately determine the number and interval of interference in the signal， and does not require any prerequisites in interference suppression， with strong robustness. Firstly， the interference interval of the echo signal was detected by using the cell averaging-constant false alarm rate detection to determine the interference occurrence section. Then， the echo signals of the radar under different virtual receiving channels were subjected to fast independent component analysis to separate out the independent components. Finally， by template matching with the interference section， the interference signal and useful signal were determined， interference cancellation was achieved， and the signal was reconstructed. Experimental results show that this algorithm has good interference suppression effect on countering radar interference. The signal-to-noise ratio of strong target 1 is improved by about 14 dB， and the signal-to-noise ratios of weak target 2 and tanget 3 are improved by about 9 dB.

Yi Language Named Entity Recognition Method Based on CR-BACC Model

WANG Chengxian1, ZHAO Qing2

2025, 0(05): 28-35. doi:10.3969/j.issn.1006-2475.2025.05.004

Asbtract ( 13 )

PDF (2145KB) ( 24 )

References | Related Articles | Metrics

This paper constructs and makes publicly available a named entity recognition dataset （YNNER） based on Yi language news text， collected the Liangshan Daily news dataset， and manually annotated the names of people， places and institutions. Considering the BiLSTM-Attention model and CNN model， global sequence and local spatial features can be extracted， and the restoration of diphthonic characters in the text can reduce the error of label recognition. This paper designs a Character Replacement BiLSTM Attention CNN conditional random field model （CR-BACC） based on character replacement. Experiments are conducted on the Chinese MSRA， People’s Daily and Yi YNNER datasets and compared with three representative algorithms. Experimental results show the effectiveness of this method in the Yi language named entity recognition task. This paper aims to promote the development of research in the field of Yi named entity recognition by providing datasets and models for the field to extend related research.

DDoS Attack Detection Method Based on Transformer Architecture

CHI Biwei1, SUN Rui2

2025, 0(05): 36-40. doi:10.3969/j.issn.1006-2475.2025.05.005

Asbtract ( 13 )

PDF (1604KB) ( 27 )

References | Related Articles | Metrics

With the rapid development of the Internet， DDoS attacks have become a major challenge in the field of network security. DDoS attacks disrupt the normal operation of target servers by controlling a large number of distributed computers to send massive amounts of malicious requests， seriously affecting the stability and security of network services. Traditional DDoS attack detection methods， such as rule-based detection， statistical methods， and machine learning approaches， often face issues such as high false positive rates and low detection efficiency when dealing with complex and dynamically changing network traffic. To address these issues， this paper proposes a Transformer-based DDoS attack detection system. The system utilizes the powerful self-attention mechanism of the Transformer model to capture long-term dependencies in network traffic， enabling more accurate identification of abnormal traffic patterns. Additionally， by incorporating positional encoding， the system can better handle temporal information and enhance the model’s ability to perceive global network traffic. Experimental results on datasets show that the Transformer-based DDoS detection model significantly outperforms comparison methods in terms of detection accuracy and recall rate， demonstrating the effectiveness of the proposed approach.

Image Encryption Method Based on Poisoning Attack Strategy

LI Zhuoqi, ZHAO Lihui

2025, 0(05): 41-47. doi:10.3969/j.issn.1006-2475.2025.05.006

Asbtract ( 12 )

PDF (1996KB) ( 23 )

References | Related Articles | Metrics

Aiming at the problem that images are prone to misuse and infringement of users’ rights， the paper proposes an image protection method based on a poisoning attack strategy. The method generates poisoned data by embedding perturbations in the image data， which significantly reduces the performance of the deep learning model using this as training data without affecting the visual quality of the original image. Using image recognition and feature extraction techniques， the dominant features of the target image category are obtained as the basis for model recognition and classification， which are added to the original dataset as perturbations， and the perturbations are constrained at both pixel and feature levels; experimental results on CIFAR-100 and ImageNet-100 show that the poisoned images generated by the poisoned attack strategy effectively reduce the classification accuracy of a variety of common deep learning models’ classification accuracy.

Survey of Application of Knowledge Graph in Field of Intelligent Manufacturing

JIANG Sulun1, 2, 3, YUAN Decheng1, GUO Qingda2, 3, LIU Jian3, YU Guangping2, 3

2025, 0(05): 48-59. doi:10.3969/j.issn.1006-2475.2025.05.007

Asbtract ( 15 )

PDF (3678KB) ( 32 )

References | Related Articles | Metrics

With the rapid promotion of new-goneration artifica intelligence， computing， and other technologies， the manufacturing field urgently needs to undergo intelligence and digitalization and upgrading. Through literature review and applied case studies， it is found that the construction of knowledge graph can promote the development of industrial intelligence， so the field of intelligent manufacturing has begun to apply knowledge graph to manage and optimize intelligent manufacturing equipment data and processes. At present， knowledge graph technology has been maturely applied in the direction of intelligent question answering， personalized recommendation， etc.， in order to explore the greater application potential of knowledge graph technology in the field of intelligent manufacturing， the current literature and application status is studied in detail and conduded. This paper firstly starts with the popular technologies such as knowledge acquisition， knowledge fusion， and knowledge reasoning involved in knowledge graphs， and then focuses on the research and analysis of several popular application directions such as industrial fault diagnosis， digital twins， human-computer collaborative interaction， and risk management based on knowledge graphs， and summarizes the general architecture， and discusses the future development trends and difficulties such as the combination with AIGC technology， and puts forward future prospects to provide reference for promoting intelligent manufacturing knowledge graphs.

Distributed System Fault Prediction Method Based on XGBoost & LightGBM

ZHANG Jun, JIANG Lin

2025, 0(05): 60-65. doi:10.3969/j.issn.1006-2475.2025.05.008

Asbtract ( 9 )

PDF (1418KB) ( 20 )

References | Related Articles | Metrics

The complexity of distributed systems often leads to node failures and other fault issues， which can reduce the system’s service capability and quality. To improve the reliability and stability of distributed systems， this paper proposes a fault prediction method for distributed systems based on XGBoost and LightGBM. Firstly， a multidimensional data preprocessing method is adopted to clarify data features， facilitating classification by the prediction model. Then， the extreme gradient boosting algorithm （XGBoost） is used to train the processed dataset， and feature selection is performed based on feature importance to enhance the model’s generalization ability and reduce overfitting. Finally， the optimized LightGBM algorithm is used for model training on the dataset. Experimental results show that the proposed method outperforms other classification models in terms of accuracy， precision， and recall. Compared to RF， XGBoost， and LightGBM models， the proposed method improves accuracy by 4.89%， 3.52%， and 1.39%， and enhances the F1 score by 5.56%， 3.91%， and 1.53%， respectively. This validates that the proposed model can be efficiently applied to distributed system fault type prediction scenarios.

Review of Development Trends of Modeling Languages and Tools for Airborne Software

CAO Guozhen1, PENG Han2, ZHANG Xiaoli2, JING Yuejuan2, HOU Yuanyuan2

2025, 0(05): 66-72. doi:10.3969/j.issn.1006-2475.2025.05.009

Asbtract ( 13 )

PDF (865KB) ( 19 )

References | Related Articles | Metrics

With the wide application of model-driven methods in the aerospace field， various modeling languages and modeling tools play increasingly important roles in the development of airborne software. On-board software modeling tools have evolved from initially graphical representation tools to industrial manufacturing software supporting software safety verification and test case generation to run-time verification. This paper classifies and expounds the development of airborne software modeling languages and tools in recent years from four aspects: the granularity of modeling elements， the coverage of the modeling process， the perspectives that the model focuses on， and the comprehensive capabilities of modeling tools. Suggestions are put forward for the development direction of our country’s airborne software modeling tools.

Cloud-PERM: Ab Initio Prediction Method for Protein Folding Simulation

XU Shengchao1, ZHOU Jipeng2

2025, 0(05): 73-78. doi:10.3969/j.issn.1006-2475.2025.05.010

Asbtract ( 14 )

PDF (1431KB) ( 25 )

References | Related Articles | Metrics

A cloud based ab initio prediction method for protein folding simulation calculation algorithm is proposed. Using a non template based ab initio prediction method， the energy function of the protein folding process is constructed by obtaining the optimal relationship between the spatial positions and energies of all atoms in the protein. The protein folding structure is predicted using protein fragment assembly technology. A lattice model is used to place the protein structure chains without overlap in the construction space of the lattice model. The PERM algorithm is used to find the lowest energy protein structure chain placement state， realize protein folding simulation calculation， divide the PERM algorithm based on the MapReduce programming model， and obtain the Cloud-PERM algorithm generated by the MapReduce programming module in Apache’s cloud computing platform Hadoop. Continuously solve the lattice model of protein folding simulation calculation， and obtain the protein folding simulation calculation result with the lowest energy. Through experimental analysis， it was found that the algorithm has a higher similarity in protein structure prediction， can achieve protein folding simulation calculation， and has strong computing power and fast speed. It can obtain the protein folding structure with the lowest energy in the same time with a large number of optimization times.

Global Path Planning for Unmanned Vehicles Based on Adaptive Artificial Potential Field Method

WANG Long, YANG Fengbao, YANG Tongyao

2025, 0(05): 79-85. doi:10.3969/j.issn.1006-2475.2025.05.011

Asbtract ( 7 )

PDF (2088KB) ( 23 )

References | Related Articles | Metrics

The artificial potential field method is commonly used in small-scale path planning tasks， but it faces issues such as low path safety and poor path quality when dealing with complex obstacles. To address these problems， an adaptive artificial potential field method is proposed. Firstly， the repulsive potential field function is improved by setting a dynamic gain coefficient to ensure target accessibility in complex environments. Secondly， virtual goal points are set adaptively according to the distribution of obstacles， globally guiding the unmanned vehicle out of local minima traps and avoiding undesirable motion states such as jitter or hesitation. Lastly， the step length is adjusted adaptively based on the safe distance between the unmanned vehicle and obstacles， reducing unnecessary evasive actions by the vehicle， thus improving path quality and ensuring safety during navigation. Experimental results show that compared with other methods， the adaptive artificial potential field method reduces the average number of planning cycles by 12.82%， the average turning angle by 58.36%， and the average path length by 7.11%. These results demonstrate that the adaptive artificial potential field method has higher path safety and better path quality than other improved methods.

DSA De-artifacting Algorithm Based on Deformation Field Registration

WANG Dongfang1, YANG Yan1, ZHANG Dong1, HAN Wenrui2, LI Mingchang2

2025, 0(05): 86-90. doi:10.3969/j.issn.1006-2475.2025.05.012

Asbtract ( 7 )

PDF (3706KB) ( 21 )

References | Related Articles | Metrics

Digital subtraction angiography （DSA） is commonly used in the diagnosis of vascular diseases. The temporal subtraction-based method requires subtraction of two frames before and after injection of contrast agent to obtain vascular images， but factors such as involuntary patient movement or unaligned equipment make artifacts exist in the subtraction results. In this paper， a DSA artifact removal algorithm based on deformation field registration is proposed to improve the imaging quality of blood vessels after subtraction and to remove the artifacts in the nonvascular region. Firstly， a mask is extracted from the subtraction image with artifacts before registration to separate the vascular region from the non-vascular region， then the background and contrast frames are unimodally aligned using a deformation field registration network， and finally the aligned deformation image and the background frames are digitally subtracted and re-imaged to obtain a vascular subtraction image with the elimination of artifacts. In the experimental results of the test set， the MSE， PSNR， SSIM， and Dice coefficients are 30.619， 33.396， 0.901， and 0.687， respectively， all of which are significantly higher than other traditional or deep learning registration methods. The results show that the method proposed in this paper is more effective in removing artifacts， and the quality of subtractive imaging has been improved.

Research Advances on 3D Object Detection Method Based on Visual Information and LiDAR for Intelligent Driving

WEI Yunsong1, 2, LI Jiaqiang1, 2, HE Chao1, 2, 3, YU Haisheng1, 2, CHEN Yanlin1, 2, ZHAO Longqing1, 2, WEI Rongkun1, 2

2025, 0(05): 91-102. doi:10.3969/j.issn.1006-2475.2025.05.013

Asbtract ( 13 )

PDF (3282KB) ( 25 )

References | Related Articles | Metrics

3D object detection based on visual information and LiDAR is one of the key technologies in intelligent driving perception and plays a crucial role in understanding complex driving scenarios. Due to the inherent limitations of single sensor and the complexity of multi-modal data， achieving high-quality 3D object detection is not a straightforward task. It requires considerring many factors， including the heterogeneity of the data and optimization. Current research work mainly focuses on data fusion processing by leveraging the complementarity of single-modal data. To advance further research in 3D object detection， this paper first reviews 3D object detection methods based on visual information and LiDAR and then reviews 3D object detection methods based on LiDAR-Camera fusion from the perspectives of temporal fusion and stage-wise fusion. In addition， commonly used datasets and evaluation metrics are introduced， followed by performance comparisons of diffrent network architectures on these datasets. The advantages and limitations of different network types are analyzed accordingly. Finally， the challenges and solutions for the 3D object detection method based on visual information and LiDAR are given.

Human Action Recognition Algorithm Based on Spatiotemporal Motion Model

XU Haining1, WANG Yankun2, 3, FAN Yong3 , 4, LUO Lina2, GUO Jing5

2025, 0(05): 103-110. doi:10.3969/j.issn.1006-2475.2025.05.014

Asbtract ( 14 )

PDF (3288KB) ( 21 )

References | Related Articles | Metrics

With the rapid development of human-computer interaction technology， efficient and accurate human action recognition techniques have demonstrated tremendous application potential in fields such as virtual reality and intelligent surveillance. However， due to the complexity and diversity of human actions， traditional recognition methods have limitations. Based on this， we propose a human action recognition algorithm that integrates a spatio-temporal motion model and deep learning to overcome these challenges. Our method converts depth video sequences into multi-angle depth video sequences through the rotation of coordinate systems， and utilizes an adaptive temporal model to segment the depth video sequences into several sub-actions. By accumulating the parts of the depth video images with large energy changes between adjacent frames， we form motion energy maps， while accumulating the parts with smaller energy changes forms static energy maps， collectively referred to as the Spatial-Temporal Motion Model （STMM）. A multi-channel convolutional neural network is introduced to extract dynamic and static features from the STMM， and Spatial Pyramid Histogram of Oriented Gradients （SPHOG） features extracted from the STMM serve as a complement to the features of the multi-channel convolutional neural network. Furthermore， we introduce adaptive moment estimation to adjust the learning rate of each parameter during neural network training， enhancing the efficiency and stability of the model training. We also introduce L2 norm regularization to reduce model complexity and prevent overfitting. Finally， we employ a fully connected neural network to classify the actions， achieving a high recognition rate on public datasets. The experimental results demonstrate that the human action recognition algorithm integrating spatio-temporal pyramid and deep learning is highly effective.

Glioma Segmentation and Classification Network Assisted by Object Detection

XU Ling1, ZHANG Dong1, WEN Shen1, HU Ping2

2025, 0(05): 111-116. doi:10.3969/j.issn.1006-2475.2025.05.015

Asbtract ( 11 )

PDF (2367KB) ( 19 )

References | Related Articles | Metrics

Glioma is a highly lethal primary intracranial malignant tumor. Preoperative non-invasive diagnosis is of great significance for the treatment and prognosis of glioma. In this paper， we propose a brain glioma segmentation and classification network assisted by object detection techniques. We proposed a brain glioma segmentation and classification network assisted by target detection technology. Based on multi-modal three-dimensional MRI， glioma regions were segmented and classified by WHO classification（Ⅱ/Ⅲ/Ⅳ）， IDH mutation status classification and 1p19q deletion status classification. The object detection technology was employed to obtain tumor region location information for auxiliary segmentation and classification， and modules such as SPP and FPN were also used to improve model performance. The model was trained on 664 glioma cases in the EGD data set. Finally， in the test set， the Dice score of glioma segmentation reached 0.88， and the classification accuracy of WHO classification， IDH mutation status and 1p19q deletion status reached 0.80， 0.72 and 0.90， respectively. Comparative experiments with the PSNet， ResNet50， and Unet+ResNet50 models demonstrate the effectiveness of our proposed model. At the same time， the ablation experiments of object detection module， SPP module and FPN module were carried out to verify the effect of the introduced module. The experimental results show that the model proposed in this paper can effectively perform the multi-task diagnosis of brain glioma before operation， and it is helpful for the treatment and prognosis of brain glioma.

SE-BCNN with Feature Recalibration for Fine-grained Conodont Identification

DENG Yuyan, HE Yueshun, HE Linlin, CHEN Jie, LI Juan, ZOU Zhiyi

2025, 0(05): 117-121. doi:10.3969/j.issn.1006-2475.2025.05.016

Asbtract ( 18 )

PDF (2510KB) ( 24 )

References | Related Articles | Metrics

Conodonts are internal shell fossils of ancient marine organisms， and their accurate identification is vital for understanding Earth’s climatic history and geological transitions. Traditional image recognition techniques primarily focus on broad category classification， which fails to meet the complex demands of fine-grained conodont classification and struggles to capture subtle yet critical feature differences. We propose a bilinear convolutional neural network（BCNN） enhanced with feature recalibration mechanisms. By integrating squeeze-and-excitation（SE） attention mechanisms and residual connections， the model’s feature extraction capability is significantly enhanced. The SE module recalibrates features by modeling inter-channel dependencies， while residual connections mitigate the vanishing gradient problem using skip connection， ensuring efficient feature transmission and reusing in deeper layers. Experimental results on a fine-grained conodont dataset demonstrate that the SE-BCNN outperforms existing methods in accuracy， precision， recall， and F1 score， achieving a classification accuracy of 89%， significantly surpassing models such as VGG16， ResNet18-BCNN， and CART.

Scene Semantic Segmentation Based on Regional Self-attention

SHI Xianwei1, FAN Xin2

2025, 0(05): 122-126. doi:10.3969/j.issn.1006-2475.2025.05.017

Asbtract ( 9 )

PDF (3201KB) ( 20 )

References | Related Articles | Metrics

The ability of computers to understand complex road environments is greatly enhanced through semantic segmentation based on scene images. In this paper， we propose a novel segmentation method that leverages local self-attention to model the long-range dependencies of different semantic objects， thereby improving the feature representation of semantic objects. The proposed approach also employs an extended convolutional strategy to mitigate the grid effect and adapt to variations in the size of segmented objects. Additionally， we utilize channel attention to determine the relative importance of each feature channel. The efficacy of proposed method is validated on the CamSeq01 and CamVid dataset， and experimental results demonstrate that our approach significantly outperforms general models in terms of segmentation performance.

Table of Content