Computer and Modernization

Flow Velocity Measurement Method Using MIMO Radar Based on 3D Point Cloud SVT Algorithm

LI Jian, SONG Yu, ZHANG Wenxin, YU Ran

2025, 0(12): 1. doi:10.3969/j.issn.1006-2475.2025.12.001

Asbtract ( 74 )

PDF (5943KB) ( 20 )

References | Related Articles | Metrics

Abstract: River surface velocity is a key parameter in hydrological monitoring, as it provides essential information for understanding hydrological conditions, regulating water volume, and preventing flood disasters. Compared with traditional contact-based flow velocity measurement methods, radar-based flow velocity measurement technology boasts advantages such as non-contact measurement (which is not affected by water body conditions) and real-time monitoring capabilities. To improve the accuracy and real-time performance of river channel flow velocity measurement, this study adopts a MIMO radar flow measurement method based on the 3D point cloud SVT algorithm. Through processing steps including projection filtering, grid partition denoising, and scale correction applied to the point cloud across three dimensions—space, velocity, and time—this method reduces noise interference and data redundancy, enhances the accuracy and stability of flow velocity measurement, and further enables the intuitive presentation of multi-point flow velocity distribution on the river channel surface. The radar can be installed in a shore-based lateral manner, which lowers the requirements for the installation environment and realizes multi-point flow velocity measurement on the river channel surface. Experimental results show that in medium-to-high flow velocity scenarios (v≥0.5 m/s), the relative error is less than 5%; in low flow velocity scenarios (0.3 m/s≤v＜0.5 m/s), the absolute error is less than 5 cm/s. In summary, this method can effectively improve the accuracy and stability of river channel flow velocity measurement, providing technical support for hydrological monitoring.

Resource Adaptive Allocation Algorithm for Multi-agent with Dynamic Demands

HU Kaiming1, ZHUANG Yi1, XU Tao1, 2

2025, 0(12): 11-18. doi:10.3969/j.issn.1006-2475.2025.12.002

Asbtract ( 57 )

PDF (2494KB) ( 8 )

References | Related Articles | Metrics

Abstract: To address the low adaptability of existing resource allocation algorithms for multi-agent systems in dynamic environments and their lack of fairness in resource distribution， this paper proposes a resource adaptive allocation algorithm for multi-agent systems oriented to dynamic demands. The proposed algorithm enhances resource utilization and fairness while optimizing allocation efficiency and system load balancing. A resource adaptive allocation model is constructed by incorporating the heterogeneity of resources and the dynamic nature of task demands in multi-agent systems. Through multi-stage optimization， the algorithm improves adaptability to dynamic environments and employs multi-objective optimization to balance resource allocation fairness and utilization. First， a fairness-based allocation strategy is used to select the set of tasks to be assigned. Then， inspired by swarm intelligence optimization， a Decentralized Search and Competitive Optimization （DSACO） algorithm is introduced to iteratively determine resource allocation schemes through multi-stage resource distribution and optimization. Finally， tasks are automatically assigned to corresponding agents based on the determined allocation schemes. Comparative simulation results demonstrate that， compared to existing algorithms， the proposed algorithm achieves not only fair resource allocation but also improves resource utilization and allocation efficiency. It exhibits strong adaptability to dynamic environments， providing an effective solution to the resource allocation challenges faced by multi-agent systems in complex scenarios.

Fault Diagnosis of Wind Turbine Gearbox Based on Convolutional Dual Channel Multilayer Perceptron Mixer and Weighted Voting Mechanism

WANG Wanglong1, XU Junyang2, SU Peng1, FU Wenlong3

2025, 0(12): 19-25. doi:10.3969/j.issn.1006-2475.2025.12.003

Asbtract ( 67 )

PDF (2637KB) ( 9 )

References | Related Articles | Metrics

Abstract: The gearbox is one of the key and vulnerable components in wind turbines， and the fault diagnosis of its health condition is of great significance to reduce operation and maintenance costs and improve cost efficiency. Therefore， a fault diagnosis method of wind turbine gearbox based on convolutional double-channel multi-layer perceptron mixer and weighted voting mechanism is proposed. Firstly， the original vibration signal is transformed into two-dimensional time-frequency image by continuous wavelet transform. Then， the local features of 2D time-frequency images are extracted by using a two-dimensional convolutional network， and the time-frequency image data is divided into non-overlapping patches. A two-channel multi-layer perceptron mixer network is constructed to extract the global features. Finally， the extracted two global feature vectors are weighted to get the final feature representation， and the final fault diagnosis result is obtained through the full connection layer classification. The test results of UConn gearbox dataset show that the proposed method has higher diagnostic performance than other traditional methods， and achieves the highest diagnostic accuracy of 100%.

Intelligent Proctoring System Based on Edge Computing

LI Jingyang1, XUE Huajian2, 3, YANG Yong1, REN Ge1

2025, 0(12): 26-31. doi:10.3969/j.issn.1006-2475.2025.12.004

Asbtract ( 63 )

PDF (2852KB) ( 10 )

References | Related Articles | Metrics

Abstract: To address the lack of remote supervision technologies, non-traceability of exam behavior, and difficulty in real-time verification of examinee identities in online remote examinations, this paper proposes an intelligent proctoring system based on edge computing. The system consists of exam terminals, teacher terminals, exam site servers, and a provincial service cluster. On the exam terminals, the CLAHE method is used to equalize image illumination before the images are transmitted to the model, with OpenVINO employed to accelerate the detection process. During the examination, the camera randomly captures desktop images containing the examinee’s face, and these images are processed by the model on the exam terminal. These images are temporarily stored on the exam site server, and after the exam, the records are sent to the provincial cluster server for post-exam verification. Experimental results demonstrate that the system has low hardware performance requirements, improves robustness to lighting variations, achieves a model accuracy of approximately 95%, and shows better fault tolerance in network performance. The system has already been deployed and applied in a remote border region.

Personalized Literature Recommendation Algorithm Based on Log Mining

ZHANG Ya, BAI Haiyan, MENG Xuyang

2025, 0(12): 32-37. doi:10.3969/j.issn.1006-2475.2025.12.005

Asbtract ( 52 )

PDF (1177KB) ( 10 )

References | Related Articles | Metrics

Abstract: In recent years, the exploration of personalized recommendation technology has increasingly occupied the core position in the field of recommendation system research. Among them, literature recommendation, as an important branch of this field, is receiving more and more attention and in-depth research. For the problems of that the current scientific and the literature recommendation algorithm in technological literature discovery system do not take into account the user’s behavior preferences and a large number of users cannot be recommended with literature due to sparse user data, this study proposes a personalized literature recommendation algorithm based on log mining. First of all, this study initializes the user behavior record log into a user-literature scoring matrix, and then in the process of in-depth analysis of user interest preferences, comprehensively considers the characteristics of global projects and local scoring information, and designs a user interest preference algorithm with the principle of Haiming’s closeness. Finally, the algorithm is fused with the weighted JMSD similarity algorithm to expand the coverage of the recommendation algorithm, and then greatly improve the accuracy of the recommendation results. The actual user behavior log data of NSTL literature service system was used in the experiment. The experimental results show that the algorithm’s overall performance is better than other baseline algorithms, with an average improvement of 0.8%, 3.2% and 1.9% in Precision, Recall and F1, which verified the effectiveness of the proposed algorithm. The research results can be applied to the literature discovery platform to further strengthen the intelligent level of document information resource service.

MPI-based Heterogeneous Computing Resource Integration and Scheduling Platform

YE Ning, FU Kang, HU Shaowen, GONG Yifeng, WANG Kang, YANG Yuxian

2025, 0(12): 38-45. doi:10.3969/j.issn.1006-2475.2025.12.006

Asbtract ( 57 )

PDF (1993KB) ( 11 )

References | Related Articles | Metrics

Abstract: Aiming to the problem that high-performance computing centers， especially small and medium-sized computing centers， are unable to undertake large-scale computing jobs due to the decentralization of heterogeneous computing resources， this paper designs and implements a heterogeneous computing resource integration and scheduling platform to realize the unified management of heterogeneous computing resources such as X86， ARM and so on， as well as collaborative computing. The platform adopts a layered fusion scheduling architecture， utilizes cluster manager server （CMS） and job manager client （JMC） to dynamically monitor the resource status， and realizes collaborative parallel computing among heterogeneous computing nodes with the help of job scheduler （JS）. Through the master-slave JMC process collaboration and MPI reduction mechanism， cross-architecture data synchronization at the physical machine level is achieved， and parallel execution of a single job on heterogeneous computing nodes at the physical machine level is realized for the first time. To address the long-tail delay effects and synchronization overhead caused by performance imbalances in heterogeneous clusters， this paper proposes a deadline-constrained minimal resource algorithm （DCMR）， which minimizes computational resource consumption while ensuring job completion deadlines are met. Test results show that the platform has almost no loss of computing performance in heterogeneous environments， and the DCMR algorithm effectively improves the utilization efficiency of heterogeneous computing resources， providing a reliable system solution to deal with heterogeneous computing environments.

Multi-view IM-NET for Fine Reconstruction of 3D Object

LIU Jianlong1, CEN Ying2, XU Bin3, JIAO Xuan4

2025, 0(12): 46-53. doi:10.3969/j.issn.1006-2475.2025.12.007

Asbtract ( 49 )

PDF (1768KB) ( 11 )

References | Related Articles | Metrics

Abstract: Implicit net (IM-NET) converts the task of reconstructing three-dimensional (3D) objects into a classification issue concerning whether sampled points in space are on or inside the object’s surface, effectively conserving computational and storage resources. However, IM-NET can only reconstruct 3D objects from a single view, and the limited amount of target information available in a single view, especially the missing information from occluded parts, leads to insufficient reconstruction accuracy. This paper extends IM-NET to adapt multi-view inputs by employing an attention module to fuse extracted features, thereby obtaining more complete target features and enhancing 3D reconstruction accuracy. Considering that the target surface generated by the Marching Cubes algorithm is not smooth enough when converting implicit expression into explicit mesh representation, this paper utilizes a mesh refinement algorithm to refine the reconstructed targets iteratively, achieving the refined reconstruction. Experimental results on the ShapNet dataset indicate that compared with single-view IM-NET and other multi-view reconstruction methods, the proposed multi-view IM-NET reconstructs 3D targets more complete and smooth, and the average intersection and union ratios of targets are significantly improved. Additionally, visualization effects show that the refined targets have richer details and smoother surfaces than those without refinement.

Construction of Knowledge Graph for Port Equipment Faults from Perspective of Bibliometrics

LI Taoying, LIU Quan, DONG Zhiyu, HAN Jiawen

2025, 0(12): 54-60. doi:10.3969/j.issn.1006-2475.2025.12.008

Asbtract ( 41 )

PDF (2592KB) ( 9 )

References | Related Articles | Metrics

Abstract: Once the port equipment malfunctions, it will seriously affect the efficiency of port production operations and cause safety hazards. Although a large number of scientific research literatures on equipment faults of port are gathered, they often have inconsistent expressions due to different focuses, which increases the difficulty for researchers and managers to timely grasp the current research status and quickly acquire knowledge. Constructing knowledge graph for port equipment faults helps to conduct knowledge queries and intelligent Q&A on port equipment faults, enhancing the efficient management and dissemination of knowledge on port equipment failures. To address this, this study adopts titles, abstracts, and keywords from Chinese literature on port equipment failures as dataset text and then employs the BIO data annotation method and the BERT-BiLSTM-CRF model to extract information about port equipment failures from the text. By introducing ontologies to formally describe entities and their relationships for solving inconsistencies and forming triplets, a knowledge graph of port equipment failures is constructed. This graph supports knowledge queries and intelligent Q&A related to port equipment failures. The construction of a crane failure knowledge graph serves as a case, validating the effectiveness of the knowledge graph construction, thereby enhancing the efficient management and dissemination of knowledge about port equipment failures.

Contrastive Learning Method for Detecting Malicious Encrypted Traffic

WU Jiahong

2025, 0(12): 61-65. doi:10.3969/j.issn.1006-2475.2025.12.009

Asbtract ( 50 )

PDF (2592KB) ( 10 )

References | Related Articles | Metrics

Abstract: To address the issue of insufficient representation capability in malicious encrypted traffic detection models， a malicious encrypted traffic detection method based on contrastive learning is proposed， with the goal of enhancing the model’s representation ability and thereby improving the detection accuracy of malicious encrypted traffic. This method diverges from traditional approaches that directly extract features from traffic data， focusing instead on learning the intrinsic representations of the data prior to feature extraction. Specifically， local and global features of encrypted traffic are extracted using a multi-scale mechanism to capture key information at different scales. Then， in the metric space of contrastive learning， the distance between encrypted traffic and the correct classification label is minimized， while the distance from the incorrect classification label is maximized by optimizing the objective function， enabling the model to better distinguish between malicious and normal encrypted traffic. After training， the model captures more discriminative features of encrypted traffic， ultimately improving detection accuracy. The experimental dataset is composed of sampling from multiple public datasets including UNSW NS 2019， CICIDS-2017， CIC-AndMal 2017， Malware Capture Facility Project Dataset， and CICIDS-2012. The results show that the method achieves 97.59% detection accuracy， exceeding comparative models， with 3.16 percentage points increase over the random forest benchmark. Furthermore， the interpretability and detection rate of the method are also improved.

KD-NeRF: Knowledge Distillation-enhanced Neural Radiance Field Reconstruction Method for Fruit Trees

TANG Hui1, LIU Xin2, HU Biwei1, LIU Fan1

2025, 0(12): 66-73. doi:10.3969/j.issn.1006-2475.2025.12.010

Asbtract ( 48 )

PDF (3685KB) ( 9 )

References | Related Articles | Metrics

Abstract: Precision agriculture urgently demands accurate three-dimensional phenotypic reconstruction of fruit trees. Traditional methods suffer from low modeling efficiency and high computational costs, while two-dimensional analysis is constrained by occlusion and a lack of depth information. Although Neural Radiance Fields (NeRF) demonstrate excellent performance, they are time-consuming and resource-intensive to train, facing significant challenges in learning complex fruit tree structures. This study proposes a knowledge distillation-enhanced NeRF method for three-dimensional fruit tree reconstruction (KD-NeRF). We construct a deep learning framework based on two-dimensional multi-view images and employ SIREN periodic activation functions to enhance high-frequency detail capture and spatial continuity through implicit regularization. A teacher-student knowledge distillation mechanism is designed to transfer deep representations to lightweight networks, improving the learning efficiency for complex structures. Experiments on diverse fruit tree datasets demonstrate that compared to NeRF, KD-NeRF achieves an 8% improvement in PSNR with 14 times ~16 times acceleration in training speed. The results of the ablation experiment confirm that the teacher-student architecture significantly improves PSNR, SSIM, and other metrics; SIREN enhances high-frequency detail representation; and their synergy produces performance gains exceeding simple additive effects. KD-NeRF addresses the efficiency bottleneck of NeRF in fruit tree reconstruction, providing efficient technical support for fruit tree phenotypic analysis, growth monitoring, and intelligent breeding, thereby advancing precision agriculture and smart orchard construction.

Camouflaged Bird Object Detection Based on Discrepancy Sense in Substation

ZHAO Xinyang1, 2, LIU Zhiyuan1, 2, ZHANG Leyi3, YIN Qiyun1, 2, LU Hongjian1, 2, LI Qingwu3

2025, 0(12): 74-80. doi:10.3969/j.issn.1006-2475.2025.12.011

Asbtract ( 63 )

PDF (7177KB) ( 10 )

References | Related Articles | Metrics

Abstract: This paper proposes a camouflaged small object detection network based on discrepancy sense in substation to address the challenge of detecting camouflaged bird objects in the environment of substations， characterized by small sizes and high similarity with surrounding backgrounds leading to low segmentation accuracy of models. The model leverages a global guidance extraction module to obtain global guidance， preserving original detailed information while enlarging the receptive field. Additionally， a boundary guidance generation module is employed to fuse features from all scales to obtain boundary guidance， thereby mitigating noise interference caused by inter-layer feature interactions. Furthermore， a dual-branch discrepancy perception module is utilized to integrate multiple guidance， progressively refining segmentation results by alternating attention between the target boundary and the surrounding background to amplify their discrepancies layer by layer. Experimental results on a self-built dataset of camouflaged small bird objects around substations demonstrate that the proposed method achieves an improvement of 2.75 percentage points in Intersection over Union compared to other camouflaged object detection algorithms， offering a reliable basis for effectively deterring camouflaged bird objects in substation.

No-reference Image Quality Assessment Based on DenseNet and Meta-learning

LIU Ziyang, JIA Huizhen, WANG Tonghan

2025, 0(12): 81-87. doi:10.3969/j.issn.1006-2475.2025.12.012

Asbtract ( 51 )

PDF (1386KB) ( 10 )

References | Related Articles | Metrics

Abstract: To address the issues of overfitting and generalization in No-reference Image Quality assessment （NR-IQA） models based on convolutional neural networks under limited datasets and complex distortion conditions， this paper utilizes meta-learning to acquire shared experiential knowledge across different tasks， thereby enhancing the model’s generalization to unknown tasks. Using DenseNet as the backbone network to extract image features， we achieve comprehensive deep supervision， improving the information flow and gradients within the network， and reducing overfitting on small sample training tasks. Additionally， a multi-head self-attention mechanism is incorporated， allowing the network to capture diverse feature information and long-range dependencies of the global image from different subspaces， enhancing the model’s learning capability. A bi-level gradient optimization method from the support set to the query set is employed to train the quality prior model on various known distortion tasks， optimizing the subsequent gradient descent process of the model parameters. Fine-tuning is performed on the target NR-IQA task， where the model can quickly adapt to unknown distortion tasks under appropriately initialized parameters. Performance and generalization tests were conducted on the authentically distorted IQA dataset LIVEC and the synthetically distorted IQA dataset KADID-10K， where the SROCC values reached 0.834 and 0.831， respectively. The results indicate that the proposed model has better learning ability and generalization compared to traditional algorithms.

Multi-experts Contrastive Learning Method for Hand Hygiene Assessment

TU Zijian1, 2, WANG Zi2, TANG Jin1, 2, 3

2025, 0(12): 88-96. doi:10.3969/j.issn.1006-2475.2025.12.013

Asbtract ( 54 )

PDF (2308KB) ( 10 )

References | Related Articles | Metrics

Abstract: Hand hygiene is one of the most effective measures for preventing hospital infections， yet compliance among healthcare workers remains relatively low. While existing computer vision-based hand hygiene evaluation methods can perform step-by-step scoring of hand hygiene action， they still struggle with accurately perceiving and assessing fine movements. To address this issue， a novel hand hygiene assessment method that employs a segmentation module with multi-experts and an action contrast evaluation module is proposed. Through a two-stage evaluation process， the method aims to enhance the segmentation and assessment of fine movements， thereby enabling a more accurate evaluation of hand hygiene action quality. Specifically， the segmentation module with multi-experts learns the characteristics of each effective hand hygiene action firstly， and performs high-precision action segmentation reasoning based on characteristic information. Secondly， the action contrast evaluation module uses a contrastive learning approach that leverages the differences between example actions and current actions to calculate the action prediction score. Ultimately， the proposed method outputs predicted scores for each effective action and computes the final predicted score through comprehensive calculation. The proposed method achieves an action segmentation accuracy of 91.4% and a correlation coefficient of 0.864 for action quality assessment on the hand hygiene assessment dataset HHA300， both superior to existing hand hygiene assessment methods. Multiple comparative and ablation experiments demonstrate the effectiveness of each module in the method， indicating that it standardizes the hand hygiene assessment process and enables effective monitoring of hand hygiene action.

Face Sketch-photo Synthesis Network Based on Attention Mechanism

YAO Li1, 2, ZHAN Bosi3, WAN Weiguo3, LUO Yitao3, YANG Yuxian4

2025, 0(12): 97-106. doi:10.3969/j.issn.1006-2475.2025.12.014

Asbtract ( 50 )

PDF (5207KB) ( 11 )

References | Related Articles | Metrics

Abstract: Face sketch-photo synthesis is an important branch of image transformation, with broad applications in digital entertainment, public security, and criminal investigation. Since face sketch images only contain grayscale information and lose most of the texture details of the face, existing methods often suffer from structural deficiencies, insufficient detail information, and color distortion in the synthesized face photos. To address these issues, this paper proposes a face sketch-photo synthesis network based on an attention mechanism. First, this paper designs a generator by combining Vision Transformer and U-Net to effectively extract global and local facial features, improving the structural integrity of the generated face photos. Additionally, an improved selective kernel attention module is constructed to enhance the model’s ability to capture fine details, enabling the generated images to retain more facial texture information. Finally, this paper designs a discriminator based on channel and pixel-wise attention to strengthen the adversarial learning capability of the generative adversarial network (GAN), reducing color distortion in the synthesized face photo images. Through subjective and objective experiments comparing with other state-of-the-art methods, the proposed approach demonstrates superior performance in both visual quality and objective metrics for face sketch-photo synthesis. On the CUHK, AR, and XM2VTS face sketch datasets, the proposed method achieves 11.6%, 6.2%, and 4.5% improvements in SSIM metric over the second-best results, respectively, proving the effectiveness of the proposed method.

Medical Image Registration Network Based on Efficient Cross-attention

HUANG Yeqin

2025, 0(12): 107-114. doi:10.3969/j.issn.1006-2475.2025.12.015

Asbtract ( 50 )

PDF (2881KB) ( 7 )

References | Related Articles | Metrics

Abstract: With the widespread application of deep learning in medical image analysis， significant progress has been made in medical image registration methods. However， existing approaches still have limitations in feature extraction and fusion， particularly in handling independent anatomical structure information between image pairs. To address this issue， this paper proposes an efficient cross-attention-based medical image registration network. The network employs two parallel branches and independently processes and fuses image features through the cross-attention mechanism， effectively improving registration accuracy. To reduce computational complexity and memory consumption， this paper introduces an efficient cross-attention mechanism that preserves global feature capturing capability while enhancing computational efficiency. Additionally， the proposed model combines Transformer and Convolutional Neural Networks （CNN）， utilizing the Transformer to capture long-range dependencies and the CNN to extract local features. This approach reduces the model’s parameter count and improves training efficiency. To evaluate the performance of the proposed model， experiments were conducted on the OASIS and BraTs2018 datasets. The proposed model achieves Dice coefficients of 0.804 and 0.732 on these two datasets， respectively， demonstrating superior registration performance compared to other methods. These experimental results indicate that the proposed model not only improves the accuracy of medical image registration but also optimizes computational efficiency， making it highly applicable in various scenarios.

Facial Color Diagnosis Classification Method Optimized by Meta-learning

LUO Guancong1, FENG Yue1, XU Hong2, QING Chuanbo1, LI Fufeng3, QIAN Peng3, LIU Huilin3

2025, 0(12): 115-122. doi:10.3969/j.issn.1006-2475.2025.12.016

Asbtract ( 52 )

PDF (1854KB) ( 11 )

Table of Content