Table of Content

30 April 2025, Volume 0 Issue 04

Previous Issue   

Gaze Estimation Model Based on Hybrid Transformer 

CHENG Zhang, LIU Dan, WANG Yanxia

2025, 0(04):  1-5.  doi:10.3969/j.issn.1006-2475.2025.04.001

Asbtract ( 54 )   PDF (1655KB) ( 44 )  

References | Related Articles | Metrics

 Combined CNN and Transformer， Transformer can gain the advantage of global feature information and improve the awareness of model context information， which can lead to improve model accuracy. A novel gaze estimation model RN-SA（ResNet-MHSA） based on a hybrid Transformer is proposed. In this model， part of the 3×3 spatial convolution layers in ResNet18 are replaced with a block composed of a 1×1 spatial convolution layer and MHSA（Multi-Head Self-Attention） layer， and the DropBlock mechanism is added to the model structure to increase the robustness of the model. Experimental results show that RN-SA model can improve the accuracy of the model while reducing the number of parameters compared with the current better model GazeTR-Hybrid， RN-SA model can improve the accuracy by 4.1% and 3.7% on EyeDiap and Gaze360 datasets， respectively， while the number of parameters is reduced by 15.8%.  Therefore， the combination of CNN and Transformer can be effectively applied to gaze estimation tasks.

Power Load Forecasting Based on TCN and Lightweight Autoformer

LI Ming, SHI Chaoshan, WEN Guihao, LUO Yonghang, TAN Yunfei

2025, 0(04):  6-11.  doi:10.3969/j.issn.1006-2475.2025.04.002

Asbtract ( 36 )   PDF (3080KB) ( 32 )  

References | Related Articles | Metrics

The accuracy of power load forecasting is crucial for energy conservation and emission reduction， and higher accuracy can enable power companies to make more reasonable plans and improve economic benefits. Although Autoformer， based on the improved Transformer architecture， has achieved good results in sequence prediction tasks， it did not fully consider the causal relationship of time when extracting temporal features， and there is too much redundant information in the attention layer， which leads to a decrease in model accuracy and memory consumption. To address these issues， this paper proposes a power load forecasting method that combines Time Convolutional Network （TCN） and an improved lightweight Autoformer model. Firstly， a time convolutional network is introduced into the Autoformer encoder to provide a larger receptive field and fully consider the causal relationship of the samples. Then， a distillation mechanism is added between the autocorrelation attention layers to reduce the number of model parameters. Finally， the results of experiment on five public datasets showed that the lightweight Autoformer combined with TCN reduced MSE and MAE by 8.95% to 32.40% and 4.91% to 15.51% respectively compared to the original model， and the prediction performance is significantly better than the other four mainstream methods， demonstrating its excellent performance.

Traffic Accident Prediction Method Based on Graph Attention and Graph Convolutional Network

ZHANG Jinsong, YUAN Yibo, MA Yuxin

2025, 0(04):  12-18.  doi:10.3969/j.issn.1006-2475.2025.04.003

Asbtract ( 25 )   PDF (1355KB) ( 28 )  

References | Related Articles | Metrics

 Traffic accidents result in significant losses to individuals and society. To enhance the accuracy of traffic accident prediction， a traffic accident prediction method based on graph attention and graph convolutional networks (GAGC) is proposed. Firstly， the method extracts complex edge feature information in the road network through an edge feature extraction module. Then， it introduces a graph attention layer to enable the network quickly focusing on nodes with frequent accidents， and uses overlapping graph attention layers to reduce information loss during feature information transmission. It also employs Dropout and Batch Normalization (BN) to balance feature importance and improve the generalization and robustness of the model. Experimental results show that GAGC achieves good results， and the model can fully consider the geospatial features in complex road networks， with better performance than five baseline models in terms of F1 index， AUC， and MAP. The ablation experiment further verifies the effectiveness and reliability of the GAGC model designed in this study.

ICS-ResNet： A Lightweight Network for Maize Leaf Disease Classification 

JI Zhengjie, WEI Linjing

2025, 0(04):  19-28.  doi:10.3969/j.issn.1006-2475.2025.04.004

Asbtract ( 28 )   PDF (7723KB) ( 27 )  

References | Related Articles | Metrics

Accurate identification of maize leaf diseases plays a crucial role in preventing crop diseases and improving maize yield. However， plant leaf images are often affected by various factors such as complex backgrounds， climate conditions， lighting， and imbalanced sample data. To enhance recognition accuracy， this study proposes a lightweight convolutional neural network named ICS-ResNet， which is based on the ResNet50 backbone network and incorporates improved spatial and channel attention modules along with depthwise separable residual structures. The residual connections in the ResNet architecture prevent gradient vanishing during deep network training. The improved channel attention module （ICA） and spatial attention module (ISA) fully leverage semantic information from different feature layers to precisely localize key network features. To reduce the number of parameters and computational costs， traditional convolution operations are replaced with depthwise separable residual structures. Additionally， a cosine annealing learning rate strategy is employed to dynamically adjust the learning rate， mitigating training instability， enhancing the model's convergence ability， and preventing it from getting trapped in local optima.Finally， experiments were conducted on the Corn dataset from PlantVillage， comparing the proposed lightweight network with six other popular networks， including CSPNet， InceptionNet_v3， EfficientNet， ShuffleNet， and MobileNet. The results demonstrate that the ICS-ResNet model achieves an accuracy of 98.87%， outperforming the other six networks by 5.03， 3.18， 1.13， 1.81， 1.13， and 0.68 percentage points， respectively. Moreover， compared to the original ResNet50， the parameter size and computational cost are reduced by 16.27 MB and 2.25 GB， respectively， significantly improving the efficiency of maize leaf disease classification.

UAV Small Target Detection Based on XMB-YOLOv5s

ZHUANG Yu, FU Xiaojin, LI Sha, WU Zheng

2025, 0(04):  29-35.  doi:10.3969/j.issn.1006-2475.2025.04.005

Asbtract ( 20 )   PDF (6112KB) ( 20 )  

References | Related Articles | Metrics

From the drone viewpoint， the detection of dense， small targets faces various shortcomings， such as low accuracy， false detection of certain targets， and missed detections. To address these issues， this paper proposes a drone-based small target detection technique using XMB-YOLOv5s. Firstly， a small target detection layer is adopted for more effective extraction and utilization of detail information within the image. Secondly， the structured embedding of BottleneckCSP and C3TR modules is used to update the C3 module to reduce computational complexity and improve training efficiency. Subsequently， the integration of the CBAM attention mechanism enables the network to better recognize and process features， thus enhancing image recognition accuracy. Finally， the Focal-EIoU Loss is employed to resolve the insensitivity of the CIoU Loss to small target detection. The experimental results indicate that， compared with traditional YOLOv5s algorithm， the XMB-YOLOv5s algorithm has increased P by 4.6 percentage points， R by 4.4 percentage points， mAP50 by 4.9 percentage points， mAP75 by 5.1 percentage points， mAP50-95 by 4 percentage points on the VisDrone2019 data set， providing a novel approach for small target detection in drone applications.

Label-independent Information Compression for Heterogeneous Graph Representation

MA Jian, WANG Yifei, MENG Li, HE Yunfei, YANG Fei

2025, 0(04):  36-41.  doi:10.3969/j.issn.1006-2475.2025.04.006

Asbtract ( 20 )   PDF (1935KB) ( 32 )  

References | Related Articles | Metrics

 The existing methods for heterogeneous graph （HG） representation are mostly based on powerful graph neural networks， which aggregate semantic information within and between meta-paths to embed nodes. However， these existing approaches overlook the heterogeneity of nodes in HG， causing irrelevant information from neighboring nodes to spread along graph structures to higher-order nodes， disturbing the HG representation. To overcome this problem， this paper proposes a heterogeneous graph representation method called Label-Independent Compression for Heterogeneous Graph （LICHGR）. The core idea of LICHGR is， under the guidance of the Information Bottleneck， to utilize the Hilbert-Schmidt Independence Criterion to restrict the propagation of label-independent information in heterogeneous graph while preserving label-dependent information as much as possible. Specifically， LICHGR constructs multi-faceted label-independent compression constraints among input features， hidden features within meta-paths， and true labels， extracting rich label-dependent information to enhance the quality of heterogeneous graph representation. Multiple experiments designed on three public datasets validate the effectiveness of LICHGR.

Graph Neural Network-based Multi-agent Reinforcement Learning for Adversarial Policy Detection Algorithm

SUN Qining1, GUI Zhiming1, LIU Yanfang2, FAN Xinxin3, LU Yunfeng4

2025, 0(04):  42-49.  doi:10.3969/j.issn.1006-2475.2025.04.007

Asbtract ( 24 )   PDF (2811KB) ( 19 )  

References | Related Articles | Metrics

In a multi-agent environment， the reinforcement learning model has security vulnerabilities in coping with adversarial attacks and is susceptible to adversarial attacks， of which adversarial policy-based adversarial attacks are more difficult to defend against because they do not directly modify the victim’s observations. To solve this problem， this paper proposes a graph neural network-based adversarial policy detection algorithm， which aims to effectively identify malicious behaviors among agents. This paper detects adversarial policy by training the graph neural network as an adversarial policy detector by employing alternative adversarial policies during the collaboration process of the agents， and calculates the trust scores of the other agent based on the local observations of the agents. The detection method in this paper provides two levels of granularity； adversarial detection at the game level detects adversarial policies with very high accuracy， and time-step level adversarial detection allows for adversarial detection at the early stage of the game and timely detection of adversarial attacks. This paper conducts a series of experiments on the StarCraft platform， whose results show that the detection method proposed in this paper can achieve an AUC value of up to 1.0 in detecting the most advanced adversarial policy-based adversarial attacks， which is better than the state-of-the-art detection methods. The detection method in this paper can detect adversarial policy faster than existing methods， and can detect the adversarial attack at the 5th time step at the earliest. Applying this paper’s detection method to adversarial defense improves the win rate of the attacked game by up to 61 percentage points. In addition experimental results show that the algorithm in this paper is highly generalizable and the detection method in this paper does not need to be trained again and can be directly used to detect observation-based adversarial attacks. Therefore， the method proposed in this paper provides an effective adversarial attack detection mechanism for reinforcement learning models in a multi-agent environment. 

UAV Path Planning Based on YOLO and PPO

ZHANG Huiyu1, LIU Lei1, YAN Dongmei2, LIANG Chengqing3

2025, 0(04):  50-55.  doi:10.3969/j.issn.1006-2475.2025.04.008

Asbtract ( 27 )   PDF (6883KB) ( 26 )  

Related Articles | Metrics

 This paper proposes an unmanned aerial vehicle path planning method based on deep reinforcement learning for complex and ever-changing three-dimensional unknown environments. This method optimizes strategies within a limited observation space to address the challenges posed by high complexity and uncertainty. Firstly， within a limited perceptual range， the YOLO network is used to extract obstacle information from the image information. Secondly， this paper designs hazard levels to address the issue of varying amounts of obstacle information at different times， and combines the extracted information from hazard levels with radar information as input to the intelligent agent. Finally， based on the proximal strategy optimization algorithm， an action selection strategy under state decomposition is designed to improve the effectiveness of drone actions. Through simulation evaluation in Gazebo， the experimental results show that compared to the proximal strategy optimization algorithm， the average reward per round has increased by 15.6 percentage points， and the average success rate has increased by 2.6 percentage points.  

A3C Based Task Offloading and Resource Allocation Algorithm for Internet of Vehicles

WU Yichuan

2025, 0(04):  56-62.  doi:10.3969/j.issn.1006-2475.2025.04.009

Asbtract ( 21 )   PDF (5730KB) ( 20 )  

References | Related Articles | Metrics

Mobile Edge Computing （MEC）， as a new technology， provides a new solution for the application of the Internet of Vehicles. However， the limited resources in the connected vehicle environment cannot meet the needs of the connected vehicle equipment， which leads to an increase in the service response time and execution energy consumption of tasks， which greatly affects the Quality of Experience （QoE） of users. In order to reduce the delay and energy consumption of task execution and improve the flexibility of algorithm deployment， this paper constructs the networked vehicle system model and proposes an asynchronous advantage actor-critic based task offloading and resource allocation strategy. The algorithm framework uses asynchronous updating to train the model， and adds time attenuation coefficient to reduce the adverse effect of backward model on global model updating. Experimental results show that the proposed algorithm can effectively improve model training efficiency and reduce task execution delay and energy consumption.

Predictive Modeling of Ash Conveying in Thermal Power Plants Based on RFECV-XGBoost and SHAP

LIU Wenxin1, XU Wenhui1, CHEN Zhaoye1, GU Haiying1, WEN Cong2, YAO Yulong2, ZENG Xi2

2025, 0(04):  63-69.  doi:10.3969/j.issn.1006-2475.2025.04.010

Asbtract ( 23 )   PDF (2754KB) ( 21 )  

References | Related Articles | Metrics

 Accurate prediction of ash transportation output in the power ash transportation system of thermal power plants is of great significance for improving the overall efficiency of power generation. At present， the pneumatic ash transportation systems in thermal power plants mainly rely on manual experience for operation. Based on this， an intelligent ash transportation prediction model based on the XGBoost （eXtreme Gradient Boosting） and SHAP （Shapley Additive Explanation） framework is proposed. Firstly， data such as air pressure and equipment temperature are acquired from the DCS （Distributed Control System） of the power plant’s ash transportation system. Secondly， to enhance the accuracy of the model predictions and prevent overfitting， RFECV （Recursive Feature Elimination with Cross-Validation） is used for feature selection. The selected feature set is then imported into the XGBoost-based intelligent ash transportation prediction model. Concurrently， the SHAP model is utilized for visual causal analysis， thereby discovering useful information from the power ash transportation data to form a knowledge base for the power plant’s ash transportation system， aiming to achieve more intelligent and efficient operation. The research results can provide data support for early warning technology for ash transportation in thermal power plants and the intelligent upgrade of ash transportation systems， which helps to energy saving and consumption reduction in power plant ash transportation systems.

Identification Method for Potential Debris Flow Basins in the Wenchuan Earthquake-Affected Area Based on CNN-KAN

ZHOU Jing1, 2, LIU Dunlong1, 2, SANG Xuejia1, 2, ZHANG Shaojie3, YANG Hongjuan3

2025, 0(04):  70-76.  doi:10.3969/j.issn.1006-2475.2025.04.011

Asbtract ( 23 )   PDF (9042KB) ( 19 )  

References | Related Articles | Metrics

 The identification of potential debris flow basins often faces challenges such as unscientific watershed division criteria， unreasonable selection of non-debris flow basins, and insufficient model accuracy. A method combining river network density with the self-organizing map（SOM）is proposed to accurately determine the optimal catchment area threshold for watershed division， with the SOM used to generate representative non-debris flow basins. A CNN-KAN model， based on an improved traditional CNN architecture, is constructed to enhance identification accuracy. Experimental results indicate that the CNN-KAN model achieves a recognition accuracy of 92.9%， outperforming multilayer perceptron， KAN， and CNN models in precision， recall, F1 score， and AUC. The identified potential debris flow basins can serve as essential computational units and focal areas for debris flow early warning in the region. 

Aircraft 4D Track Tracking Algorithm Based on Cloud Platform and Eye Movement Fixation Point Data

CHEN Peng

2025, 0(04):  77-82.  doi:10.3969/j.issn.1006-2475.2025.04.012

Asbtract ( 14 )   PDF (1872KB) ( 20 )  

References | Related Articles | Metrics

A 4D aircraft track tracking algorithm based on cloud platform and eye movement fixation point data is proposed to track aircraft tracks in real time， detect and correct potential safety hazards in time， and reduce the occurrence of flight accidents. The 4D aircraft track tracking architecture is built with the help of cloud platform， which includes five levels of data acquisition， transmission， storage， processing and application services. Infrared cameras and sensors in the acquisition layer are used to collect pilot eye movement data and aircraft flight data. Using the transport layer， these data are transmitted wirelessly to the cloud storage area in the storage layer. In the data processing layer， the aircraft flight data with multiple information sources is integrated. In the application service layer， the trajectory information of the pilot’s fixation point is obtained through the operation of pupil positioning and fixation point calibration， and it is regarded as the important supplementary information of aircraft 4D track tracking. A 4D flight path prediction model of aircraft based on deep neural network is constructed. The integrated flight data of aircraft and the pilot’s gaze point trajectory information are taken as the input of the model， and the output result of the model is the 4D flight path prediction result of aircraft. Markov distance is used to compare the forecast results with the preset route， and the control parameters of the aircraft are adjusted according to the comparison results to realize the 4D track tracking. Experimental results show that this method can track aircraft 4D track accurately and has a good performance in differential flatness.

AGP Calculation Methods in UAV Imagery Based on Image Segmentation 

LI Kai, JIN Yunpeng, LI Haiyang, KONG Shasha, YANG Peng, FANG Chengwu, HUANG Xiangjie, HAN Yaosheng, LI Chunmei

2025, 0(04):  83-88.  doi:10.3969/j.issn.1006-2475.2025.04.013

Asbtract ( 25 )   PDF (9030KB) ( 32 )  

References | Related Articles | Metrics

Grassland degradation is a critical issue in the Three Rivers Source Region that cannot be overlooked. Employing deep learning techniques for the evaluating of grassland degradation in the Three Rivers Source Region is a pivotal step towards intelligent grassland assessment. However， a challenge in semantic segmentation lies in the potential inconsistency of altitudes in UAV-captured imagery， which can lead to discrepancies between computed proportions of poisonous weed cover and actual conditions， consequently introducing errors in grassland degradation assessment. This study proposes a method to calculate the Actual Ground Proportion （AGP） for both known and unknown heights of captured grassland images. For images with known heights， we select to utilize the captured altitude for AGP calculation and then map images of varying altitudes to a common height for coverage computation. For images with unknown heights， we train a sorrel instance segmentation model to calculate AGP based on instance segmentation results， followed by coverage computation. Experimental restlts demonstrate that， in comparison to direct coverage calculation， the use of instance segmentation reduces the error from 2.7% to 0.39%. This approach holds significant importance in enhancing the accuracy of intelligent grassland degradation assessment.

Infrared and Visible Image Fusion Based on Twin Axial-attention and Dual-discriminator Generative Adversarial Network

WANG Lidan1, ZHAO Huaici2, PAN Duotao1, FANG Jian2, YUAN Decheng1

2025, 0(04):  89-95.  doi:10.3969/j.issn.1006-2475.2025.04.014

Asbtract ( 21 )   PDF (5722KB) ( 23 )  

References | Related Articles | Metrics

For the same scene， the fused image of infrared and visible can preserve the thermal radiation information of the foreground target and the background texture details at the same time， and the description is more comprehensive and accurate. However， many classical fusion algorithms based on deep learning usually have the defects of insufficient information retention and unbalanced feature fusion. To solve these problems， an image fusion algorithm based on twin axial-attention and dual-discriminator generating adversarial network is proposed. The generator uses a double-dense convolutional network as a multi-scale feature extractor and introduces spatially enhanced branch and twin axial attention to capture local information and long-range dependencies. The adversarial game between the dual discriminator and the generator is constructed， and the retention degree of differential features is balanced by restricting the similarity between the two source images and the fusion image. The perceptual loss function based on pre-trained VGG19 can overcome the problem of losing high-level features such as semantic-level features. The experimental results on the TNO dataset show that the proposed method achieves prominent fusion results with clear textures and has significant improvements in both subjective and objective evaluation metrics compared to other classical algorithms， demonstrating its advancement.

Cigarette Laser Code Recognition Method Based on DBNet and CRNN Fusion Model

MA Qi1, WEN Yudong1, LIANG Shangrong2, WANG Ke2

2025, 0(04):  96-102.  doi:10.3969/j.issn.1006-2475.2025.04.015

Asbtract ( 27 )   PDF (1875KB) ( 23 )  

References | Related Articles | Metrics

Cigarette laser code recognition is an important means of tobacco inspection work. Aiming at the problem of low detection and recognition rate due to the complex background of cigarette code， this paper proposes a cigarette code recognition method based on DBNet and CRNN. First， the DBNet model is used to detect the smoke code region， and the accurate positioning and extraction of the cigarette code is realized by introducing differentiable binarization. Then， the CRNN model is used to identify the features of the cigarette code region image after positioning and cropping processing， and the spatial and temporal features are obtained by combining the VGG network and the deep Bi-LSTM network to realize the accurate identification of the cigarette code. Experimental result has an accuracy rate of 91.9% in detection， 83.4% in recognition， the testing accuracy after deploying the mobile APP is 83.0%， which shows that the method proposed in this paper can provide accurate cigarette laser code recognition effect.

Occupational Pneumoconiosis Screening Based on HA-Net Model 

WANG Jiale, SONG Wenai, FU Lizhen

2025, 0(04):  103-110.  doi:10.3969/j.issn.1006-2475.2025.04.016

Asbtract ( 23 )   PDF (4286KB) ( 20 )  

References | Related Articles | Metrics

 By combining deep learning methods and attention mechanisms， this study aims to improve the accuracy and efficiency of screening for occupational pneumoconiosis based on digital radiography. An improved deep learning model， hybrid attention network （HA-Net）， is proposed， which integrates squeeze-and-excitation block （SEB） and coordinate attention block （CAB） to enhance feature representation capabilities. SEB extracts inter-channel relationship information through global average pooling， uses fully connected layers to adjust channel weights， and multiplies the adjusted weights with the original input feature maps to strengthen important features. CAB captures spatial information through global pooling in both horizontal and vertical directions， then generates attention weights via 1×1 convolution and channel restoration， which are subsequently multiplied with the feature maps processed by SEB. Finally， these components are integrated into the ResNet50V2 model to distinguish between pneumoconiosis and non-pneumoconiosis images and accurately screen suspected cases. Experimental results show that the proposed model performs excellently in the task of screening occupational pneumoconiosis with high accuracy. It can reliably detect pneumoconiosis cases and also demonstrates high precision and sensitivity in identifying suspected cases.

Cross Border E-commerce Data Security Sharing Model Based on Zero Trust and Blockchain

LI Xiaomeng1, 2, JIANG Rong1, LIANG Zhihong3, YU Yimin1

2025, 0(04):  111-118.  doi:10.3969/j.issn.1006-2475.2025.04.017

Asbtract ( 22 )   PDF (8326KB) ( 24 )  

References | Related Articles | Metrics

With the rapid development of cross-border e-commerce industry， the explosive growth of related data in cross-border e-commerce has made data security issues increasingly prominent. Traditional cross-border e-commerce platforms often face issues of data leakage， tampering， and untrustworthiness， leading to a decrease in the trust of merchants and users to the platform. To address this issue， this paper proposes a cross-border e-commerce data security sharing model based on zero trust and blockchain （EDSM-BZT）. Firstly， a zero trust based multi factor identity authentication scheme is designed， which combines time and knowledge factors to dynamically verify each visitor. Secondly， a trust evaluation strategy is added to the scheme to calculate the trust value of the accessing subject and achieve more fine-grained control over user access behavior. Finally， an on chain and off chain collaborative storage sub model in the model is built to ensure the secure transmission and storage of cross-border e-commerce data. The experimental results show that the model can achieve dynamic control of access authorization and efficient storage of cross-border e-commerce data， with good performance and scalability.

Construction and Migration Method of Zero Trust Architecture for Marketing System

LI Zhihao, ZHAO Cong, WU You, CHEN Zechun, HE hang, DONG Chongchong

2025, 0(04):  119-126.  doi:10.3969/j.issn.1006-2475.2025.04.018

Asbtract ( 23 )   PDF (4187KB) ( 23 )  

References | Related Articles | Metrics

With the rise of cloud services， IoT， and other technologies， the boundaries in traditional network architecture have become more complicated and hazier. Firstly， adhering to the principle of "never trust， always verify"， a network security architecture based on zero-trust multi-dimensional dynamic verification is proposed to better adapt to the borderless trend of today’s network and improve the protection ability of sensitive data and business systems. Secondly， in view of the problem that most researches on zero trust are still in the theoretical stage， this paper proposes migration methods for enterprises to implement zero trust， such as business migration and workflow allocation. Finally， the security performance of the zero-trust architecture is analyzed to verify its effectiveness. The zero-trust architecture for marketing system effectively improves the security performance of the system， solves the shortcoming of the traditional boundary-based network architecture that is vulnerable to lateral attack after being breached by the attacker， and provides a low-risk implementation and migration method， which supports the smooth landing of the theory.