Table of Content

30 January 2025, Volume 43 Issue 1

Previous Issue   

Special Issue on Computer Application

Disease Prediction via Capsule Network and Causal Reasoning

SUN Mingchen, JIN Hui, WANG Ying

2025, 43(1):  1-19.  doi:10.3969/j.issn.0255-8297.2025.01.001

Asbtract ( 181 )   PDF (2777KB) ( 69 )  

References | Related Articles | Metrics

Existing deep learning-based disease prediction models are predominantly datadriven, leading to a high dependency on the sample size and the coverage of disease types in the training dataset. The current methods for disease prediction have the following limitations: 1) When the model is trained on a limited range of disease types, its performance deteriorates significantly and may produce incorrect predictions for rare diseases. 2) The training data may contain features that are irrelevant or have weak correlations with the prediction target. This noise may prevent the model from making stable and reliable predictions, thus failing to meet the practical needs of high safety and reliability required in medical applications. To address these issues, this paper proposes a disease prediction model named CausalCap, which integrates capsule networks with causal inference. Specifically, we obtain the causal effects and relationships between clinical features and disease labels, and construct a causal graph of clinical features. The causal graph is then pruned to delete false nodes with no causal relationships to the disease labels, only retaining key nodes that truly influence the occurrence of the disease, resulting in a refined disease causal graph. Finally, hierarchical graph capsule neural network (HGCN) classifies the disease causal graph for disease prediction. Extensive evaluations on six public datasets demonstrate that CausalCap achieves an average improvement of 2.50% in ACC and 6.46% in F1 metrics compared to the suboptimal methods.

Entity Relationship Extraction Framework Based on Pre-trained Large Language Model and Its Application

WEI Wei, JIN Chenggong, YANG Long, ZHOU Mo, MENG Xiangzhu, FENG Hui

2025, 43(1):  20-34.  doi:10.3969/j.issn.0255-8297.2025.01.002

Asbtract ( 245 )   PDF (1446KB) ( 251 )  

References | Related Articles | Metrics

Entity relationship extraction is a crucial foundation for building large-scale knowledge graphs and domain-specific datasets. This paper proposes an entity relationship extraction framework based on pre-trained large language models (PLLM-RE) for relation extraction in circular economy policies. Within this framework, entity recognition of circular economy policy texts is performed based on the model RoBERTa. Subsequently, the bidirectional encoder representation from Transformers (BERT) is employed for entity relation extraction, facilitating the construction of a knowledge graph in the field of circular economic policies. Experimental results demonstrate the framework outperforms the baseline models including BiLSTM-ATT, PCNN, BERT and ALBERT in task of entity relationship extraction for circular economy policies. These findings validate the adaptability and superiority of the proposed framework, providing new ideas for information mining and policy analysis in the field of circular economy resources in the future.

Avalanche Prediction and Prevention Strategies Integrating Machine Learning and Dynamic Model Optimization

JIN Yongchao, WANG Zhijian, JIA Huishuang, DU Yuntian, HU Xinting, CHEN Xuebin

2025, 43(1):  35-50.  doi:10.3969/j.issn.0255-8297.2025.01.003

Asbtract ( 150 )   PDF (4878KB) ( 58 )  

References | Related Articles | Metrics

Blasting is an effective method to prevent avalanches, but it is difficult to determine the appropriate blasting time, blasting location and blasting energy. This study began by collecting and preprocessing indicator data about avalanches. Then we conducted exploratory data analysis on the data, and found a strong seasonal pattern in avalanche occurrences. Using 80% of the data as the training set and 20% as the test set, the support vector machine, random forest and perceptron neural network models were established, with parameters optimized using the Bayesian optimization algorithm. The results showed that the perceptron neural network achieved the highest accuracy. Subsequently, the three models were integrated according to the loss degree, and three integration strategies were compared. The results showed that the highest accuracy of the SVM-RF-MLP model was 0.952. A basic blasting energy model was then developed, taking into account the changes in mountain height and snow layer density over time. Using historical data, a dynamic avalanche stability blasting energy model was built to identify the distribution patterns in snow layer stability. The data is simulated and verified, and a three-dimensional mountain visualization analysis is performed to obtain the optimal blasting timing, blasting location and blasting energy.

Improved Daily Average 2 m Temperature Correction Method Based on U-Net

WANG Binglun, FANG Wei

2025, 43(1):  51-65.  doi:10.3969/j.issn.0255-8297.2025.01.004

Asbtract ( 90 )   PDF (13115KB) ( 31 )  

References | Related Articles | Metrics

In response to the limitations of the widely utilized deep learning model U-Net, which is unable to adequately learn spatial features and suffers from the loss of image detail information, the S-CUnet 3+ model has been proposed. S-CUnet 3+ enhances U-Net in two ways: firstly, it integrates the original model with the Swin Transformer, enabling it to learn global features of images, and secondly, it introduces multi-scale connection operations. The model also adopts pre-training and fine-tuning strategies to correct multiple forecast lead times simultaneously. Experimental results of correcting daily average 2 m temperature forecasts across seven lead times show that the S-CUnet 3+ model has a significant correction effect for all lead times, with the best correction effect at the 24-hour lead time. The mean absolute error and root mean square error are reduced by 50.64% and 49.25%, respectively. Moreover, S-CUnet 3+ outperforms seven existing correction methods: anomaly numerical-correction with observations, quantile regression, ridge regression, U-Net, CU-Net, Dense-CUnet, and RA-UNet.

A Semantic Segmentation Network Based on Lightweight Convolutional Modules

LIAN Xiaofeng, KANG Maomao, TAN Li, WANG Yanli

2025, 43(1):  66-79.  doi:10.3969/j.issn.0255-8297.2025.01.005

Asbtract ( 127 )   PDF (5717KB) ( 107 )  

References | Related Articles | Metrics

Semantic simultaneous localization and mapping augmented with deep learning provides an effective solution for handling dynamic scenes. However, this technology still faces challenges of high computational resource consumption and model complexity. To address these issues, this paper proposes a lightweight semantic segmentation network based on improvements to BlendMask. Firstly, a lightweight Ghost-depthwise separable convolution with efficient channel attention block (GDS-ECA) module is designed. This module replaces a few convolution operations in Ghost convolution with depthwise separable convolution to reduce parameters and computational load, while incorporating an attention mechanism to enhance feature representation capabilities. Secondly, a bottleneck GDS-ECA attention transformer network (BGTNet) is proposed, which applies GDS-ECA convolution to the neck module’s convolution layers to improve feature extraction precision. Additionally, traditional convolutions in the feature pyramid network (FPN) are replaced with GDS-ECA convolutions, creating a lightweight FPN (L-FPN). Combined with BGTNet, this forms the Backbone of the proposed semantic segmentation network. Finally, experiments on the COCO dataset validate the improvements, demonstrating a 7.3 ms reduction in processing time per image, and a 1.5% improvement in average precision.

Lightweight and Efficient Image Steganography Based on Convolutional Neural Network

DUAN Xintao, BAI Luwei, XU Kaiou, ZHANG Meng, BAO Mengru, WU Yinhang, QIN Chuan

2025, 43(1):  80-93.  doi:10.3969/j.issn.0255-8297.2025.01.006

Asbtract ( 144 )   PDF (11678KB) ( 38 )  

References | Related Articles | Metrics

Deep learning-based image steganography often faces challenges due to the large number of model parameters and high computational demands. In response, a lightweight and efficient image steganography method is proposed. Firstly Ghost module is introduced into the encoder and decoder, which reduces the number of parameters and computational complexity. Furthermore, a multi-scale feature fusion module is designed to capture complex relationships within multidimensional data. In addition, a novel hybrid loss function is proposed, which can improve image steganography quality without modifying the model. Experimental results show that the proposed method achieves a peak signal-to-noise ratio (PSNR) of 47.59 dB on a 256×256 pixel image. Compared with the current best image steganography method, the steganography quality is improved by 1.7 dB, the number of parameters is reduced by 77%, and number of computations by 91%. These results confirm that the proposed method effectively enhances steganography quality with reduced number of parameters and computational complexity of the model, achieving a lightweight and efficient model.

Lightweight Fault Diagnosis Model Based on Parallel Optimized CBAM

JIA Zhiyang, XU Zhao, LENG Yanmei, WEN Xin, GONG Haoyu

2025, 43(1):  94-109.  doi:10.3969/j.issn.0255-8297.2025.01.007

Asbtract ( 135 )   PDF (6520KB) ( 67 )  

References | Related Articles | Metrics

In engineering practice, the performance of fault diagnosis models is affected by factors such as strong noise interference, limited sample sizes, and high model complexity, which poses challenges to the application of existing data-driven intelligent models for equipment diagnosis. To address these issues, this paper proposes a lightweight model, PCSA-Net, based on a parallel optimized convolutional block attention module (CBAM). First, a multi-scale signal feature extractor (SFE) is used to convert the input sensor signal into a feature map. Then, the traditional CBAM is optimized through the development of a collaborative attention block, the design of a learnable layer scaling strategy, and the parallelization of perceptual data features. Additionally, a PW-Pool dimension reduction module is introduced by combining point convolution with average pooling layers to reduce the number of model parameters. The channel feature vector of the feature map is then integrated to obtain the final diagnosis result. Finally, the proposed model is validated using two datasets containing common bearing faults. Experimental results show that in the small sample bearing fault diagnosis (BFD) task, the proposed model outperforms the existing mainstream fault diagnosis framework in terms of lightness and robustness, and meets the practical needs of bearing fault detection in real-world applications.

Recommendation Algorithm Based on Relative Trust Enhancement

CHENG Jiayi, CHEN Lingjiao, WU Yuezhong

2025, 43(1):  110-122.  doi:10.3969/j.issn.0255-8297.2025.01.008

Asbtract ( 104 )   PDF (751KB) ( 23 )  

References | Related Articles | Metrics

Socialized recommendation has become one of the hot research topics in recent years. Introducing users’ social relationships into the recommendation algorithm based on their historical behavior can alleviate the problems of data sparsity and cold start faced by recommendation systems. This paper proposes a relative trust enhancement recommendation algorithm based on the CosRA (RTECosRA). In the bipartite “user-object” network, this algorithm allocates resources based on the CosRA similarity index and introduces users’ trust relationships during the resource allocation process. It adjusts the resource values obtained by trusted users, thereby increasing the recommendation rate of the items selected by trusted users. The results on the FriendFeed and Epinions datasets show that compared with the baseline algorithms, the RTECosRA algorithm has improvements in both accuracy and diversity. Moreover, by incorporating trust relationships, it expands the recommendable range for users and alleviates the cold start problem to a certain extent.

A Diffusion Map Recommendation Model Based on Multi-hop Mechanism

LIU Jianing, ZHANG Sijia, ZHANG Zhenglong, WANG Yihan, AN Zongshi

2025, 43(1):  123-136.  doi:10.3969/j.issn.0255-8297.2025.01.009

Asbtract ( 117 )   PDF (1273KB) ( 52 )  

References | Related Articles | Metrics

To address the challenges of high-order modeling and insufficient user feature modeling in knowledge graph-based recommendation systems, a diffusion map recommendation model based on multi-hop mechanism (MultiHop-GDN) is proposed. This model mines high-order semantic information from the knowledge graph through an end-to-end method, covering three parts: knowledge graph construction, feature extraction network design and multi-hop diffusion model development. The knowledge graph is constructed using user and project attributes, enabling in-depth analysis of information such as user interests, preferences, and historical behaviors to build user portraits and interest models. A feature extraction network is introduced to capture deep semantic information and obtain prediction values through the calculation of this model. Comparative experiments on two public datasets show that MultiHop-GDN effectively achieves high-level modeling of both users and projects, outperforming other representative models in recommendation effects.

Video-Based Facial Feature Computation Methods

WANG Yingxiao, YANG Yanhong, TAN Yunfeng

2025, 43(1):  137-153.  doi:10.3969/j.issn.0255-8297.2025.01.010

Asbtract ( 136 )   PDF (730KB) ( 176 )  

References | Related Articles | Metrics

This paper presents a review of research in video face recognition conducted over the past five years. It provides a comparative analysis of the facial feature computation methods, categorizing them into traditional approaches, deep learning techniques, and feature aggregation/fusion methods. Traditional feature extraction methods include linear and nonlinear approaches, while deep learning methods include spatial and temporal feature extraction techniques. Feature aggregation and fusion methods integrate multiple feature sources and fuse features from different time periods to improve recognition performance. At the end of each subsection, this paper also provides a unified analysis of the algorithms used in the literature, highlighting their advantages, evaluation metrics, and applications. Through this research, we aim to provide more reliable and efficient solutions for practical applications of video face recognition systems and promote further advancements in this field.

Bird Action Recognition Based on Multiple Excitation and Pyramid Split Attention

DENG Shuchong, CHEN Aibin, DAI Zijian

2025, 43(1):  154-168.  doi:10.3969/j.issn.0255-8297.2025.01.011

Asbtract ( 99 )   PDF (7521KB) ( 35 )  

References | Related Articles | Metrics

Aiming at the problem of low recognition accuracy and high misclassification rate of traditional action recognition methods in dealing with complex bird action patterns, an enhanced deep learning model is proposed. The model integrates a multiple-excitation module and pyramid split attention to improve 3D residual networks, aiming to improve both the accuracy and efficiency of bird action recognition. The inter-frame difference method is utilized to effectively reduce the computational burden while preserving critical spatio-temporal information, thereby improving the recognition accuracy. The introduction of a multiple-excitation module improves the original residual block so that the model can accurately capture subtle motion action features, which solves ambiguities in recognizing complex dynamic actions of birds. Additionally, the original 3D convolutional layer is replaced with 3D pyramid split attention to achieve effective capture of bird action features at different scales. Experiments conducted on a self-built bird action video dataset demonstrate a high recognition accuracy of 90.48%, which significantly outperforms the baseline model and other existing popular action recognition networks. These results confirm that the model can effectively handle the complex bird action recognition task.

Benchmarking of Spiking Neural Networks and Performance Evaluation of Neuromorphic Training Frameworks

HU Wangxin, CHENG Yingchao, HE Yulin, HUANG Zhexue, CAI Zhanchuan

2025, 43(1):  169-182.  doi:10.3969/j.issn.0255-8297.2025.01.012

Asbtract ( 133 )   PDF (805KB) ( 145 )  

References | Related Articles | Metrics

With the growing interest in spiking neural networks (SNNs), the development of open-source neuromorphic training frameworks has also accelerated. However, there is currently a lack of systematic guidelines for selecting these frameworks. To address this issue, this paper proposes a benchmarking method for SNNs based on image classification tasks. This method designs a convolutional neural network and a fully connected deep neural networks to evaluate two SNN training approaches: direct training with surrogate gradient backpropagation and conversion from artificial neural networks (ANNs) to SNNs. Based on the MNIST, Fashion-MNIST, and CIFAR-10 benchmark image datasets, the performance comparisons of various neuromorphic training frameworks are conducted using evaluation metrics such as training time and classification accuracy. Experimental results indicate that the neuromorphic training framework SpikingJelly outperforms others in terms of both training time and classification accuracy in direct SNN training, while the Lava framework achieves the highest classification accuracy in ANN-to-SNN conversion training.

An Exponential Moving Average-Based Mechanism for IoT Edge Device Selection

WU Tong, YUAN Peiyan

2025, 43(1):  183-194.  doi:10.3969/j.issn.0255-8297.2025.01.013

Asbtract ( 87 )   PDF (1416KB) ( 25 )  

References | Related Articles | Metrics

In small cell networks, the number of user equipment (UE) often exceeds the number of small base stations (SBS), with each SBS capable of serving multiple UE and each UE covered by multiple SBS. It is a significant challenge to decide the optimal SBS to serve each UE in small cell networks. Traditional coordinated multiple point (CoMP) operations, which allocate resources based on proportionate fairness, fail to consider the dynamic nature of the system, leading to suboptimal resource utilization. Therefore, this paper proposes a CoMP operation method based on exponential moving average (EMA). The proposed method prioritizes all UE and SBS at each time slot and increases the allocation weight for devices with high priority in the next time slot, enabling high-weight UE to receive services from more SBS and ultimately achieving dynamic resource adjustment. Experimental results demonstrate that the EMA-based CoMP operation significantly enhances system peak time and overall throughput efficiency while reducing the system’s dropout rate. Furthermore, it provides a better device selection mechanism for the internet of things edge system.