Table of Content

28 March 2025, Volume 0 Issue 03

Previous Issue   

Construction of Depression Recognition Model Based on Multi-Feature Fusion

HOU Menghan, WEI Changfa

2025, 0(03):  1-5.  doi:10.3969/j.issn.1006-2475.2025.03.001

Asbtract ( 82 )   PDF (1144KB) ( 93 )  

References | Related Articles | Metrics

 In recent years， depression has become the primary problem of global mental health burden. In order to identify it， this paper proposes a depression recognition model combining BERT， BiLSTM and ConvNeXt. Firstly， the BERT model is used to generate feature vectors with rich semantics. Secondly， the BiLSTM， and ConvNeXt model is used to obtain the context information and the local features of the text， respectively. Thirdly， to alleviate the loss of semantic information in the feature extraction process， the context and local learned by BiLSTM and ConvNeXt models are fused through residual connections. Finally， depression is recognized according to the fused feature information. The experimental results show that the proposed model improves the accuracy， recall and F1 value compared with other deep learning models， which  can effectively extract the depression features of the text and improve the accuracy of depression recognition.

Improved YOLOv8s Algorithm Based on GiraffeDet for Transmission Line Icing Detection

TANG Rui1, WU Jianchao1, CHEN Jianbo1, CHAI Jiang1, WANG Qian1, HE Yuchen2

2025, 0(03):  6-11.  doi:10.3969/j.issn.1006-2475.2025.03.002

Asbtract ( 53 )   PDF (4197KB) ( 56 )  

References | Related Articles | Metrics

The icing of transmission lines can greatly impact the safety and stability of the power grid system. Due to the distribution of transmission lines in mountainous areas， forest areas， and unmanned open areas， workers cannot obtain on-site information in the event of damage such as rain， snow， and freezing. To accurately identify the icing situation of transmission lines in complex environments such as mountainous and uninhabited areas， this paper proposes an improved YOLOv8s-based detection method. Firstly， SIoU is adopted as the loss function to improve the training speed and accuracy of the model.  Secondly， by replacing some ordinary convolutions with dual convolutions， the information exchange between different channels is enhanced， effectively improving the efficiency of feature extraction， thereby further accelerating the convergence speed of the model. Finally， the GiraffeDet network structure is introduced to replace the original network structure and utilizes multi-scale information and the global context of feature map to make the model perform better in detecting small targets and complex scenes， improving the accuracy and robustness of detection. The experimental results show that compared with YOLOv8s， the improved method meets certain requirements for accuracy， reduces the model size by 7.3 MB， and significantly improves speed.

Multilevel Joint Graph Embedding for Lipophilic Molecular Classification

CAO Lu, DING Cangfeng, MA Lerong, YAN Zhaoyao, YOU Hao

2025, 0(03):  12-21.  doi:10.3969/j.issn.1006-2475.2025.03.003

Asbtract ( 36 )   PDF (1652KB) ( 60 )  

References | Related Articles | Metrics

 Classification of lipophilic molecules is an important area of research in the fields of bioinformatics and chemistry， where the goal is to efficiently classify molecules in terms of lipophilicity on the basis of their chemical structure and functional characteristics. However， due to the complex and diverse properties of lipophilic molecules， the traditional graph neural network classification methods fail to effectively extract the hierarchical relationships within the molecule and fully consider the structural information of the molecule when dealing with this type of problem， which results in the loss of information about the key atoms and the lack of global structural information. To address the above problems， a Multilevel Joint Graph Embedding Network （Mul-JoG） is proposed. Mul-JoG fuses Graph Transformer and graph pooling strategies to construct network layers， by concatenating the outputs of different network layers， and each layer fuses the information from all previous layers to form a multi-level joint graph embedding network. By obtaining the topological structure of molecules from multiple perspectives， the network captures the global information and interactions of molecules， effectively modeling the complex structure of molecules， and realizing the accurate classification of lipophilic molecules. The experimental results on the drug molecule dataset show that Mul-JoG achieved 97.96% and 92.94% in AUC and ACC， respectively. Compared with the benchmark method， the AUC and ACC is improved by 1.53 and 3.07 percentage points， respectively. The results showed that Mul-JoG is able to accurately classify lipophilic molecules.

Speech Enhancement Algorithm Based on Parallel Cascaded Time-frequency Conformer Generative Adversarial Network

WANG Zeyu, HAN Jianning, HAO Guodong, YANG Run

2025, 0(03):  22-28.  doi:10.3969/j.issn.1006-2475.2025.03.004

Asbtract ( 47 )   PDF (5158KB) ( 44 )  

References | Related Articles | Metrics

 Generative adversarial networks continuously improve network mapping capabilities through the adversarial training mechanism， giving them powerful noise reduction capabilities and are widely used in the field of speech enhancement. In order to solve the problem that the existing generative adversarial network speech enhancement methods do not fully utilize the time-frequency correlation and global correlation in the speech feature sequence and have poor denoising performance， this paper proposes a parallel cascaded time-frequency Conformer generative adversarial network for single channel speech enhancement. Firstly， the parallel cascaded time-frequency Conformer models the sequential features of time and frequency in the speech spectrogram， extracting local and global solicitations in the time domain and frequency domain for generator learning. Then， the two Decoder paths are used to learn the speech spectrogram with the amplitude mask of the noisy speech and the spectrogram of the clean speech respectively to fuse the output of the two paths to obtain the generated speech. Finally， an indicator discriminator is used to evaluate the relevant evaluation index scores of the speech generated by the generator， and the generator generation is improved through adversarial training. The quality of the voice is verified on the public dataset VoiceBank+Demand.

Review of Large Language Model Question Answering Systems for International Event Analysis

LEI Jiyue, SU Peng, NIE Yun, LIN Chuan

2025, 0(03):  29-37.  doi:10.3969/j.issn.1006-2475.2025.03.005

Asbtract ( 46 )   PDF (984KB) ( 57 )  

References | Related Articles | Metrics

 As the core focus of current artificial intelligence research， large language models have demonstrated strong cross-domain understanding and generation capabilities. They are widely used in many fields including event analysis， and promote the innovation and development of intelligent question answering system with its excellent performance. Although large language models show strong processing ability in general Q&A scenarios， they still face challenges in dealing with international events with deep professional backgrounds and high dynamics that affect international relations. In recent years， many scholars have focused on domain-specific large language models and quantitative analysis system for international relations， but there are few literature reviews on the cross-field of the combination of the two. In order to provide a comprehensive framework for developers and researchers of large language model question answering systems for international event analysis. Firstly， starting from the early international event analysis system， combined with the actual needs of international event analysis， the applicability of various general large language models in this task is analyzed. Secondly， by referring to the successful cases in the fields of finance， education， medicine， law， etc.， the strategy of constructing the domain-specific large-model question-and-answer system for international event analysis is extracted. In addition， the open data set resources closely related to the task are systematically combed. Finally， the current bottleneck and the future development direction are deeply analyzed.

Method for Underwater Robot Cage Inspection Control Based on Model Predictive Control

ZHANG Jiaxu, LIU Xiaoyang, HONG Shengcheng, SHENG Yifan, WANG Mingyang

2025, 0(03):  38-44.  doi:10.3969/j.issn.1006-2475.2025.03.006

Asbtract ( 38 )   PDF (3523KB) ( 49 )  

References | Related Articles | Metrics

Aiming at the high cost and low efficiency of manual periodic inspection of aquaculture cages during the aquaculture process， a method based on MPC for ROV （Cable Remote Control Underwater Robot） aquaculture cage inspection trajectory tracking control is proposed. Firstly， the physical and kinematic constraints of the ROV based on actual conditions are included， and combining the ROV kinematic and dynamic models， a speed controlled MPC controller is designed. Using the traditional PID （Proposal Integration Difference） control algorithm as the baseline， a model-free PID controller is designed. Secondly， the simulation experiment of two-dimensional horizontal track and three-dimensional space track tracking in cage aquaculture environment is carried out and compared. Finally， the experimental results show that the proposed method has advantages such as good trajectory tracking performance， stable ROV operation， and small fluctuations. The method proposed in this paper for solving the low efficiency of aquaculture cage inspection provides an advanced inspection solution for the aquaculture industry.

Specific Emitter Identification Method Based on Multi-source Unsupervised Domain Adaptation

ZHANG Taotao, XIE Jun, QIAO Pingjuan

2025, 0(03):  45-51.  doi:10.3969/j.issn.1006-2475.2025.03.007

Asbtract ( 40 )   PDF (2965KB) ( 47 )  

References | Related Articles | Metrics

 Due to the influence of the transmission environment and the change in the working state of the radiation source equipment， the channel noise of the signal to be identified and the training signal will be different， which will lead to the decrease of the recognition accuracy of the trained model. In order to solve this problem， most studies use single-source unsupervised domain adaptation method to use labeled samples under specific noise for the learning of unlabeled samples under the target noise to be identified. On the one hand， the labeled data collected in the actual situation may come from multiple source domains. On the other hand， the target domain can usually be regarded as a combination of multiple source domains. In order to explore the specific emitter recognition method based on multi-source unsupervised domain adaptation， a multi-source unsupervised domain adaptation method based on prototype alignment and contrast learning is proposed， which fully learns and utilizes the semantic structure information in the domain. Firstly， a prototype alignment method of multiple source domains and target domains is used to learn the feature representation of multiple source domains and a new pseudo-label strategy is designed. Then， this paper designs a weighted intra-domain sample-to-prototype comparative learning method to increase intra-class compactness and inter-class distinguishability. The experimental results on public datasets show that the proposed method achieves the best results in tasks with target domains of 4 db and 8 db， and the accuracy rates are 94.1 % and 97.4 %， respectively， which are 2.4 and 1.2 percentage points higher than the existing methods， indicating the effectiveness of the proposed method.

One-stage Semi-supervised Object Detection by Reusing Unreliable Pseudo-labels

SHAO Yeqin1, WANG Haiquan2, ZHOU Kunyang3, GUO Yudi2, SHI Quan1

2025, 0(03):  52-59.  doi:10.3969/j.issn.1006-2475.2025.03.008

Asbtract ( 40 )   PDF (1927KB) ( 33 )  

References | Related Articles | Metrics

The key to semi-supervised object detection methods is to assign pseudo labels to the targets of unlabeled data. To guarantee the quality of pseudo-labels， the semi-supervised object detection methods usually use a confidence threshold to filter low-quality pseudo-labels， which will cause most pseudo-labels to be removed due to their low confidence. Contrastive learning is used to reuse most of low-confidence unreliable pseudo labels for boosting the performance of semi-supervised object detection method. Specifically， the pseudo-labels are divided into reliable and unreliable ones according to the prediction confidence. Besides the reliable pseudo-labels， the unreliable pseudo-labels are exploited as negative samples for model training of contrast learning. To balance the number of unreliable pseudo-labels between different classes， a memory module is designed to store the unreliable pseudo-labels of different batches in the training process. The experimental results show that the mAP of the improved semi-supervised method on COCO data set is 13.6%， 23.0%， and 27.5% with the labeling ratio of 1%， 5%， and 10%， which is better than the existing semi-supervised learning methods. On the COCO-additional data set， the mAP of the improved semi-supervised method reaches 44.7%， which is 4.5 percentage points higher than supervised learning.

Classification Method of EEG Signals for Depression Based on Multi-Scale Dynamic Convolution and Attention Mechanism

LI Haoran1, HE Wenxue1, XU Jiazhen1, YANG Banghua2

2025, 0(03):  60-65.  doi:10.3969/j.issn.1006-2475.2025.03.009

Asbtract ( 42 )   PDF (1635KB) ( 45 )  

References | Related Articles | Metrics

Depression is a serious mental disorder that negatively affects the patient’s quality of life and social functioning. In order to explore an electroencephalogram-based classification method for depression to improve the accuracy of early diagnosis of depression， this paper designes a deep learning model called MDATCNet， which exploits a multi-scale dynamic convolution module capturing the rich features of signals in both spatial and frequency dimensions. To further enhance the representation of the model， this paper integrates the multi-head self-attention mechanism， which allows the model to adaptively focus on the features that are most helpful for decision-making. Then， the time convolutional layer is responsible for mining the time series patterns in the time series data. Finally， the features are passed to a Softmax classifier to classify EEG signals. The feasibility of the model is evaluated on the public depression dataset using the ten-fold cross-validation method， and the recognition accuracy， sensitivity and specificity of the method based on MDATCNet in EEG can achieve 94.71%， 99.37%， and 90.34%， respectively， and the experimental results show that the proposed model can effectively help the early diagnosis of depression.

Collaborative Interference Strategy for Secure Communication in Power IoT Based on Improved Genetic Algorithm

MAO Jiahui1, ZHU Xueqiong2, HU Chengbo2, TANG Peiyao1

2025, 0(03):  66-70.  doi:10.3969/j.issn.1006-2475.2025.03.010

Asbtract ( 37 )   PDF (1314KB) ( 36 )  

References | Related Articles | Metrics

 As wireless communication is more and more widely used in power IoT， the security of wireless communication in power IoT is getting increasing attention. In power IoT， secure transmission rate and power consumption are the key factors to guarantee the security of wireless communication. In order to guarantee the security of wireless communication system in power IoT， a physical layer secure communication method based on cluster collaborative interference strategy is proposed. Firstly， in the substation scenario， the eavesdropping nodes are outside the substation for eavesdropping， and multiple sensor nodes within the cluster of aggregation nodes use the received orthogonal code pieces to assist in the interference during the mutual communication between the sink nodes in the station. Secondly， since the sensor nodes are battery-powered， the secure transmission energy consumption is taken as the optimization target by considering the secure communication rate and energy consumption comprehensively. Finally， an improved genetic algorithm is used to learn an optimized collaborative jamming strategy. The simulation results show that the proposed algorithm can accelerate the convergence speed and reduce the secure transmission energy consumption compared with the traditional genetic algorithm and node random selection algorithm.

sORF-BERT： A Method on Identifying Coding sORFs Based on Pre-trained Models

BIAN Xinye1, XIE Dongmei1, WANG Ziling1, QU Zhijian1, YU Jiafeng2

2025, 0(03):  71-77.  doi:10.3969/j.issn.1006-2475.2025.03.011

Asbtract ( 38 )   PDF (3041KB) ( 41 )  

Related Articles | Metrics

 Small open reading frames （sORFs）， which are open reading frames in the genome that do not exceed 300 bases in length， are identified as crucial for maintaining cellular metabolic balance and fundamental physiological functions of organisms. To excavate the deep characteristics of sORF sequences and to enhance the accuracy of cross-species prediction of coding and non-coding sORFs， a sORF-BERT neural network model is proposed. This model integrates DNABERT pre-training with a data blending encoding strategy and introduces the CAL module to learn multi-scale features of sORFs. Analyses are conducted on prokaryotic genomes， as well as human， mouse， arabidopsis， and escherichia coli datasets. After pre-training and fine-tuning， the sORF-BERT model can effectively capture the rich biological features of sORF sequences and utilize the CAL to better learn sORF features across different scales. Cross-species experimental comparisons of sORF-BERT with six published advanced methods， including CPPred， DeepCPP， CNCI， CPPred-sORF， MiPiped， and PsORFs， reveal that sORF-BERT improves performance across five independent test datasets. Compared to the second-ranked PsORFs， sORF-BERT shows increases of 0.42 ~ 18.72 percentage points in ACC and 1.08 ~ 11.75 percentage points in MCC， thereby demonstrating the superiority of this method in predicting coding sORFs and its potential to advance basic biological research.

Multi-scale Feature Image Defogging Algorithm Based on Content-guided Attention Fusion

PU Yaya, WANG Yanbo, SU Yongdong, XU Zhongcheng

2025, 0(03):  78-85.  doi:10.3969/j.issn.1006-2475.2025.03.012

Asbtract ( 42 )   PDF (2746KB) ( 55 )  

References | Related Articles | Metrics

 Aiming at the problems of color distortion and detail blur in current defogging methods， a multi-scale feature image defogging algorithm based on content-guided attention fusion is proposed with encoder-decoder network architecture. Firstly， multi-scale feature extraction module is used to encode， and three parallel expanded convolutions with different scales and squeeze and excitation attention are designed to enlarge the receptor field， extract features of different scales， and improve feature utilization. Secondly， in the decoder， the content-guided attention fusion module is designed to dynamically improve different weights for the deep and the shallow features to retain more effective feature information. Finally， pyramid scene parsing network is introduced to improve the ability of global information acquisition. The experimental results show that compared with other algorithms， the proposed algorithm improves 26.13% and 6.39% on the peak signal-to-noise ratio and structural similarity of SOTS datasets， respectively. The entropy and average gradient of the real fog datasets are increased by 3.27% and 21.09% respectively. The proposed algorithm improves the problem of defog incompleteness and detail blur.

Remote Sensing Image Classification Based on Multi-scale Feature Extraction

LUO Hao, LI Xianfeng

2025, 0(03):  86-92.  doi:10.3969/j.issn.1006-2475.2025.03.013

Asbtract ( 46 )   PDF (1491KB) ( 47 )  

References | Related Articles | Metrics

In order to improve the accuracy of convolutional neural network in the classification task of remote sensing image， a method based on multi-scale feature extraction for classification is proposed. Aiming at the problems that the scale difference of target objects in remote sensing images is large and ordinary convolutional extraction tends to produce redundant features， a multi-scale hybrid convolution module with ordinary convolution and atrous convolution is proposed， which can effectively enhance the feature extraction ability of the model. In terms of feature fusion， a feature cross fusion module is proposed， which can effectively fuse the semantic information of each branch and make full use of the information between various scale features for deep fusion and interaction. Aiming at the problem of complex land cover information in remote sensing images， a parallel attention module is proposed to apply attention mechanisms to each branch， which makes it pay more attention to the key parts of the image and ignore the redundant information. Compared with the existing methods， the classification performance of the proposed method is significantly improved. On the data sets WHU-RS19、RSSCN7 and SIRI-WHU，the overall accuracy of the proposed method reaches 98.19%、94.18% and 96.37%， respectively.

Ultrasonic Image Segmentation of Thyroid Nodules Based on HA-UNet++

ZHU Yongtian, TIAN Fei, DONG Baoliang

2025, 0(03):  93-98.  doi:10.3969/j.issn.1006-2475.2025.03.014

Asbtract ( 47 )   PDF (2690KB) ( 51 )  

References | Related Articles | Metrics

Thyroid disease is one of the most frequently diagnosed nodular lesions in adult population， and it’s incidence is increasing year by year. With the development of artificial intelligence technology， the automatic diagnosis of thyroid ultrasound images using computer vision technology can significantly improve the accuracy and efficiency of diagnosis. However， most image segmentation methods based on deep learning， limited by the size of receptive field， cannot focus on the important features of the image in time and extract them effectively， resulting in low segmentation accuracy. In order to solve the above problems， a new deep learning network model HA-UNet ++ （Hybrid Dilated Convolution-Attention-UNet++） is adopted in this paper to segment ultrasonic images of thyroid nodules. HA-UNet ++ improves backbone network structure at each stage of encoding path. At the same time， hybrid dilated convolution is added to the convolution blocks with three layers of convolution in the network， and attention mechanism is added to each convolution block， so that it can quickly predict the enhanced thyroid nodule data set. On this basis， thyroid nodules are labeled and segmented.

Tongue Constitution Classification Method Based on Deep Learning

XIE Haiqing, LING Jiaqi, YI Xinbo

2025, 0(03):  99-105.  doi:10.3969/j.issn.1006-2475.2025.03.015

Asbtract ( 46 )   PDF (5379KB) ( 48 )  

References | Related Articles | Metrics

In response to the minimal inter-class differences in tongue images and the insufficient feature extraction by traditional networks， this paper constructs datasets for tongue image semantic segmentation and classification and conducts data preprocessing. Based on RepVGG network algorithm design and optimization， a multi-feature fusion tongue constitution classification network MTSNet based on convolutional neural network is proposed. MTSNet employs a multi-scale feature pyramid and combine high-level and low-level semantic information learned by the network to enhance the network’s representational capabilities. The addition of squeeze-excitation convolutional layers in the RepBlock module enables the network to focus more on information-rich features. The experimental results show that MTSNet significantly enhances classification performance across nine types of tongue constitutions， and its accuracy is 32.11 percentage points higher than that of AlexNet， 22.37 percentage points higher than that of SVM， and 17.68 percentage points higher than that of Resnet-18. Compared with the unoptimized RepVGG network， MTSNet achieves improvements of 9.90 percentage points in accuracy， 14.01 percentage points in macro-averaging， 9.90 percentage points in micro-averaging， and 11.09 percentage points in weighted-averaging. This tongue constitution， classification method provides scientific basis for users’ health management and has good reference application for traditional Chinese medicine’s adjunctive treatment and scientific research.

AMDFF-Net： Adaptive Multi-dimensional Feature Fusion Network for Tiny Object Detection

LIU Yaokai, REN Dejun, LIU Chongyi, LU Yudong

2025, 0(03):  106-112.  doi:10.3969/j.issn.1006-2475.2025.03.016

Asbtract ( 45 )   PDF (3369KB) ( 44 )  

References | Related Articles | Metrics

 Tiny object detection is a huge challenge in object detection research because tiny objects take up fewer pixels in the image， which results in a lack of feature information. To address this issue， an adaptive multi-dimensional feature fusion network （AMDFF-Net） for tiny target detection is designed to improve the accuracy of tiny object detection. Firstly， by integrating pooling layers and attention mechanisms， this paper constructs a pooling attention module， enabling the model to achieve a larger receptive field to enable self-adaptive and long-range correlations in self-attention. Secondly， an adaptive selection multi-dimensional feature fusion（ASMFF） module is designed， and an adaptive multi-dimensional feature pyramid network is designed based on the ASMFF module. This network adaptively fuses image features at different scales to enhance the information about tiny objects. To verify the performance and generalization of the model， experiments are conducted on the VisDrone2019， AI-TOD， and TinyPerson datasets. The experimental results show that AMDFF-Net improves the accuracy of tiny target detection， and the effectiveness of the proposed model in tiny target detection is verified by comparing with other mainstream algorithms.

PRNG Algorithm Based on Improved Piecewise Logistic Mapping

HUA Man1, LI Jingchang1, LI Yanling2

2025, 0(03):  113-118.  doi:10.3969/j.issn.1006-2475.2025.03.017

Asbtract ( 42 )   PDF (2987KB) ( 40 )  

References | Related Articles | Metrics

 The proposed piecewise Logistic mapping improves some limitations of the classical Logistic mapping and presents a new avenue for research in chaotic flow cryptographic design due to its high complexity and strong security. In order to improve the security risks of piecewise Logistic chaotic mapping， such as non-uniform distribution of chaotic sequences and small key space， an improved method of piecewise Logistic mapping based on feedback adjustment of initial values and control parameters is proposed. This method incorporates a dynamic system parameter mapping function to achieve a more uniform probability density of output state value distribution. Based on the optimized piecewise Logistic chaotic mapping， the PRNG algorithm is redesigned， and the Bit-level digital image encryption experiment is analyzed. The experimental results show that the chaotic sequence generated by the algorithm is more prominent， more evenly distributed， and has better randomness， which has a wide application prospect in stream cipher algorithm design.

Network Intrusion Detection Method Based on Convolutional Neural Networks with convLSTM

ZHANG Yue, GUO Zixin, HUANG Yibin, YAN Tao

2025, 0(03):  119-126.  doi:10.3969/j.issn.1006-2475.2025.03.018

Asbtract ( 56 )   PDF (1102KB) ( 57 )  

References | Related Articles | Metrics

 In the field of network intrusion detection， machine learning methods that manually extract features in feature engineering are generally used， but the manual feature extraction method is prone to losing important feature information； In addition， different types of attack traffic play different roles in detection， and existing algorithms generally suffer from important information loss and low accuracy in identifying attack types. A hybrid algorithm based on Convolutional Long-Short Term Memory （convLSTM） and Convolutional Neural Networks （CNN） is proposed for anomaly traffic detection in response to the aforementioned issues， Which directly use the payload of network traffic as data samples without manual extraction of complex traffic features， fully explores the structural features of traffic， extracts temporal and spatial features， and generates accurate intrusion detection feature vectors. The experimental results show that on the CIC-ISDS2017 dataset， the accuracy of the hybrid algorithm convLSTM-CNN in network intrusion detection reaches 99.39%. Compared with the simple DNN， SVM， LSTM， GRU-CNN and other models， it has a higher accuracy and lower false alarm rate， indicating that the algorithm improves the efficiency of abnormal traffic detection.