最新刊期
卷 24 ， 期 3 ， 2025
- “在高等教育领域，生成式AI正深刻改变学生学习和创新能力培养。研究发现，20.55%的学生科研任务中使用了生成式AI，低年级学生依赖度更高。现有检测工具对编辑后的AI生成内容鲁棒性不足，需优化科研任务设计和改进检测技术。”摘要：Generative AI is profoundly changing the landscape of higher education， bringing severe challenges to the cultivation of students' learning and innovation abilities. To this end， we focus on the application and prevention of generative AI in student research tasks， adopt empirical analysis methods， and use mature automated detection tools to evaluate the research reports of computer science graduate students in two Double First Class universities. The results indicate that the use of generative AI in student research tasks reaches 20.55%， demonstrating a certain degree of universality； The highest AI usage rate was found in the literature review， reaching 33.33%， and no traces of AI were detected in the experimental report； The dependence of lower grade students on generative AI is significantly higher than that of higher grade students， with usage rates of 35.29% and 7.69%， respectively. In addition， existing detection tools exhibit significant vulnerability when facing AI generated content that has been manually edited， especially when dealing with finely edited articles， which increases the risk of detection evasion due to insufficient robustness. The above findings highlight the importance of optimizing research task design and the urgency of improving AI detection technology. These measures are of great significance for strengthening the research ability cultivation of graduate students and enhancing their research innovation level.关键词：generative AI;AI-generated content detection;automatic detection tools;higher education;graduate research assignments0|0|0更新时间：2025-04-10
-
Transformer Neural Network Photovoltaic Prediction Algorithm Based on Singular Spectrum Analysis AI导读
“在光伏发电预测领域，研究人员提出了SSA-Trans算法，有效降低了预测误差，为电网安全稳定运行提供保障。”摘要：Accurate photovoltaic power generation prediction can provide guarantees for the safe and stable operation of the power grid. A Transformer neural network photovoltaic prediction algorithm SSA Trans based on singular spectrum analysis is proposed to address the problems of overly simple preprocessing methods and insufficient efficiency in identifying periodic patterns in current photovoltaic prediction algorithms. This algorithm introduces singular spectrum analysis technology in data processing， and reconstructs the sequence after removing the noise sequence that has a significant impact on the solar irradiance time series. A Transformer network prediction model is established for the reconstructed sequence， and the timestamp of the sequence is position encoded and used as the network's feature input along with weather data. Using a sliding window to input the divided data into the Transformer model for training and then for prediction. Comparing the predictive performance of the proposed algorithm with the other existing algorithms on three publicly available datasets， the results showed that the normalized average absolute error of the proposed algorithm decreased by approximately 31.94%， 20.37%， and 14.07%， respectively. Meanwhile， the ablation experiment confirmed the effectiveness of singular spectrum analysis and sliding window technique.关键词：photovoltaic forecasting;singular spectrum analysis;position encoding;Transformer0|0|0更新时间：2025-04-10 - “在环保领域，电动汽车路径规划问题备受关注。专家提出改进蚁群算法，有效平衡搜索多样性与收敛性，提高解的精度和收敛速度。”摘要：Due to environmental considerations， electric vehicles are becoming increasingly popular in contemporary society， and their path planning problems have also been widely studied. Vehicle path planning is a well-known NP hard problem， and tram path planning also needs to consider the charging problem. To this end， an improved ant colony algorithm is proposed， and a random competition based ant colony pheromone update strategy is designed. Some ants are randomly selected from the ant colony， paired together， and a better updated pheromone matrix is selected from each pair of ants， thus maintaining a good balance between search diversity and convergence； At the same time， local search strategies are added in the later stages of iteration to improve the accuracy of solving the optimal solution. The experimental results show that compared with traditional ant colony algorithm， the improved ant colony algorithm has more advantages in solution accuracy and convergence speed.关键词：ant colony optimization;random competition;routing planning;combinatorial optimization;pheromone update10|23|0更新时间：2025-04-10
- “在癫痫诊断领域，研究者提出了一种新模型，自动提取fMRI信息，准确率达99.1%，为癫痫发作自动检测提供新思路。”摘要：Epilepsy is a common central nervous system brain disease， and scalp EEG is the gold standard for diagnosing epilepsy. However， the mechanism of epileptic discharge， triggering factors， and the mechanism of brain recovery after discharge are still not fully understood. Therefore， a model MSC BiLSTM combining multi-scale convolution with bilinear long short-term memory network is proposed to automatically extract effective information from functional magnetic resonance imaging （fMRI）， while comprehensively capturing spatiotemporal and temporal dynamic features， and analyzing the spatiotemporal features obtained by the brain during pre discharge induction and post discharge recovery processes at the mesoscale. The experimental results show that the model achieves an accuracy of 99.1% in epilepsy detection on the EEG fMRI synchronous acquisition dataset. By analyzing the visualization results of the model， the reasons for the phase differences of seven major brain networks in the epileptic seizure and non seizure periods were studied. The research results provide ideas for automatic detection of epileptic seizures.关键词：electroencephalography;functional magnetic resonance imaging;multi-scale convolution;BiLSTM;epilepsy detection0|0|0更新时间：2025-04-10
- “在基因表达调控领域，研究者提出了一种结合双路时间卷积网络和长短距离融合学习机制的模型，有效预测DNA与转录因子的结合位点，为深入理解基因表达及调控机理提供新思路。”摘要：Accurate identification of transcription factor binding sites （TFBSs） is crucial for understanding gene expression and regulatory mechanisms. Convolutional Neural Network （CNN） and Long Short-Term Memory （LSTM） models have significantly improved accuracy in this task compared to traditional machine learning approaches. However， CNNs specialize in learning local spatial features but ignore long-distance dependencies in DNA sequences， while LSTMs are proficient in learning sequential relationships but computationally inefficient due to a lack of parallel computing ability. This paper proposes a novel dual-path sequential network integrating long-short distance fusion learning to address the above issues. In terms of structure， this paper employs the Temporal Convolutional Network （TCN） as the feature extractor that supports sequential modeling and parallel processing. The dual-path structure can learn complementary DNA features， improving the learning stability. In terms of features， this paper leverages the context information modeling capability of TCN， and designs a long-short distance fusion learning strategy to strengthen the feature representation for prediction. The experiment results on 165 ChIP-seq datasets show that our method outperforms the popular deep learning based methods. This study introduces a valuable framework for TFBSs prediction by combining sequential features with different distance dependency information.关键词：transcription factor;binding site;sequence data processing;temporal convolutional network;feature fusion0|0|0更新时间：2025-04-10
- “在温室气体排放预测领域，专家提出了基于CNN-LSTM混合神经网络模型的农田温室气体排放预测方法，为提高预测精度提供解决方案。”摘要：In order to fully explore the temporal characteristics in greenhouse gas emission data and improve the accuracy of greenhouse gas emission prediction， a prediction method of farmland greenhouse gas emissions based on CNN-LSTM hybrid neural network model is proposed . Based on the field experiment in Wenxian County， Longnan City， Gansu Province， the influencing factors such as temperature， water content and total nitrogen measured in the experiment were used as the input variables of the neural network， and the emission of soil greenhouse gases was used as the output variables， and a hybrid model of convolutional neural network and long short-term memory neural network for greenhouse gas prediction was established. The results show that the correlation coefficients （R2（CO2）=0.924 2， R2（CH4）=0.955 6， R2（N2O）=0.964 2）， root mean square error （RMSE（CO2）=0.012 6， RMSE（CH4）=0.015 3， RMSE（N2O）=0.033 0） and mean absolute error （MAE（CO2）=0.015 2， MAE（CH4）=0.033 0） and mean absolute error （MAE（CO2）=0.0152， MAE（CH4）=0.011 5 and MAE（N2O）=0.027 0） were higher than those of the BP artificial neural network model and the LSTM long short-term memory network model， indicating that the CNN-LSTM hybrid neural network model was more suitable for predicting greenhouse gas emissions from farmland.关键词：agricultural soil;greenhouse gas emission;convolutional neural network;long and short term memory network0|0|0更新时间：2025-04-10
- “在处理类别不平衡数据集领域，BRPE算法通过特征采样和下采样优化，有效提升F1和AUC值。”摘要：In many real-world problems， the datasets are typically imbalanced which probably degenerate the learning algorithm. To handle these skewed datasets， there are many class imbalance learning methods are proposed， especially ensemble methods due to their efficiency. While most of these ensemble methods mainly focus on the level of samples and neglect the features aspect. And conventional random sampling method do not pay enough attention to the boundary which always contain hard classified samples. Propose an ensemble sampling method named BRPE to overcome this deficiency. BRPE firstly samples a feature subset； then down-sample majority class instances via its closest euclidean distance to minority class samples to create a balanced random patch as training subset； then trains a base learner using each of subsets， and finally obtains the output combined of these learners. Experiments on both 10 synthetic datasets and 8 real-world datasets show that BRPE can achieve higher F1 and AUC values than other four existing ensemble methods for class imbalance.关键词：imbalanced data;class imbalance learning;ensemble algorithm;weighted sampling;random patch0|0|0更新时间：2025-04-10
Artificial Intelligence
- “在软件开发领域，专家提出了一种基于自然语言处理的文本需求自动生成需求模型的新方法，为提高软件开发效率和质量提供解决方案。”摘要：During the requirements phase of software development， ambiguity and vagueness in textual requirements can lead to inconsistencies between subsequent software artifacts and the requirements description. This makes it challenging to ensure a smooth software development process and the quality of the final product. To address the inconsistency between textual requirements and the generated requirements model （UML class diagram）， we propose a new method based on natural language processing for automatically generating requirements models from textual requirements. This method extracts information from textual requirements， analyzes the natural language text， converts it into a formal model， and ultimately generates a requirements model. A prototype tool was developed to implement consistency checking of textual requirements and automatic generation of requirements models. Case studies show that the proposed method helps software development teams better understand textual requirements， reduce incon-sistencies， and thereby improve the efficiency and quality of the software development process.关键词：textual requirements;software development;natural language processing;inconsistency;requirements model;class diagram0|0|0更新时间：2025-04-10
- “在基础设施薄弱场景下，专家构建无线自组织网络服务功能链编排模型，提出优化算法，降低时延12%。”摘要：In order to solve the problem that a single mobile node is difficult to provide the complete computing capability and functions required for the task under the scenario of weak infrastructure， a general model of service function chain orchestration for wireless Ad-Hoc networks under infrastructure-less environment is constructed， which comprehensively takes into account the global functional requirements and performance requirements of the service function chain. A service function chain scheduling algorithm based on heuristic algorithm is proposed to realize the optimized combination of heterogeneous functions of each node， and the performance of the algorithm is improved by dynamically adjusting the weights of the heuristic function. Simulation results show that the proposed algorithm reduces the average end-to-end delay by 12% and takes less time compared to other algorithms.关键词：wireless Ad-Hoc networks;infrastructure-less;functional heterogeneity;service function chain0|0|0更新时间：2025-04-10
- “在群智感知领域，专家提出了基于信任的工人选择策略，有效提升数据可信度和服务质量，降低成本。”摘要：In the CrowdSensing（CS） networks， the presence of malicious workers who engage in deceptive practices， such as false reporting or data fabrication， to gain additional rewards is an unavoidable concern， thereby affecting the service quality of the system. To address this issue， a Trust Based Workers Selection Scheme （BTWS） is proposed， which adopts a proactive trust evaluation approach. This scheme leverages trustworthy worker-reported data as reference samples for comparing the trustworthiness of other workers and dynamically evaluating their trust levels. The primary objective is to exclude malicious workers and selectively engage reliable workers for task execution to ensure data trustworthiness， consequently enhancing application service quality. Experimental results demonstrate that compared to the Cost Minimization Scheme Based on Passive Trust Evaluation（CMPT） ， the proposed scheme achieves a 1.9 times improvement in trust difference， accompanied by a 9% increase in the proportion of outstanding workers and a 63.8% cost reduction. Compared to the Quality Maximization scheme based on Trust UAV-Assisted（QMUT） ， the proposed scheme achieves an 8.7% increase in the ratio of outstanding workers and a 63% cost reduction with the same level of trust difference. In essence， the proposed scheme ensures data credibility and achieves application service quality optimization with low cost.关键词：CrowdSensing（CS）;workers selection;trust evaluation0|0|0更新时间：2025-04-10
Computer Software and Theory
- “在推荐系统领域，研究者提出了MCPD模型，通过预训练和去噪图卷积网络提升推荐准确性。”摘要：Currently， recommendation systems based on Graph Convolutional Networks （GCNs） commonly suffer from issues such as noise， low training efficiency， and inability to select appropriate loss functions for effective joint optimization. To this end， a multi task learning recommendation model MCPD is proposed， which combines pre training and denoising graph convolutional networks. The graph convolutional network focuses on the collaborative signals between high-order neighbors to generate more accurate user and item embeddings. Firstly， pre training is conducted on both the user and the project using bidirectional attention to improve the convergence speed and training time efficiency of the model. Secondly， a neighbor edge denoising autoencoder model is designed to combine traditional graph convolutional networks with attention mechanisms in the neighbor edge denoising task to identify noisy edges. The embedding is encoded and decoded using a denoising autoencoder DAE to reduce noise. Finally， select the cosine contrastive loss function with the best performance， and combine multi task learning to jointly optimize bidirectional attention aggregation pre training， neighbor edge denoising， and denoising autoencoder to ensure model recommendation accuracy. Experiments on three standard datasets showed that the Recall and NDCG metrics of the MCPD model reached 7.10， 6.00， 19.09 and 5.85， 4.82， 15.75， respectively， outperforming other baselines. In terms of recommendation accuracy， it has significant advantages compared to GCN based recommendation systems.关键词：recommender system;graph convolutional networks;collaborative filtering;denoising;multi-task learning0|0|0更新时间：2025-04-10
- “在论文相似性研究领域，专家提出了一种基于研究领域及其关联的量化方法，为论文相似性评估提供了新方案。”摘要：Every year， a large number of papers spanning different fields are published， some of which delve into specific areas for in-depth analysis， while others involve interdisciplinary fields. Although the research areas of these papers provide guidelines for understanding their topics， many researchers mainly rely on natural language processing techniques to evaluate the similarity of papers， ignoring the inherent importance of the research areas. To this end， an innovative method for quantifying paper similarity is proposed based on the utilization of research fields and their interrelationships. Firstly， describe a paper using a set of 4-level vectors constructed from relevant research fields； Next， introduce a correlation matrix to clarify the interrelationships between these fields， and finally evaluate the similarity between the paper representation vector and other paper representation vectors to calculate the paper similarity. The experiment shows that the proposed method has significant advantages in the dataset of open academic graph， providing a new method for the similarity of research papers in the field.关键词：field of study;paper similarity;relevance matrix;paper representation49|327|0更新时间：2025-04-10
- “在华中农业大学学术造假事件中，通过B站评论挖掘和情感分析，揭示了公众对正义、学术关注和理性的多元情感倾向。”摘要：This study conducts topic mining and sentiment analysis on Bilibili video comments related to the event of students from Huazhong Agricultural University jointly reporting Professor Huang Ruofei for academic misconduct. Data collection involves scraping 4 405 comments from Bilibili using Python's Requests library， followed by text preprocessing steps such as segmentation with jieba， stop word removal， custom dictionary creation， and TF-IDF text vectorization. The Latent Dirichlet Allocation （LDA） topic model is utilized to model the comments， determining the optimal number of topics as through perplexity and pyLDAvis visualization results， and identifying hot topics. The results of topic mining are categorized into three datasets， and sentiment analysis is performed using the snowNLP library to classify each comment into positive， negative， or neutral sentiment categories. Statistical analysis and visualization showcase the distribution of different sentiment categories and tendencies. The study reveals multiple hot topics including justice and effort， public concern over academic misconduct， and rationality and courage. The sentiment analysis uncovers diverse sentiment tendencies among the public， with the majority supporting and praising the students' reporting behavior while expressing concerns and anger about the academic environment. However， the study acknowledges certain limitations and suggests future research expand the scope by collecting and analyzing data from more social media platforms to obtain a more comprehensive understanding of public opinions and sentiment feedback. Additionally， integrating quantitative and qualitative methods could further explore the underlying reasons and mechanisms behind academic misconduct events.关键词：academic misconduct;LDA;snowNLP;topic mining;sentiment analysis4|0|0更新时间：2025-04-10
Big Data and Information Processing
- “在现代互联网环境中，网络流量分类技术至关重要。专家提出引入数据增强技术，有效提高恶意流量检测准确性和效率，为网络流量分类任务提供新解决方案。”摘要：Network traffic classification plays a crucial role in modern internet environments， essential for understanding traffic sources， identifying abnormal patterns， and detecting network attacks. However， with the widespread use of encrypted traffic and anonymous network technologies， traditional rule-based methods and deep packet inspection （DPI） approaches face growing challenges. In recent years， machine learning and deep learning methods have provided new perspectives for network traffic classification， but their reliance on large labeled datasets and the issue of data imbalance limit the effectiveness of model training. To address these problems， this paper introduces data augmentation techniques during the fine-tuning phase of NetMamba. By synthesizing traffic samples， the approach expands the dataset， balances class distributions， and improves the model's generalization capability. Experimental results show that data augmentation can effectively enhance the accuracy and efficiency of malicious traffic detection， reduce labeling costs， and prevent overfitting. The proposed method offers a novel solution for network traffic classification， especially in scenarios with limited data and class imbalance， making it particularly valuable in enhancing the detection of malicious traffic.关键词：network traffic classification;encrypted traffic;pre-training;Mamba;data augmentation0|0|0更新时间：2025-04-10
- “在物联网漏洞挖掘领域，专家建立了基于指针分析的自动化框架，显著提升了漏洞检出率和时间效率。”摘要：IoT vulnerability mining mainly targets binary programs with unknown source code， but there is a significant amount of manual auditing work that urgently requires a highly automated process for guidance. In the field of static analysis technology， pointer analysis， as an underlying technology， has shown the potential to adapt to various application scenarios with its highly automated analysis process and excellent results. By leveraging the advantages of pointer analysis and relying on the disassembly platform Ghidra， the introduced P-code is encapsulated to form PIR； Then， based on PIR， we designed pointer analysis algorithms and taint analysis algorithms that meet the requirements of vulnerability mining， and ultimately implemented an extensible analysis framework. The performance test results for CWE78 vulnerability detection show that the proposed framework correctly detects most vulnerabilities. Compared with existing vulnerability analysis tools， the vulnerability detection rate has increased by 86.2% and the time efficiency has increased by 38.7%. This framework not only verifies known vulnerabilities， but also has the ability to discover new vulnerabilities.关键词：vulnerability mining;P-code;pointer analysis;taint analysis0|0|0更新时间：2025-04-10
- “在网络安全领域，专家提出了基于深度聚类与对比学习的网络入侵检测模型，有效提升了检测实时性，为网络安全防护提供了新方案。”摘要：In the Internet era， network malicious intrusion events are increasing rapidly， and the importance of network intrusion detection is becoming increasingly prominent. In order to improve the real-time performance of intrusion detection and solve the problem of insufficient labeled attack samples， a network intrusion detection model based on deep clustering and contrastive learning is proposed. Firstly， the modified VGGNet and LSTM networks are used to extract local and global features， respectively. The clustering optimization model is optimized through examples and comparisons， and abnormal data is detected based on clustering methods. Secondly， taking the raw data packets of network traffic as input， the network traffic data is enhanced through optimized data augmentation methods to achieve contrastive learning. Experiments and comparative studies on the CIC-IDS-2017 dataset have shown that the proposed model outperforms other similar methods， and the modified VGGN is more suitable for processing network traffic data. The introduction of LSTM and contrastive learning can both improve model performance.关键词：contrastive learning;deep clustering;convolutional neural network;network intrusion detection;anomaly detection0|0|0更新时间：2025-04-10
Network Communication and Information Security
- “在驾驶疲劳研究领域，专家提出了基于注意力机制的多尺度卷积神经网络，有效区分不同被试的疲劳特征，为开发高效疲劳检测技术提供新思路。”摘要：Assessments based on electroencephalographic （EEG） signals are considered to be one of the most predictive and reliable methods for driving fatigue research. However， research in this area still has a low classification power in cross-subject studies and there are no uniform standards for classification metrics， and there are no recognised and validated methods for fatigue detection. Many methods simply extract spatiotemporal features and process the data directly with cumbersome parameter adjustments， ignoring differences in EEG performance across subjects. In addition， the contribution of different EEG channels to driving fatigue detection is also ignored. In order to solve the above problems， a multiscale convolutional neural network （MSCNN） based on the attention mechanism is proposed. The network includes a multiscale feature extraction layer， a feature recognition layer and a compressive classification layer. In the multiscale feature extraction layer， the network is able to automatically extract effective features from EEG signals. In the feature recognition layer， the effective features extracted in the previous layer are filtered and the data classification is finally completed by the compressed classification layer. Validated by using the publicly available SEED-VIG and self-made Simulated Fatigue Driving （SFDE） datasets， the experimental results show that the classification accuracies of the MSCNN on SEED-VIG and SFDE data for the mixed experiments are 91.36% and 92.06%， and the classification accuracies across the subjects are 75.54% and 76.52%， which are higher than those of the current state-of-the-art methods. In addition， by analysing the attentional activation weights and brain topology maps of the model， we investigate the contribution of different brain regions to the classification detection as well as the trend of different frequency bands between the two states. In conclusion， this study provides new research ideas for the application of brain-computer interfaces in driving fatigue research and promotes the further development of this field towards practical applications.关键词：EEG;fatigue driving detection;attention mechanism;multiscale0|0|0更新时间：2025-04-10
- “在自动驾驶3D目标检测领域，研究者提出了一种基于激光雷达点云自适应下采样和点注意力特征聚合的新方法，有效提升了对点数较少目标物体的检测精度。”摘要：Point cloud down-sampling and feature extraction are indispensable steps in the LiDAR point clouds-based 3D object detection methods for automatic driving. However， the traditional down-sampling methods are indiscriminate for all data points in the point cloud， which makes the points on the object evenly diluted， and the object with fewer points may lose all its points. At the same time， the current point cloud feature extraction method only retains the maximum value of the same dimension in the feature， which can't make full use of the point cloud feature. To address these issues， proposes a 3D object detection method based on LiDAR point cloud adaptive down-sampling and point attention feature aggregation. This method acquires category information of data points by utilizing a learning-based adaptive down-sampling operation， thus realizing categorical down-sampling of LiDAR point clouds. At the same time， the point attention feature aggregation method is used to aggregate the features of sampling points according to the attention coefficient between points， which can make full use of the feature information of the point cloud. In KITTI data set， the detection accuracy of cars， pedestrians and cyclists reached 78.49%、44.82% and 63.59% respectively， which verified the effectiveness of this method. Compared with other methods， the detection accuracy of the object with fewer points is significantly improved.关键词：3D object detection;LiDAR point cloud;adaptive down-sampling;point attention0|0|0更新时间：2025-04-10
- “在美术艺考评分领域，专家提出了基于深度学习的MFCFNet模型，实现了素描图像自动分档评分，准确度达到90.84%，效果最优。”摘要：The process of grading and grading works in art exams is a complex and tedious task. How to achieve fair and just grading， intelligent automation of the process， and accurate grading is an important research topic. To this end， a multi-scale feature cross fusion sketch image automatic grading network MFCFNet based on deep learning is proposed. Firstly， ResNet-34 is used as the backbone network to extract multi-scale features using a feature pyramid structure embedded with point by point group convolution and channel rearrangement； Next， the features are fused through the feature cross fusion module， and finally the fused features are input into the classifier to obtain the final automatic classification result. The experiments on the SCNU-400 sketch dataset and the self-made Sketch Portrait dataset show that the proposed model improves classification accuracy and F1 score by 2.82% and 2.8% respectively compared to the ResNet-34 network. Compared with the current mainstream methods， the classification accuracy reaches 90.84%， which is the best and has strong generalization ability.关键词：convolutional neural network;multi-scale;feature cross-fusion;sketch image;automatic grading0|0|0更新时间：2025-04-10
- “在微小孔径图像检测领域，FMGSA超分辨率方法显著提高了测量精度，为解决孔径尺寸测量问题提供可靠方案。”摘要：The task of detecting small aperture images includes detecting the size， shape， surface features， and positional accuracy of holes， among which aperture size measurement is one of the crucial and high-precision tasks. In the process of image measurement， problems such as low image quality are common due to high cost of image acquisition systems and limited acquisition environments. In recent years， significant progress has been made in super-resolution technology based on reference images， especially in restoring high-frequency details， which provides an effective way to improve image quality. This article proposes an innovative super-resolution method to improve image measurement accuracy， also known as FMGSA. This method uses fast feature matching from coarse to fine and a global self attention mechanism to find the mapping relationship between reference image features and low resolution image features. The experiment proved that the average error of the small aperture measured by the measurement system with the introduction of FMGSA method is only 0.004 mm， and the correct detection rate is 92.98%. Compared with other super-resolution methods， the FMGSA method based on reference images shows the highest measurement accuracy， and this study provides a reliable and efficient solution for measuring small aperture images.关键词：super resolution reconstruction;image measurement;deep learning;micro aperture measurement0|0|0更新时间：2025-04-10
- “在点云语义分割领域，研究者提出了一种基于点云加权聚合的新方法，显著提升了网络拟合能力和分割精度。”摘要：Point cloud data， with its intrinsic disorder and irregular shape， presents significant challenges to traditional semantic segmentation algorithms that typically rely on a static feature aggregation approach， leading to limited adaptability to diverse data distributions. To tackle this issue， we introduce a novel semantic segmentation technique that leverages weighted aggregation specifically tailored for point cloud data. This method starts by deploying a weighted aggregation module that dynamically generates weighting coefficients for neighboring points， facilitating an adaptive and weighted combination of local features that substantially improves the representation of neighborhood characteristics. Moreover， to deepen and optimize feature extraction， an advanced inverse residual module （weighted invresMLP）， which incorporates distance-based weighting， is developed. Integrating these components， we establish a comprehensive end-to-end framework for point cloud semantic segmentation named the Weighted Local Aggregation Neural Network （WLA-Net）. After extensive experiments on large-scale public datasets S3DIS and ScanNet， it is found that the proposed method famously improves the fitting ability of the network with higher accuracy compared to other methods.关键词：point cloud data;semantic segmentation;feature aggregation;weighted invresMLP0|0|0更新时间：2025-04-10
- “在三维目标检测领域，研究者基于VirConv算法提出改进方法，优化特征提取网络，提升模型效率和泛化能力，降低误检率。”摘要：To address the challenges of high false detection rates and poor performance in detecting distant and small objects in current point cloud 3D object detection algorithms， this paper proposes an improved 3D object detection method based on the VirConv algorithm. We design a Bird's Eye View （BEV） feature extraction network called C-ECVNet， which optimizes the point cloud encoding network of VirConv. This network introduces the ECVBlock module to enhance object features， enabling more precise extraction of spatial structure information from the original point cloud. Additionally， it incorporates a channel self-attention mechanism to capture hierarchical attention between channels， thereby improving model efficiency and generalization ability， while enhancing feature extraction capabilities. Experimental results on the KITTI test set demonstrate that our algorithm exhibits stronger robustness and lower false detection rates when processing complex environments， distant targets， and small objects.关键词：point cloud;3D object detection;VirConv;feature enhancement0|0|0更新时间：2025-04-10
- “在人脸识别领域，研究者提出了FFMapNet，通过自注意力和特征融合提升模板反演重建质量，实验显示优于现有方法。”摘要：Aiming at the problem that the face template feature data of the template inversion and reconstruction method in face recognition system is single， the template vector is not enough to capture diversity， and the generated image is not rich in details， a new face template mapping network （FFMapNet） is proposed. Firstly， the self attention feature extraction module （SAFEM） is used in conjunction with the deconvolution layer to upsample the initial facial template features and introduce high-dimensional channel features； Secondly， utilizing the global information modeling function of self attention mechanism， features are extracted and enhanced from template feature inversion； Finally， the feature fusion module （FFM） is used to further process and fuse the enhanced features， improving the consistency of the latent space （LS） mapping from the face template to the face generator network， and generating more expressive feature maps. Template inversion attacks were conducted on state-of-the-art facial recognition systems on MOBIO， LFW， and AgeDB datasets， and experimental results showed that the proposed method outperformed existing methods in terms of reconstruction quality and attack success rate.关键词：face recognition;template inversion;blackbox attack;self-attention mechanism;feature fusion0|0|0更新时间：2025-04-10
- “MCI-YOLOv8s算法在太阳能电池板表面缺陷检测领域取得新进展，提高了检测精度和速度，为工业应用提供新思路。”摘要：Aiming at the problems of high missed detection rate and poor detection accuracy in traditional object detection methods for surface defect detection of solar panels， an improved YOLOv8s solar panel surface defect detection algorithm MCI-YOLOv8s is proposed. Firstly， multi-scale dilation attention is introduced into the YOLOv8s model to more effectively focus on the important feature information of defects at different scales in solar cells； Secondly， replacing the original detection head of YOLOv8s with a more lightweight distributed focus detection head can improve its feature extraction ability while reducing the number of model parameters； Finally， replace the original loss function with the Inner SIOU loss function to solve the problem of slow model convergence speed. The experiment shows that the mAP@0.5 of improved model is 93.40%， which is 3.80% higher than the original benchmark model. The experimental results conducted on the NEU-DET dataset of steel strip surface defects confirm that the improved model has good generalization performance. Improving the speed and accuracy of model detection meets industrial requirements， providing ideas for real-time detection of surface defects in solar panels.关键词：YOLOv8s;multi-scale dilated attention;concentrated layerwise localization attention head;Inner-SIoU;solar panel;defect detection0|0|0更新时间：2025-04-10
- “在航拍图像目标检测领域，YOLO-SC2算法通过改进网络结构和损失函数，提升了小目标检测精度，为解决航拍图像检测问题提供新方案。”摘要：The widespread issue in aerial imagery of small target sizes， variable scales， and complex backgrounds leads to low detection accuracy with the YOLO series algorithms. To address this， this paper proposes the YOLO-SC2 algorithm， which is based on the YOLOv5s architecture. The algorithm incorporates a fine-grained convolutional module and integrates a Transformer-based C3TR layer into the network structure. The C2F module replaces the C3 module to enhance the extraction of target feature information. Additionally， the Focal-EIoU loss function is replaced to improve the accuracy of small target detection， and the Soft-NMS algorithm is employed for this purpose. Finally， a decoupled head fusion is introduced to enhance detection accuracy under multiple target scenarios. Experimental results on the VisDrone2019 public dataset demonstrate that the proposed model achieves improvements of 6.2%， 8.2%， and 8.4% in the P， mAP_0.5， and mAP_0.5：0.95 metrics， respectively， compared to the original model. Comparative analysis with other algorithms validates the effectiveness of the improved algorithm presented in this study.关键词：Yolov5s;fine-grained convolutional module;C3TR;loss function;soft non-maximum suppression （NMS） algorithm;decoupled heads1|0|0更新时间：2025-04-10
- “在行人重识别领域，专家提出了基于Res2Net-Transformer的多尺度特征融合算法，有效提升了识别效果。”摘要：To accurately recognize pedestrian images ， a person re-identification algorithm based on Res2Net-Transformer for multi-scale feature fusion is proposed. This method consists of a global feature extraction module， a deep aggregation module， and a feature alignment module. In the global feature extraction module， the Res2Net module is introduced into the ResNet50 network， enabling the network to extract more fine-grained features. The multi-scale deep aggregation module achieves the recursive aggregation of multi-scale features. The feature alignment module is used to mitigate the recognition impact caused by feature misalignment. Comparison with existing methods， this method approach demonstrates better robustness on Market1501、DukeMTMC-reID and MSMT17 dataset， it yields superior results in pedestrian re-identification.关键词：person re-identification;image recognition;feature alignment;multi-scale feature fusion0|0|0更新时间：2025-04-10
Graphic and Image Processing
- “在机器学习教育领域，专家提出了基于大模型的教学改革方案，旨在培养新时代需求的专业人才。”摘要：With the rapid development of machine learning technology， cultivating machine learning professionals with practical abilities has become an urgent task in the current education field. The widespread application of big model technology provides an opportunity for machine learning practical courses to cultivate professional talents that meet the needs of the new era. To this end， a reform plan for machine learning practical course teaching based on large models is proposed， aiming to break the constraints of traditional teaching modes， deepen students' understanding of machine learning theory， and cultivate machine learning professionals with innovative thinking and practical application skills. Firstly， adjust and optimize the teaching content to make the curriculum more closely aligned with practical application needs； Secondly， reform teaching methods to enhance students' learning initiative and practical skills； Finally， by driving practical projects， we aim to cultivate students' practical application abilities， providing reference and inspiration for cultivating new engineering talents.关键词：large model;machine learning practical course;teaching content reform;teaching method reform;practical ability2|0|0更新时间：2025-04-10
- “在计算机、信息通信专业领域，探究式教学方式有效提升学生工程思维，为学生全面发展提供保障。”摘要：Computer networking is an important foundational course for majors such as computer science and information communication. The traditional teaching mode of this course has the problem of insufficient cultivation of students' engineering thinking. To this end， with the goal of cultivating students' independent thinking ability， practical operation ability， analysis and problem-solving ability， and teamwork ability， an inquiry based teaching method that emphasizes the cultivation of engineering thinking is proposed. Through rich classroom design and the introduction of engineering cases， vivid teaching forms and strict teaching process control methods are adopted to improve students' participation. At the same time， a more comprehensive course assessment plan is implemented to form a teaching system based on theoretical knowledge and engineering practice as a means to comprehensively enhance students' engineering thinking. The teaching practice results show that this teaching method effectively improves the quality of teaching and provides strong guarantees for the comprehensive development of students.关键词：computer networks;engineering thinking;inquiry-based teaching;process control1|0|0更新时间：2025-04-10
- “在新工科背景下，重塑软件开发核心课程群，推行多层次PBL人才培养体系，有效提升学生综合能力与素养。”摘要：Under the background of new engineering disciplines， it is necessary to cultivate software development engineers with innovative thinking， practical skills， and lifelong learning abilities. In response to prominent teaching issues such as outdated course content updates， insufficient awareness of software development processes， and inadequate cross disciplinary engineering training， we have reshaped the core software development curriculum group， centered on student development， and implemented a multi-level PBL talent cultivation system that combines problem-based learning （PbBL）， program-based learning （PgBL）， and project-based learning （PjBL）. We have gradually implemented the teaching reform policy of "learning by doing， learning by doing， and learning by creating"， established the connection between curriculum and course， in class and out of class， and implemented an integrated development model of "teaching research student technology activities"， effectively improving students' comprehensive abilities and literacy. The teaching achievements are significant， and it is an innovative practice of talent cultivation reform under the background of new engineering.关键词：problem-based learning;program-based learning;project-based learning;student-centered;emerging engineering education2|0|0更新时间：2025-04-10
Computer Talents and Education
- Technical support is provided by Beijing Founder electronics co., LTD 鄂ICP备18021494号-4
鄂公网安备42010602004949号 - It is recommended to read the content of this site in Chrome&IE9+. Please switch to extreme mode in browser 360.
- Cookies We use cookies to help provide and enhance our service and tailor content. By continuing, you agree to the use of cookies.
0