最新刊期
卷 23 ， 期 7 ， 2024
- “在车辆共乘领域，专家构建了可持续发展运营模型，提出了多模式实时车辆共乘匹配方法，为平台决策者提供有益参考。”摘要：Vehicle sharing can allocate idle transportation resources， improve economic efficiency， create environmental advantages and social welfare. A sustainable development operation model is constructed to address the issue of real-time carpooling， introducing a dual mode of dedicated and ride hailing， and considering the triple bottom line of sustainability， which includes important indicators such as economic benefits， environmental pollution， and social responsibility. A multi-mode real-time vehicle carpooling matching method based on sustainable development is proposed. Based on the comprehensive consideration of the triple bottom line principle， a routing dictionary mechanism is adopted to generate the optimal route between private car owners and passengers， select the optimal carpooling combination， and use the non dominated sorting genetic algorithm NSGA-III to obtain the optimal solution for the proposed multi-objective model. The experimental results show that when only maximizing economic goals， the average deterioration of environmental indicators is about 32.7%， and the average decrease of social indicators is about 9.8%； However， sacrificing only a small amount of economic indicators can achieve a significant improvement in environmental and social indicators. Therefore， environmental and social indicators should be included in the target decision values. Ignoring them and only optimizing profits will lead to an increase in carbon footprint and a deterioration in social responsibility. Exploring the correlation between economic benefits， environmental pollution， and social responsibility， as well as the changing trends under different influencing factors， can provide useful references for decision-makers of vehicle sharing platforms.关键词：urban transportation;ride-sharing matching;sustainable development;genetic algorithm;Pareto frontier190|70|0更新时间：2024-07-18
- “在滑坡易发性评价领域，机器学习模型挖掘数据信息，建立联系，寻找规律，为评价提供重要手段。”摘要：Landslide susceptibility evaluation is the necessary basis for landslide hazard evaluation and risk evaluation. Machine learning model can mine potential information in massive landslide disaster data， establish data connection， and find the essential law behind the landslide phenomenon， which is one of the important means of current landslide susceptibility evaluation research. Based on the literature retrieval platform， this paper introduces the application status of different machine learning models in the field of landslide susceptibility evaluation， and discusses the advantages and disadvantages of machine learning models in the application of landslide susceptibility evaluation and the application situation， and puts forward the following understanding：①the coupled machine learning model is significantly better than the uncoupled single machine learning model in landslide susceptibility evaluation， and the improved machine learning model is also significantly better than the unimproved single machine learning model in landslide susceptibility evaluation；②In the future， it is important to explore the mechanism of action， process and the interpretability of results of the "dark box" model algorithm， improve the accuracy and accuracy of landslide susceptibility evaluation， improve the shortcomings of model application research， and realize the nested coupling with artificial intelligence；③The characteristics and application of machine learning model are derived from its construction principle and run through the whole process of model application as inherent characteristics. Although machine learning model has received wide attention， people's understanding of it is still not deep enough. No consistent views and unified evaluation criteria have been formed yet， which needs further research.关键词：machine learning;landslide;geological disasters;landslide susceptibility assessment132|173|0更新时间：2024-07-18
- “在金融领域，一种新的财务报表命名实体识别方法被提出，有效提升了信息提取效率。该方法在中文命名实体识别和金融领域命名实体识别中表现优异。”摘要：Financial statements play an important role in the development planning of enterprises in the financial field， but extracting effective information from the statements still heavily relies on manual labor. To this end， a named entity recognition method for financial statements is proposed that integrates key information and entity boundary information to improve the efficiency of extracting effective information from financial statements. Firstly， a convolutional gating unit consisting of two convolutional layers， self attention mechanism， and gating mechanism is used to extract local features from the encoder's output， screen key information， and guide entity recognition； Then， the entity boundary prediction module is used to integrate the entity boundary information into the long sequence semantic features with sentence dependency relationships； Finally， the key information and the long sequence semantic features fused with entity boundary information are input into the conditional random field layer to extract the dependencies between adjacent labels that meet the entity labeling rules and obtain the global optimal label sequence. The experiment shows that the F1 values of the proposed model on the Resume and MSRA datasets are 95.75% and 94.92%， respectively， which are better than all comparison models， proving the effectiveness of this method in Chinese named entity recognition； The accuracy， recall， and F1 values on the financial report publication dataset are 87.93%， 92.45%， and 90.13%， respectively. Compared with the baseline model， the model performs better and can effectively identify named entities in the financial field.关键词：finance;name entity recognition;convolutional gating units;entity boundary prediction;conditional random fields80|30|0更新时间：2024-07-18
- “在智能辅助诊断领域，专家提出了基于深度学习的腰椎骨结构诊断方法框架，显著提升了诊断准确率和特异性。”摘要：Aiming at the problem of difficulty in intelligent assisted diagnosis of complex lumbar spine bone structure， a deep learning based computer-aided diagnostic method framework is proposed to assist in the diagnosis of lumbar disc herniation （LDH）. Firstly， a Resblock module is added to the encoding and decoding process of U-Net while preserving the skip connections of U-Net to enhance feature transfer in the target area， reduce feature loss， and accelerate model convergence speed. Secondly， the minimum envelope rectangle method is used to locate the center of the vertebrae， and ROI of appropriate size is cropped on the sagittal plane of the vertebrae based on the positioning， achieving fully automatic ROI acquisition. Finally， in the Xception network， average pooling is used instead of Flatten operation， and BN layer， Droupout layer， and dynamic learning rate are added to improve the speed and accuracy of the model. Regarding the MRI case of lumbar disc herniation in a certain hospital in Shanghai， after evaluating and training the classification model evaluation criteria， it was found that the proposed framework had a diagnostic accuracy of 94.46% for ACC， 94.60% for specific SPE， 97.09% for sensitivity SEN， and 94.32% for accuracy PRE. Compared with previous studies， this has been improved， which is of great significance for promoting the clinical application of computer-aided diagnosis.关键词：deep learning;U-Net;diagnosis of lumbar disc herniation;Xception;segmentation of vertebrae;intelligent aided diagnosis72|49|0更新时间：2024-07-18
-
Framework For Community Conflict Resoultion Based on Collaborative Small and Large Language Models AI导读
“在社区矛盾调解领域，专家提出了基于大小语言模型协同的调解框架，有效降低成本，提升调解质量。”摘要：The large language models（LLM） are increasingly being applied in people's lives due to their remarkable ability in contextual learning and causal reasoning. Corresponding to the capabilities of LLM， mediators in community conflict resolution require a certain level of discernment and a neutral perspective to mediate conflicts after understanding them thoroughly. Therefore， LLM can to some extent alleviate the issues of insufficient human resources， high mediation difficulty， and lack of credibility in existing community conflict resolution systems. However， the cost of invoking LLM limits their usage in communities. This paper proposes a framework based on the collaboration between large and small language models. The framework utilizes freely available small language models to generate conflict summaries and is divided into two approaches： human-machine diversion and human-machine collaboration， based on the mediator's level of involvement. Experiments demonstrate that this framework can reduce the cost while approaching the mediation quality provided by human mediators.关键词：large language model;small language model;human-machine separation;human-machine collaboration;conflict resolution44|105|0更新时间：2024-07-18 - “在人工智能领域，专家提出了一个对孪生数字人语言和行为授权的框架，通过人脸识别和区块链加密生成数字签名，有效对视频进行真人核对、信息验证和物理信息保存授权。”摘要：Due to the rapid development of AIGC （artificial intelligence generated content）， a large number of generated videos have emerged on the web， but generative AI may generate a lot of false information， which could be seriously misleading to the public， and unscrupulous people could use the low cost and efficiency to fake texts， images， or even videos to commit fraud， intimidation， defamation， etc. A framework for authorising the language and behaviour of twin digital humans is proposed. The framework generates digital signatures by combining face recognition and encryption in the blockchain， and finally generates digital certificates to store the information for authorisation.The method effectively generates a video of the twin digital person for real person verification， message validation and physical information retention authorisation. It has been experimentally verified that the framework is fully capable of achieving authorisation and identification of the statements and actions posted within specific slices of time and space through content detection of the videos.关键词：twin digital people;authorization framework;speech encryption;digital certificates;generative artificial intelligence44|37|0更新时间：2024-07-18
- “在图书馆领域，研究者提出了基于LDA与双向GRU神经网络的借阅主题热度预测模型，为图书馆阅读推广活动提供重要参考。”摘要：The analysis of book borrowing theme can mine read borrowing preferences and reading rules of readers. By using the prediction model of borrowing theme heat， it can predict the change trend of borrowing theme heat， which is of great significance for libraries to carry out reading promotion activities. In order to solve the problem of book borrowing topic extraction and topic heat prediction， this paper proposes a borrowing topic heat prediction model based on LDA and bidirectional GRU neural network. The algorithm extracts the borrowing book features and borrowing topics of readers in different time periods through LDA algorithm. On the basis of calculating the heat of borrowing topics in different time periods and constructing the data set of borrowing topic heat sequence， a topic heat prediction model based on bidirectional GRU neural network is constructed to predict the change trend of future topic heat， and the experimental evaluation is carried out on the paper literature borrowing record data set of Xiamen University Library. The simulation results show that the model can accurately obtain the relationship between borrowing topics and keywords， and compared with algorithms such as machine learning， the model can effectively reduce the prediction error of borrowing topics.关键词：heat prediction;borrowing topics discover;deep learning;bidirectional gated recurrent unit;latent Dirichlet allocation21|22|0更新时间：2024-07-18
- “在工业领域，提出了基于子图嵌入的实体链接方法LPBS，通过强化学习获取上下文集合并编码，利用双塔模型和交叉注意力机制融合特征，有效提升了链接预测性能。”摘要：Link prediction aims to predict missing fact triplets in the knowledge graph query process， and is commonly used in tasks such as intelligent question answering and information retrieval. However， due to the large number of nodes and relationships in the knowledge graph， encoding the entire graph requires significant resources， and the encoding method of graph embedding lacks the semantic information inherent in the query sentence， resulting in unsatisfactory link prediction results. To this end， a subgraph embedding based entity linking method LPBS is proposed. Based on reinforcement learning models， relevant strategies are designed to obtain the upper and lower text sets of predicted link paths and merge them for input encoding. Then， the embedding features of query sentences and subgraphs are obtained through a dual tower model based on multi head self attention mechanism. Finally， the quantitative features are fused through cross attention mechanism to obtain the predicted distribution of each node. Testing on a self built industrial dataset found that the proposed method achieved an MMR of 0.362， Hits@1 reached 0.313 and demonstrated the effectiveness of the model through ablation experiments.关键词：link prediction;reinforcement learning;multi-head self-attention mechanism;double-tower model;cross-attention mechanism18|67|0更新时间：2024-03-21
- “在社交网络领域，ORSH算法通过衡量节点特征，有效提升影响力传播范围，相较于NCSH算法平均提高5.4%。”摘要：Maximizing influence is one of the hot topics in the field of social network research， which aims to maximize the spread of influence by selecting a small number of seed nodes. Traditional heuristic algorithms often only focus on a single feature of a node， ignoring the combination of multiple network centrality indicators， and are greatly influenced by network structure， which can easily lead to the phenomenon of "rich club". To this end， a maximum influence algorithm ORSH based on neighbor overlap ratio and structural holes is proposed， which measures whether a node has the characteristics to become a seed node through two indicators： neighbor overlap ratio and structural hole properties. Experiments were conducted on six real network datasets， and it was found that the influence propagation range of this algorithm was increased by an average of 5.4% compared to the NCSH algorithm based on node coverage and structural holes， indicating that the ORSH algorithm can effectively select the most influential nodes.关键词：social network;influence maximization;overlap ratio of neighbors;structural hole;heuristic algorithm35|74|0更新时间：2024-07-18
- “在节能减排领域，专家分析集中供热温控方法，建立模糊控制PID调节策略，有效降低能耗，助力“双碳”目标实现。”摘要：With the continuous promotion and efforts of the national "dual carbon" policy， energy conservation and emission reduction issues are receiving increasing attention； As the most common winter heating mode at present， the municipal centralized hydronics in cities and towns consumes huge energy； It is of great significance to reduce heating energy consumption while ensuring user comfort； The study analyzed common household temperature control methods for centralized heating， fully explored the trend of changes in actual temperature and temperature control panel setting temperature errors， improved the grey model prediction of historical error sequences， obtained prediction errors， and used both actual and prediction errors as inputs for fuzzy control. Fuzzy rules were established to enable the fuzzy controller to output adjustment parameters， jointly adjust PID control， and reduce errors， Make the original control strategy respond in advance， adjust the system towards balance， reduce process energy consumption， and achieve the goal of energy conservation and emission reduction； Through MATLAB simulation experiments， the results show that the method responds in a timely manner， achieves energy-saving under the premise of comfort， and contributes to the achievement of the "dual carbon" goal.关键词：double carbon;grey model;fuzzy control;MATLAB simulation24|88|0更新时间：2024-07-18
- “在路径规划领域，改进的Informed-RRT*算法通过双向贪婪搜索、自适应步长和懒惰采样，有效提升了算法的优化效率和收敛速度，为复杂环境路径规划提供解决方案。”摘要：An improved Informed-RRT* algorithm is proposed to address the problems of blindness， slow convergence and low optimization efficiency of the Informed-RRT* algorithm in path planning. First， a two-way greedy search is introduced when finding the initial path， which speeds up the initial path finding rate. Then， adaptive step size is introduced in the tree growth process instead of fixed step size for growth， so that the algorithm can find better paths in the face of different environments. Finally， lazy sampling is used instead of the original random sampling to remove the useless nodes before the algorithm is processed， which reduces the operational pressure of the algorithm and also speeds up the convergence of the algorithm. The experimental results show that the optimized algorithm can quickly find a better path in the face of the complex environment.关键词：path planning;Informed-RRT*;adaptive step size;mobile robot83|63|0更新时间：2024-07-18
- “OMHuman虚拟数字人解决方案在人工智能与元宇宙应用领域取得突破，提升了人机交互体验，为虚拟主持人等场景提供高效解决方案。”摘要：Virtual digital humans are the intersection of artificial intelligence and metaverse applications， involving technology fields such as control engines， natural language processing， 3D graphics rendering， speech recognition and synthesis， and require multi-level collaborative design of software and hardware stacks. To this end， a loosely coupled control engine is proposed for the OMHuman virtual digital human solution based on the all-in-one machine deployment mode. It uses an independent graphics card to achieve graphics rendering and implements artificial intelligence model inference on the Intel OpenVINO computing engine through self-developed algorithms. This solves the shortcomings of traditional solutions in voice action collaborative control and other aspects， while also taking into account the end user experience， development costs， and deployment costs. Comparative tests have shown that the reasoning performance of the OMHuman virtual digital human model is 2-3 times that of traditional engines， and the graphics rendering efficiency is twice that of core graphics cards. It can meet human-computer interaction needs in a natural way and has been successfully applied in scenarios such as virtual hosts and intelligent data analysts.关键词：virtual digital human;artificial intelligence;all-in-one-box;control engine;natural language processing;graphics rendering68|77|0更新时间：2024-07-18
- “在企业级应用领域，基于Ceph和Kubernetes构建的云原生Elasticsearch平台，解决了资源利用率低和运维效率问题，具有高可用性、数据持久化等优势，性能表现良好，具有较好的运维和经济效益。”摘要：In order to solve the problems of Elasticsearch in enterprise-level application scenarios， such as low utilization and O&M efficiency， data persistence of containers and diversity of resource requirements， an enterprise-grade and cloud-native Elasticsearch platform is built based on Ceph and Kubernetes. The architectural high availability， data persistence， containerized deployment and standardized resource delivery of the Elasticsearch platform are designed crucially. The advantages of the architecture are qualitatively analyzed and the performance is quantitatively tested with the traditional deployment architecture. The experimental results show that the platform has obvious advantages in architecture design， good performance， good O&M and economic benefits in enterprise-level applications.关键词：Elasticsearch;cloud-native;distributed storage;distributed computing;containerization32|39|0更新时间：2024-07-18
- “最新室内甲醛监测系统，实现多房间在线监测，数据实时显示，稳定性高，具有实用价值。”摘要：Aiming at the problem that the traditional indoor formaldehyde detector cannot monitor the change of formaldehyde concentration at multiple measuring points at the same time， this study designs an indoor formaldehyde monitoring system based on wireless network. The monitoring system includes three parts： acquisition terminal， coordinator and upper computer. The acquisition terminal obtains environmental change data through formaldehyde sensor and temperature and humidity sensor， and sends it to the coordinator by ZigBee wireless networking. After receiving the data， the coordinator writes it into OLED screen and transmits it to the upper computer through serial port， and uses python script to insert real-time data into MySQL database. The upper computer reads and displays the data in MySQL library on the web page based on Apache server and PHP language. After testing， the online monitoring of formaldehyde concentration in indoor multi-rooms is realized. The coordinator OLED screen and browser page display data changes in real time， and the local or cloud monitoring records can be viewed through the computer and mobile terminal. Experiments show that the system has high stability， simplicity and flexibility， and has high practical value and application prospect.关键词：Zigbee;formaldehyde concentration;monitoring system design;database39|72|0更新时间：2024-07-18
- “新闻推荐领域取得新进展，专家提出融合知识图谱和用户长短期兴趣的新闻推荐模型，为提高新闻推荐准确率提供解决方案。”摘要：Aiming at the problem that the existing research on news recommendation systems has ignored the use of external knowledge entities to mine the potential knowledge level relationships between news， and has not combined users' short-term preferences for news recommendation， this paper proposes a news recommendation algorithm that combines knowledge graphs and users long-term and short-term interests. The model consists of three parts： a news semantic encoder， a user interest encoder and a click predictor. In the news semantic encoder， in addition to using the news's own title， introduction， and category information to learn the news semantic representation， it also uses the news title and the knowledge entities mentioned in the introduction are combined with the WikiData knowledge graph to construct a knowledge subgraph， and learn the potential knowledge-level connections between news from the knowledge subgraph. In the user interest encoder， the attention mechanism is used to extract the user's long-term interest from the user's historical click news sequence， and the GRU network is used to learn the user's short-term preference， and then the user's long-term interest and short-term preference are combined to construct the user's comprehensive interest representation. Comparative experiments and ablation experiments were carried out on the MIND-small dataset， the KGLS model improved by 2.92% compared with the most advanced baseline model on the AUC index reflecting the accuracy of the model.关键词：recommendation system;news recommendation;knowledge graph;short-term and long-term interests;GRU network38|124|0更新时间：2024-04-03
- “在在线社交网络信息传播演化趋势分析领域，专家构建改进聚集无标度网络，采用单参数囚徒困境博弈模型和费米规则更新策略，分析影响网络中合作者比例演化的因素，为解决信息传播合作困境提供解决方案。”摘要：Evolutionary game theory combines game theory with dynamic evolution process and pursues dynamic balance in the process of evolution. This method better realizes the analysis of the evolution trend of online social network information dissemination. In order to study the evolution dynamics of information dissemination in online social networks， and analyze the reasons and ways of reaching a certain equilibrium state， first of all， combined with the characteristics of online social networks， an improved clustering scale-free network is constructed. The interaction between users adopts a single-parameter prisoner's dilemma game model， and the Fermi rule update strategy. Secondly， by analyzing the factors that affect the evolution of the proportion of partners in the network through data simulation， it is concluded that the network aggregation and the weight of game returns will promote the cooperative behavior in the network， while the temptation of betrayal and noise factors will inhibit it. Finally， combined with the real public opinion communication data of Sina Weibo， this paper analyzes the information communication of the real network.The results show that the proportion of partners is the result of the interaction of game dynamics and network topology， and it is necessary to adopt appropriate incentive mechanism to solve the cooperative dilemma of information transmission in online social networks.关键词：network evolution game;clustering;scale-free network;prisoner's dilemma game;Fermi rule28|159|0更新时间：2024-07-18
- “在推荐系统领域，专家提出了混合动态负采样模型，有效提升了系统稳定性和准确性。”摘要：Negative samples have a significant impact on collaborative filtering recommendation tasks， and high-quality negative samples can help models accurately describe user profiles. A hybrid dynamic negative sampling model is proposed based on the idea of difficult negative samples to address the existing problems of false negative samples and high computational complexity. Firstly， the range and sequence of negative samples for each user are determined through dynamic negative sampling methods and service recommendation models； Then quickly sample a large number of difficult and negative sample candidates for each user； Next， using a hybrid approach， assemble the sampled negative sample set into a difficult negative sample to expand the perceptual domain and incorporate more information； Finally， an attention mechanism is introduced to guide the fusion of negative samples， thereby improving system stability. Comparative experiments with baseline models on publicly available datasets in Alibaba， Yelp2018， and Amazon have shown that the proposed model outperforms existing baseline models under multiple evaluation metrics， demonstrating the effectiveness of the model.关键词：collaborative filtering;negative sampling;hard negative sample;recommendation system;dynamic negative sampling34|193|0更新时间：2024-07-18
- “最新研究进展显示，专家提出结合Prophet和LightGBM模型，有效预测PM2.5浓度变化趋势，为空气质量管理提供解决方案。”摘要：In recent years， the issue of PM2.5 pollution has become increasingly prominent， causing serious impacts on people's physical health and environmental quality. Therefore， establishing an accurate PM2.5 concentration prediction model is of great significance for pollution prevention and air quality management. A combined model combining Prophet model and LightGBM model is proposed to address the nonlinear， high noise， and non-stationary characteristics of PM2.5 time series. In order to verify the effectiveness of the model， the Prophet LightGBM model and four other prediction models were compared and analyzed with PM2.5 concentration data in Lanzhou City as an example， as well as their prediction effects in different seasons. The results showed that the Prophet LightGBM model was more accurate in predicting the trend of PM2.5 concentration changes compared to the comparative model. The RMSE value reached 6.557， the MAE value reached 4.543， and the MAPE value reached 14.344%. It showed better performance in predicting accuracy and stability in summer and autumn， with the RMSE value reaching 3.155， the MAE value reaching 2.169， and the MAPE value reaching 9.4% when the RMSE value was optimal.关键词：PM2.5 concentration prediction;prophet model;LightGBM model;composite model88|64|0更新时间：2024-07-18
-
Research on Feature Extraction of Mobile Phone Middle Frame Based on Binocular Structured Light AI导读
“在手机自动点胶领域，专家设计了基于双目结构光的手机中框特征提取系统，提升了点胶效率与精确度，为高效自动点胶打下了良好基础。”摘要：Path extraction in automatic dispensing operations on mobile phones commonly uses a combination of line structured light and two-dimensional image analysis methods. Line structured light has high accuracy but slow speed， and two-dimensional image analysis methods require a lot of time to debug and have poor universality. To improve the efficiency and accuracy of dispensing glue on mobile phone frames， a feature extraction system for mobile phone frames based on binocular structured light is designed. The system uses a four step phase-shifting method to encode qualified Lei codes， and introduces a robust Gray code binarization method to improve accuracy while filtering the background. After polar correction and decoding， the three-dimensional point cloud coordinates are calculated based on phase matching of the decoding information. After preprocessing the point cloud model， a point cloud contour extraction method combining coarse and fine extraction is used. RANSAC is used to fit the contour lines， and the dispensing path is obtained by offsetting along the normal vector. The experimental results show that， while ensuring accuracy， the three-dimensional imaging time of the system is 26.1% of that of the line laser method， and the total processing time is 56.5% of that of the line structured light combined with two-dimensional image analysis method， indicating better robustness，which has a fast extraction speed and stable extraction effect， laying a good foundation for efficient automatic dispensing.关键词：automatic dispensing;structured light;point cloud;feature extraction32|42|0更新时间：2024-07-18 - “在图像描述生成领域，研究者提出了改进Transformer的多尺度模型，能生成更准确的语言描述。”摘要：The Transformer model is widely used in image description generation tasks， but it has the following problems： ① relying on complex neural networks for image preprocessing； ② Self attention has a quadratic computational complexity； ③ Masked Self Attention lacks image guidance information. To this end， an improved Transformer based multi-scale image description generation model is proposed. Firstly， the image is divided into multi-scale image blocks to obtain multi-level image features， which are then linearly mapped as input to the Transformer， avoiding the steps of complex neural network preprocessing and improving model training and inference speed； Then， linear complexity memory attention is used in the encoder to learn the prior knowledge of the entire dataset through learnable shared memory units and explore potential correlations between samples； Finally， visual guided attention is introduced into the decoder， using visual features as auxiliary information to guide the decoder in generating semantic descriptions that better match the image content. The test results on the COCO 2014 dataset show that compared to the base model， the improved model has improved scores on CIDEr， METEOR， ROUGE， and SPICE indicators by 2.6， 0.7， 0.4， and 0.7， respectively. The multi-scale image description generation model based on improved Transformer can generate more accurate language descriptions.关键词：image captioning;Transformer model;memory attention;multi-scale image;self-attention65|89|0更新时间：2024-07-18
-
Unsupervised Low Illumination Image Enhancement Based on Channel Attention and Illumination Weights AI导读
“VARRNet无监督低照度图像增强技术，有效提升亮度、保留细节、消除伪影和光晕。”摘要：Some existing unsupervised low light image enhancement methods may reduce the brightness of highlights in areas with insufficient image exposure， resulting in artifacts in the enhanced image； A single TV loss cannot distinguish the details of the lighting feature map， and it will also ignore areas with prominent differences in brightness at the edges of the lighting feature map， leading to the occurrence of halo phenomena. To this end， a unsupervised low light image enhancement method VARRNet based on channel attention and lighting weight is proposed. Firstly， VARRNet converts images into HSV space and combines V space with Retinex theory to avoid information loss； Secondly， in order to prevent the generation of artifacts during the brightness enhancement process， a brightness estimation network was designed to introduce channel attention ECA to allocate the weights of input feature maps， in order to restore the brightness of underexposed areas and effectively maintain the brightness of highlight areas； Finally， in the brightness estimation network， TV loss and lighting component weight are combined to preserve the rich detail information of the enhanced feature map and eliminate halos at strong edges. Compared with five popular low light image enhancement methods， VARRNet achieved better visualization results in brightness enhancement， detail preservation， color restoration， artifact suppression， and halo removal.关键词：unsupervised learning;Retinex;low illumination image enhancement;channel attention;lighting smoothness30|52|0更新时间：2024-03-21 - “在生活垃圾分类领域，EW_DCGAN通过改进DCGAN数据增强技术，有效提升了图像质量并增强了数据集，为垃圾分类研究提供了新方向。”摘要：To address the issues of low image quality and uneven class distribution in the current field of household waste classification datasets， a garbage image generation method based on improved DCGAN data augmentation （EW-DCGAN） is proposed. Firstly， redesign the network structure of DCGAN and adjust the size of the output image of the generator to 128 × 128 pixels； Then， the loss function BCE Loss is replaced with a loss function with Wasserstein distance， and a gradient penalty term is introduced to enhance the discriminative ability of the model discriminator； Finally， the ECA attention mechanism is added to the model generator to better cope with the interference of invalid information in the image， thereby efficiently extracting useful features. The experiment shows that the image quality generated using EW-DCGAN is higher， and the FID value decreases significantly compared to images generated only using DCGAN. It can expand and enhance the dataset in the field of garbage classification. The comparison of ResNet， MobileNet， and EfficientNet neural networks based on transfer learning on the pre enhanced and post enhanced datasets showed that the accuracy of the models improved by 7.09%， 5.34%， and 4.8%， respectively， compared to the original dataset.关键词：deep convolutional generative adversarial network;garbage classification;data augmentation;Wasserstein distance59|25|0更新时间：2024-07-18
- “在图像处理领域，研究者提出了基于经典Sobel算子的八方向自适应阈值Sobel算子边缘检测算法，并在FPGA芯片上实现，有效提升了边缘检测效果和实时性。”摘要：In order to solve the problem of insufficient details in capturing image edge information using the classic Sobel operator edge detection， and to address the problems of poor real-time performance and low speed of traditional software based image processing techniques， an eight direction adaptive threshold Sobel operator edge detection algorithm based on the classic Sobel operator is proposed， and it is run on an FPGA chip. To evaluate the performance of the algorithm， OV7725 camera was used to obtain video image information， and then secondary processing was performed through grayscale conversion， Gaussian filtering， and edge detection algorithms. The experiment shows that the improved Sobel operator edge detection algorithm has a significant improvement in edge detection performance compared to the classical Sobel operator， and can more accurately capture the detailed information of image edges. The running speed on FPGA can meet the requirements of real-time performance， providing a feasible solution for real-time image processing in the FPGA field.关键词：FPGA;image processing;Sobel operator;edge detection26|28|0更新时间：2024-03-21
- “在智能洒水车领域，研究人员提出了改进型YOLOv7-Tiny植物检测算法，有效提升了检测精度和速度。”摘要：Detection and recognition of urban roadside plants is a key technology for intelligent sprinkler vehicles. A modified YOLOv7 Tiny plant detection algorithm is proposed to address the issues of small target missed detection and occlusion in roadside vegetation image detection. When creating a dataset， the original dataset is obtained using camera realistic shooting and image crawler crawling methods， manually annotated using LabelImg， and the dataset is expanded using Mosaic data augmentation methods. To achieve both accuracy and high detection speed， the YOLOv7 Tiny network is first used as the baseline， and a parameter free SimAM attention mechanism is introduced in the Head part of the network to focus on more important feature information without increasing model complexity； Then， in the Head section of the network， ACmix is replaced with some traditional convolutions to achieve more efficient feature fusion； Finally， in the algorithm， SIOU is used to replace the CIOU of the original YOLOv7 Tiny network model to optimize the loss function， reducing the degree of freedom of the loss function and improving network robustness. The experiment shows that the average accuracy of the improved algorithm on the test set is average mAP@50：95 reaches 67.2%， which is 3.1% higher than the YOLOv7 Tiny algorithm. It has high detection accuracy while ensuring the lightweight of the model， and can meet the accuracy and speed requirements of lightweight plant detection for intelligent sprinklers..关键词：target detection;lightweight network;attention mechanism;YOLOv7-Tiny;SIOU35|23|0更新时间：2024-03-21
- “在胶质瘤手术中，基于多个图像目标的三维注册方法定位精度更高，为增强现实系统应用于神经肿瘤学提供参考。”摘要：In glioma resection surgery， doctors inevitably need to frequently move or rotate their heads to observe the surgical area. To evaluate the positioning accuracy of navigation systems based on augmented reality technology in glioma surgery， a three-dimensional registration method based on multiple image targets is proposed. The tester wears HoloLens2 and overlays its 3D virtual model on a real head model， measuring the virtual and real spatial coordinates of the reference markers on the scalp and glioma from different observation angles. Compare the proposed method with traditional interactive label free registration methods and analyze the impact of the two methods on positioning accuracy under different observation angles. The results showed that the average scalp localization error and glioma localization error of the registration method based on multiple image targets were （2.20±0.59） mm and （2.10±0.72） mm， respectively； The average scalp localization error and glioma localization error based on interactive unlabeled registration method are （4.52±1.77） mm and （4.43±1.69） mm， respectively， with the former having higher localization accuracy. The 3D registration method based on multiple image targets provides a reference for the application of augmented reality systems in neurooncology.关键词：augmented reality;glioma;3D registration;HoloLens2;positioning error30|30|0更新时间：2024-07-18
- “在地方高校计算机研究生创新能力培养领域，专家分析了培养条件和学科特点，提出了五个方面的培养策略，并验证了相关因素对创新能力的影响，为后续工作提供了建议。”摘要：To solve the problem of insufficient high-quality resources and poor quality of student sources in the process of cultivating innovative abilities for computer graduate students in local higher education institutions. Based on the analysis of the conditions and characteristics of graduate education in computer science in local universities， this paper proposes to explore and reflect on the cultivation of innovation ability in computer science graduate students from five aspects： accumulation of basic knowledge， internal academic exchange and cooperation， external research environment support， personal potential characteristics， and academic achievement verification. Through the research results， the impact of relevant factors on the cultivation of innovation ability in graduate students is further confirmed， and suggestions for subsequent work are proposed.关键词：local university;computer science;graduate student education;innovation ability training64|48|0更新时间：2024-07-18
- “据最新报道，ChatGPT技术在英语学习领域的应用研究取得进展，为相关技术发展提供参考。”摘要：ChatGPT is an artificial intelligence dialogue system based on a large language model. Its excellent natural language understanding and text generation abilities have laid the technical foundation for English learning， and the application of this technology to serve English learning has gradually become a hot topic of exploration. To this end， the technical characteristics of ChatGPT and the possibility of combining English learning disciplines were analyzed， and specific examples were used to explore the application mode of ChatGPT in English learning. Then， the advantages and potential issues of applying ChatGPT and related technologies to English learning were discussed， and corresponding countermeasures were proposed to provide reference and inspiration for English learning researchers supported by ChatGPT technology.关键词：ChatGPT;application model;English learning;large model21|32|0更新时间：2024-03-21
- “在程序设计课程思政教学领域，专家提出了融入中国传统文化的“五位一体”教学模式，为解决思政融入难题提供新方案。”摘要：Aiming at the difficulties in integrating ideological and political education and the lack of systematic solutions for programming course， this paper takes "Java programming" as an example and proposes a "five in one" case based ideological and political teaching model that integrates traditional Chinese culture . This model can be summarized as "one spirit， two line combination， three links， four dimensions， and five modules". The proposed model has achieved good results in the implementation of the Java programming course in the past three years， providing reference for the implementation of ideological and political education in other programming courses.关键词：Java programming;curriculum ideology and politics;teaching reform;five in one;traditional culture59|30|0更新时间：2024-07-18
- 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