In geometry processing,symmetry research benefits from global geo-metric features of complete shapes,but the shape of an object captured in real-world applications is often incomplete due to the limited sensor resoluti...In geometry processing,symmetry research benefits from global geo-metric features of complete shapes,but the shape of an object captured in real-world applications is often incomplete due to the limited sensor resolution,single viewpoint,and occlusion.Different from the existing works predicting symmetry from the complete shape,we propose a learning approach for symmetry predic-tion based on a single RGB-D image.Instead of directly predicting the symmetry from incomplete shapes,our method consists of two modules,i.e.,the multi-mod-al feature fusion module and the detection-by-reconstruction module.Firstly,we build a channel-transformer network(CTN)to extract cross-fusion features from the RGB-D as the multi-modal feature fusion module,which helps us aggregate features from the color and the depth separately.Then,our self-reconstruction net-work based on a 3D variational auto-encoder(3D-VAE)takes the global geo-metric features as input,followed by a prediction symmetry network to detect the symmetry.Our experiments are conducted on three public datasets:ShapeNet,YCB,and ScanNet,we demonstrate that our method can produce reliable and accurate results.展开更多
In recent years,with the increase in the price of cryptocurrencies,the number of malicious cryptomining software has increased significantly.With their powerful spreading ability,cryptomining malware can unknowingly o...In recent years,with the increase in the price of cryptocurrencies,the number of malicious cryptomining software has increased significantly.With their powerful spreading ability,cryptomining malware can unknowingly occupy our resources,harm our interests,and damage more legitimate assets.However,although current traditional rule-based malware detection methods have a low false alarm rate,they have a relatively low detection rate when faced with a large volume of emerging malware.Even though common machine learning-based or deep learning-based methods have certain ability to learn and detect unknown malware,the characteristics they learn are single and independent,and cannot be learned adaptively.Aiming at the above problems,we propose a deep learning model with multi-input of multi-modal features,which can simultaneously accept digital features and image features on different dimensions.The model in turn includes parallel learning of three sub-models and ensemble learning of another specific sub-model.The four sub-models can be processed in parallel on different devices and can be further applied to edge computing environments.The model can adaptively learn multi-modal features and output prediction results.The detection rate of our model is as high as 97.01%and the false alarm rate is only 0.63%.The experimental results prove the advantage and effectiveness of the proposed method.展开更多
Laser cleaning is a highly nonlinear physical process for solving poor single-modal(e.g., acoustic or vision)detection performance and low inter-information utilization. In this study, a multi-modal feature fusion net...Laser cleaning is a highly nonlinear physical process for solving poor single-modal(e.g., acoustic or vision)detection performance and low inter-information utilization. In this study, a multi-modal feature fusion network model was constructed based on a laser paint removal experiment. The alignment of heterogeneous data under different modals was solved by combining the piecewise aggregate approximation and gramian angular field. Moreover, the attention mechanism was introduced to optimize the dual-path network and dense connection network, enabling the sampling characteristics to be extracted and integrated. Consequently, the multi-modal discriminant detection of laser paint removal was realized. According to the experimental results, the verification accuracy of the constructed model on the experimental dataset was 99.17%, which is 5.77% higher than the optimal single-modal detection results of the laser paint removal. The feature extraction network was optimized by the attention mechanism, and the model accuracy was increased by 3.3%. Results verify the improved classification performance of the constructed multi-modal feature fusion model in detecting laser paint removal, the effective integration of acoustic data and visual image data, and the accurate detection of laser paint removal.展开更多
In order to solve difficult detection of far and hard objects due to the sparseness and insufficient semantic information of LiDAR point cloud,a 3D object detection network with multi-modal data adaptive fusion is pro...In order to solve difficult detection of far and hard objects due to the sparseness and insufficient semantic information of LiDAR point cloud,a 3D object detection network with multi-modal data adaptive fusion is proposed,which makes use of multi-neighborhood information of voxel and image information.Firstly,design an improved ResNet that maintains the structure information of far and hard objects in low-resolution feature maps,which is more suitable for detection task.Meanwhile,semantema of each image feature map is enhanced by semantic information from all subsequent feature maps.Secondly,extract multi-neighborhood context information with different receptive field sizes to make up for the defect of sparseness of point cloud which improves the ability of voxel features to represent the spatial structure and semantic information of objects.Finally,propose a multi-modal feature adaptive fusion strategy which uses learnable weights to express the contribution of different modal features to the detection task,and voxel attention further enhances the fused feature expression of effective target objects.The experimental results on the KITTI benchmark show that this method outperforms VoxelNet with remarkable margins,i.e.increasing the AP by 8.78%and 5.49%on medium and hard difficulty levels.Meanwhile,our method achieves greater detection performance compared with many mainstream multi-modal methods,i.e.outperforming the AP by 1%compared with that of MVX-Net on medium and hard difficulty levels.展开更多
Deception detection plays a crucial role in criminal investigation.Videos contain a wealth of information regarding apparent and physiological changes in individuals,and thus can serve as an effective means of decepti...Deception detection plays a crucial role in criminal investigation.Videos contain a wealth of information regarding apparent and physiological changes in individuals,and thus can serve as an effective means of deception detection.In this paper,we investigate video-based deception detection considering both apparent visual features such as eye gaze,head pose and facial action unit(AU),and non-contact heart rate detected by remote photoplethysmography(rPPG)technique.Multiple wrapper-based feature selection methods combined with the K-nearest neighbor(KNN)and support vector machine(SVM)classifiers are employed to screen the most effective features for deception detection.We evaluate the performance of the proposed method on both a self-collected physiological-assisted visual deception detection(PV3D)dataset and a public bag-oflies(BOL)dataset.Experimental results demonstrate that the SVM classifier with symbiotic organisms search(SOS)feature selection yields the best overall performance,with an area under the curve(AUC)of 83.27%and accuracy(ACC)of 83.33%for PV3D,and an AUC of 71.18%and ACC of 70.33%for BOL.This demonstrates the stability and effectiveness of the proposed method in video-based deception detection tasks.展开更多
The complexity of cancer frequently results in diverse therapeutic responses among patients with the same cancer type undergoing identical treatments.Additionally,the development of anti-cancer drugs faces significant...The complexity of cancer frequently results in diverse therapeutic responses among patients with the same cancer type undergoing identical treatments.Additionally,the development of anti-cancer drugs faces significant challenges due to extended timelines and high attrition rates during the process.Advances in machine learning,combined with the availability of extensive drug research databases,have facilitated the emergence of computational approaches designed to predict drug responses more accurately.However,the noise and uncertainty introduced by incomplete records across different biological data sources pose challenges,ultimately constraining the predictive capabilities of these models.To overcome these limitations,this study introduces a novel multi-modal learning-based drug response prediction framework,DSGPred.By constructing dynamic sample graphs,DSGPred accounts for missing modality types and quantities within individual samples,enabling a more granular understanding of data completeness.The framework integrates multi-modal heterogeneous graph convolutional networks with advanced fusion modules to deeply capture and synthesize diverse drug and cell line features.Additionally,DSGPred employs an interaction-focused feature extraction module to learn dual interaction modes,thereby enhancing the richness of drug-cell line interaction embeddings.Experimental evaluations on the benchmark and independent datasets indicate that DSGPred consistently surpasses existing methods in predictive performance.Furthermore,tests involving previously unseen drugs and cell lines validate DSGPred’s generalization ability.Practical applicability is further underscored through case studies,highlighting its utility in real-world scenarios,and offering robust predictions and insights for drug response prediction and personalized therapy design.The codes for DSGPred are available at https://github.com/zhc940702/DSGPred/.展开更多
[Objective]Accurate prediction of tomato growth height is crucial for optimizing production environments in smart farming.However,current prediction methods predominantly rely on empirical,mechanistic,or learning-base...[Objective]Accurate prediction of tomato growth height is crucial for optimizing production environments in smart farming.However,current prediction methods predominantly rely on empirical,mechanistic,or learning-based models that utilize either images data or environmental data.These methods fail to fully leverage multi-modal data to capture the diverse aspects of plant growth comprehensively.[Methods]To address this limitation,a two-stage phenotypic feature extraction(PFE)model based on deep learning algorithm of recurrent neural network(RNN)and long short-term memory(LSTM)was developed.The model integrated environment and plant information to provide a holistic understanding of the growth process,emploied phenotypic and temporal feature extractors to comprehensively capture both types of features,enabled a deeper understanding of the interaction between tomato plants and their environment,ultimately leading to highly accurate predictions of growth height.[Results and Discussions]The experimental results showed the model's ef‐fectiveness:When predicting the next two days based on the past five days,the PFE-based RNN and LSTM models achieved mean absolute percentage error(MAPE)of 0.81%and 0.40%,respectively,which were significantly lower than the 8.00%MAPE of the large language model(LLM)and 6.72%MAPE of the Transformer-based model.In longer-term predictions,the 10-day prediction for 4 days ahead and the 30-day prediction for 12 days ahead,the PFE-RNN model continued to outperform the other two baseline models,with MAPE of 2.66%and 14.05%,respectively.[Conclusions]The proposed method,which leverages phenotypic-temporal collaboration,shows great potential for intelligent,data-driven management of tomato cultivation,making it a promising approach for enhancing the efficiency and precision of smart tomato planting management.展开更多
Visible and infrared(RGB-IR)fusion object detection plays an important role in security,disaster relief,etc.In recent years,deep-learning-based RGB-IR fusion detection methods have been developing rapidly,but still st...Visible and infrared(RGB-IR)fusion object detection plays an important role in security,disaster relief,etc.In recent years,deep-learning-based RGB-IR fusion detection methods have been developing rapidly,but still struggle to deal with the complex and changing scenarios captured by drones,mainly due to two reasons:(A)RGB-IR fusion detectors are susceptible to inferior inputs that degrade performance and stability.(B)RGB-IR fusion detectors are susceptible to redundant features that reduce accuracy and efficiency.In this paper,an innovative RGB-IR fusion detection framework based on global-local feature optimization,named GLFDet,is proposed to improve the detection performance and efficiency of drone-captured objects.The key components of GLFDet include a Global Feature Optimization(GFO)module,a Local Feature Optimization(LFO)module and a Channel Separation Fusion(CSF)module.Specifically,GFO calculates the information content of the input image from the frequency domain and optimizes the features holistically.Then,LFO dynamically selects high-value features and filters out low-value features before fusion,which significantly improves the efficiency of fusion.Finally,CSF fuses the RGB and IR features across the corresponding channels,which avoids the rearrangement of the channel relationships and enhances the model stability.Extensive experimental results show that the proposed method achieves the best performance on three popular RGB-IR datasets Drone Vehicle,VEDAI,and LLVIP.In addition,GLFDet is more lightweight than other comparable models,making it more appealing to edge devices such as drones.The code is available at https://github.com/lao chen330/GLFDet.展开更多
Gait recognition is a key biometric for long-distance identification,yet its performance is severely degraded by real-world challenges such as varying clothing,carrying conditions,and changing viewpoints.While combini...Gait recognition is a key biometric for long-distance identification,yet its performance is severely degraded by real-world challenges such as varying clothing,carrying conditions,and changing viewpoints.While combining silhouette and skeleton data is a promising direction,effectively fusing these heterogeneous modalities and adaptively weighting their contributions in response to diverse conditions remains a central problem.This paper introduces GaitMAFF,a novelMulti-modal Adaptive Feature Fusion Network,to address this challenge.Our approach first transforms discrete skeleton joints into a dense SkeletonMap representation to align with silhouettes,then employs an attention-based module to dynamically learn the fusion weights between the two modalities.These fused features are processed by a powerful spatio-temporal backbone withWeighted Global-Local Feature FusionModules(WFFM)to learn a discriminative representation.Extensive experiments on the challenging CCPG and Gait3D datasets show that GaitMAFF achieves state-of-the-art performance,with an average Rank-1 accuracy of 84.6%on CCPG and 58.7%on Gait3D.These results demonstrate that our adaptive fusion strategy effectively integrates complementary multimodal information,significantly enhancing gait recognition robustness and accuracy in complex scenes and providing a practical solution for real-world applications.展开更多
Titanium-based semiconductors are known for their high chemical stability and suitable band gap widths.However,the conventional experimental screening methods are inefficient due to the wide variety of materials.To sp...Titanium-based semiconductors are known for their high chemical stability and suitable band gap widths.However,the conventional experimental screening methods are inefficient due to the wide variety of materials.To speed up the selection process,this work focuses on interpretable feature learning and band gap prediction for titanium-based semiconductors.First,titanium compounds were selected from the Materials Project database by machine learning,and elemental features were extracted using the Magpie descriptors.Then,principal component analysis(PCA)was applied to reduce the data dimensionality,creating a representative dataset.Meantime,heatmaps and SHAP(SHapley Additive exPlanations)methods were used to demonstrate the influence of key features such as electronegativity,covalent radius,period number,and unit cell volume on the bandgap,understanding the relationship between the material’s properties and performance.After comparing different machine learning models,including Random Forest(RF),Support Vector Machines(SVM),Linear Regression(LR),and Gradient Boosting Regression(GBR),the RF was found to be the most accurate for band gap prediction.Finally,the model performance was improved through parameter tuning,showing high accuracy.These findings provide strong data support and design guidance for the development of materials in fields like photocatalysis and solar cells.展开更多
Automated essay scoring(AES)systems have gained significant importance in educational settings,offering a scalable,efficient,and objective method for evaluating student essays.However,developing AES systems for Arabic...Automated essay scoring(AES)systems have gained significant importance in educational settings,offering a scalable,efficient,and objective method for evaluating student essays.However,developing AES systems for Arabic poses distinct challenges due to the language’s complex morphology,diglossia,and the scarcity of annotated datasets.This paper presents a hybrid approach to Arabic AES by combining text-based,vector-based,and embeddingbased similarity measures to improve essay scoring accuracy while minimizing the training data required.Using a large Arabic essay dataset categorized into thematic groups,the study conducted four experiments to evaluate the impact of feature selection,data size,and model performance.Experiment 1 established a baseline using a non-machine learning approach,selecting top-N correlated features to predict essay scores.The subsequent experiments employed 5-fold cross-validation.Experiment 2 showed that combining embedding-based,text-based,and vector-based features in a Random Forest(RF)model achieved an R2 of 88.92%and an accuracy of 83.3%within a 0.5-point tolerance.Experiment 3 further refined the feature selection process,demonstrating that 19 correlated features yielded optimal results,improving R2 to 88.95%.In Experiment 4,an optimal data efficiency training approach was introduced,where training data portions increased from 5%to 50%.The study found that using just 10%of the data achieved near-peak performance,with an R2 of 85.49%,emphasizing an effective trade-off between performance and computational costs.These findings highlight the potential of the hybrid approach for developing scalable Arabic AES systems,especially in low-resource environments,addressing linguistic challenges while ensuring efficient data usage.展开更多
Image-based face pose estimation tries to estimate the facial direction with 2D images.It provides important information for many face recognition applications.However,it is a difficult task due to complex conditions ...Image-based face pose estimation tries to estimate the facial direction with 2D images.It provides important information for many face recognition applications.However,it is a difficult task due to complex conditions and appearances.Deep learning method used in this field has the disadvantage of ignoring the natural structures of human faces.To solve this problem,a framework is proposed in this paper to estimate face poses with regression,which is based on deep learning and multi-modal feature loss(M2FL).Different from current loss functions using only a single type of features,the descriptive power was improved by combining multiple image features.To achieve it,hypergraph-based manifold regularization was applied.In this way,the loss of face pose estimation was reduced.Experimental results on commonly-used benchmark datasets demonstrate the performance of M2FL.展开更多
Deep learning has made significant progress in the field of oriented object detection for remote sensing images.However,existing methods still face challenges when dealing with difficult tasks such as multi-scale targ...Deep learning has made significant progress in the field of oriented object detection for remote sensing images.However,existing methods still face challenges when dealing with difficult tasks such as multi-scale targets,complex backgrounds,and small objects in remote sensing.Maintaining model lightweight to address resource constraints in remote sensing scenarios while improving task completion for remote sensing tasks remains a research hotspot.Therefore,we propose an enhanced multi-scale feature extraction lightweight network EM-YOLO based on the YOLOv8s architecture,specifically optimized for the characteristics of large target scale variations,diverse orientations,and numerous small objects in remote sensing images.Our innovations lie in two main aspects:First,a dynamic snake convolution(DSC)is introduced into the backbone network to enhance the model’s feature extraction capability for oriented targets.Second,an innovative focusing-diffusion module is designed in the feature fusion neck to effectively integrate multi-scale feature information.Finally,we introduce Layer-Adaptive Sparsity for magnitude-based Pruning(LASP)method to perform lightweight network pruning to better complete tasks in resource-constrained scenarios.Experimental results on the lightweight platform Orin demonstrate that the proposed method significantly outperforms the original YOLOv8s model in oriented remote sensing object detection tasks,and achieves comparable or superior performance to state-of-the-art methods on three authoritative remote sensing datasets(DOTA v1.0,DOTA v1.5,and HRSC2016).展开更多
Discriminative region localization and efficient feature encoding are crucial for fine-grained object recognition.However,existing data augmentation methods struggle to accurately locate discriminative regions in comp...Discriminative region localization and efficient feature encoding are crucial for fine-grained object recognition.However,existing data augmentation methods struggle to accurately locate discriminative regions in complex backgrounds,small target objects,and limited training data,leading to poor recognition.Fine-grained images exhibit“small inter-class differences,”and while second-order feature encoding enhances discrimination,it often requires dual Convolutional Neural Networks(CNN),increasing training time and complexity.This study proposes a model integrating discriminative region localization and efficient second-order feature encoding.By ranking feature map channels via a fully connected layer,it selects high-importance channels to generate an enhanced map,accurately locating discriminative regions.Cropping and erasing augmentations further refine recognition.To improve efficiency,a novel second-order feature encoding module generates an attention map from the fourth convolutional group of Residual Network 50 layers(ResNet-50)and multiplies it with features from the fifth group,producing second-order features while reducing dimensionality and training time.Experiments on Caltech-University of California,San Diego Birds-200-2011(CUB-200-2011),Stanford Car,and Fine-Grained Visual Classification of Aircraft(FGVC Aircraft)datasets show state-of-the-art accuracy of 88.9%,94.7%,and 93.3%,respectively.展开更多
Accurate purchase prediction in e-commerce critically depends on the quality of behavioral features.This paper proposes a layered and interpretable feature engineering framework that organizes user signals into three ...Accurate purchase prediction in e-commerce critically depends on the quality of behavioral features.This paper proposes a layered and interpretable feature engineering framework that organizes user signals into three layers:Basic,Conversion&Stability(efficiency and volatility across actions),and Advanced Interactions&Activity(crossbehavior synergies and intensity).Using real Taobao(Alibaba’s primary e-commerce platform)logs(57,976 records for 10,203 users;25 November–03 December 2017),we conducted a hierarchical,layer-wise evaluation that holds data splits and hyperparameters fixed while varying only the feature set to quantify each layer’s marginal contribution.Across logistic regression(LR),decision tree,random forest,XGBoost,and CatBoost models with stratified 5-fold cross-validation,the performance improvedmonotonically fromBasic to Conversion&Stability to Advanced features.With LR,F1 increased from 0.613(Basic)to 0.962(Advanced);boosted models achieved high discrimination(0.995 AUC Score)and an F1 score up to 0.983.Calibration and precision–recall analyses indicated strong ranking quality and acknowledged potential dataset and period biases given the short(9-day)window.By making feature contributions measurable and reproducible,the framework complements model-centric advances and offers a transparent blueprint for production-grade behavioralmodeling.The code and processed artifacts are publicly available,and future work will extend the validation to longer,seasonal datasets and hybrid approaches that combine automated feature learning with domain-driven design.展开更多
Autism spectrum disorder(AsD)is a highly heterogeneous neurodevelopmental disorder.Early diagnosis and intervention are crucial for improving outcomes.Traditional single-modality diagnostic methods are subjective,limi...Autism spectrum disorder(AsD)is a highly heterogeneous neurodevelopmental disorder.Early diagnosis and intervention are crucial for improving outcomes.Traditional single-modality diagnostic methods are subjective,limited,and struggle to reveal the underlying pathological mechanisms.In contrast,multimodal data analysis integrates behavioral,physiological,and neuroimaging information with advanced machine-learning and deeplearning algorithms to overcome these limitations.In this review,we surveyed the recent pediatric AsD literature,highlighting artificial intelligence-driven diagnostic techniques,multimodal data fusion strategies,and emerging trends in ASD assessment.We surveyed studies that integrated two or more modalities and summarized the fusion levels,learning paradigms,tasks,datasets,and metrics.Multimodal approaches outperform singlemodality baselines in classification,severity estimation,and subtyping by leveraging complementary information and reducing modality-specific biases.Multimodal approaches significantly enhance diagnostic accuracy and comprehensiveness,enabling early screening of AsD,symptom subtyping,severity assessment,and personalized interventions.Advances in multimodal fusion techniques have promoted progress in precision medicine for the treatment of ASD.展开更多
In multi-modal emotion recognition,excessive reliance on historical context often impedes the detection of emotional shifts,while modality heterogeneity and unimodal noise limit recognition performance.Existing method...In multi-modal emotion recognition,excessive reliance on historical context often impedes the detection of emotional shifts,while modality heterogeneity and unimodal noise limit recognition performance.Existing methods struggle to dynamically adjust cross-modal complementary strength to optimize fusion quality and lack effective mechanisms to model the dynamic evolution of emotions.To address these issues,we propose a multi-level dynamic gating and emotion transfer framework for multi-modal emotion recognition.A dynamic gating mechanism is applied across unimodal encoding,cross-modal alignment,and emotion transfer modeling,substantially improving noise robustness and feature alignment.First,we construct a unimodal encoder based on gated recurrent units and feature-selection gating to suppress intra-modal noise and enhance contextual representation.Second,we design a gated-attention crossmodal encoder that dynamically calibrates the complementary contributions of visual and audio modalities to the dominant textual features and eliminates redundant information.Finally,we introduce a gated enhanced emotion transfer module that explicitly models the temporal dependence of emotional evolution in dialogues via transfer gating and optimizes continuity modeling with a comparative learning loss.Experimental results demonstrate that the proposed method outperforms state-of-the-art models on the public MELD and IEMOCAP datasets.展开更多
The Financial Technology(FinTech)sector has witnessed rapid growth,resulting in increasingly complex and high-volume digital transactions.Although this expansion improves efficiency and accessibility,it also introduce...The Financial Technology(FinTech)sector has witnessed rapid growth,resulting in increasingly complex and high-volume digital transactions.Although this expansion improves efficiency and accessibility,it also introduces significant vulnerabilities,including fraud,money laundering,and market manipulation.Traditional anomaly detection techniques often fail to capture the relational and dynamic characteristics of financial data.Graph Neural Networks(GNNs),capable of modeling intricate interdependencies among entities,have emerged as a powerful framework for detecting subtle and sophisticated anomalies.However,the high-dimensionality and inherent noise of FinTech datasets demand robust feature selection strategies to improve model scalability,performance,and interpretability.This paper presents a comprehensive survey of GNN-based approaches for anomaly detection in FinTech,with an emphasis on the synergistic role of feature selection.We examine the theoretical foundations of GNNs,review state-of-the-art feature selection techniques,analyze their integration with GNNs,and categorize prevalent anomaly types in FinTech applications.In addition,we discuss practical implementation challenges,highlight representative case studies,and propose future research directions to advance the field of graph-based anomaly detection in financial systems.展开更多
Real-time multi-person pose estimation(MPE)built upon neural network architectures aims to simultaneously detect multiple human instances and regress joint coordinates in dynamic scenes.However,due to factors such as ...Real-time multi-person pose estimation(MPE)built upon neural network architectures aims to simultaneously detect multiple human instances and regress joint coordinates in dynamic scenes.However,due to factors such as high model complexity and limited expression of keypoint information,both the efficiency and accuracy of real-time MPE remain to be improved.To mitigate the adverse impacts caused by the aforementioned issues,this work develops FSEM-Pose,a real-time MPE model rooted in the YOLOv10 framework.In detail,first,FSEM-Pose upgrades the backbone module of the baseline network by introducing the Feature Shuffling-Convolution(FS-Conv),which effectively reduces the backbone size while maximizing the retention of spatial information from the input image.Second,FSEM-Pose incorporates a Feature Saliency Enhancement Module(FSEM)to strengthen the feature encoding of human keypoints,thereby improving the accuracy of pose estimation.Finally,FSEM-Pose further enhances inference efficiency via a lightweight optimization of the head using shared convolutional layers.Our method achieves competitive results across multiple accuracy and efficiency metrics on the MS COCO 2017 and CrowdPose datasets.While being lightweight in design,it improves average precision(AP)by 2.1%and 2.5%,respectively.展开更多
Accurate detection of smoke and fire sources is critical for early fire warning and environmental monitoring.However,conventional detection approaches are highly susceptible to noise,illumination variations,and comple...Accurate detection of smoke and fire sources is critical for early fire warning and environmental monitoring.However,conventional detection approaches are highly susceptible to noise,illumination variations,and complex environmental conditions,which often reduce detection accuracy and real-time performance.To address these limitations,we propose Lightweight and Precise YOLO(LP-YOLO),a high-precision detection framework that integrates a self-attention mechanism with a feature pyramid,built upon YOLOv8.First,to overcome the restricted receptive field and parameter redundancy of conventional Convolutional Neural Networks(CNNs),we design an enhanced backbone based on Wavelet Convolutions(WTConv),which expands the receptive field through multifrequency convolutional processing.Second,a Bidirectional Feature Pyramid Network(BiFPN)is employed to achieve bidirectional feature fusion,enhancing the representation of smoke features across scales.Third,to mitigate the challenge of ambiguous object boundaries,we introduce the Frequency-aware Feature Fusion(FreqFusion)module,in which the Adaptive Low-Pass Filter(ALPF)reduces intra-class inconsistencies,the offset generator refines boundary localization,and the Adaptive High-Pass Filter(AHPF)recovers high-frequency details lost during down-sampling.Experimental evaluations demonstrate that LP-YOLO significantly outperforms the baseline YOLOv8,achieving an improvement of 9.3%in mAP@50 and 9.2%in F1-score.Moreover,the model is 56.6%and 32.4%smaller than YOLOv7-tiny and EfficientDet,respectively,while maintaining real-time inference speed at 238 frames per second(FPS).Validation on multiple benchmark datasets,including D-Fire,FIRESENSE,and BoWFire,further confirms its robustness and generalization ability,with detection accuracy consistently exceeding 82%.These results highlight the potential of LP-YOLO as a practical solution with high accuracy,robustness,and real-time performance for smoke and fire source detection.展开更多
文摘In geometry processing,symmetry research benefits from global geo-metric features of complete shapes,but the shape of an object captured in real-world applications is often incomplete due to the limited sensor resolution,single viewpoint,and occlusion.Different from the existing works predicting symmetry from the complete shape,we propose a learning approach for symmetry predic-tion based on a single RGB-D image.Instead of directly predicting the symmetry from incomplete shapes,our method consists of two modules,i.e.,the multi-mod-al feature fusion module and the detection-by-reconstruction module.Firstly,we build a channel-transformer network(CTN)to extract cross-fusion features from the RGB-D as the multi-modal feature fusion module,which helps us aggregate features from the color and the depth separately.Then,our self-reconstruction net-work based on a 3D variational auto-encoder(3D-VAE)takes the global geo-metric features as input,followed by a prediction symmetry network to detect the symmetry.Our experiments are conducted on three public datasets:ShapeNet,YCB,and ScanNet,we demonstrate that our method can produce reliable and accurate results.
基金supported by the Key Research and Development Program of Shandong Province(Soft Science Project)(2020RKB01364).
文摘In recent years,with the increase in the price of cryptocurrencies,the number of malicious cryptomining software has increased significantly.With their powerful spreading ability,cryptomining malware can unknowingly occupy our resources,harm our interests,and damage more legitimate assets.However,although current traditional rule-based malware detection methods have a low false alarm rate,they have a relatively low detection rate when faced with a large volume of emerging malware.Even though common machine learning-based or deep learning-based methods have certain ability to learn and detect unknown malware,the characteristics they learn are single and independent,and cannot be learned adaptively.Aiming at the above problems,we propose a deep learning model with multi-input of multi-modal features,which can simultaneously accept digital features and image features on different dimensions.The model in turn includes parallel learning of three sub-models and ensemble learning of another specific sub-model.The four sub-models can be processed in parallel on different devices and can be further applied to edge computing environments.The model can adaptively learn multi-modal features and output prediction results.The detection rate of our model is as high as 97.01%and the false alarm rate is only 0.63%.The experimental results prove the advantage and effectiveness of the proposed method.
基金Project(51875491) supported by the National Natural Science Foundation of ChinaProject(2021T3069) supported by the Fujian Science and Technology Plan STS Project,China。
文摘Laser cleaning is a highly nonlinear physical process for solving poor single-modal(e.g., acoustic or vision)detection performance and low inter-information utilization. In this study, a multi-modal feature fusion network model was constructed based on a laser paint removal experiment. The alignment of heterogeneous data under different modals was solved by combining the piecewise aggregate approximation and gramian angular field. Moreover, the attention mechanism was introduced to optimize the dual-path network and dense connection network, enabling the sampling characteristics to be extracted and integrated. Consequently, the multi-modal discriminant detection of laser paint removal was realized. According to the experimental results, the verification accuracy of the constructed model on the experimental dataset was 99.17%, which is 5.77% higher than the optimal single-modal detection results of the laser paint removal. The feature extraction network was optimized by the attention mechanism, and the model accuracy was increased by 3.3%. Results verify the improved classification performance of the constructed multi-modal feature fusion model in detecting laser paint removal, the effective integration of acoustic data and visual image data, and the accurate detection of laser paint removal.
基金National Youth Natural Science Foundation of China(No.61806006)Innovation Program for Graduate of Jiangsu Province(No.KYLX160-781)Jiangsu University Superior Discipline Construction Project。
文摘In order to solve difficult detection of far and hard objects due to the sparseness and insufficient semantic information of LiDAR point cloud,a 3D object detection network with multi-modal data adaptive fusion is proposed,which makes use of multi-neighborhood information of voxel and image information.Firstly,design an improved ResNet that maintains the structure information of far and hard objects in low-resolution feature maps,which is more suitable for detection task.Meanwhile,semantema of each image feature map is enhanced by semantic information from all subsequent feature maps.Secondly,extract multi-neighborhood context information with different receptive field sizes to make up for the defect of sparseness of point cloud which improves the ability of voxel features to represent the spatial structure and semantic information of objects.Finally,propose a multi-modal feature adaptive fusion strategy which uses learnable weights to express the contribution of different modal features to the detection task,and voxel attention further enhances the fused feature expression of effective target objects.The experimental results on the KITTI benchmark show that this method outperforms VoxelNet with remarkable margins,i.e.increasing the AP by 8.78%and 5.49%on medium and hard difficulty levels.Meanwhile,our method achieves greater detection performance compared with many mainstream multi-modal methods,i.e.outperforming the AP by 1%compared with that of MVX-Net on medium and hard difficulty levels.
基金National Natural Science Foundation of China(No.62271186)Anhui Key Project of Research and Development Plan(No.202104d07020005)。
文摘Deception detection plays a crucial role in criminal investigation.Videos contain a wealth of information regarding apparent and physiological changes in individuals,and thus can serve as an effective means of deception detection.In this paper,we investigate video-based deception detection considering both apparent visual features such as eye gaze,head pose and facial action unit(AU),and non-contact heart rate detected by remote photoplethysmography(rPPG)technique.Multiple wrapper-based feature selection methods combined with the K-nearest neighbor(KNN)and support vector machine(SVM)classifiers are employed to screen the most effective features for deception detection.We evaluate the performance of the proposed method on both a self-collected physiological-assisted visual deception detection(PV3D)dataset and a public bag-oflies(BOL)dataset.Experimental results demonstrate that the SVM classifier with symbiotic organisms search(SOS)feature selection yields the best overall performance,with an area under the curve(AUC)of 83.27%and accuracy(ACC)of 83.33%for PV3D,and an AUC of 71.18%and ACC of 70.33%for BOL.This demonstrates the stability and effectiveness of the proposed method in video-based deception detection tasks.
基金supported by the National Natural Science Foundation of China(Nos.62472202,U22A2041,and 62272309)the National Funded Postdoctoral Program of China(No.G2C20233162)the Science and Technology Major Project of Changsha(No.kh2402004).
文摘The complexity of cancer frequently results in diverse therapeutic responses among patients with the same cancer type undergoing identical treatments.Additionally,the development of anti-cancer drugs faces significant challenges due to extended timelines and high attrition rates during the process.Advances in machine learning,combined with the availability of extensive drug research databases,have facilitated the emergence of computational approaches designed to predict drug responses more accurately.However,the noise and uncertainty introduced by incomplete records across different biological data sources pose challenges,ultimately constraining the predictive capabilities of these models.To overcome these limitations,this study introduces a novel multi-modal learning-based drug response prediction framework,DSGPred.By constructing dynamic sample graphs,DSGPred accounts for missing modality types and quantities within individual samples,enabling a more granular understanding of data completeness.The framework integrates multi-modal heterogeneous graph convolutional networks with advanced fusion modules to deeply capture and synthesize diverse drug and cell line features.Additionally,DSGPred employs an interaction-focused feature extraction module to learn dual interaction modes,thereby enhancing the richness of drug-cell line interaction embeddings.Experimental evaluations on the benchmark and independent datasets indicate that DSGPred consistently surpasses existing methods in predictive performance.Furthermore,tests involving previously unseen drugs and cell lines validate DSGPred’s generalization ability.Practical applicability is further underscored through case studies,highlighting its utility in real-world scenarios,and offering robust predictions and insights for drug response prediction and personalized therapy design.The codes for DSGPred are available at https://github.com/zhc940702/DSGPred/.
文摘[Objective]Accurate prediction of tomato growth height is crucial for optimizing production environments in smart farming.However,current prediction methods predominantly rely on empirical,mechanistic,or learning-based models that utilize either images data or environmental data.These methods fail to fully leverage multi-modal data to capture the diverse aspects of plant growth comprehensively.[Methods]To address this limitation,a two-stage phenotypic feature extraction(PFE)model based on deep learning algorithm of recurrent neural network(RNN)and long short-term memory(LSTM)was developed.The model integrated environment and plant information to provide a holistic understanding of the growth process,emploied phenotypic and temporal feature extractors to comprehensively capture both types of features,enabled a deeper understanding of the interaction between tomato plants and their environment,ultimately leading to highly accurate predictions of growth height.[Results and Discussions]The experimental results showed the model's ef‐fectiveness:When predicting the next two days based on the past five days,the PFE-based RNN and LSTM models achieved mean absolute percentage error(MAPE)of 0.81%and 0.40%,respectively,which were significantly lower than the 8.00%MAPE of the large language model(LLM)and 6.72%MAPE of the Transformer-based model.In longer-term predictions,the 10-day prediction for 4 days ahead and the 30-day prediction for 12 days ahead,the PFE-RNN model continued to outperform the other two baseline models,with MAPE of 2.66%and 14.05%,respectively.[Conclusions]The proposed method,which leverages phenotypic-temporal collaboration,shows great potential for intelligent,data-driven management of tomato cultivation,making it a promising approach for enhancing the efficiency and precision of smart tomato planting management.
基金supported by the National Natural Science Foundation of China(No.62276204)the Fundamental Research Funds for the Central Universities,China(No.YJSJ24011)+1 种基金the Natural Science Basic Research Program of Shaanxi,China(Nos.2022JM-340 and 2023-JC-QN-0710)the China Postdoctoral Science Foundation(Nos.2020T130494 and 2018M633470)。
文摘Visible and infrared(RGB-IR)fusion object detection plays an important role in security,disaster relief,etc.In recent years,deep-learning-based RGB-IR fusion detection methods have been developing rapidly,but still struggle to deal with the complex and changing scenarios captured by drones,mainly due to two reasons:(A)RGB-IR fusion detectors are susceptible to inferior inputs that degrade performance and stability.(B)RGB-IR fusion detectors are susceptible to redundant features that reduce accuracy and efficiency.In this paper,an innovative RGB-IR fusion detection framework based on global-local feature optimization,named GLFDet,is proposed to improve the detection performance and efficiency of drone-captured objects.The key components of GLFDet include a Global Feature Optimization(GFO)module,a Local Feature Optimization(LFO)module and a Channel Separation Fusion(CSF)module.Specifically,GFO calculates the information content of the input image from the frequency domain and optimizes the features holistically.Then,LFO dynamically selects high-value features and filters out low-value features before fusion,which significantly improves the efficiency of fusion.Finally,CSF fuses the RGB and IR features across the corresponding channels,which avoids the rearrangement of the channel relationships and enhances the model stability.Extensive experimental results show that the proposed method achieves the best performance on three popular RGB-IR datasets Drone Vehicle,VEDAI,and LLVIP.In addition,GLFDet is more lightweight than other comparable models,making it more appealing to edge devices such as drones.The code is available at https://github.com/lao chen330/GLFDet.
基金funded by the Natural Science Foundation of Chongqing Municipality,grant number CSTB2022NSCQ-MSX0503.
文摘Gait recognition is a key biometric for long-distance identification,yet its performance is severely degraded by real-world challenges such as varying clothing,carrying conditions,and changing viewpoints.While combining silhouette and skeleton data is a promising direction,effectively fusing these heterogeneous modalities and adaptively weighting their contributions in response to diverse conditions remains a central problem.This paper introduces GaitMAFF,a novelMulti-modal Adaptive Feature Fusion Network,to address this challenge.Our approach first transforms discrete skeleton joints into a dense SkeletonMap representation to align with silhouettes,then employs an attention-based module to dynamically learn the fusion weights between the two modalities.These fused features are processed by a powerful spatio-temporal backbone withWeighted Global-Local Feature FusionModules(WFFM)to learn a discriminative representation.Extensive experiments on the challenging CCPG and Gait3D datasets show that GaitMAFF achieves state-of-the-art performance,with an average Rank-1 accuracy of 84.6%on CCPG and 58.7%on Gait3D.These results demonstrate that our adaptive fusion strategy effectively integrates complementary multimodal information,significantly enhancing gait recognition robustness and accuracy in complex scenes and providing a practical solution for real-world applications.
文摘Titanium-based semiconductors are known for their high chemical stability and suitable band gap widths.However,the conventional experimental screening methods are inefficient due to the wide variety of materials.To speed up the selection process,this work focuses on interpretable feature learning and band gap prediction for titanium-based semiconductors.First,titanium compounds were selected from the Materials Project database by machine learning,and elemental features were extracted using the Magpie descriptors.Then,principal component analysis(PCA)was applied to reduce the data dimensionality,creating a representative dataset.Meantime,heatmaps and SHAP(SHapley Additive exPlanations)methods were used to demonstrate the influence of key features such as electronegativity,covalent radius,period number,and unit cell volume on the bandgap,understanding the relationship between the material’s properties and performance.After comparing different machine learning models,including Random Forest(RF),Support Vector Machines(SVM),Linear Regression(LR),and Gradient Boosting Regression(GBR),the RF was found to be the most accurate for band gap prediction.Finally,the model performance was improved through parameter tuning,showing high accuracy.These findings provide strong data support and design guidance for the development of materials in fields like photocatalysis and solar cells.
基金funded by Deanship of Graduate studies and Scientific Research at Jouf University under grant No.(DGSSR-2024-02-01264).
文摘Automated essay scoring(AES)systems have gained significant importance in educational settings,offering a scalable,efficient,and objective method for evaluating student essays.However,developing AES systems for Arabic poses distinct challenges due to the language’s complex morphology,diglossia,and the scarcity of annotated datasets.This paper presents a hybrid approach to Arabic AES by combining text-based,vector-based,and embeddingbased similarity measures to improve essay scoring accuracy while minimizing the training data required.Using a large Arabic essay dataset categorized into thematic groups,the study conducted four experiments to evaluate the impact of feature selection,data size,and model performance.Experiment 1 established a baseline using a non-machine learning approach,selecting top-N correlated features to predict essay scores.The subsequent experiments employed 5-fold cross-validation.Experiment 2 showed that combining embedding-based,text-based,and vector-based features in a Random Forest(RF)model achieved an R2 of 88.92%and an accuracy of 83.3%within a 0.5-point tolerance.Experiment 3 further refined the feature selection process,demonstrating that 19 correlated features yielded optimal results,improving R2 to 88.95%.In Experiment 4,an optimal data efficiency training approach was introduced,where training data portions increased from 5%to 50%.The study found that using just 10%of the data achieved near-peak performance,with an R2 of 85.49%,emphasizing an effective trade-off between performance and computational costs.These findings highlight the potential of the hybrid approach for developing scalable Arabic AES systems,especially in low-resource environments,addressing linguistic challenges while ensuring efficient data usage.
基金the National Natural Science Foundation of China(61871464 and 61836002)the Fujian Provincial Natural Science Foundation of China(2018J01573)+1 种基金the Foundation of Fujian Educational Committee(JAT160357)Distinguished Young Scientific Research Talents Plan in Universities of Fujian Province and the Program for New Century Excellent Talents in University of Fujian Province.
文摘Image-based face pose estimation tries to estimate the facial direction with 2D images.It provides important information for many face recognition applications.However,it is a difficult task due to complex conditions and appearances.Deep learning method used in this field has the disadvantage of ignoring the natural structures of human faces.To solve this problem,a framework is proposed in this paper to estimate face poses with regression,which is based on deep learning and multi-modal feature loss(M2FL).Different from current loss functions using only a single type of features,the descriptive power was improved by combining multiple image features.To achieve it,hypergraph-based manifold regularization was applied.In this way,the loss of face pose estimation was reduced.Experimental results on commonly-used benchmark datasets demonstrate the performance of M2FL.
基金funded by the Hainan Province Science and Technology Special Fund under Grant ZDYF2024GXJS292.
文摘Deep learning has made significant progress in the field of oriented object detection for remote sensing images.However,existing methods still face challenges when dealing with difficult tasks such as multi-scale targets,complex backgrounds,and small objects in remote sensing.Maintaining model lightweight to address resource constraints in remote sensing scenarios while improving task completion for remote sensing tasks remains a research hotspot.Therefore,we propose an enhanced multi-scale feature extraction lightweight network EM-YOLO based on the YOLOv8s architecture,specifically optimized for the characteristics of large target scale variations,diverse orientations,and numerous small objects in remote sensing images.Our innovations lie in two main aspects:First,a dynamic snake convolution(DSC)is introduced into the backbone network to enhance the model’s feature extraction capability for oriented targets.Second,an innovative focusing-diffusion module is designed in the feature fusion neck to effectively integrate multi-scale feature information.Finally,we introduce Layer-Adaptive Sparsity for magnitude-based Pruning(LASP)method to perform lightweight network pruning to better complete tasks in resource-constrained scenarios.Experimental results on the lightweight platform Orin demonstrate that the proposed method significantly outperforms the original YOLOv8s model in oriented remote sensing object detection tasks,and achieves comparable or superior performance to state-of-the-art methods on three authoritative remote sensing datasets(DOTA v1.0,DOTA v1.5,and HRSC2016).
基金supported,in part,by the National Nature Science Foundation of China under Grant 62272236,62376128 and 62306139the Natural Science Foundation of Jiangsu Province under Grant BK20201136,BK20191401.
文摘Discriminative region localization and efficient feature encoding are crucial for fine-grained object recognition.However,existing data augmentation methods struggle to accurately locate discriminative regions in complex backgrounds,small target objects,and limited training data,leading to poor recognition.Fine-grained images exhibit“small inter-class differences,”and while second-order feature encoding enhances discrimination,it often requires dual Convolutional Neural Networks(CNN),increasing training time and complexity.This study proposes a model integrating discriminative region localization and efficient second-order feature encoding.By ranking feature map channels via a fully connected layer,it selects high-importance channels to generate an enhanced map,accurately locating discriminative regions.Cropping and erasing augmentations further refine recognition.To improve efficiency,a novel second-order feature encoding module generates an attention map from the fourth convolutional group of Residual Network 50 layers(ResNet-50)and multiplies it with features from the fifth group,producing second-order features while reducing dimensionality and training time.Experiments on Caltech-University of California,San Diego Birds-200-2011(CUB-200-2011),Stanford Car,and Fine-Grained Visual Classification of Aircraft(FGVC Aircraft)datasets show state-of-the-art accuracy of 88.9%,94.7%,and 93.3%,respectively.
基金supported by the research fund of Hanyang University(HY-202500000001616).
文摘Accurate purchase prediction in e-commerce critically depends on the quality of behavioral features.This paper proposes a layered and interpretable feature engineering framework that organizes user signals into three layers:Basic,Conversion&Stability(efficiency and volatility across actions),and Advanced Interactions&Activity(crossbehavior synergies and intensity).Using real Taobao(Alibaba’s primary e-commerce platform)logs(57,976 records for 10,203 users;25 November–03 December 2017),we conducted a hierarchical,layer-wise evaluation that holds data splits and hyperparameters fixed while varying only the feature set to quantify each layer’s marginal contribution.Across logistic regression(LR),decision tree,random forest,XGBoost,and CatBoost models with stratified 5-fold cross-validation,the performance improvedmonotonically fromBasic to Conversion&Stability to Advanced features.With LR,F1 increased from 0.613(Basic)to 0.962(Advanced);boosted models achieved high discrimination(0.995 AUC Score)and an F1 score up to 0.983.Calibration and precision–recall analyses indicated strong ranking quality and acknowledged potential dataset and period biases given the short(9-day)window.By making feature contributions measurable and reproducible,the framework complements model-centric advances and offers a transparent blueprint for production-grade behavioralmodeling.The code and processed artifacts are publicly available,and future work will extend the validation to longer,seasonal datasets and hybrid approaches that combine automated feature learning with domain-driven design.
基金supported by the National Key Research and Development Program of China(Research Grant Number:2023YFC3603600).
文摘Autism spectrum disorder(AsD)is a highly heterogeneous neurodevelopmental disorder.Early diagnosis and intervention are crucial for improving outcomes.Traditional single-modality diagnostic methods are subjective,limited,and struggle to reveal the underlying pathological mechanisms.In contrast,multimodal data analysis integrates behavioral,physiological,and neuroimaging information with advanced machine-learning and deeplearning algorithms to overcome these limitations.In this review,we surveyed the recent pediatric AsD literature,highlighting artificial intelligence-driven diagnostic techniques,multimodal data fusion strategies,and emerging trends in ASD assessment.We surveyed studies that integrated two or more modalities and summarized the fusion levels,learning paradigms,tasks,datasets,and metrics.Multimodal approaches outperform singlemodality baselines in classification,severity estimation,and subtyping by leveraging complementary information and reducing modality-specific biases.Multimodal approaches significantly enhance diagnostic accuracy and comprehensiveness,enabling early screening of AsD,symptom subtyping,severity assessment,and personalized interventions.Advances in multimodal fusion techniques have promoted progress in precision medicine for the treatment of ASD.
基金funded by“the Fanying Special Program of the National Natural Science Foundation of China,grant number 62341307”“the Scientific research project of Jiangxi Provincial Department of Education,grant number GJJ200839”“theDoctoral startup fund of JiangxiUniversity of Technology,grant number 205200100402”.
文摘In multi-modal emotion recognition,excessive reliance on historical context often impedes the detection of emotional shifts,while modality heterogeneity and unimodal noise limit recognition performance.Existing methods struggle to dynamically adjust cross-modal complementary strength to optimize fusion quality and lack effective mechanisms to model the dynamic evolution of emotions.To address these issues,we propose a multi-level dynamic gating and emotion transfer framework for multi-modal emotion recognition.A dynamic gating mechanism is applied across unimodal encoding,cross-modal alignment,and emotion transfer modeling,substantially improving noise robustness and feature alignment.First,we construct a unimodal encoder based on gated recurrent units and feature-selection gating to suppress intra-modal noise and enhance contextual representation.Second,we design a gated-attention crossmodal encoder that dynamically calibrates the complementary contributions of visual and audio modalities to the dominant textual features and eliminates redundant information.Finally,we introduce a gated enhanced emotion transfer module that explicitly models the temporal dependence of emotional evolution in dialogues via transfer gating and optimizes continuity modeling with a comparative learning loss.Experimental results demonstrate that the proposed method outperforms state-of-the-art models on the public MELD and IEMOCAP datasets.
基金supported by Ho Chi Minh City Open University,Vietnam under grant number E2024.02.1CD and Suan Sunandha Rajabhat University,Thailand.
文摘The Financial Technology(FinTech)sector has witnessed rapid growth,resulting in increasingly complex and high-volume digital transactions.Although this expansion improves efficiency and accessibility,it also introduces significant vulnerabilities,including fraud,money laundering,and market manipulation.Traditional anomaly detection techniques often fail to capture the relational and dynamic characteristics of financial data.Graph Neural Networks(GNNs),capable of modeling intricate interdependencies among entities,have emerged as a powerful framework for detecting subtle and sophisticated anomalies.However,the high-dimensionality and inherent noise of FinTech datasets demand robust feature selection strategies to improve model scalability,performance,and interpretability.This paper presents a comprehensive survey of GNN-based approaches for anomaly detection in FinTech,with an emphasis on the synergistic role of feature selection.We examine the theoretical foundations of GNNs,review state-of-the-art feature selection techniques,analyze their integration with GNNs,and categorize prevalent anomaly types in FinTech applications.In addition,we discuss practical implementation challenges,highlight representative case studies,and propose future research directions to advance the field of graph-based anomaly detection in financial systems.
基金supported by the Talent Startup Program of Huangshan University under Grant No.2025xkjq003Additional partial funding was gratefully received from the Scientific Research Project of the Anhui Provincial Department of Education under Grant No.2025AHGXZK40303.
文摘Real-time multi-person pose estimation(MPE)built upon neural network architectures aims to simultaneously detect multiple human instances and regress joint coordinates in dynamic scenes.However,due to factors such as high model complexity and limited expression of keypoint information,both the efficiency and accuracy of real-time MPE remain to be improved.To mitigate the adverse impacts caused by the aforementioned issues,this work develops FSEM-Pose,a real-time MPE model rooted in the YOLOv10 framework.In detail,first,FSEM-Pose upgrades the backbone module of the baseline network by introducing the Feature Shuffling-Convolution(FS-Conv),which effectively reduces the backbone size while maximizing the retention of spatial information from the input image.Second,FSEM-Pose incorporates a Feature Saliency Enhancement Module(FSEM)to strengthen the feature encoding of human keypoints,thereby improving the accuracy of pose estimation.Finally,FSEM-Pose further enhances inference efficiency via a lightweight optimization of the head using shared convolutional layers.Our method achieves competitive results across multiple accuracy and efficiency metrics on the MS COCO 2017 and CrowdPose datasets.While being lightweight in design,it improves average precision(AP)by 2.1%and 2.5%,respectively.
基金supported by the National Natural Science Foundation of China(No.62203163)the Scientific Research Project of Hunan Provincial Education Department(No.24A0519)+1 种基金the Hunan Provincial Natural Science Foundation(No.2025JJ60407)the Postgraduate Scientific Research Innovation Project of Hunan Province(No.CX2024100).
文摘Accurate detection of smoke and fire sources is critical for early fire warning and environmental monitoring.However,conventional detection approaches are highly susceptible to noise,illumination variations,and complex environmental conditions,which often reduce detection accuracy and real-time performance.To address these limitations,we propose Lightweight and Precise YOLO(LP-YOLO),a high-precision detection framework that integrates a self-attention mechanism with a feature pyramid,built upon YOLOv8.First,to overcome the restricted receptive field and parameter redundancy of conventional Convolutional Neural Networks(CNNs),we design an enhanced backbone based on Wavelet Convolutions(WTConv),which expands the receptive field through multifrequency convolutional processing.Second,a Bidirectional Feature Pyramid Network(BiFPN)is employed to achieve bidirectional feature fusion,enhancing the representation of smoke features across scales.Third,to mitigate the challenge of ambiguous object boundaries,we introduce the Frequency-aware Feature Fusion(FreqFusion)module,in which the Adaptive Low-Pass Filter(ALPF)reduces intra-class inconsistencies,the offset generator refines boundary localization,and the Adaptive High-Pass Filter(AHPF)recovers high-frequency details lost during down-sampling.Experimental evaluations demonstrate that LP-YOLO significantly outperforms the baseline YOLOv8,achieving an improvement of 9.3%in mAP@50 and 9.2%in F1-score.Moreover,the model is 56.6%and 32.4%smaller than YOLOv7-tiny and EfficientDet,respectively,while maintaining real-time inference speed at 238 frames per second(FPS).Validation on multiple benchmark datasets,including D-Fire,FIRESENSE,and BoWFire,further confirms its robustness and generalization ability,with detection accuracy consistently exceeding 82%.These results highlight the potential of LP-YOLO as a practical solution with high accuracy,robustness,and real-time performance for smoke and fire source detection.
