As the number and complexity of sensors in autonomous vehicles continue to rise,multimodal fusionbased object detection algorithms are increasingly being used to detect 3D environmental information,significantly advan...As the number and complexity of sensors in autonomous vehicles continue to rise,multimodal fusionbased object detection algorithms are increasingly being used to detect 3D environmental information,significantly advancing the development of perception technology in autonomous driving.To further promote the development of fusion algorithms and improve detection performance,this paper discusses the advantages and recent advancements of multimodal fusion-based object detection algorithms.Starting fromsingle-modal sensor detection,the paper provides a detailed overview of typical sensors used in autonomous driving and introduces object detection methods based on images and point clouds.For image-based detection methods,they are categorized into monocular detection and binocular detection based on different input types.For point cloud-based detection methods,they are classified into projection-based,voxel-based,point cluster-based,pillar-based,and graph structure-based approaches based on the technical pathways for processing point cloud features.Additionally,multimodal fusion algorithms are divided into Camera-LiDAR fusion,Camera-Radar fusion,Camera-LiDAR-Radar fusion,and other sensor fusion methods based on the types of sensors involved.Furthermore,the paper identifies five key future research directions in this field,aiming to provide insights for researchers engaged in multimodal fusion-based object detection algorithms and to encourage broader attention to the research and application of multimodal fusion-based object detection.展开更多
Objective To develop a non-invasive predictive model for coronary artery stenosis severity based on adaptive multi-modal integration of traditional Chinese and western medicine data.Methods Clinical indicators,echocar...Objective To develop a non-invasive predictive model for coronary artery stenosis severity based on adaptive multi-modal integration of traditional Chinese and western medicine data.Methods Clinical indicators,echocardiographic data,traditional Chinese medicine(TCM)tongue manifestations,and facial features were collected from patients who underwent coro-nary computed tomography angiography(CTA)in the Cardiac Care Unit(CCU)of Shanghai Tenth People's Hospital between May 1,2023 and May 1,2024.An adaptive weighted multi-modal data fusion(AWMDF)model based on deep learning was constructed to predict the severity of coronary artery stenosis.The model was evaluated using metrics including accura-cy,precision,recall,F1 score,and the area under the receiver operating characteristic(ROC)curve(AUC).Further performance assessment was conducted through comparisons with six ensemble machine learning methods,data ablation,model component ablation,and various decision-level fusion strategies.Results A total of 158 patients were included in the study.The AWMDF model achieved ex-cellent predictive performance(AUC=0.973,accuracy=0.937,precision=0.937,recall=0.929,and F1 score=0.933).Compared with model ablation,data ablation experiments,and various traditional machine learning models,the AWMDF model demonstrated superior per-formance.Moreover,the adaptive weighting strategy outperformed alternative approaches,including simple weighting,averaging,voting,and fixed-weight schemes.Conclusion The AWMDF model demonstrates potential clinical value in the non-invasive prediction of coronary artery disease and could serve as a tool for clinical decision support.展开更多
Multi-modal fusion technology gradually become a fundamental task in many fields,such as autonomous driving,smart healthcare,sentiment analysis,and human-computer interaction.It is rapidly becoming the dominant resear...Multi-modal fusion technology gradually become a fundamental task in many fields,such as autonomous driving,smart healthcare,sentiment analysis,and human-computer interaction.It is rapidly becoming the dominant research due to its powerful perception and judgment capabilities.Under complex scenes,multi-modal fusion technology utilizes the complementary characteristics of multiple data streams to fuse different data types and achieve more accurate predictions.However,achieving outstanding performance is challenging because of equipment performance limitations,missing information,and data noise.This paper comprehensively reviews existing methods based onmulti-modal fusion techniques and completes a detailed and in-depth analysis.According to the data fusion stage,multi-modal fusion has four primary methods:early fusion,deep fusion,late fusion,and hybrid fusion.The paper surveys the three majormulti-modal fusion technologies that can significantly enhance the effect of data fusion and further explore the applications of multi-modal fusion technology in various fields.Finally,it discusses the challenges and explores potential research opportunities.Multi-modal tasks still need intensive study because of data heterogeneity and quality.Preserving complementary information and eliminating redundant information between modalities is critical in multi-modal technology.Invalid data fusion methods may introduce extra noise and lead to worse results.This paper provides a comprehensive and detailed summary in response to these challenges.展开更多
Current works of environmental perception for connected autonomous electrified vehicles(CAEVs)mainly focus on the object detection task in good weather and illumination conditions,they often perform poorly in adverse ...Current works of environmental perception for connected autonomous electrified vehicles(CAEVs)mainly focus on the object detection task in good weather and illumination conditions,they often perform poorly in adverse scenarios and have a vague scene parsing ability.This paper aims to develop an end-to-end sharpening mixture of experts(SMoE)fusion framework to improve the robustness and accuracy of the perception systems for CAEVs in complex illumination and weather conditions.Three original contributions make our work distinctive from the existing relevant literature.The Complex KITTI dataset is introduced which consists of 7481 pairs of modified KITTI RGB images and the generated LiDAR dense depth maps,and this dataset is fine annotated in instance-level with the proposed semi-automatic annotation method.The SMoE fusion approach is devised to adaptively learn the robust kernels from complementary modalities.Comprehensive comparative experiments are implemented,and the results show that the proposed SMoE framework yield significant improvements over the other fusion techniques in adverse environmental conditions.This research proposes a SMoE fusion framework to improve the scene parsing ability of the perception systems for CAEVs in adverse conditions.展开更多
The application of fusion technology is of considerable importance in the field of multi-modal viewport prediction.The latest attention-based fusion methods have been shown to perform well in prediction accuracy.Howev...The application of fusion technology is of considerable importance in the field of multi-modal viewport prediction.The latest attention-based fusion methods have been shown to perform well in prediction accuracy.However,these methods fail to account for the differential density of information among the three modalities involved in viewport prediction-trajectory,visual,and audio.Visual and audio modalities present primitive signal information,while trajectory modality shows advanced time-series information.In this paper,a viewport prediction framework based on a Modality Diversity-Aware(MDA)fusion network is proposed to achieve multi-modal feature interaction.Firstly,we designed a fusion module to promote the combination of visual and auditory modalities,augmenting their efficacy as advanced complementary features.Subsequently,we utilize cross-modal attention to enable reinforced integration of visual-audio fused information and trajectory features.Our method addresses the issue of differing information densities among the three modalities,ensuring a fair and effective interaction between them.To evaluate the efficacy of the proposed approach,we conducted experiments on a widely-used public dataset.Experiments demonstrate that our approach predicts accurate viewport areas with a significant decrease in model parameters.展开更多
Multi-modal Named Entity Recognition(MNER)aims to better identify meaningful textual entities by integrating information from images.Previous work has focused on extracting visual semantics at a fine-grained level,or ...Multi-modal Named Entity Recognition(MNER)aims to better identify meaningful textual entities by integrating information from images.Previous work has focused on extracting visual semantics at a fine-grained level,or obtaining entity related external knowledge from knowledge bases or Large Language Models(LLMs).However,these approaches ignore the poor semantic correlation between visual and textual modalities in MNER datasets and do not explore different multi-modal fusion approaches.In this paper,we present MMAVK,a multi-modal named entity recognition model with auxiliary visual knowledge and word-level fusion,which aims to leverage the Multi-modal Large Language Model(MLLM)as an implicit knowledge base.It also extracts vision-based auxiliary knowledge from the image formore accurate and effective recognition.Specifically,we propose vision-based auxiliary knowledge generation,which guides the MLLM to extract external knowledge exclusively derived from images to aid entity recognition by designing target-specific prompts,thus avoiding redundant recognition and cognitive confusion caused by the simultaneous processing of image-text pairs.Furthermore,we employ a word-level multi-modal fusion mechanism to fuse the extracted external knowledge with each word-embedding embedded from the transformerbased encoder.Extensive experimental results demonstrate that MMAVK outperforms or equals the state-of-the-art methods on the two classical MNER datasets,even when the largemodels employed have significantly fewer parameters than other baselines.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
[Objectives]This study was conducted to achieve rapid and accurate detection of protein content in rice with a particle size of 1.0 mm.[Methods]A multi-model fusion strategy was proposed on the basis of Stacking ensem...[Objectives]This study was conducted to achieve rapid and accurate detection of protein content in rice with a particle size of 1.0 mm.[Methods]A multi-model fusion strategy was proposed on the basis of Stacking ensemble learning.A base learner pool was constructed,containing Partial Least Squares(PLS),Support Vector Machine(SVM),Deep Extreme Learning Machine(DELM),Random Forest(RF),Gradient Boosting Decision Tree(GBDT),and Multilayer Perceptron(MLP).PLS,DELM,and Linear Regression(LR)were used as meta-learner candidates.Employing integer coding technology,systematic dynamic combinations of base learners and meta-learners were generated,resulting in a total of 40 non-repetitive fusion models.The optimal combination was selected through a comprehensive evaluation based on multiple assessment indicators.[Results]The combination"PLS-DELM-MLP-LR"(code 1367)achieved coefficients of determination of 0.9732 and 0.9780 on the validation set and independent test set,respectively,with relative root mean square errors of 2.35%and 2.36%,and residual predictive deviations of 6.1075 and 6.7479,respectively.[Conclusions]The Stacking fusion model significantly enhances the predictive accuracy and robustness of spectral quantitative analysis,providing an efficient and feasible solution for modeling complex agricultural product spectral data.展开更多
Background:Medical imaging advancements are constrained by fundamental trade-offs between acquisition speed,radiation dose,and image quality,forcing clinicians to work with noisy,incomplete data.Existing reconstructio...Background:Medical imaging advancements are constrained by fundamental trade-offs between acquisition speed,radiation dose,and image quality,forcing clinicians to work with noisy,incomplete data.Existing reconstruction methods either compromise on accuracy with iterative algorithms or suffer from limited generalizability with task-specific deep learning approaches.Methods:We present LDM-PIR,a lightweight physics-conditioned diffusion multi-model for medical image reconstruction that addresses key challenges in magnetic resonance imaging(MRI),CT,and low-photon imaging.Unlike traditional iterative methods,which are computationally expensive,or task-specific deep learning approaches lacking generalizability,integrates three innovations.A physics-conditioned diffusion framework that embeds acquisition operators(Fourier/Radon transforms)and noise models directly into the reconstruction process.A multi-model architecture that unifies denoising,inpainting,and super-resolution via shared weight conditioning.A lightweight design(2.1M parameters)enabling rapid inference(0.8s/image on GPU).Through self-supervised fine-tuning with measurement consistency losses adapts to new imaging modalities using fewer annotated samples.Results:Achieves state-of-the-art performance on fastMRI(peak signal-to-noise ratio(PSNR):34.04 for single-coil/31.50 for multi-coil)and Lung Image Database Consortium and Image Database Resource Initiative(28.83 PSNR under Poisson noise).Clinical evaluations demonstrate superior preservation of anatomical structures,with SSIM improvements of 8.8%for single-coil and 4.36%for multi-coil MRI over uDPIR.Conclusion:It offers a flexible,efficient,and scalable solution for medical image reconstruction,addressing the challenges of noise,undersampling,and modality generalization.The model’s lightweight design allows for rapid inference,while its self-supervised fine-tuning capability minimizes reliance on large annotated datasets,making it suitable for real-world clinical applications.展开更多
Retinal diseases are a serious threat to human visual health and their early diagnosis is crucial.Currently,most of the retinal disease diagnostic algorithms are based on a single imaging modality of fundus color phot...Retinal diseases are a serious threat to human visual health and their early diagnosis is crucial.Currently,most of the retinal disease diagnostic algorithms are based on a single imaging modality of fundus color photography(FCP)or optical coherence tomography(OCT).These methods can only reflect retinal diseases to a certain extent,ignoring the speci ficity of modalities between different imaging modalities.In this research,a newmulti-scale feature fusion network(MSFF-Net)model for multi-modal retinal image diagnosis is proposed.The MSFF-Net model employs a dualbranch architecture design,enabling efficient learning and extraction of multi-modal feature information related to retinal diseases from CFP and OCT images.MSFF-Net improves disease diagnosis by combining multi-scale features of CFP and OCT images.When evaluated on challenging datasets,the model achieved an accuracy of 95.00%and an F1-score of 95.24%for retinal disease diagnosis.Even under low-quality dataset conditions,it maintained robust performance,with diagnostic accuracy and F1-scores of 71.50%and 71.73%,respectively.In addition,the MSFFNet model outperformed eight state-of-the-art single and multi-modal models in the comparison experiments.The proposed MSFF-Net model provides ophthalmologists with a more accurate and efficient diagnostic pathway that helps them detect and treat retinal diseases earlier.展开更多
To address the challenge of missing modal information in entity alignment and to mitigate information loss or bias arising frommodal heterogeneity during fusion,while also capturing shared information acrossmodalities...To address the challenge of missing modal information in entity alignment and to mitigate information loss or bias arising frommodal heterogeneity during fusion,while also capturing shared information acrossmodalities,this paper proposes a Multi-modal Pre-synergistic Entity Alignmentmodel based on Cross-modalMutual Information Strategy Optimization(MPSEA).The model first employs independent encoders to process multi-modal features,including text,images,and numerical values.Next,a multi-modal pre-synergistic fusion mechanism integrates graph structural and visual modal features into the textual modality as preparatory information.This pre-fusion strategy enables unified perception of heterogeneous modalities at the model’s initial stage,reducing discrepancies during the fusion process.Finally,using cross-modal deep perception reinforcement learning,the model achieves adaptive multilevel feature fusion between modalities,supporting learningmore effective alignment strategies.Extensive experiments on multiple public datasets show that the MPSEA method achieves gains of up to 7% in Hits@1 and 8.2% in MRR on the FBDB15K dataset,and up to 9.1% in Hits@1 and 7.7% in MRR on the FBYG15K dataset,compared to existing state-of-the-art methods.These results confirm the effectiveness of the proposed model.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
Integrating multiple medical imaging techniques,including Magnetic Resonance Imaging(MRI),Computed Tomography,Positron Emission Tomography(PET),and ultrasound,provides a comprehensive view of the patient health status...Integrating multiple medical imaging techniques,including Magnetic Resonance Imaging(MRI),Computed Tomography,Positron Emission Tomography(PET),and ultrasound,provides a comprehensive view of the patient health status.Each of these methods contributes unique diagnostic insights,enhancing the overall assessment of patient condition.Nevertheless,the amalgamation of data from multiple modalities presents difficulties due to disparities in resolution,data collection methods,and noise levels.While traditional models like Convolutional Neural Networks(CNNs)excel in single-modality tasks,they struggle to handle multi-modal complexities,lacking the capacity to model global relationships.This research presents a novel approach for examining multi-modal medical imagery using a transformer-based system.The framework employs self-attention and cross-attention mechanisms to synchronize and integrate features across various modalities.Additionally,it shows resilience to variations in noise and image quality,making it adaptable for real-time clinical use.To address the computational hurdles linked to transformer models,particularly in real-time clinical applications in resource-constrained environments,several optimization techniques have been integrated to boost scalability and efficiency.Initially,a streamlined transformer architecture was adopted to minimize the computational load while maintaining model effectiveness.Methods such as model pruning,quantization,and knowledge distillation have been applied to reduce the parameter count and enhance the inference speed.Furthermore,efficient attention mechanisms such as linear or sparse attention were employed to alleviate the substantial memory and processing requirements of traditional self-attention operations.For further deployment optimization,researchers have implemented hardware-aware acceleration strategies,including the use of TensorRT and ONNX-based model compression,to ensure efficient execution on edge devices.These optimizations allow the approach to function effectively in real-time clinical settings,ensuring viability even in environments with limited resources.Future research directions include integrating non-imaging data to facilitate personalized treatment and enhancing computational efficiency for implementation in resource-limited environments.This study highlights the transformative potential of transformer models in multi-modal medical imaging,offering improvements in diagnostic accuracy and patient care outcomes.展开更多
The primary objective of Chinese spelling correction(CSC)is to detect and correct erroneous characters in Chinese text,which can result from various factors,such as inaccuracies in pinyin representation,character rese...The primary objective of Chinese spelling correction(CSC)is to detect and correct erroneous characters in Chinese text,which can result from various factors,such as inaccuracies in pinyin representation,character resemblance,and semantic discrepancies.However,existing methods often struggle to fully address these types of errors,impacting the overall correction accuracy.This paper introduces a multi-modal feature encoder designed to efficiently extract features from three distinct modalities:pinyin,semantics,and character morphology.Unlike previous methods that rely on direct fusion or fixed-weight summation to integrate multi-modal information,our approach employs a multi-head attention mechanism to focuse more on relevant modal information while dis-regarding less pertinent data.To prevent issues such as gradient explosion or vanishing,the model incorporates a residual connection of the original text vector for fine-tuning.This approach ensures robust model performance by maintaining essential linguistic details throughout the correction process.Experimental evaluations on the SIGHAN benchmark dataset demonstrate that the pro-posed model outperforms baseline approaches across various metrics and datasets,confirming its effectiveness and feasibility.展开更多
Multi-modal knowledge graph completion(MMKGC)aims to complete missing entities or relations in multi-modal knowledge graphs,thereby discovering more previously unknown triples.Due to the continuous growth of data and ...Multi-modal knowledge graph completion(MMKGC)aims to complete missing entities or relations in multi-modal knowledge graphs,thereby discovering more previously unknown triples.Due to the continuous growth of data and knowledge and the limitations of data sources,the visual knowledge within the knowledge graphs is generally of low quality,and some entities suffer from the issue of missing visual modality.Nevertheless,previous studies of MMKGC have primarily focused on how to facilitate modality interaction and fusion while neglecting the problems of low modality quality and modality missing.In this case,mainstream MMKGC models only use pre-trained visual encoders to extract features and transfer the semantic information to the joint embeddings through modal fusion,which inevitably suffers from problems such as error propagation and increased uncertainty.To address these problems,we propose a Multi-modal knowledge graph Completion model based on Super-resolution and Detailed Description Generation(MMCSD).Specifically,we leverage a pre-trained residual network to enhance the resolution and improve the quality of the visual modality.Moreover,we design multi-level visual semantic extraction and entity description generation,thereby further extracting entity semantics from structural triples and visual images.Meanwhile,we train a variational multi-modal auto-encoder and utilize a pre-trained multi-modal language model to complement the missing visual features.We conducted experiments on FB15K-237 and DB13K,and the results showed that MMCSD can effectively perform MMKGC and achieve state-of-the-art performance.展开更多
Background Gesture recognition has attracted significant attention because of its wide range of potential applications.Although multi-modal gesture recognition has made significant progress in recent years,a popular m...Background Gesture recognition has attracted significant attention because of its wide range of potential applications.Although multi-modal gesture recognition has made significant progress in recent years,a popular method still is simply fusing prediction scores at the end of each branch,which often ignores complementary features among different modalities in the early stage and does not fuse the complementary features into a more discriminative feature.Methods This paper proposes an Adaptive Cross-modal Weighting(ACmW)scheme to exploit complementarity features from RGB-D data in this study.The scheme learns relations among different modalities by combining the features of different data streams.The proposed ACmW module contains two key functions:(1)fusing complementary features from multiple streams through an adaptive one-dimensional convolution;and(2)modeling the correlation of multi-stream complementary features in the time dimension.Through the effective combination of these two functional modules,the proposed ACmW can automatically analyze the relationship between the complementary features from different streams,and can fuse them in the spatial and temporal dimensions.Results Extensive experiments validate the effectiveness of the proposed method,and show that our method outperforms state-of-the-art methods on IsoGD and NVGesture.展开更多
Seeing through dense occlusions and reconstructing scene images is an important but challenging task.Traditional framebased image de-occlusion methods may lead to fatal errors when facing extremely dense occlusions du...Seeing through dense occlusions and reconstructing scene images is an important but challenging task.Traditional framebased image de-occlusion methods may lead to fatal errors when facing extremely dense occlusions due to the lack of valid information available from the limited input occluded frames.Event cameras are bio-inspired vision sensors that record the brightness changes at each pixel asynchronously with high temporal resolution.However,synthesizing images solely from event streams is ill-posed since only the brightness changes are recorded in the event stream,and the initial brightness is unknown.In this paper,we propose an event-enhanced multi-modal fusion hybrid network for image de-occlusion,which uses event streams to provide complete scene information and frames to provide color and texture information.An event stream encoder based on the spiking neural network(SNN)is proposed to encode and denoise the event stream efficiently.A comparison loss is proposed to generate clearer results.Experimental results on a largescale event-based and frame-based image de-occlusion dataset demonstrate that our proposed method achieves state-of-the-art performance.展开更多
Power Shell has been widely deployed in fileless malware and advanced persistent threat(APT)attacks due to its high stealthiness and live-off-theland technique.However,existing works mainly focus on deobfuscation and ...Power Shell has been widely deployed in fileless malware and advanced persistent threat(APT)attacks due to its high stealthiness and live-off-theland technique.However,existing works mainly focus on deobfuscation and malicious detection,lacking the malicious Power Shell families classification and behavior analysis.Moreover,the state-of-the-art methods fail to capture fine-grained features and semantic relationships,resulting in low robustness and accuracy.To this end,we propose Power Detector,a novel malicious Power Shell script detector based on multimodal semantic fusion and deep learning.Specifically,we design four feature extraction methods to extract key features from character,token,abstract syntax tree(AST),and semantic knowledge graph.Then,we intelligently design four embeddings(i.e.,Char2Vec,Token2Vec,AST2Vec,and Rela2Vec) and construct a multi-modal fusion algorithm to concatenate feature vectors from different views.Finally,we propose a combined model based on transformer and CNN-Bi LSTM to implement Power Shell family detection.Our experiments with five types of Power Shell attacks show that PowerDetector can accurately detect various obfuscated and stealth PowerShell scripts,with a 0.9402 precision,a 0.9358 recall,and a 0.9374 F1-score.Furthermore,through singlemodal and multi-modal comparison experiments,we demonstrate that PowerDetector’s multi-modal embedding and deep learning model can achieve better accuracy and even identify more unknown attacks.展开更多
For the analysis of spinal and disc diseases,automated tissue segmentation of the lumbar spine is vital.Due to the continuous and concentrated location of the target,the abundance of edge features,and individual diffe...For the analysis of spinal and disc diseases,automated tissue segmentation of the lumbar spine is vital.Due to the continuous and concentrated location of the target,the abundance of edge features,and individual differences,conventional automatic segmentation methods perform poorly.Since the success of deep learning in the segmentation of medical images has been shown in the past few years,it has been applied to this task in a number of ways.The multi-scale and multi-modal features of lumbar tissues,however,are rarely explored by methodologies of deep learning.Because of the inadequacies in medical images availability,it is crucial to effectively fuse various modes of data collection for model training to alleviate the problem of insufficient samples.In this paper,we propose a novel multi-modality hierarchical fusion network(MHFN)for improving lumbar spine segmentation by learning robust feature representations from multi-modality magnetic resonance images.An adaptive group fusion module(AGFM)is introduced in this paper to fuse features from various modes to extract cross-modality features that could be valuable.Furthermore,to combine features from low to high levels of cross-modality,we design a hierarchical fusion structure based on AGFM.Compared to the other feature fusion methods,AGFM is more effective based on experimental results on multi-modality MR images of the lumbar spine.To further enhance segmentation accuracy,we compare our network with baseline fusion structures.Compared to the baseline fusion structures(input-level:76.27%,layer-level:78.10%,decision-level:79.14%),our network was able to segment fractured vertebrae more accurately(85.05%).展开更多
The multi-modal characteristics of mineral particles play a pivotal role in enhancing the classification accuracy,which is critical for obtaining a profound understanding of the Earth's composition and ensuring ef...The multi-modal characteristics of mineral particles play a pivotal role in enhancing the classification accuracy,which is critical for obtaining a profound understanding of the Earth's composition and ensuring effective exploitation utilization of its resources.However,the existing methods for classifying mineral particles do not fully utilize these multi-modal features,thereby limiting the classification accuracy.Furthermore,when conventional multi-modal image classification methods are applied to planepolarized and cross-polarized sequence images of mineral particles,they encounter issues such as information loss,misaligned features,and challenges in spatiotemporal feature extraction.To address these challenges,we propose a multi-modal mineral particle polarization image classification network(MMGC-Net)for precise mineral particle classification.Initially,MMGC-Net employs a two-dimensional(2D)backbone network with shared parameters to extract features from two types of polarized images to ensure feature alignment.Subsequently,a cross-polarized intra-modal feature fusion module is designed to refine the spatiotemporal features from the extracted features of the cross-polarized sequence images.Ultimately,the inter-modal feature fusion module integrates the two types of modal features to enhance the classification precision.Quantitative and qualitative experimental results indicate that when compared with the current state-of-the-art multi-modal image classification methods,MMGC-Net demonstrates marked superiority in terms of mineral particle multi-modal feature learning and four classification evaluation metrics.It also demonstrates better stability than the existing models.展开更多
基金funded by the Yangtze River Delta Science and Technology Innovation Community Joint Research Project(2023CSJGG1600)the Natural Science Foundation of Anhui Province(2208085MF173)Wuhu“ChiZhu Light”Major Science and Technology Project(2023ZD01,2023ZD03).
文摘As the number and complexity of sensors in autonomous vehicles continue to rise,multimodal fusionbased object detection algorithms are increasingly being used to detect 3D environmental information,significantly advancing the development of perception technology in autonomous driving.To further promote the development of fusion algorithms and improve detection performance,this paper discusses the advantages and recent advancements of multimodal fusion-based object detection algorithms.Starting fromsingle-modal sensor detection,the paper provides a detailed overview of typical sensors used in autonomous driving and introduces object detection methods based on images and point clouds.For image-based detection methods,they are categorized into monocular detection and binocular detection based on different input types.For point cloud-based detection methods,they are classified into projection-based,voxel-based,point cluster-based,pillar-based,and graph structure-based approaches based on the technical pathways for processing point cloud features.Additionally,multimodal fusion algorithms are divided into Camera-LiDAR fusion,Camera-Radar fusion,Camera-LiDAR-Radar fusion,and other sensor fusion methods based on the types of sensors involved.Furthermore,the paper identifies five key future research directions in this field,aiming to provide insights for researchers engaged in multimodal fusion-based object detection algorithms and to encourage broader attention to the research and application of multimodal fusion-based object detection.
基金Construction Program of the Key Discipline of State Administration of Traditional Chinese Medicine of China(ZYYZDXK-2023069)Research Project of Shanghai Municipal Health Commission (2024QN018)Shanghai University of Traditional Chinese Medicine Science and Technology Development Program (23KFL005)。
文摘Objective To develop a non-invasive predictive model for coronary artery stenosis severity based on adaptive multi-modal integration of traditional Chinese and western medicine data.Methods Clinical indicators,echocardiographic data,traditional Chinese medicine(TCM)tongue manifestations,and facial features were collected from patients who underwent coro-nary computed tomography angiography(CTA)in the Cardiac Care Unit(CCU)of Shanghai Tenth People's Hospital between May 1,2023 and May 1,2024.An adaptive weighted multi-modal data fusion(AWMDF)model based on deep learning was constructed to predict the severity of coronary artery stenosis.The model was evaluated using metrics including accura-cy,precision,recall,F1 score,and the area under the receiver operating characteristic(ROC)curve(AUC).Further performance assessment was conducted through comparisons with six ensemble machine learning methods,data ablation,model component ablation,and various decision-level fusion strategies.Results A total of 158 patients were included in the study.The AWMDF model achieved ex-cellent predictive performance(AUC=0.973,accuracy=0.937,precision=0.937,recall=0.929,and F1 score=0.933).Compared with model ablation,data ablation experiments,and various traditional machine learning models,the AWMDF model demonstrated superior per-formance.Moreover,the adaptive weighting strategy outperformed alternative approaches,including simple weighting,averaging,voting,and fixed-weight schemes.Conclusion The AWMDF model demonstrates potential clinical value in the non-invasive prediction of coronary artery disease and could serve as a tool for clinical decision support.
基金supported by the Natural Science Foundation of Liaoning Province(Grant No.2023-MSBA-070)the National Natural Science Foundation of China(Grant No.62302086).
文摘Multi-modal fusion technology gradually become a fundamental task in many fields,such as autonomous driving,smart healthcare,sentiment analysis,and human-computer interaction.It is rapidly becoming the dominant research due to its powerful perception and judgment capabilities.Under complex scenes,multi-modal fusion technology utilizes the complementary characteristics of multiple data streams to fuse different data types and achieve more accurate predictions.However,achieving outstanding performance is challenging because of equipment performance limitations,missing information,and data noise.This paper comprehensively reviews existing methods based onmulti-modal fusion techniques and completes a detailed and in-depth analysis.According to the data fusion stage,multi-modal fusion has four primary methods:early fusion,deep fusion,late fusion,and hybrid fusion.The paper surveys the three majormulti-modal fusion technologies that can significantly enhance the effect of data fusion and further explore the applications of multi-modal fusion technology in various fields.Finally,it discusses the challenges and explores potential research opportunities.Multi-modal tasks still need intensive study because of data heterogeneity and quality.Preserving complementary information and eliminating redundant information between modalities is critical in multi-modal technology.Invalid data fusion methods may introduce extra noise and lead to worse results.This paper provides a comprehensive and detailed summary in response to these challenges.
基金Supported by National Natural Science Foundation of China(Grant Nos.51975118,52025121,51975103,51905095)National Natural Science Foundation of Jiangsu Province(Grant No.BK20180401).
文摘Current works of environmental perception for connected autonomous electrified vehicles(CAEVs)mainly focus on the object detection task in good weather and illumination conditions,they often perform poorly in adverse scenarios and have a vague scene parsing ability.This paper aims to develop an end-to-end sharpening mixture of experts(SMoE)fusion framework to improve the robustness and accuracy of the perception systems for CAEVs in complex illumination and weather conditions.Three original contributions make our work distinctive from the existing relevant literature.The Complex KITTI dataset is introduced which consists of 7481 pairs of modified KITTI RGB images and the generated LiDAR dense depth maps,and this dataset is fine annotated in instance-level with the proposed semi-automatic annotation method.The SMoE fusion approach is devised to adaptively learn the robust kernels from complementary modalities.Comprehensive comparative experiments are implemented,and the results show that the proposed SMoE framework yield significant improvements over the other fusion techniques in adverse environmental conditions.This research proposes a SMoE fusion framework to improve the scene parsing ability of the perception systems for CAEVs in adverse conditions.
文摘The application of fusion technology is of considerable importance in the field of multi-modal viewport prediction.The latest attention-based fusion methods have been shown to perform well in prediction accuracy.However,these methods fail to account for the differential density of information among the three modalities involved in viewport prediction-trajectory,visual,and audio.Visual and audio modalities present primitive signal information,while trajectory modality shows advanced time-series information.In this paper,a viewport prediction framework based on a Modality Diversity-Aware(MDA)fusion network is proposed to achieve multi-modal feature interaction.Firstly,we designed a fusion module to promote the combination of visual and auditory modalities,augmenting their efficacy as advanced complementary features.Subsequently,we utilize cross-modal attention to enable reinforced integration of visual-audio fused information and trajectory features.Our method addresses the issue of differing information densities among the three modalities,ensuring a fair and effective interaction between them.To evaluate the efficacy of the proposed approach,we conducted experiments on a widely-used public dataset.Experiments demonstrate that our approach predicts accurate viewport areas with a significant decrease in model parameters.
基金funded by Research Project,grant number BHQ090003000X03.
文摘Multi-modal Named Entity Recognition(MNER)aims to better identify meaningful textual entities by integrating information from images.Previous work has focused on extracting visual semantics at a fine-grained level,or obtaining entity related external knowledge from knowledge bases or Large Language Models(LLMs).However,these approaches ignore the poor semantic correlation between visual and textual modalities in MNER datasets and do not explore different multi-modal fusion approaches.In this paper,we present MMAVK,a multi-modal named entity recognition model with auxiliary visual knowledge and word-level fusion,which aims to leverage the Multi-modal Large Language Model(MLLM)as an implicit knowledge base.It also extracts vision-based auxiliary knowledge from the image formore accurate and effective recognition.Specifically,we propose vision-based auxiliary knowledge generation,which guides the MLLM to extract external knowledge exclusively derived from images to aid entity recognition by designing target-specific prompts,thus avoiding redundant recognition and cognitive confusion caused by the simultaneous processing of image-text pairs.Furthermore,we employ a word-level multi-modal fusion mechanism to fuse the extracted external knowledge with each word-embedding embedded from the transformerbased encoder.Extensive experimental results demonstrate that MMAVK outperforms or equals the state-of-the-art methods on the two classical MNER datasets,even when the largemodels employed have significantly fewer parameters than other baselines.
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.
文摘[Objectives]This study was conducted to achieve rapid and accurate detection of protein content in rice with a particle size of 1.0 mm.[Methods]A multi-model fusion strategy was proposed on the basis of Stacking ensemble learning.A base learner pool was constructed,containing Partial Least Squares(PLS),Support Vector Machine(SVM),Deep Extreme Learning Machine(DELM),Random Forest(RF),Gradient Boosting Decision Tree(GBDT),and Multilayer Perceptron(MLP).PLS,DELM,and Linear Regression(LR)were used as meta-learner candidates.Employing integer coding technology,systematic dynamic combinations of base learners and meta-learners were generated,resulting in a total of 40 non-repetitive fusion models.The optimal combination was selected through a comprehensive evaluation based on multiple assessment indicators.[Results]The combination"PLS-DELM-MLP-LR"(code 1367)achieved coefficients of determination of 0.9732 and 0.9780 on the validation set and independent test set,respectively,with relative root mean square errors of 2.35%and 2.36%,and residual predictive deviations of 6.1075 and 6.7479,respectively.[Conclusions]The Stacking fusion model significantly enhances the predictive accuracy and robustness of spectral quantitative analysis,providing an efficient and feasible solution for modeling complex agricultural product spectral data.
文摘Background:Medical imaging advancements are constrained by fundamental trade-offs between acquisition speed,radiation dose,and image quality,forcing clinicians to work with noisy,incomplete data.Existing reconstruction methods either compromise on accuracy with iterative algorithms or suffer from limited generalizability with task-specific deep learning approaches.Methods:We present LDM-PIR,a lightweight physics-conditioned diffusion multi-model for medical image reconstruction that addresses key challenges in magnetic resonance imaging(MRI),CT,and low-photon imaging.Unlike traditional iterative methods,which are computationally expensive,or task-specific deep learning approaches lacking generalizability,integrates three innovations.A physics-conditioned diffusion framework that embeds acquisition operators(Fourier/Radon transforms)and noise models directly into the reconstruction process.A multi-model architecture that unifies denoising,inpainting,and super-resolution via shared weight conditioning.A lightweight design(2.1M parameters)enabling rapid inference(0.8s/image on GPU).Through self-supervised fine-tuning with measurement consistency losses adapts to new imaging modalities using fewer annotated samples.Results:Achieves state-of-the-art performance on fastMRI(peak signal-to-noise ratio(PSNR):34.04 for single-coil/31.50 for multi-coil)and Lung Image Database Consortium and Image Database Resource Initiative(28.83 PSNR under Poisson noise).Clinical evaluations demonstrate superior preservation of anatomical structures,with SSIM improvements of 8.8%for single-coil and 4.36%for multi-coil MRI over uDPIR.Conclusion:It offers a flexible,efficient,and scalable solution for medical image reconstruction,addressing the challenges of noise,undersampling,and modality generalization.The model’s lightweight design allows for rapid inference,while its self-supervised fine-tuning capability minimizes reliance on large annotated datasets,making it suitable for real-world clinical applications.
基金supported by the National Natural Science Foundation of China(Nos.82472104 and U24B2053)the Natural Science Basic Research Program of Shaanxi(No.2025JC-JCQN-023)+2 种基金the Key Core Technology Research and Development of Shaanxi(No.2024QY2-GJHX-03)the Innovation Capability Support Program of Shaanxi(Program No.2023-CX-TD-54)the Xidian University Specially Funded Project for Interdisciplinary Exploration(No.TZJHF202510).
文摘Retinal diseases are a serious threat to human visual health and their early diagnosis is crucial.Currently,most of the retinal disease diagnostic algorithms are based on a single imaging modality of fundus color photography(FCP)or optical coherence tomography(OCT).These methods can only reflect retinal diseases to a certain extent,ignoring the speci ficity of modalities between different imaging modalities.In this research,a newmulti-scale feature fusion network(MSFF-Net)model for multi-modal retinal image diagnosis is proposed.The MSFF-Net model employs a dualbranch architecture design,enabling efficient learning and extraction of multi-modal feature information related to retinal diseases from CFP and OCT images.MSFF-Net improves disease diagnosis by combining multi-scale features of CFP and OCT images.When evaluated on challenging datasets,the model achieved an accuracy of 95.00%and an F1-score of 95.24%for retinal disease diagnosis.Even under low-quality dataset conditions,it maintained robust performance,with diagnostic accuracy and F1-scores of 71.50%and 71.73%,respectively.In addition,the MSFFNet model outperformed eight state-of-the-art single and multi-modal models in the comparison experiments.The proposed MSFF-Net model provides ophthalmologists with a more accurate and efficient diagnostic pathway that helps them detect and treat retinal diseases earlier.
基金partially supported by the National Natural Science Foundation of China under Grants 62471493 and 62402257(for conceptualization and investigation)partially supported by the Natural Science Foundation of Shandong Province,China under Grants ZR2023LZH017,ZR2024MF066,and 2023QF025(for formal analysis and validation)+1 种基金partially supported by the Open Foundation of Key Laboratory of Computing Power Network and Information Security,Ministry of Education,Qilu University of Technology(Shandong Academy of Sciences)under Grant 2023ZD010(for methodology and model design)partially supported by the Russian Science Foundation(RSF)Project under Grant 22-71-10095-P(for validation and results verification).
文摘To address the challenge of missing modal information in entity alignment and to mitigate information loss or bias arising frommodal heterogeneity during fusion,while also capturing shared information acrossmodalities,this paper proposes a Multi-modal Pre-synergistic Entity Alignmentmodel based on Cross-modalMutual Information Strategy Optimization(MPSEA).The model first employs independent encoders to process multi-modal features,including text,images,and numerical values.Next,a multi-modal pre-synergistic fusion mechanism integrates graph structural and visual modal features into the textual modality as preparatory information.This pre-fusion strategy enables unified perception of heterogeneous modalities at the model’s initial stage,reducing discrepancies during the fusion process.Finally,using cross-modal deep perception reinforcement learning,the model achieves adaptive multilevel feature fusion between modalities,supporting learningmore effective alignment strategies.Extensive experiments on multiple public datasets show that the MPSEA method achieves gains of up to 7% in Hits@1 and 8.2% in MRR on the FBDB15K dataset,and up to 9.1% in Hits@1 and 7.7% in MRR on the FBYG15K dataset,compared to existing state-of-the-art methods.These results confirm the effectiveness of the proposed model.
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supported by the Deanship of Research and Graduate Studies at King Khalid University under Small Research Project grant number RGP1/139/45.
文摘Integrating multiple medical imaging techniques,including Magnetic Resonance Imaging(MRI),Computed Tomography,Positron Emission Tomography(PET),and ultrasound,provides a comprehensive view of the patient health status.Each of these methods contributes unique diagnostic insights,enhancing the overall assessment of patient condition.Nevertheless,the amalgamation of data from multiple modalities presents difficulties due to disparities in resolution,data collection methods,and noise levels.While traditional models like Convolutional Neural Networks(CNNs)excel in single-modality tasks,they struggle to handle multi-modal complexities,lacking the capacity to model global relationships.This research presents a novel approach for examining multi-modal medical imagery using a transformer-based system.The framework employs self-attention and cross-attention mechanisms to synchronize and integrate features across various modalities.Additionally,it shows resilience to variations in noise and image quality,making it adaptable for real-time clinical use.To address the computational hurdles linked to transformer models,particularly in real-time clinical applications in resource-constrained environments,several optimization techniques have been integrated to boost scalability and efficiency.Initially,a streamlined transformer architecture was adopted to minimize the computational load while maintaining model effectiveness.Methods such as model pruning,quantization,and knowledge distillation have been applied to reduce the parameter count and enhance the inference speed.Furthermore,efficient attention mechanisms such as linear or sparse attention were employed to alleviate the substantial memory and processing requirements of traditional self-attention operations.For further deployment optimization,researchers have implemented hardware-aware acceleration strategies,including the use of TensorRT and ONNX-based model compression,to ensure efficient execution on edge devices.These optimizations allow the approach to function effectively in real-time clinical settings,ensuring viability even in environments with limited resources.Future research directions include integrating non-imaging data to facilitate personalized treatment and enhancing computational efficiency for implementation in resource-limited environments.This study highlights the transformative potential of transformer models in multi-modal medical imaging,offering improvements in diagnostic accuracy and patient care outcomes.
基金Supported by the National Natural Science Foundation of China(No.61472256,61170277)the Hujiang Foundation(No.A14006).
文摘The primary objective of Chinese spelling correction(CSC)is to detect and correct erroneous characters in Chinese text,which can result from various factors,such as inaccuracies in pinyin representation,character resemblance,and semantic discrepancies.However,existing methods often struggle to fully address these types of errors,impacting the overall correction accuracy.This paper introduces a multi-modal feature encoder designed to efficiently extract features from three distinct modalities:pinyin,semantics,and character morphology.Unlike previous methods that rely on direct fusion or fixed-weight summation to integrate multi-modal information,our approach employs a multi-head attention mechanism to focuse more on relevant modal information while dis-regarding less pertinent data.To prevent issues such as gradient explosion or vanishing,the model incorporates a residual connection of the original text vector for fine-tuning.This approach ensures robust model performance by maintaining essential linguistic details throughout the correction process.Experimental evaluations on the SIGHAN benchmark dataset demonstrate that the pro-posed model outperforms baseline approaches across various metrics and datasets,confirming its effectiveness and feasibility.
基金funded by Research Project,grant number BHQ090003000X03。
文摘Multi-modal knowledge graph completion(MMKGC)aims to complete missing entities or relations in multi-modal knowledge graphs,thereby discovering more previously unknown triples.Due to the continuous growth of data and knowledge and the limitations of data sources,the visual knowledge within the knowledge graphs is generally of low quality,and some entities suffer from the issue of missing visual modality.Nevertheless,previous studies of MMKGC have primarily focused on how to facilitate modality interaction and fusion while neglecting the problems of low modality quality and modality missing.In this case,mainstream MMKGC models only use pre-trained visual encoders to extract features and transfer the semantic information to the joint embeddings through modal fusion,which inevitably suffers from problems such as error propagation and increased uncertainty.To address these problems,we propose a Multi-modal knowledge graph Completion model based on Super-resolution and Detailed Description Generation(MMCSD).Specifically,we leverage a pre-trained residual network to enhance the resolution and improve the quality of the visual modality.Moreover,we design multi-level visual semantic extraction and entity description generation,thereby further extracting entity semantics from structural triples and visual images.Meanwhile,we train a variational multi-modal auto-encoder and utilize a pre-trained multi-modal language model to complement the missing visual features.We conducted experiments on FB15K-237 and DB13K,and the results showed that MMCSD can effectively perform MMKGC and achieve state-of-the-art performance.
基金the Chinese National Natural Science Foundation Projects(61961160704,61876179)the Key Project of the General Logistics Department(ASW17C001)the Science and Technology Development Fund of Macao(0010/2019/AFJ,0025/2019/AKP).
文摘Background Gesture recognition has attracted significant attention because of its wide range of potential applications.Although multi-modal gesture recognition has made significant progress in recent years,a popular method still is simply fusing prediction scores at the end of each branch,which often ignores complementary features among different modalities in the early stage and does not fuse the complementary features into a more discriminative feature.Methods This paper proposes an Adaptive Cross-modal Weighting(ACmW)scheme to exploit complementarity features from RGB-D data in this study.The scheme learns relations among different modalities by combining the features of different data streams.The proposed ACmW module contains two key functions:(1)fusing complementary features from multiple streams through an adaptive one-dimensional convolution;and(2)modeling the correlation of multi-stream complementary features in the time dimension.Through the effective combination of these two functional modules,the proposed ACmW can automatically analyze the relationship between the complementary features from different streams,and can fuse them in the spatial and temporal dimensions.Results Extensive experiments validate the effectiveness of the proposed method,and show that our method outperforms state-of-the-art methods on IsoGD and NVGesture.
基金supported by National Natural Science Funds of China (Nos. 62088102 and 62021002)Beijing Natural Science Foundation, China (No. 4222025)
文摘Seeing through dense occlusions and reconstructing scene images is an important but challenging task.Traditional framebased image de-occlusion methods may lead to fatal errors when facing extremely dense occlusions due to the lack of valid information available from the limited input occluded frames.Event cameras are bio-inspired vision sensors that record the brightness changes at each pixel asynchronously with high temporal resolution.However,synthesizing images solely from event streams is ill-posed since only the brightness changes are recorded in the event stream,and the initial brightness is unknown.In this paper,we propose an event-enhanced multi-modal fusion hybrid network for image de-occlusion,which uses event streams to provide complete scene information and frames to provide color and texture information.An event stream encoder based on the spiking neural network(SNN)is proposed to encode and denoise the event stream efficiently.A comparison loss is proposed to generate clearer results.Experimental results on a largescale event-based and frame-based image de-occlusion dataset demonstrate that our proposed method achieves state-of-the-art performance.
基金This work was supported by National Natural Science Foundation of China(No.62172308,No.U1626107,No.61972297,No.62172144,and No.62062019).
文摘Power Shell has been widely deployed in fileless malware and advanced persistent threat(APT)attacks due to its high stealthiness and live-off-theland technique.However,existing works mainly focus on deobfuscation and malicious detection,lacking the malicious Power Shell families classification and behavior analysis.Moreover,the state-of-the-art methods fail to capture fine-grained features and semantic relationships,resulting in low robustness and accuracy.To this end,we propose Power Detector,a novel malicious Power Shell script detector based on multimodal semantic fusion and deep learning.Specifically,we design four feature extraction methods to extract key features from character,token,abstract syntax tree(AST),and semantic knowledge graph.Then,we intelligently design four embeddings(i.e.,Char2Vec,Token2Vec,AST2Vec,and Rela2Vec) and construct a multi-modal fusion algorithm to concatenate feature vectors from different views.Finally,we propose a combined model based on transformer and CNN-Bi LSTM to implement Power Shell family detection.Our experiments with five types of Power Shell attacks show that PowerDetector can accurately detect various obfuscated and stealth PowerShell scripts,with a 0.9402 precision,a 0.9358 recall,and a 0.9374 F1-score.Furthermore,through singlemodal and multi-modal comparison experiments,we demonstrate that PowerDetector’s multi-modal embedding and deep learning model can achieve better accuracy and even identify more unknown attacks.
基金supported in part by the Technology Innovation 2030 under Grant 2022ZD0211700.
文摘For the analysis of spinal and disc diseases,automated tissue segmentation of the lumbar spine is vital.Due to the continuous and concentrated location of the target,the abundance of edge features,and individual differences,conventional automatic segmentation methods perform poorly.Since the success of deep learning in the segmentation of medical images has been shown in the past few years,it has been applied to this task in a number of ways.The multi-scale and multi-modal features of lumbar tissues,however,are rarely explored by methodologies of deep learning.Because of the inadequacies in medical images availability,it is crucial to effectively fuse various modes of data collection for model training to alleviate the problem of insufficient samples.In this paper,we propose a novel multi-modality hierarchical fusion network(MHFN)for improving lumbar spine segmentation by learning robust feature representations from multi-modality magnetic resonance images.An adaptive group fusion module(AGFM)is introduced in this paper to fuse features from various modes to extract cross-modality features that could be valuable.Furthermore,to combine features from low to high levels of cross-modality,we design a hierarchical fusion structure based on AGFM.Compared to the other feature fusion methods,AGFM is more effective based on experimental results on multi-modality MR images of the lumbar spine.To further enhance segmentation accuracy,we compare our network with baseline fusion structures.Compared to the baseline fusion structures(input-level:76.27%,layer-level:78.10%,decision-level:79.14%),our network was able to segment fractured vertebrae more accurately(85.05%).
基金supported by the National Natural Science Foundation of China(Grant Nos.62071315 and 62271336).
文摘The multi-modal characteristics of mineral particles play a pivotal role in enhancing the classification accuracy,which is critical for obtaining a profound understanding of the Earth's composition and ensuring effective exploitation utilization of its resources.However,the existing methods for classifying mineral particles do not fully utilize these multi-modal features,thereby limiting the classification accuracy.Furthermore,when conventional multi-modal image classification methods are applied to planepolarized and cross-polarized sequence images of mineral particles,they encounter issues such as information loss,misaligned features,and challenges in spatiotemporal feature extraction.To address these challenges,we propose a multi-modal mineral particle polarization image classification network(MMGC-Net)for precise mineral particle classification.Initially,MMGC-Net employs a two-dimensional(2D)backbone network with shared parameters to extract features from two types of polarized images to ensure feature alignment.Subsequently,a cross-polarized intra-modal feature fusion module is designed to refine the spatiotemporal features from the extracted features of the cross-polarized sequence images.Ultimately,the inter-modal feature fusion module integrates the two types of modal features to enhance the classification precision.Quantitative and qualitative experimental results indicate that when compared with the current state-of-the-art multi-modal image classification methods,MMGC-Net demonstrates marked superiority in terms of mineral particle multi-modal feature learning and four classification evaluation metrics.It also demonstrates better stability than the existing models.
