Underwater pipeline inspection plays a vital role in the proactive maintenance and management of critical marine infrastructure and subaquatic systems.However,the inspection of underwater pipelines presents a challeng...Underwater pipeline inspection plays a vital role in the proactive maintenance and management of critical marine infrastructure and subaquatic systems.However,the inspection of underwater pipelines presents a challenge due to factors such as light scattering,absorption,restricted visibility,and ambient noise.The advancement of deep learning has introduced powerful techniques for processing large amounts of unstructured and imperfect data collected from underwater environments.This study evaluated the efficacy of the You Only Look Once(YOLO)algorithm,a real-time object detection and localization model based on convolutional neural networks,in identifying and classifying various types of pipeline defects in underwater settings.YOLOv8,the latest evolution in the YOLO family,integrates advanced capabilities,such as anchor-free detection,a cross-stage partial network backbone for efficient feature extraction,and a feature pyramid network+path aggregation network neck for robust multi-scale object detection,which make it particularly well-suited for complex underwater environments.Due to the lack of suitable open-access datasets for underwater pipeline defects,a custom dataset was captured using a remotely operated vehicle in a controlled environment.This application has the following assets available for use.Extensive experimentation demonstrated that YOLOv8 X-Large consistently outperformed other models in terms of pipe defect detection and classification and achieved a strong balance between precision and recall in identifying pipeline cracks,rust,corners,defective welds,flanges,tapes,and holes.This research establishes the baseline performance of YOLOv8 for underwater defect detection and showcases its potential to enhance the reliability and efficiency of pipeline inspection tasks in challenging underwater environments.展开更多
Artificial Intelligence(AI)in healthcare enables predicting diabetes using data-driven methods instead of the traditional ways of screening the disease,which include hemoglobin A1c(HbA1c),oral glucose tolerance test(O...Artificial Intelligence(AI)in healthcare enables predicting diabetes using data-driven methods instead of the traditional ways of screening the disease,which include hemoglobin A1c(HbA1c),oral glucose tolerance test(OGTT),and fasting plasma glucose(FPG)screening techniques,which are invasive and limited in scale.Machine learning(ML)and deep neural network(DNN)models that use large datasets to learn the complex,nonlinear feature interactions,but the conventional ML algorithms are data sensitive and often show unstable predictive accuracy.Conversely,DNN models are more robust,though the ability to reach a high accuracy rate consistently on heterogeneous datasets is still an open challenge.For predicting diabetes,this work proposed a hybrid DNN approach by integrating a bidirectional long short-term memory(BiLSTM)network with a bidirectional gated recurrent unit(BiGRU).A robust DL model,developed by combining various datasets with weighted coefficients,dense operations in the connection of deep layers,and the output aggregation using batch normalization and dropout functions to avoid overfitting.The goal of this hybrid model is better generalization and consistency among various datasets,which facilitates the effective management and early intervention.The proposed DNN model exhibits an excellent predictive performance as compared to the state-of-the-art and baseline ML and DNN models for diabetes prediction tasks.The robust performance indicates the possible usefulness of DL-based models in the development of disease prediction in healthcare and other areas that demand high-quality analytics.展开更多
Accurate streamflow prediction is essential for flood warning,reservoir operation,irrigation scheduling,hydropower planning,and sustainable water management,yet remains challenging due to the complexity of hydrologica...Accurate streamflow prediction is essential for flood warning,reservoir operation,irrigation scheduling,hydropower planning,and sustainable water management,yet remains challenging due to the complexity of hydrological processes.Although data-driven models often outperform conventional physics-based hydrological modelling approaches,their real-world deployment is limited by cost,infrastructure demands,and the interdisciplinary expertise required.To bridge this gap,this study developed QPred,a regional,lightweight,cost-effective,web-delivered application for daily streamflow forecasting.The study executed an end-to-end workflow,from field data acquisition to accessible web-based deployment for on-demand forecasting.High-resolution rainfall data were recorded with tippingbucket gauges and loggers,while river water depth in the Aglar and Paligaad watersheds was converted to discharge using site-specific rating curves,resulting in a daily dataset of precipitation,river water level and discharge.Four DL architectures were trained,including vanilla Long Short-Term Memory(LSTM),stacked LSTM,bidirectional LSTM,and Gated Recurrent Unit(GRU),and evaluated using Nash-Sutcliffe Efficiency(NSE),Coefficient of Determination(R2),Root-Mean-Square-Error-Standard-Deviation Ratio(RSR),and Percentage Bias(PBIAS)metrics.Performance was watershed-specific,as the vanilla LSTM demonstrated the best generalisation for the Aglar watershed(R2=0.88,NSE=0.82,RMSE=0.12 during validation),while the GRU achieved the highest validation accuracy in Paligaad(R2=0.88,NSE=0.88,RMSE=0.49).All models achieved satisfactory to excellent performance during calibration(R2>0.91,NSE>0.91 for both watersheds),demonstrating strong capability to capture streamflow dynamics.The highest performing models were selected and embedded into the QPred application.QPred was developed as a lightweight web pipeline,utilising Google Colab as the primary execution environment,Flask as the backend inference framework,Google Drive for artefact storage,andNgrok for secureHTTPS tunnelling.Auser-friendly front end utilises range sliders(bounded by observed minima and maxima)to gather inputs and provides discharge data along with metadata,thereby enhancing transparency.This work demonstrates that accurate,context-aware deep learningmodels can be delivered through low-cost,web-based platforms,providing a reproducible and scalable pipeline for hydrological applications in other watersheds and for practitioners.展开更多
Human Activity Recognition(HAR)is a novel area for computer vision.It has a great impact on healthcare,smart environments,and surveillance while is able to automatically detect human behavior.It plays a vital role in ...Human Activity Recognition(HAR)is a novel area for computer vision.It has a great impact on healthcare,smart environments,and surveillance while is able to automatically detect human behavior.It plays a vital role in many applications,such as smart home,healthcare,human computer interaction,sports analysis,and especially,intelligent surveillance.In this paper,we propose a robust and efficient HAR system by leveraging deep learning paradigms,including pre-trained models,CNN architectures,and their average-weighted fusion.However,due to the diversity of human actions and various environmental influences,as well as a lack of data and resources,achieving high recognition accuracy remain elusive.In this work,a weighted average ensemble technique is employed to fuse three deep learning models:EfficientNet,ResNet50,and a custom CNN.The results of this study indicate that using a weighted average ensemble strategy for developing more effective HAR models may be a promising idea for detection and classification of human activities.Experiments by using the benchmark dataset proved that the proposed weighted ensemble approach outperformed existing approaches in terms of accuracy and other key performance measures.The combined average-weighted ensemble of pre-trained and CNN models obtained an accuracy of 98%,compared to 97%,96%,and 95%for the customized CNN,EfficientNet,and ResNet50 models,respectively.展开更多
To achieve the potential performance gain of massive multiple-input multiple-output(MIMO)systems,base stations(BS)require downlink channel state information(CSI)fed back by users to execute beamforming design,especial...To achieve the potential performance gain of massive multiple-input multiple-output(MIMO)systems,base stations(BS)require downlink channel state information(CSI)fed back by users to execute beamforming design,especially in the frequency division duplex(FDD)systems.However,due to the enormous number of antennas in massive MIMO systems,the feedback overhead of downlink CSI acquisition is extremely large.To address this issue,deep learning(DL)techniques have been introduced to de velop high-accuracy feedback strategies under limited backhaul constraints.In this paper,we provide an overview of DL-based CSI compression and feedback approaches in massive MIMO systems.Specifically,we introduce the conventional CSI compression and feedback schemes and the existing problems.Besides,we elaborate on various DL techniques employed in CSI compression from the perspective of network architecture and analyze the advantages of different techniques.We also enumerate the applications of DL-based methods for solving practical challenges in CSI compression and feedback.In addition,we brief the remaining issues in deep CSI compression and indicate potential directions in future wireless networks.展开更多
Wearable sensors integrated with deep learning techniques have the potential to revolutionize seamless human-machine interfaces for real-time health monitoring,clinical diagnosis,and robotic applications.Nevertheless,...Wearable sensors integrated with deep learning techniques have the potential to revolutionize seamless human-machine interfaces for real-time health monitoring,clinical diagnosis,and robotic applications.Nevertheless,it remains a critical challenge to simultaneously achieve desirable mechanical and electrical performance along with biocompatibility,adhesion,self-healing,and environmental robustness with excellent sensing metrics.Herein,we report a multifunctional,anti-freezing,selfadhesive,and self-healable organogel pressure sensor composed of cobalt nanoparticle encapsulated nitrogen-doped carbon nanotubes(CoN CNT)embedded in a polyvinyl alcohol-gelatin(PVA/GLE)matrix.Fabricated using a binary solvent system of water and ethylene glycol(EG),the CoN CNT/PVA/GLE organogel exhibits excellent flexibility,biocompatibility,and temperature tolerance with remarkable environmental stability.Electrochemical impedance spectroscopy confirms near-stable performance across a broad humidity range(40%-95%RH).Freeze-tolerant conductivity under sub-zero conditions(-20℃)is attributed to the synergistic role of CoN CNT and EG,preserving mobility and network integrity.The Co N CNT/PVA/GLE organogel sensor exhibits high sensitivity of 5.75 k Pa^(-1)in the detection range from 0 to 20 k Pa,ideal for subtle biomechanical motion detection.A smart human-machine interface for English letter recognition using deep learning achieved 98%accuracy.The organogel sensor utility was extended to detect human gestures like finger bending,wrist motion,and throat vibration during speech.展开更多
This paper proposes a deep learning-based 3D LiDAR perception framework designed for applications such as autonomous robots and vehicles.To address the high dependency on large-scale annotated data—an inherent limita...This paper proposes a deep learning-based 3D LiDAR perception framework designed for applications such as autonomous robots and vehicles.To address the high dependency on large-scale annotated data—an inherent limitation of deep learning models—this study introduces a hybrid perception architecture that incorporates expertdriven LiDAR processing techniques into the deep neural network.Traditional 3DLiDAR processingmethods typically remove ground planes and apply distance-or density-based clustering for object detection.In this work,such expert knowledge is encoded as feature-level inputs and fused with the deep network,therebymitigating the data dependency issue of conventional learning-based approaches.Specifically,the proposedmethod combines two expert algorithms—Patchwork++for ground segmentation and DBSCAN for clustering—with a PointPillars-based LiDAR detection network.We design four hybrid versions of the network depending on the stage and method of integrating expert features into the feature map of the deep model.Among these,Version 4 incorporates a modified neck structure in PointPillars and introduces a new Cluster 2D Pseudo-Map Branch that utilizes cluster-level pseudo-images generated from Patchwork++and DBSCAN.This version achieved a+3.88%improvement mean Average Precision(mAP)compared to the baseline PointPillars.The results demonstrate that embedding expert-based perception logic into deep neural architectures can effectively enhance performance and reduce dependency on extensive training datasets,offering a promising direction for robust 3D LiDAR object detection in real-world scenarios.展开更多
Background:Studies have shown that heart rate variability(HRV)is a predictor of the prognosis of cardiovascular diseases.Contact heartbeat monitoring equipment is widely used,especially in hospitals,and benefits from ...Background:Studies have shown that heart rate variability(HRV)is a predictor of the prognosis of cardiovascular diseases.Contact heartbeat monitoring equipment is widely used,especially in hospitals,and benefits from the rapidity and accuracy of the detection of physiological health indicators.However,long-term contact with equipment has many adverse effects.The purpose of this study was to improve the accuracy of HRV detection via noncontact equipment,thus enabling HRV to be assessed in various scenarios.Methods:A novel deep learning approach was proposed for measuring heartbeats through camera videos.First,we performed facial segmentation and divided the face into 16 grid cells with different light balance scores.After the trend is filtered by the Hamming window,a transformer-based neural network is used to further filter the signal.Finally,heart rate(HR)and HRV are estimated.Results:We used 1 million synthetic data points for pretraining and a public dataset in combination with a dataset that we constructed for task training.The final results were obtained on a test dataset that we constructed.The accuracy for HR with a low light balance score(0.867-0.983)was greater than that with a high score(0.667-0.750).Our method had higher accuracy in estimating HR than traditional filtering methods(0.167-0.417)and state-of-the-art neural network filtering methods(0.783-0.917)did.The root mean square error of the HRV from the time domain was the lowest,and the correlation index score was the highest for the HRV from the frequency domain estimated by our method compared with those estimated by two neural networks.Conclusions:Light balance,large sample training,and two-stage training can improve the accuracy of HRV estimation.展开更多
Nondestructive measurement technology of phenotype can provide substantial phenotypic data support for applications such as seedling breeding,management,and quality testing.The current method of measuring seedling phe...Nondestructive measurement technology of phenotype can provide substantial phenotypic data support for applications such as seedling breeding,management,and quality testing.The current method of measuring seedling phenotypes mainly relies on manual measurement which is inefficient,subjective and destroys samples.Therefore,the paper proposes a nondestructive measurement method for the canopy phenotype of the watermelon plug seedlings based on deep learning.The Azure Kinect was used to shoot canopy color images,depth images,and RGB-D images of the watermelon plug seedlings.The Mask-RCNN network was used to classify,segment,and count the canopy leaves of the watermelon plug seedlings.To reduce the error of leaf area measurement caused by mutual occlusion of leaves,the leaves were repaired by CycleGAN,and the depth images were restored by image processing.Then,the Delaunay triangulation was adopted to measure the leaf area in the leaf point cloud.The YOLOX target detection network was used to identify the growing point position of each seedling on the plug tray.Then the depth differences between the growing point and the upper surface of the plug tray were calculated to obtain plant height.The experiment results show that the nondestructive measurement algorithm proposed in this paper achieves good measurement performance for the watermelon plug seedlings from the 1 true-leaf to 3 true-leaf stages.The average relative error of measurement is 2.33%for the number of true leaves,4.59%for the number of cotyledons,8.37%for the leaf area,and 3.27%for the plant height.The experiment results demonstrate that the proposed algorithm in this paper provides an effective solution for the nondestructive measurement of the canopy phenotype of the plug seedlings.展开更多
The stator of the maglev track plays a crucial role in the operation of the maglev system.Currently,the efficiency of maglev track inspection is limited by several factors,including the large span of elevated structur...The stator of the maglev track plays a crucial role in the operation of the maglev system.Currently,the efficiency of maglev track inspection is limited by several factors,including the large span of elevated structures,manual visual inspection,short inspection window times,and limited GPS positioning accuracy.To address these issues,this paper proposes a deep learning-based method for detecting and locating stator surface damage.This study establishes a maglev track stator surface image dataset,trains different object detection models,and compares their performance.Ultimately,YOLO and ByteTrack object tracking algorithms were chosen as the basic framework and enhanced to achieve automatic identification of high-speed maglev track stator surface damage images and track and count stator surface localization feature images.By matching the identified damaged images with their corresponding stator segment and beam segment sequence numbers,the location of the damage is pinpointed to the corresponding stator segment,enabling rapid and accurate identification and localization of complex damage to the maglev track stator surface.展开更多
The exponential growth of video content has driven significant advancements in video summarization techniques in recent years. Breakthroughs in deep learning have been particularly transformative, enabling more effect...The exponential growth of video content has driven significant advancements in video summarization techniques in recent years. Breakthroughs in deep learning have been particularly transformative, enabling more effective detection of key information and creating new possibilities for video synopsis. To summarize recent progress and accelerate research in this field,this paper provides a comprehensive review of deep learningbased video summarization methods developed over the past decade. We begin by examining the research landscape of video abstraction technologies and identifying core challenges in video summarization. Subsequently, we systematically analyze prevailing deep learning frameworks and methodologies employed in current video summarization systems, offering researchers a clear roadmap of the field's evolution. Unlike previous review works,we first classify research papers based on the structural hierarchy of the video(from frame-level to shot-level to video-level),then further categorize them according to the summary backbone model(feature extraction and spatiotemporal modeling).This approach provides a more systematic and hierarchical organization of the documents. Following this comprehensive review,we summarize the benchmark datasets and evaluation metrics commonly employed in the field. Finally, we analyze persistent challenges and propose insightful directions for future research,providing a forward-looking perspective on video summarization technologies. This systematic literature review is of great reference value to new researchers exploring the fields of deep learning and video summarization.展开更多
Underground engineering projects such as deep tunnel excavation often encounter rockburst disasters accompanied by numerous microseismic events.Rapid interpretation of microseismic signals is crucial for the timely id...Underground engineering projects such as deep tunnel excavation often encounter rockburst disasters accompanied by numerous microseismic events.Rapid interpretation of microseismic signals is crucial for the timely identification of rockbursts.However,conventional processing encompasses multi-step workflows,including classification,denoising,picking,locating,and computational analysis,coupled with manual intervention,which collectively compromise the reliability of early warnings.To address these challenges,this study innovatively proposes the“microseismic stethoscope"-a multi-task machine learning and deep learning model designed for the automated processing of massive microseismic signals.This model efficiently extracts three key parameters that are necessary for recognizing rockburst disasters:rupture location,microseismic energy,and moment magnitude.Specifically,the model extracts raw waveform features from three dedicated sub-networks:a classifier for source zone classification,and two regressors for microseismic energy and moment magnitude estimation.This model demonstrates superior efficiency compared to traditional processing and semi-automated processing,reducing per-event processing time from 0.71 s to 0.49 s to merely 0.036 s.It concurrently achieves 98%accuracy in source zone classification,with microseismic energy and moment magnitude estimation errors of 0.13 and 0.05,respectively.This model has been well applied and validated in the Daxiagu Tunnel case in Sichuan,China.The application results indicate that the model is as accurate as traditional methods in determining source parameters,and thus can be used to identify potential geomechanical processes of rockburst disasters.By enhancing the signal processing reliability of microseismic events,the proposed model in this study presents a significant advancement in the identification of rockburst disasters.展开更多
Microscopy imaging is fundamental in analyzing bacterial morphology and dynamics,offering critical insights into bacterial physiology and pathogenicity.Image segmentation techniques enable quantitative analysis of bac...Microscopy imaging is fundamental in analyzing bacterial morphology and dynamics,offering critical insights into bacterial physiology and pathogenicity.Image segmentation techniques enable quantitative analysis of bacterial structures,facilitating precise measurement of morphological variations and population behaviors at single-cell resolution.This paper reviews advancements in bacterial image segmentation,emphasizing the shift from traditional thresholding and watershed methods to deep learning-driven approaches.Convolutional neural networks(CNNs),U-Net architectures,and three-dimensional(3D)frameworks excel at segmenting dense biofilms and resolving antibiotic-induced morphological changes.These methods combine automated feature extraction with physics-informed postprocessing.Despite progress,challenges persist in computational efficiency,cross-species generalizability,and integration with multimodal experimental workflows.Future progress will depend on improving model robustness across species and imaging modalities,integrating multimodal data for phenotype-function mapping,and developing standard pipelines that link computational tools with clinical diagnostics.These innovations will expand microbial phenotyping beyond structural analysis,enabling deeper insights into bacterial physiology and ecological interactions.展开更多
This study proposed a deep learning-based nanoindentation simulation method to address the challenge of obtaining the mechanical parameters of rock-forming minerals and the complexity of regression analysis.This appro...This study proposed a deep learning-based nanoindentation simulation method to address the challenge of obtaining the mechanical parameters of rock-forming minerals and the complexity of regression analysis.This approach enables the accurate assessment of rock-forming minerals'mechanical parameters.A material database of nanoindentation load-depth(P-h)curves was generated using the material point method(MPM)to characterize the mechanical behavior of major rock-forming minerals(quartz,albite,and muscovite)in sandstone.We used Bayesian hyperparameter optimization to determine the optimal hyperparameters for training a deep neural network(DNN).The trained DNN model accurately predicted the material parameters of rock-forming minerals using experimental nanoindentation P-h data.Numerical simulations of the uniaxial compression of heterogeneous sandstones were conducted using the predicted parameters to assess the sandstones’macro-mechanical characteristics.The research findings provide new insights into the fundamental mechanical behavior of heterogeneous rock materials.展开更多
Deep learning and fatigue life prediction remain focal research areas in rail vehicle engineering.This study addresses the vibration fatigue of wheelset lifting lug in Chengdu Metro Line 1 bogies,aiming to develop a f...Deep learning and fatigue life prediction remain focal research areas in rail vehicle engineering.This study addresses the vibration fatigue of wheelset lifting lug in Chengdu Metro Line 1 bogies,aiming to develop a fatigue life prediction method for critical bogie components using deep learning models and measured track load spectra.Extensive field tests on Chengdu Metro Line 1 were conducted to acquire acceleration and stress response data of the wheelset lifting lug,generating training samples for the neural network system.Component stress responses were calculated via time-domain track acceleration and validated against in-situ stress measurements.Results show that neural network-fitted dynamic stress values exhibit excellent consistency with measured data,with errors constrained within 5%.This study validates the proposed small-sample deep learning approach as an effective and accurate solution for fatigue life prediction of critical bogie components under operational load conditions.展开更多
As deep learning(DL)models are increasingly deployed in sensitive domains(e.g.,healthcare),concerns over privacy and security have intensified.Conventional penetration testing frameworks,such asOWASP and NIST,are effe...As deep learning(DL)models are increasingly deployed in sensitive domains(e.g.,healthcare),concerns over privacy and security have intensified.Conventional penetration testing frameworks,such asOWASP and NIST,are effective for traditional networks and applications but lack the capabilities to address DL-specific threats,such asmodel inversion,membership inference,and adversarial attacks.This review provides a comprehensive analysis of penetration testing for the privacy of DL models,examining the shortfalls of existing frameworks,tools,and testing methodologies.Through systematic evaluation of existing literature and empirical analysis,we identify three major contributions:(i)a critical assessment of traditional penetration testing frameworks’inadequacies when applied to DL-specific privacy vulnerabilities,(ii)a comprehensive evaluation of state-of-the-art privacy-preserving methods and their integration with penetration testing workflows,and(iii)the development of a structured framework that combines reconnaissance,threat modeling,exploitation,and post-exploitation phases specifically tailored for DL privacy assessment.Moreover,this review evaluates popular solutions such as IBMAdversarial Robustness Toolbox and TensorFlowPrivacy,alongside privacy-preserving techniques(e.g.,Differential Privacy,Homomorphic Encryption,and Federated Learning),which we systematically analyze through comparative studies of their effectiveness,computational overhead,and practical deployment constraints.While these techniques offer promising safeguards,their adoption is hindered by accuracy loss,performance overheads,and the rapid evolution of attack strategies.Our findings reveal that no single existing solution provides comprehensive protection,which leads us to propose a hybrid approach that strategically combines multiple privacy-preserving mechanisms.The findings of this survey underscore an urgent need for automated,regulationcompliant penetration testing frameworks specifically tailored to DL systems.We argue for hybrid privacy solutions that combinemultiple protectivemechanisms to ensure bothmodel accuracy and privacy.Building on our analysis,we present actionable recommendations for developing adaptive penetration testing strategies that incorporate automated vulnerability assessment,continuous monitoring,and regulatory compliance verification.展开更多
With the continuous development of artificial intelligence and computer vision technology,numerous deep learning-based lane line detection methods have emerged.DeepLabv3+,as a classic semantic segmentation model,has f...With the continuous development of artificial intelligence and computer vision technology,numerous deep learning-based lane line detection methods have emerged.DeepLabv3+,as a classic semantic segmentation model,has found widespread application in the field of lane line detection.However,the accuracy of lane line segmentation is often compromised by factors such as changes in lighting conditions,occlusions,and wear and tear on the lane lines.Additionally,DeepLabv3+suffers from high memory consumption and challenges in deployment on embedded platforms.To address these issues,this paper proposes a lane line detection method for complex road scenes based on DeepLabv3+and MobileNetV4(MNv4).First,the lightweight MNv4 is adopted as the backbone network,and the standard convolutions in ASPP are replaced with depthwise separable convolutions.Second,a polarization attention mechanism is introduced after the ASPP module to enhance the model’s generalization capability.Finally,the Simple Linear Iterative Clustering(SLIC)superpixel segmentation algorithmis employed to preserve lane line edge information.MNv4-DeepLabv3+was tested on the TuSimple and CULane datasets.On the TuSimple dataset,theMean Intersection over Union(MIoU)and Mean Pixel Accuracy(mPA)improved by 1.01%and 7.49%,respectively.On the CULane dataset,MIoU andmPA increased by 3.33%and 7.74%,respectively.Thenumber of parameters decreased from 54.84 to 3.19 M.Experimental results demonstrate that MNv4-DeepLabv3+significantly optimizes model parameter count and enhances segmentation accuracy.展开更多
To counteract small sample size,severe class imbalance and high feature redundancy in 90-day mRS prediction after stroke,this study proposes a four-stage pipeline-“ADASYN re-sampling→clinical+statistical feature scr...To counteract small sample size,severe class imbalance and high feature redundancy in 90-day mRS prediction after stroke,this study proposes a four-stage pipeline-“ADASYN re-sampling→clinical+statistical feature screening→dimensionality reduction→5-fold cross-validation”-and benchmark composite deep-learning architectures.ADASYN first balances the minority classes in the original feature space.Next,a tri-level filter(clinical domain knowledge,variance threshold,mutual information)removes clinically meaningless or redundant variables,after which PCA compresses the remaining features while preserving critical neurological signatures(e.g.,brain-herniation history).Four hybrid CNN-RNN models are trained and compared under strict 5-fold cross-validation;the optimal ensemble yields stable,clinically interpretable probabilities that can support individualized rehabilitation planning.展开更多
This study aimed to integrate Monte Carlo(MC)simulation with deep learning(DL)-based denoising techniques to achieve fast and accurate prediction of high-quality electronic portal imaging device(EPID)transmission dose...This study aimed to integrate Monte Carlo(MC)simulation with deep learning(DL)-based denoising techniques to achieve fast and accurate prediction of high-quality electronic portal imaging device(EPID)transmission dose(TD)for patientspecific quality assurance(PSQA).A total of 100 lung cases were used to obtain the noisy EPID TD by the ARCHER MC code under four kinds of particle numbers(1×10^(6),1×10^(7),1×10^(8)and 1×10^(9)),and the original EPID TD was denoised by the SUNet neural network.The denoised EPID TD was assessed both qualitatively and quantitatively using the structural similarity(SSIM),peak signal-to-noise ratio(PSNR),and gamma passing rate(GPR)with respect to 1×10^(9)as a reference.The computation times for both the MC simulation and DL-based denoising were recorded.As the number of particles increased,both the quality of the noisy EPID TD and computation time increased significantly(1×10^(6):1.12 s,1×10^(7):1.72 s,1×10^(8):8.62 s,and 1×10^(9):73.89 s).In contrast,the DL-based denoising time remained at 0.13-0.16 s.The denoised EPID TD shows a smoother visual appearance and profile curves,but differences between 1×10^(6)and 1×10^(9)still remain.SSIM improves from 0.61 to 0.95 for 1×10^(6),0.70 to 0.96 for 1×10^(7),and 0.90 to 0.97 for 1×10^(8).PSNR increases by>20%for 1×10^(6)and 1×10^(7),and>10%for 1×10^(8).GPR improves from 48.47%to 89.10%for 1×10^(6),61.04%to 94.35%for 1×10^(7),and 91.88%to 99.55%for 1×10^(8).The method that combines MC simulation with DL-based denoising for EPID TD generation can accelerate TD prediction and maintain high accuracy,offering a promising solution for efficient PSQA.展开更多
Deep learning methods have achieved significant progress in solving partial differential equations.However,when applied to the widely used anisotropic scattering neutron transport equations in reactor engineering,thes...Deep learning methods have achieved significant progress in solving partial differential equations.However,when applied to the widely used anisotropic scattering neutron transport equations in reactor engineering,these encounter significant challenges.To address this issue,this study introduces a multi-antiderivative transformation alternating iterative deep learning method(M-AIM).This method transforms the integral terms of the scattering and fission sources in the transport equation into multiple antiderivative functions corresponding to the integrand,converts the differential-integral form of the transport equation into an exact differential equation,and establishes the necessary constraints for a unique solution.The M-AIM uses multiple deep neural networks to map the unknown angular flux density of transport equations and represents various forms of antiderivative functions.It constructs the corresponding weighted loss functions.By alternating iterative training with deep learning methods applied to these neural networks,the loss is reduced gradually.When the loss decreases to a preset minimum,the neural network approaches a numerical solution for both angular flux density and antiderivative functions.This paper presents a numerical verification of geometries such as flat plates and spheres.It verifies the validity of the theoretical framework and associated methods.The study contributes to the development of novel technical approaches for applying deep learning to solve anisotropic scattering neutron transport equations in reactor engineering.展开更多
文摘Underwater pipeline inspection plays a vital role in the proactive maintenance and management of critical marine infrastructure and subaquatic systems.However,the inspection of underwater pipelines presents a challenge due to factors such as light scattering,absorption,restricted visibility,and ambient noise.The advancement of deep learning has introduced powerful techniques for processing large amounts of unstructured and imperfect data collected from underwater environments.This study evaluated the efficacy of the You Only Look Once(YOLO)algorithm,a real-time object detection and localization model based on convolutional neural networks,in identifying and classifying various types of pipeline defects in underwater settings.YOLOv8,the latest evolution in the YOLO family,integrates advanced capabilities,such as anchor-free detection,a cross-stage partial network backbone for efficient feature extraction,and a feature pyramid network+path aggregation network neck for robust multi-scale object detection,which make it particularly well-suited for complex underwater environments.Due to the lack of suitable open-access datasets for underwater pipeline defects,a custom dataset was captured using a remotely operated vehicle in a controlled environment.This application has the following assets available for use.Extensive experimentation demonstrated that YOLOv8 X-Large consistently outperformed other models in terms of pipe defect detection and classification and achieved a strong balance between precision and recall in identifying pipeline cracks,rust,corners,defective welds,flanges,tapes,and holes.This research establishes the baseline performance of YOLOv8 for underwater defect detection and showcases its potential to enhance the reliability and efficiency of pipeline inspection tasks in challenging underwater environments.
基金supported by the School of Digital Science,Universiti Brunei Darussalam,Brunei.
文摘Artificial Intelligence(AI)in healthcare enables predicting diabetes using data-driven methods instead of the traditional ways of screening the disease,which include hemoglobin A1c(HbA1c),oral glucose tolerance test(OGTT),and fasting plasma glucose(FPG)screening techniques,which are invasive and limited in scale.Machine learning(ML)and deep neural network(DNN)models that use large datasets to learn the complex,nonlinear feature interactions,but the conventional ML algorithms are data sensitive and often show unstable predictive accuracy.Conversely,DNN models are more robust,though the ability to reach a high accuracy rate consistently on heterogeneous datasets is still an open challenge.For predicting diabetes,this work proposed a hybrid DNN approach by integrating a bidirectional long short-term memory(BiLSTM)network with a bidirectional gated recurrent unit(BiGRU).A robust DL model,developed by combining various datasets with weighted coefficients,dense operations in the connection of deep layers,and the output aggregation using batch normalization and dropout functions to avoid overfitting.The goal of this hybrid model is better generalization and consistency among various datasets,which facilitates the effective management and early intervention.The proposed DNN model exhibits an excellent predictive performance as compared to the state-of-the-art and baseline ML and DNN models for diabetes prediction tasks.The robust performance indicates the possible usefulness of DL-based models in the development of disease prediction in healthcare and other areas that demand high-quality analytics.
文摘Accurate streamflow prediction is essential for flood warning,reservoir operation,irrigation scheduling,hydropower planning,and sustainable water management,yet remains challenging due to the complexity of hydrological processes.Although data-driven models often outperform conventional physics-based hydrological modelling approaches,their real-world deployment is limited by cost,infrastructure demands,and the interdisciplinary expertise required.To bridge this gap,this study developed QPred,a regional,lightweight,cost-effective,web-delivered application for daily streamflow forecasting.The study executed an end-to-end workflow,from field data acquisition to accessible web-based deployment for on-demand forecasting.High-resolution rainfall data were recorded with tippingbucket gauges and loggers,while river water depth in the Aglar and Paligaad watersheds was converted to discharge using site-specific rating curves,resulting in a daily dataset of precipitation,river water level and discharge.Four DL architectures were trained,including vanilla Long Short-Term Memory(LSTM),stacked LSTM,bidirectional LSTM,and Gated Recurrent Unit(GRU),and evaluated using Nash-Sutcliffe Efficiency(NSE),Coefficient of Determination(R2),Root-Mean-Square-Error-Standard-Deviation Ratio(RSR),and Percentage Bias(PBIAS)metrics.Performance was watershed-specific,as the vanilla LSTM demonstrated the best generalisation for the Aglar watershed(R2=0.88,NSE=0.82,RMSE=0.12 during validation),while the GRU achieved the highest validation accuracy in Paligaad(R2=0.88,NSE=0.88,RMSE=0.49).All models achieved satisfactory to excellent performance during calibration(R2>0.91,NSE>0.91 for both watersheds),demonstrating strong capability to capture streamflow dynamics.The highest performing models were selected and embedded into the QPred application.QPred was developed as a lightweight web pipeline,utilising Google Colab as the primary execution environment,Flask as the backend inference framework,Google Drive for artefact storage,andNgrok for secureHTTPS tunnelling.Auser-friendly front end utilises range sliders(bounded by observed minima and maxima)to gather inputs and provides discharge data along with metadata,thereby enhancing transparency.This work demonstrates that accurate,context-aware deep learningmodels can be delivered through low-cost,web-based platforms,providing a reproducible and scalable pipeline for hydrological applications in other watersheds and for practitioners.
基金supported by Princess Nourah bint Abdulrahman University Researchers Supporting Project number(PNURSP2026R765),Princess Nourah bint Abdulrahman University,Riyadh,Saudi Arabia.
文摘Human Activity Recognition(HAR)is a novel area for computer vision.It has a great impact on healthcare,smart environments,and surveillance while is able to automatically detect human behavior.It plays a vital role in many applications,such as smart home,healthcare,human computer interaction,sports analysis,and especially,intelligent surveillance.In this paper,we propose a robust and efficient HAR system by leveraging deep learning paradigms,including pre-trained models,CNN architectures,and their average-weighted fusion.However,due to the diversity of human actions and various environmental influences,as well as a lack of data and resources,achieving high recognition accuracy remain elusive.In this work,a weighted average ensemble technique is employed to fuse three deep learning models:EfficientNet,ResNet50,and a custom CNN.The results of this study indicate that using a weighted average ensemble strategy for developing more effective HAR models may be a promising idea for detection and classification of human activities.Experiments by using the benchmark dataset proved that the proposed weighted ensemble approach outperformed existing approaches in terms of accuracy and other key performance measures.The combined average-weighted ensemble of pre-trained and CNN models obtained an accuracy of 98%,compared to 97%,96%,and 95%for the customized CNN,EfficientNet,and ResNet50 models,respectively.
基金supported by ZTE Industry-University-Institute Cooperation Funds under Grant No.IA20240319003the NSFC under Grant No.62571112。
文摘To achieve the potential performance gain of massive multiple-input multiple-output(MIMO)systems,base stations(BS)require downlink channel state information(CSI)fed back by users to execute beamforming design,especially in the frequency division duplex(FDD)systems.However,due to the enormous number of antennas in massive MIMO systems,the feedback overhead of downlink CSI acquisition is extremely large.To address this issue,deep learning(DL)techniques have been introduced to de velop high-accuracy feedback strategies under limited backhaul constraints.In this paper,we provide an overview of DL-based CSI compression and feedback approaches in massive MIMO systems.Specifically,we introduce the conventional CSI compression and feedback schemes and the existing problems.Besides,we elaborate on various DL techniques employed in CSI compression from the perspective of network architecture and analyze the advantages of different techniques.We also enumerate the applications of DL-based methods for solving practical challenges in CSI compression and feedback.In addition,we brief the remaining issues in deep CSI compression and indicate potential directions in future wireless networks.
基金supported by the Basic Science Research Program(2023R1A2C3004336,RS-202300243807)&Regional Leading Research Center(RS-202400405278)through the National Research Foundation of Korea(NRF)grant funded by the Korea Government(MSIT)。
文摘Wearable sensors integrated with deep learning techniques have the potential to revolutionize seamless human-machine interfaces for real-time health monitoring,clinical diagnosis,and robotic applications.Nevertheless,it remains a critical challenge to simultaneously achieve desirable mechanical and electrical performance along with biocompatibility,adhesion,self-healing,and environmental robustness with excellent sensing metrics.Herein,we report a multifunctional,anti-freezing,selfadhesive,and self-healable organogel pressure sensor composed of cobalt nanoparticle encapsulated nitrogen-doped carbon nanotubes(CoN CNT)embedded in a polyvinyl alcohol-gelatin(PVA/GLE)matrix.Fabricated using a binary solvent system of water and ethylene glycol(EG),the CoN CNT/PVA/GLE organogel exhibits excellent flexibility,biocompatibility,and temperature tolerance with remarkable environmental stability.Electrochemical impedance spectroscopy confirms near-stable performance across a broad humidity range(40%-95%RH).Freeze-tolerant conductivity under sub-zero conditions(-20℃)is attributed to the synergistic role of CoN CNT and EG,preserving mobility and network integrity.The Co N CNT/PVA/GLE organogel sensor exhibits high sensitivity of 5.75 k Pa^(-1)in the detection range from 0 to 20 k Pa,ideal for subtle biomechanical motion detection.A smart human-machine interface for English letter recognition using deep learning achieved 98%accuracy.The organogel sensor utility was extended to detect human gestures like finger bending,wrist motion,and throat vibration during speech.
基金supported by Basic Science Research Program through the National Research Foundation of Korea(NRF)funded by the Ministry of Education(RS-2023-00245084)by Korea Institute for Advancement of Technology(KIAT)grant funded by the Korea Government(MOTIE)(RS-2024-00415938,HRD Program for Industrial Innovation)and Soonchunhyang University.
文摘This paper proposes a deep learning-based 3D LiDAR perception framework designed for applications such as autonomous robots and vehicles.To address the high dependency on large-scale annotated data—an inherent limitation of deep learning models—this study introduces a hybrid perception architecture that incorporates expertdriven LiDAR processing techniques into the deep neural network.Traditional 3DLiDAR processingmethods typically remove ground planes and apply distance-or density-based clustering for object detection.In this work,such expert knowledge is encoded as feature-level inputs and fused with the deep network,therebymitigating the data dependency issue of conventional learning-based approaches.Specifically,the proposedmethod combines two expert algorithms—Patchwork++for ground segmentation and DBSCAN for clustering—with a PointPillars-based LiDAR detection network.We design four hybrid versions of the network depending on the stage and method of integrating expert features into the feature map of the deep model.Among these,Version 4 incorporates a modified neck structure in PointPillars and introduces a new Cluster 2D Pseudo-Map Branch that utilizes cluster-level pseudo-images generated from Patchwork++and DBSCAN.This version achieved a+3.88%improvement mean Average Precision(mAP)compared to the baseline PointPillars.The results demonstrate that embedding expert-based perception logic into deep neural architectures can effectively enhance performance and reduce dependency on extensive training datasets,offering a promising direction for robust 3D LiDAR object detection in real-world scenarios.
基金National Natural Science Foundation of China,Grant/Award Number:72204169Department of Science and Technology of Sichuan Province,Grant/Award Number:2021YFS0393。
文摘Background:Studies have shown that heart rate variability(HRV)is a predictor of the prognosis of cardiovascular diseases.Contact heartbeat monitoring equipment is widely used,especially in hospitals,and benefits from the rapidity and accuracy of the detection of physiological health indicators.However,long-term contact with equipment has many adverse effects.The purpose of this study was to improve the accuracy of HRV detection via noncontact equipment,thus enabling HRV to be assessed in various scenarios.Methods:A novel deep learning approach was proposed for measuring heartbeats through camera videos.First,we performed facial segmentation and divided the face into 16 grid cells with different light balance scores.After the trend is filtered by the Hamming window,a transformer-based neural network is used to further filter the signal.Finally,heart rate(HR)and HRV are estimated.Results:We used 1 million synthetic data points for pretraining and a public dataset in combination with a dataset that we constructed for task training.The final results were obtained on a test dataset that we constructed.The accuracy for HR with a low light balance score(0.867-0.983)was greater than that with a high score(0.667-0.750).Our method had higher accuracy in estimating HR than traditional filtering methods(0.167-0.417)and state-of-the-art neural network filtering methods(0.783-0.917)did.The root mean square error of the HRV from the time domain was the lowest,and the correlation index score was the highest for the HRV from the frequency domain estimated by our method compared with those estimated by two neural networks.Conclusions:Light balance,large sample training,and two-stage training can improve the accuracy of HRV estimation.
基金funded by the National Key Research and Development Program of China(Grant No.2019YFD1001900)the HZAU-AGIS Cooperation Fund(Grant No.SZYJY2022006).
文摘Nondestructive measurement technology of phenotype can provide substantial phenotypic data support for applications such as seedling breeding,management,and quality testing.The current method of measuring seedling phenotypes mainly relies on manual measurement which is inefficient,subjective and destroys samples.Therefore,the paper proposes a nondestructive measurement method for the canopy phenotype of the watermelon plug seedlings based on deep learning.The Azure Kinect was used to shoot canopy color images,depth images,and RGB-D images of the watermelon plug seedlings.The Mask-RCNN network was used to classify,segment,and count the canopy leaves of the watermelon plug seedlings.To reduce the error of leaf area measurement caused by mutual occlusion of leaves,the leaves were repaired by CycleGAN,and the depth images were restored by image processing.Then,the Delaunay triangulation was adopted to measure the leaf area in the leaf point cloud.The YOLOX target detection network was used to identify the growing point position of each seedling on the plug tray.Then the depth differences between the growing point and the upper surface of the plug tray were calculated to obtain plant height.The experiment results show that the nondestructive measurement algorithm proposed in this paper achieves good measurement performance for the watermelon plug seedlings from the 1 true-leaf to 3 true-leaf stages.The average relative error of measurement is 2.33%for the number of true leaves,4.59%for the number of cotyledons,8.37%for the leaf area,and 3.27%for the plant height.The experiment results demonstrate that the proposed algorithm in this paper provides an effective solution for the nondestructive measurement of the canopy phenotype of the plug seedlings.
基金supported in part by the National Natural Science Foundation of China under Grant 52432012in part by the Shanghai Science and Technology Project with 25ZR1402508。
文摘The stator of the maglev track plays a crucial role in the operation of the maglev system.Currently,the efficiency of maglev track inspection is limited by several factors,including the large span of elevated structures,manual visual inspection,short inspection window times,and limited GPS positioning accuracy.To address these issues,this paper proposes a deep learning-based method for detecting and locating stator surface damage.This study establishes a maglev track stator surface image dataset,trains different object detection models,and compares their performance.Ultimately,YOLO and ByteTrack object tracking algorithms were chosen as the basic framework and enhanced to achieve automatic identification of high-speed maglev track stator surface damage images and track and count stator surface localization feature images.By matching the identified damaged images with their corresponding stator segment and beam segment sequence numbers,the location of the damage is pinpointed to the corresponding stator segment,enabling rapid and accurate identification and localization of complex damage to the maglev track stator surface.
基金supported by UKRI(EP/Z000025/1)Horizon Europe Programme under the MSCA grant for the ACMod project(101130271)。
文摘The exponential growth of video content has driven significant advancements in video summarization techniques in recent years. Breakthroughs in deep learning have been particularly transformative, enabling more effective detection of key information and creating new possibilities for video synopsis. To summarize recent progress and accelerate research in this field,this paper provides a comprehensive review of deep learningbased video summarization methods developed over the past decade. We begin by examining the research landscape of video abstraction technologies and identifying core challenges in video summarization. Subsequently, we systematically analyze prevailing deep learning frameworks and methodologies employed in current video summarization systems, offering researchers a clear roadmap of the field's evolution. Unlike previous review works,we first classify research papers based on the structural hierarchy of the video(from frame-level to shot-level to video-level),then further categorize them according to the summary backbone model(feature extraction and spatiotemporal modeling).This approach provides a more systematic and hierarchical organization of the documents. Following this comprehensive review,we summarize the benchmark datasets and evaluation metrics commonly employed in the field. Finally, we analyze persistent challenges and propose insightful directions for future research,providing a forward-looking perspective on video summarization technologies. This systematic literature review is of great reference value to new researchers exploring the fields of deep learning and video summarization.
基金supported by the National Natural Science Foundation of China(Grant Nos.42130719 and 42177173)the Doctoral Direct Train Project of Chongqing Natural Science Foundation(Grant No.CSTB2023NSCQ-BSX0029).
文摘Underground engineering projects such as deep tunnel excavation often encounter rockburst disasters accompanied by numerous microseismic events.Rapid interpretation of microseismic signals is crucial for the timely identification of rockbursts.However,conventional processing encompasses multi-step workflows,including classification,denoising,picking,locating,and computational analysis,coupled with manual intervention,which collectively compromise the reliability of early warnings.To address these challenges,this study innovatively proposes the“microseismic stethoscope"-a multi-task machine learning and deep learning model designed for the automated processing of massive microseismic signals.This model efficiently extracts three key parameters that are necessary for recognizing rockburst disasters:rupture location,microseismic energy,and moment magnitude.Specifically,the model extracts raw waveform features from three dedicated sub-networks:a classifier for source zone classification,and two regressors for microseismic energy and moment magnitude estimation.This model demonstrates superior efficiency compared to traditional processing and semi-automated processing,reducing per-event processing time from 0.71 s to 0.49 s to merely 0.036 s.It concurrently achieves 98%accuracy in source zone classification,with microseismic energy and moment magnitude estimation errors of 0.13 and 0.05,respectively.This model has been well applied and validated in the Daxiagu Tunnel case in Sichuan,China.The application results indicate that the model is as accurate as traditional methods in determining source parameters,and thus can be used to identify potential geomechanical processes of rockburst disasters.By enhancing the signal processing reliability of microseismic events,the proposed model in this study presents a significant advancement in the identification of rockburst disasters.
基金financially supported by the Open Project Program of Wuhan National Laboratory for Optoelectronics(No.2022WNLOKF009)the National Natural Science Foundation of China(No.62475216)+2 种基金the Key Research and Development Program of Shaanxi(No.2024GH-ZDXM-37)the Fujian Provincial Natural Science Foundation of China(No.2024J01060)the Startup Program of XMU,and the Fundamental Research Funds for the Central Universities.
文摘Microscopy imaging is fundamental in analyzing bacterial morphology and dynamics,offering critical insights into bacterial physiology and pathogenicity.Image segmentation techniques enable quantitative analysis of bacterial structures,facilitating precise measurement of morphological variations and population behaviors at single-cell resolution.This paper reviews advancements in bacterial image segmentation,emphasizing the shift from traditional thresholding and watershed methods to deep learning-driven approaches.Convolutional neural networks(CNNs),U-Net architectures,and three-dimensional(3D)frameworks excel at segmenting dense biofilms and resolving antibiotic-induced morphological changes.These methods combine automated feature extraction with physics-informed postprocessing.Despite progress,challenges persist in computational efficiency,cross-species generalizability,and integration with multimodal experimental workflows.Future progress will depend on improving model robustness across species and imaging modalities,integrating multimodal data for phenotype-function mapping,and developing standard pipelines that link computational tools with clinical diagnostics.These innovations will expand microbial phenotyping beyond structural analysis,enabling deeper insights into bacterial physiology and ecological interactions.
基金supported by the National Key Research and Development Program of China(Grant no.2023YFC3009005)the Chongqing Technology Innovation and Application Development Special Key Project(Grant no.CSTB2022TIAD-KPX0135)the Fundamental Research Funds for the Central Universities(Grant no.2023CDJKYJH068).
文摘This study proposed a deep learning-based nanoindentation simulation method to address the challenge of obtaining the mechanical parameters of rock-forming minerals and the complexity of regression analysis.This approach enables the accurate assessment of rock-forming minerals'mechanical parameters.A material database of nanoindentation load-depth(P-h)curves was generated using the material point method(MPM)to characterize the mechanical behavior of major rock-forming minerals(quartz,albite,and muscovite)in sandstone.We used Bayesian hyperparameter optimization to determine the optimal hyperparameters for training a deep neural network(DNN).The trained DNN model accurately predicted the material parameters of rock-forming minerals using experimental nanoindentation P-h data.Numerical simulations of the uniaxial compression of heterogeneous sandstones were conducted using the predicted parameters to assess the sandstones’macro-mechanical characteristics.The research findings provide new insights into the fundamental mechanical behavior of heterogeneous rock materials.
基金supported by the CRRC Original Technology TenYear Cultivation Program(Grant No.2022CYY007)。
文摘Deep learning and fatigue life prediction remain focal research areas in rail vehicle engineering.This study addresses the vibration fatigue of wheelset lifting lug in Chengdu Metro Line 1 bogies,aiming to develop a fatigue life prediction method for critical bogie components using deep learning models and measured track load spectra.Extensive field tests on Chengdu Metro Line 1 were conducted to acquire acceleration and stress response data of the wheelset lifting lug,generating training samples for the neural network system.Component stress responses were calculated via time-domain track acceleration and validated against in-situ stress measurements.Results show that neural network-fitted dynamic stress values exhibit excellent consistency with measured data,with errors constrained within 5%.This study validates the proposed small-sample deep learning approach as an effective and accurate solution for fatigue life prediction of critical bogie components under operational load conditions.
基金supported in part by the Tianjin Natural Science Foundation Project(24JCZDJC01000)the Fundamental Research Funds for the Central Universities of China(No.3122025091).
文摘As deep learning(DL)models are increasingly deployed in sensitive domains(e.g.,healthcare),concerns over privacy and security have intensified.Conventional penetration testing frameworks,such asOWASP and NIST,are effective for traditional networks and applications but lack the capabilities to address DL-specific threats,such asmodel inversion,membership inference,and adversarial attacks.This review provides a comprehensive analysis of penetration testing for the privacy of DL models,examining the shortfalls of existing frameworks,tools,and testing methodologies.Through systematic evaluation of existing literature and empirical analysis,we identify three major contributions:(i)a critical assessment of traditional penetration testing frameworks’inadequacies when applied to DL-specific privacy vulnerabilities,(ii)a comprehensive evaluation of state-of-the-art privacy-preserving methods and their integration with penetration testing workflows,and(iii)the development of a structured framework that combines reconnaissance,threat modeling,exploitation,and post-exploitation phases specifically tailored for DL privacy assessment.Moreover,this review evaluates popular solutions such as IBMAdversarial Robustness Toolbox and TensorFlowPrivacy,alongside privacy-preserving techniques(e.g.,Differential Privacy,Homomorphic Encryption,and Federated Learning),which we systematically analyze through comparative studies of their effectiveness,computational overhead,and practical deployment constraints.While these techniques offer promising safeguards,their adoption is hindered by accuracy loss,performance overheads,and the rapid evolution of attack strategies.Our findings reveal that no single existing solution provides comprehensive protection,which leads us to propose a hybrid approach that strategically combines multiple privacy-preserving mechanisms.The findings of this survey underscore an urgent need for automated,regulationcompliant penetration testing frameworks specifically tailored to DL systems.We argue for hybrid privacy solutions that combinemultiple protectivemechanisms to ensure bothmodel accuracy and privacy.Building on our analysis,we present actionable recommendations for developing adaptive penetration testing strategies that incorporate automated vulnerability assessment,continuous monitoring,and regulatory compliance verification.
文摘With the continuous development of artificial intelligence and computer vision technology,numerous deep learning-based lane line detection methods have emerged.DeepLabv3+,as a classic semantic segmentation model,has found widespread application in the field of lane line detection.However,the accuracy of lane line segmentation is often compromised by factors such as changes in lighting conditions,occlusions,and wear and tear on the lane lines.Additionally,DeepLabv3+suffers from high memory consumption and challenges in deployment on embedded platforms.To address these issues,this paper proposes a lane line detection method for complex road scenes based on DeepLabv3+and MobileNetV4(MNv4).First,the lightweight MNv4 is adopted as the backbone network,and the standard convolutions in ASPP are replaced with depthwise separable convolutions.Second,a polarization attention mechanism is introduced after the ASPP module to enhance the model’s generalization capability.Finally,the Simple Linear Iterative Clustering(SLIC)superpixel segmentation algorithmis employed to preserve lane line edge information.MNv4-DeepLabv3+was tested on the TuSimple and CULane datasets.On the TuSimple dataset,theMean Intersection over Union(MIoU)and Mean Pixel Accuracy(mPA)improved by 1.01%and 7.49%,respectively.On the CULane dataset,MIoU andmPA increased by 3.33%and 7.74%,respectively.Thenumber of parameters decreased from 54.84 to 3.19 M.Experimental results demonstrate that MNv4-DeepLabv3+significantly optimizes model parameter count and enhances segmentation accuracy.
基金Shanghai University of Engineering Science Undergraduate Innovation Training Program(Project No.:cx2521005)。
文摘To counteract small sample size,severe class imbalance and high feature redundancy in 90-day mRS prediction after stroke,this study proposes a four-stage pipeline-“ADASYN re-sampling→clinical+statistical feature screening→dimensionality reduction→5-fold cross-validation”-and benchmark composite deep-learning architectures.ADASYN first balances the minority classes in the original feature space.Next,a tri-level filter(clinical domain knowledge,variance threshold,mutual information)removes clinically meaningless or redundant variables,after which PCA compresses the remaining features while preserving critical neurological signatures(e.g.,brain-herniation history).Four hybrid CNN-RNN models are trained and compared under strict 5-fold cross-validation;the optimal ensemble yields stable,clinically interpretable probabilities that can support individualized rehabilitation planning.
基金supported by National Key R&D Program of China(No.2022YFC2404604)Chongqing Research Institution Performance Incentive Guidance Special Project(No.CSTB2023JXJL-YFX0080)Chongqing Medical Scientific Research Project(Joint project of Chongqing Health Commission and Science and Technology Bureau)(No.2022DBXM005)。
文摘This study aimed to integrate Monte Carlo(MC)simulation with deep learning(DL)-based denoising techniques to achieve fast and accurate prediction of high-quality electronic portal imaging device(EPID)transmission dose(TD)for patientspecific quality assurance(PSQA).A total of 100 lung cases were used to obtain the noisy EPID TD by the ARCHER MC code under four kinds of particle numbers(1×10^(6),1×10^(7),1×10^(8)and 1×10^(9)),and the original EPID TD was denoised by the SUNet neural network.The denoised EPID TD was assessed both qualitatively and quantitatively using the structural similarity(SSIM),peak signal-to-noise ratio(PSNR),and gamma passing rate(GPR)with respect to 1×10^(9)as a reference.The computation times for both the MC simulation and DL-based denoising were recorded.As the number of particles increased,both the quality of the noisy EPID TD and computation time increased significantly(1×10^(6):1.12 s,1×10^(7):1.72 s,1×10^(8):8.62 s,and 1×10^(9):73.89 s).In contrast,the DL-based denoising time remained at 0.13-0.16 s.The denoised EPID TD shows a smoother visual appearance and profile curves,but differences between 1×10^(6)and 1×10^(9)still remain.SSIM improves from 0.61 to 0.95 for 1×10^(6),0.70 to 0.96 for 1×10^(7),and 0.90 to 0.97 for 1×10^(8).PSNR increases by>20%for 1×10^(6)and 1×10^(7),and>10%for 1×10^(8).GPR improves from 48.47%to 89.10%for 1×10^(6),61.04%to 94.35%for 1×10^(7),and 91.88%to 99.55%for 1×10^(8).The method that combines MC simulation with DL-based denoising for EPID TD generation can accelerate TD prediction and maintain high accuracy,offering a promising solution for efficient PSQA.
基金supported by the National Natural Science Foundation of China(No.12575189)。
文摘Deep learning methods have achieved significant progress in solving partial differential equations.However,when applied to the widely used anisotropic scattering neutron transport equations in reactor engineering,these encounter significant challenges.To address this issue,this study introduces a multi-antiderivative transformation alternating iterative deep learning method(M-AIM).This method transforms the integral terms of the scattering and fission sources in the transport equation into multiple antiderivative functions corresponding to the integrand,converts the differential-integral form of the transport equation into an exact differential equation,and establishes the necessary constraints for a unique solution.The M-AIM uses multiple deep neural networks to map the unknown angular flux density of transport equations and represents various forms of antiderivative functions.It constructs the corresponding weighted loss functions.By alternating iterative training with deep learning methods applied to these neural networks,the loss is reduced gradually.When the loss decreases to a preset minimum,the neural network approaches a numerical solution for both angular flux density and antiderivative functions.This paper presents a numerical verification of geometries such as flat plates and spheres.It verifies the validity of the theoretical framework and associated methods.The study contributes to the development of novel technical approaches for applying deep learning to solve anisotropic scattering neutron transport equations in reactor engineering.
