Weakly Supervised Semantic Segmentation(WSSS),which relies only on image-level labels,has attracted significant attention for its cost-effectiveness and scalability.Existing methods mainly enhance inter-class distinct...Weakly Supervised Semantic Segmentation(WSSS),which relies only on image-level labels,has attracted significant attention for its cost-effectiveness and scalability.Existing methods mainly enhance inter-class distinctions and employ data augmentation to mitigate semantic ambiguity and reduce spurious activations.However,they often neglect the complex contextual dependencies among image patches,resulting in incomplete local representations and limited segmentation accuracy.To address these issues,we propose the Context Patch Fusion with Class Token Enhancement(CPF-CTE)framework,which exploits contextual relations among patches to enrich feature repre-sentations and improve segmentation.At its core,the Contextual-Fusion Bidirectional Long Short-Term Memory(CF-BiLSTM)module captures spatial dependencies between patches and enables bidirectional information flow,yield-ing a more comprehensive understanding of spatial correlations.This strengthens feature learning and segmentation robustness.Moreover,we introduce learnable class tokens that dynamically encode and refine class-specific semantics,enhancing discriminative capability.By effectively integrating spatial and semantic cues,CPF-CTE produces richer and more accurate representations of image content.Extensive experiments on PASCAL VOC 2012 and MS COCO 2014 validate that CPF-CTE consistently surpasses prior WSSS methods.展开更多
In today's connected world,the generation of massive streaming data across diverse domains has become commonplace.In the presence of concept drift,class imbalance,label scarcity,and new class emergence,these chall...In today's connected world,the generation of massive streaming data across diverse domains has become commonplace.In the presence of concept drift,class imbalance,label scarcity,and new class emergence,these challenges jointly degrade representation stability,bias learning toward outdated distributions,and reduce the resilience and reliability of detection in dynamic environments.This paper proposes a streaming classincremental learning(SCIL)framework to address these issues.The SCIL framework integrates an autoencoder(AE)with a multi-layer perceptron for multi-class prediction,employs a dual-loss strategy(classification and reconstruction)for prediction and new class detection,uses corrected pseudo-labels for online training,manages classes with queues,and applies oversampling to handle imbalance.The rationale behind the method's structure is elucidated through ablation studies,and a comprehensive experimental evaluation is performed using both real-world and synthetic datasets that feature class imbalance,incremental classes,and concept drifts.Our results demonstrate that SCIL outperforms strong baselines and state-of-the-art methods.In line with our commitment to Open Science,we make our code and datasets available to the community.展开更多
Most Convolutional Neural Network(CNN)interpretation techniques visualize only the dominant cues that the model relies on,but there is no guarantee that these represent all the evidence the model uses for classificati...Most Convolutional Neural Network(CNN)interpretation techniques visualize only the dominant cues that the model relies on,but there is no guarantee that these represent all the evidence the model uses for classification.This limitation becomes critical when hidden secondary cues—potentially more meaningful than the visualized ones—remain undiscovered.This study introduces CasCAM(Cascaded Class Activation Mapping)to address this fundamental limitation through counterfactual reasoning.By asking“if this dominant cue were absent,what other evidence would the model use?”,CasCAM progressively masks the most salient features and systematically uncovers the hierarchy of classification evidence hidden beneath them.Experimental results demonstrate that CasCAM effectively discovers the full spectrum of reasoning evidence and can be universally applied with nine existing interpretation methods.展开更多
Accurate prediction of flood events is important for flood control and risk management.Machine learning techniques contributed greatly to advances in flood predictions,and existing studies mainly focused on predicting...Accurate prediction of flood events is important for flood control and risk management.Machine learning techniques contributed greatly to advances in flood predictions,and existing studies mainly focused on predicting flood resource variables using single or hybrid machine learning techniques.However,class-based flood predictions have rarely been investigated,which can aid in quickly diagnosing comprehensive flood characteristics and proposing targeted management strategies.This study proposed a prediction approach of flood regime metrics and event classes coupling machine learning algorithms with clustering-deduced membership degrees.Five algorithms were adopted for this exploration.Results showed that the class membership degrees accurately determined event classes with class hit rates up to 100%,compared with the four classes clustered from nine regime metrics.The nonlinear algorithms(Multiple Linear Regression,Random Forest,and least squares-Support Vector Machine)outperformed the linear techniques(Multiple Linear Regression and Stepwise Regression)in predicting flood regime metrics.The proposed approach well predicted flood event classes with average class hit rates of 66.0%-85.4%and 47.2%-76.0%in calibration and validation periods,respectively,particularly for the slow and late flood events.The predictive capability of the proposed prediction approach for flood regime metrics and classes was considerably stronger than that of hydrological modeling approach.展开更多
In image analysis,high-precision semantic segmentation predominantly relies on supervised learning.Despite significant advancements driven by deep learning techniques,challenges such as class imbalance and dynamic per...In image analysis,high-precision semantic segmentation predominantly relies on supervised learning.Despite significant advancements driven by deep learning techniques,challenges such as class imbalance and dynamic performance evaluation persist.Traditional weighting methods,often based on pre-statistical class counting,tend to overemphasize certain classes while neglecting others,particularly rare sample categories.Approaches like focal loss and other rare-sample segmentation techniques introduce multiple hyperparameters that require manual tuning,leading to increased experimental costs due to their instability.This paper proposes a novel CAWASeg framework to address these limitations.Our approach leverages Grad-CAM technology to generate class activation maps,identifying key feature regions that the model focuses on during decision-making.We introduce a Comprehensive Segmentation Performance Score(CSPS)to dynamically evaluate model performance by converting these activation maps into pseudo mask and comparing them with Ground Truth.Additionally,we design two adaptive weights for each class:a Basic Weight(BW)and a Ratio Weight(RW),which the model adjusts during training based on real-time feedback.Extensive experiments on the COCO-Stuff,CityScapes,and ADE20k datasets demonstrate that our CAWASeg framework significantly improves segmentation performance for rare sample categories while enhancing overall segmentation accuracy.The proposed method offers a robust and efficient solution for addressing class imbalance in semantic segmentation tasks.展开更多
Legal case classification involves the categorization of legal documents into predefined categories,which facilitates legal information retrieval and case management.However,real-world legal datasets often suffer from...Legal case classification involves the categorization of legal documents into predefined categories,which facilitates legal information retrieval and case management.However,real-world legal datasets often suffer from class imbalances due to the uneven distribution of case types across legal domains.This leads to biased model performance,in the form of high accuracy for overrepresented categories and underperformance for minority classes.To address this issue,in this study,we propose a data augmentation method that masks unimportant terms within a document selectively while preserving key terms fromthe perspective of the legal domain.This approach enhances data diversity and improves the generalization capability of conventional models.Our experiments demonstrate consistent improvements achieved by the proposed augmentation strategy in terms of accuracy and F1 score across all models,validating the effectiveness of the proposed method in legal case classification.展开更多
Natural gas hydrate in Class Ⅰ reservoirs holds significant commercial potential,as demonstrated by production trials in the South China Sea.However,experimental studies have focused largely on Class Ⅲ systems,with ...Natural gas hydrate in Class Ⅰ reservoirs holds significant commercial potential,as demonstrated by production trials in the South China Sea.However,experimental studies have focused largely on Class Ⅲ systems,with Class Ⅰ/Ⅱ reservoirs remaining underrepresented due to the difficulties in simulating the geothermal gradient and interlayer interactions.This study investigates depressurization performance across all three classes using a novel 360°rotatable reactor with segmented temperature control,enabling precise simulation of reservoir conditions.Results reveal:(i)Class Ⅰ shows two-stage gas production,with 50%from early free gas enabling rapid depressurization,followed by dissociated gas dominance.They achieve 38.4%-78.3%higher cumulative production and superior gas-to-water ratios due to efficient energy use.(ii)The free gas layer in Class Ⅰ accelerates pressure and heat transfer.Class Ⅱ’s water layer provides sensible heat but causes water blocking,impairing heat flow.Class Ⅲ exhibits rapid initial dissociation but a quick decline without fluid support.(iii)Low temperature,low hydrate saturation,and high production pressure collectively reduce efficiency by increasing flow resistance,limiting gas supply,and reducing dissociation drive.Over-depressurization risks hydrate reformation and ice blockage.This work bridges experimental gaps for Class Ⅰ/Ⅱ reservoirs,offering key insights for optimizing recovery.展开更多
Given the growing number of vehicle accidents caused by unintended acceleration and braking failure,verifying Sudden Unintended Acceleration(SUA)incidents has become a persistent challenge.A central issue of debate is...Given the growing number of vehicle accidents caused by unintended acceleration and braking failure,verifying Sudden Unintended Acceleration(SUA)incidents has become a persistent challenge.A central issue of debate is whether such events stem frommechanical malfunctions or driver pedalmisapplications.However,existing verification procedures implemented by vehiclemanufacturers often involve closed tests after vehicle recalls;thus raising ongoing concerns about reliability and transparency.Consequently,there is a growing need for a user-driven framework that enables independent data acquisition and verification.Although previous studies have addressed SUA detection using deep learning,few have explored howclass granularity optimization affects power efficiency and inference performance in real-time Edge AI systems.To address this problem,this work presents a cloud-assisted artificial intelligence(AI)solution for the reliable verification of SUA occurrences.The proposed system integrates multimodal sensor streams including camera-based foot images,On-Board Diagnostics II(OBD-II)signals,and six-axismeasurements to determine whether the brake pedal was actually engaged at themoment of a suspected SUA.Beyond image acquisition,convolutional neural network(CNN)models perform real-time inference to classify the driver’s pedal operation states with the resulting outputs transmitted and archived in the cloud.A dedicated dataset of brake and accelerator pedal images was collected from 15 vehicles produced by 6 domestic and international manufacturers.Using this dataset,transfer learning techniques were applied to compare and analyze model performance and generalization as the CNN class granularity varied from coarse to fine levels.Furthermore,classification performance was evaluated in terms of latency and power efficiency under different class configurations.The experimental results demonstrated that the proposed solution identified the driver’s pedal behavior accurately and promptly,with the two-class model achieving the highest F1-score and accuracy among all granularity settings.展开更多
文摘Weakly Supervised Semantic Segmentation(WSSS),which relies only on image-level labels,has attracted significant attention for its cost-effectiveness and scalability.Existing methods mainly enhance inter-class distinctions and employ data augmentation to mitigate semantic ambiguity and reduce spurious activations.However,they often neglect the complex contextual dependencies among image patches,resulting in incomplete local representations and limited segmentation accuracy.To address these issues,we propose the Context Patch Fusion with Class Token Enhancement(CPF-CTE)framework,which exploits contextual relations among patches to enrich feature repre-sentations and improve segmentation.At its core,the Contextual-Fusion Bidirectional Long Short-Term Memory(CF-BiLSTM)module captures spatial dependencies between patches and enables bidirectional information flow,yield-ing a more comprehensive understanding of spatial correlations.This strengthens feature learning and segmentation robustness.Moreover,we introduce learnable class tokens that dynamically encode and refine class-specific semantics,enhancing discriminative capability.By effectively integrating spatial and semantic cues,CPF-CTE produces richer and more accurate representations of image content.Extensive experiments on PASCAL VOC 2012 and MS COCO 2014 validate that CPF-CTE consistently surpasses prior WSSS methods.
基金supported by the European Research Council(ERC)under Grant Agreement No.951424(Water-Futures)by the Republic of Cyprus through the Deputy Ministry of Research,Innovation and Digital Policy.
文摘In today's connected world,the generation of massive streaming data across diverse domains has become commonplace.In the presence of concept drift,class imbalance,label scarcity,and new class emergence,these challenges jointly degrade representation stability,bias learning toward outdated distributions,and reduce the resilience and reliability of detection in dynamic environments.This paper proposes a streaming classincremental learning(SCIL)framework to address these issues.The SCIL framework integrates an autoencoder(AE)with a multi-layer perceptron for multi-class prediction,employs a dual-loss strategy(classification and reconstruction)for prediction and new class detection,uses corrected pseudo-labels for online training,manages classes with queues,and applies oversampling to handle imbalance.The rationale behind the method's structure is elucidated through ablation studies,and a comprehensive experimental evaluation is performed using both real-world and synthetic datasets that feature class imbalance,incremental classes,and concept drifts.Our results demonstrate that SCIL outperforms strong baselines and state-of-the-art methods.In line with our commitment to Open Science,we make our code and datasets available to the community.
基金supported by the Basic Science Research Program through the National Research Foundation of Korea(NRF),funded by the Ministry of Education(RS-2023-00249743).
文摘Most Convolutional Neural Network(CNN)interpretation techniques visualize only the dominant cues that the model relies on,but there is no guarantee that these represent all the evidence the model uses for classification.This limitation becomes critical when hidden secondary cues—potentially more meaningful than the visualized ones—remain undiscovered.This study introduces CasCAM(Cascaded Class Activation Mapping)to address this fundamental limitation through counterfactual reasoning.By asking“if this dominant cue were absent,what other evidence would the model use?”,CasCAM progressively masks the most salient features and systematically uncovers the hierarchy of classification evidence hidden beneath them.Experimental results demonstrate that CasCAM effectively discovers the full spectrum of reasoning evidence and can be universally applied with nine existing interpretation methods.
基金National Key Research and Development Program of China,No.2023YFC3006704National Natural Science Foundation of China,No.42171047CAS-CSIRO Partnership Joint Project of 2024,No.177GJHZ2023097MI。
文摘Accurate prediction of flood events is important for flood control and risk management.Machine learning techniques contributed greatly to advances in flood predictions,and existing studies mainly focused on predicting flood resource variables using single or hybrid machine learning techniques.However,class-based flood predictions have rarely been investigated,which can aid in quickly diagnosing comprehensive flood characteristics and proposing targeted management strategies.This study proposed a prediction approach of flood regime metrics and event classes coupling machine learning algorithms with clustering-deduced membership degrees.Five algorithms were adopted for this exploration.Results showed that the class membership degrees accurately determined event classes with class hit rates up to 100%,compared with the four classes clustered from nine regime metrics.The nonlinear algorithms(Multiple Linear Regression,Random Forest,and least squares-Support Vector Machine)outperformed the linear techniques(Multiple Linear Regression and Stepwise Regression)in predicting flood regime metrics.The proposed approach well predicted flood event classes with average class hit rates of 66.0%-85.4%and 47.2%-76.0%in calibration and validation periods,respectively,particularly for the slow and late flood events.The predictive capability of the proposed prediction approach for flood regime metrics and classes was considerably stronger than that of hydrological modeling approach.
基金supported by the Funds for Central-Guided Local Science and Technology Development(Grant No.202407AC110005)Key Technologies for the Construction of a Whole-Process Intelligent Service System for Neuroendocrine Neoplasm.Supported by 2023 Opening Research Fund of Yunnan Key Laboratory of Digital Communications(YNJTKFB-20230686,YNKLDC-KFKT-202304).
文摘In image analysis,high-precision semantic segmentation predominantly relies on supervised learning.Despite significant advancements driven by deep learning techniques,challenges such as class imbalance and dynamic performance evaluation persist.Traditional weighting methods,often based on pre-statistical class counting,tend to overemphasize certain classes while neglecting others,particularly rare sample categories.Approaches like focal loss and other rare-sample segmentation techniques introduce multiple hyperparameters that require manual tuning,leading to increased experimental costs due to their instability.This paper proposes a novel CAWASeg framework to address these limitations.Our approach leverages Grad-CAM technology to generate class activation maps,identifying key feature regions that the model focuses on during decision-making.We introduce a Comprehensive Segmentation Performance Score(CSPS)to dynamically evaluate model performance by converting these activation maps into pseudo mask and comparing them with Ground Truth.Additionally,we design two adaptive weights for each class:a Basic Weight(BW)and a Ratio Weight(RW),which the model adjusts during training based on real-time feedback.Extensive experiments on the COCO-Stuff,CityScapes,and ADE20k datasets demonstrate that our CAWASeg framework significantly improves segmentation performance for rare sample categories while enhancing overall segmentation accuracy.The proposed method offers a robust and efficient solution for addressing class imbalance in semantic segmentation tasks.
基金supported by the Institute of Information&Communications Technology Planning&Evaluation(IITP)grant funded by the Korea government(MSIT)[RS-2021-II211341,Artificial Intelligence Graduate School Program(Chung-Ang University)],and by the Chung-Ang University Graduate Research Scholarship in 2024.
文摘Legal case classification involves the categorization of legal documents into predefined categories,which facilitates legal information retrieval and case management.However,real-world legal datasets often suffer from class imbalances due to the uneven distribution of case types across legal domains.This leads to biased model performance,in the form of high accuracy for overrepresented categories and underperformance for minority classes.To address this issue,in this study,we propose a data augmentation method that masks unimportant terms within a document selectively while preserving key terms fromthe perspective of the legal domain.This approach enhances data diversity and improves the generalization capability of conventional models.Our experiments demonstrate consistent improvements achieved by the proposed augmentation strategy in terms of accuracy and F1 score across all models,validating the effectiveness of the proposed method in legal case classification.
基金partially funded by Shenzhen Science and Technology Program(No.JCYJ20240813112038050)the National Natural Science Foundation of China(No.52404059)+1 种基金the Economy Trade and Information Commission of Shenzhen Municipality,China(No.HYCYPT20140507010002)the Key Program of Marine Economy Development(Six Marine Industries)Special Foundation of the Department of Natural Resources of Guangdong Province,China(No.GDOE[2021]55).
文摘Natural gas hydrate in Class Ⅰ reservoirs holds significant commercial potential,as demonstrated by production trials in the South China Sea.However,experimental studies have focused largely on Class Ⅲ systems,with Class Ⅰ/Ⅱ reservoirs remaining underrepresented due to the difficulties in simulating the geothermal gradient and interlayer interactions.This study investigates depressurization performance across all three classes using a novel 360°rotatable reactor with segmented temperature control,enabling precise simulation of reservoir conditions.Results reveal:(i)Class Ⅰ shows two-stage gas production,with 50%from early free gas enabling rapid depressurization,followed by dissociated gas dominance.They achieve 38.4%-78.3%higher cumulative production and superior gas-to-water ratios due to efficient energy use.(ii)The free gas layer in Class Ⅰ accelerates pressure and heat transfer.Class Ⅱ’s water layer provides sensible heat but causes water blocking,impairing heat flow.Class Ⅲ exhibits rapid initial dissociation but a quick decline without fluid support.(iii)Low temperature,low hydrate saturation,and high production pressure collectively reduce efficiency by increasing flow resistance,limiting gas supply,and reducing dissociation drive.Over-depressurization risks hydrate reformation and ice blockage.This work bridges experimental gaps for Class Ⅰ/Ⅱ reservoirs,offering key insights for optimizing recovery.
基金supported by Basic Science Research Program to Research Institute for Basic Sciences(RIBS)of Jeju National University through the National Research Foundation of Korea(NRF)funded by the Ministry of Education(RS-2019-NR040080)This research was also carried out with the support of the Jeju RISE Center,funded by the Ministry of Education and Jeju Special Self-Governing Province in 2025,as part of the“Regional Innovation System&Education(RISE):Glocal University 30”initiative.
文摘Given the growing number of vehicle accidents caused by unintended acceleration and braking failure,verifying Sudden Unintended Acceleration(SUA)incidents has become a persistent challenge.A central issue of debate is whether such events stem frommechanical malfunctions or driver pedalmisapplications.However,existing verification procedures implemented by vehiclemanufacturers often involve closed tests after vehicle recalls;thus raising ongoing concerns about reliability and transparency.Consequently,there is a growing need for a user-driven framework that enables independent data acquisition and verification.Although previous studies have addressed SUA detection using deep learning,few have explored howclass granularity optimization affects power efficiency and inference performance in real-time Edge AI systems.To address this problem,this work presents a cloud-assisted artificial intelligence(AI)solution for the reliable verification of SUA occurrences.The proposed system integrates multimodal sensor streams including camera-based foot images,On-Board Diagnostics II(OBD-II)signals,and six-axismeasurements to determine whether the brake pedal was actually engaged at themoment of a suspected SUA.Beyond image acquisition,convolutional neural network(CNN)models perform real-time inference to classify the driver’s pedal operation states with the resulting outputs transmitted and archived in the cloud.A dedicated dataset of brake and accelerator pedal images was collected from 15 vehicles produced by 6 domestic and international manufacturers.Using this dataset,transfer learning techniques were applied to compare and analyze model performance and generalization as the CNN class granularity varied from coarse to fine levels.Furthermore,classification performance was evaluated in terms of latency and power efficiency under different class configurations.The experimental results demonstrated that the proposed solution identified the driver’s pedal behavior accurately and promptly,with the two-class model achieving the highest F1-score and accuracy among all granularity settings.
