Pub Date : 2025-02-22DOI: 10.1016/j.knosys.2025.113175
Chang Hu, Yihong Dong, Shoubo Peng
Due to the complexity and incompleteness of cognitive tests, as well as subjective biases in humans, using functional magnetic resonance imaging (fMRI) data for accurate diagnosis of psychiatric disease is a challenging task. In addition, existing contrastive methods are also limited by graph augmentation and negative sampling methods in population-based classification. In order to improve the representation learning and classification of fMRI under limited labeled data, we propose a new contrastive self-supervised learning method based on spectral augmentation, namely Multiscale Spectral Augmentation for Graph Contrastive Learning (MSA-GCL) for fMRI Analysis. Concretely, we adopt a two-stage spectral augmentation method by initialization and fine-tuning to mine features of multimodal data. This approach effectively addresses the limitations faced by models that solely rely on coarse-grained spectral augmentation, which leads to weak robustness and limited generalization on medical datasets. Besides, we add a semantic module to fully utilize non-imaging data. Our method is tested on ABIDE I and ADHD-200 datasets, demonstrating superior performance in diagnosis of autism spectrum disorders(ASD) and attention deficit and hyperactivity disorder(ADHD).
{"title":"Multiscale Spectral Augmentation for Graph Contrastive Learning for fMRI analysis to diagnose psychiatric disease","authors":"Chang Hu, Yihong Dong, Shoubo Peng","doi":"10.1016/j.knosys.2025.113175","DOIUrl":"10.1016/j.knosys.2025.113175","url":null,"abstract":"<div><div>Due to the complexity and incompleteness of cognitive tests, as well as subjective biases in humans, using functional magnetic resonance imaging (fMRI) data for accurate diagnosis of psychiatric disease is a challenging task. In addition, existing contrastive methods are also limited by graph augmentation and negative sampling methods in population-based classification. In order to improve the representation learning and classification of fMRI under limited labeled data, we propose a new contrastive self-supervised learning method based on spectral augmentation, namely Multiscale Spectral Augmentation for Graph Contrastive Learning (MSA-GCL) for fMRI Analysis. Concretely, we adopt a two-stage spectral augmentation method by initialization and fine-tuning to mine features of multimodal data. This approach effectively addresses the limitations faced by models that solely rely on coarse-grained spectral augmentation, which leads to weak robustness and limited generalization on medical datasets. Besides, we add a semantic module to fully utilize non-imaging data. Our method is tested on ABIDE I and ADHD-200 datasets, demonstrating superior performance in diagnosis of autism spectrum disorders(ASD) and attention deficit and hyperactivity disorder(ADHD).</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"314 ","pages":"Article 113175"},"PeriodicalIF":7.2,"publicationDate":"2025-02-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143474489","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-02-21DOI: 10.1016/j.knosys.2025.113184
Yuao Zhang , Shuya Ke , Jing Li , Weihua Liu , Jueliang Hu , Kaixiang Yang
The broad learning system (BLS) is a recently developed neural network framework recognized for its efficiency and effectiveness in handling high-dimensional data with a flat network architecture. However, traditional BLS models are highly sensitive to outliers and noisy data, which can significantly degrade performance. While incorporating the -norm loss function enhances robustness against outliers, it often compromises performance on clean datasets. To address this limitation, we propose the Huber-type robust broad learning system (HR-BLS), which integrates the Huber loss function into BLS, effectively combining the strengths of both -norm and -norm loss functions to achieve balanced robustness against data anomalies. Moreover, the elastic-net regularization is included to simultaneously enhance model stability and promote sparsity. To effectively manage large-scale and distributed data, we extend HR-BLS by introducing the distributed Huber-type robust broad learning system (DHR-BLS). Given the non-differentiability of the -norm, traditional gradient-based optimization methods are insufficient. Therefore, we adopt the alternating direction method of multipliers (ADMM) to train, ensuring convergence through the use of appropriate constraints. Experimental results on both synthetic and benchmark datasets show that HR-BLS outperforms traditional BLS and other state-of-the-art robust learning methods in terms of accuracy and robustness. Furthermore, DHR-BLS demonstrates exceptional scalability and effectiveness, making it suitable for distributed learning environments.
{"title":"DHR-BLS: A Huber-type robust broad learning system with its distributed version","authors":"Yuao Zhang , Shuya Ke , Jing Li , Weihua Liu , Jueliang Hu , Kaixiang Yang","doi":"10.1016/j.knosys.2025.113184","DOIUrl":"10.1016/j.knosys.2025.113184","url":null,"abstract":"<div><div>The broad learning system (BLS) is a recently developed neural network framework recognized for its efficiency and effectiveness in handling high-dimensional data with a flat network architecture. However, traditional BLS models are highly sensitive to outliers and noisy data, which can significantly degrade performance. While incorporating the <span><math><msub><mrow><mi>ℓ</mi></mrow><mrow><mn>1</mn></mrow></msub></math></span>-norm loss function enhances robustness against outliers, it often compromises performance on clean datasets. To address this limitation, we propose the Huber-type robust broad learning system (HR-BLS), which integrates the Huber loss function into BLS, effectively combining the strengths of both <span><math><msub><mrow><mi>ℓ</mi></mrow><mrow><mn>1</mn></mrow></msub></math></span>-norm and <span><math><msub><mrow><mi>ℓ</mi></mrow><mrow><mn>2</mn></mrow></msub></math></span>-norm loss functions to achieve balanced robustness against data anomalies. Moreover, the elastic-net regularization is included to simultaneously enhance model stability and promote sparsity. To effectively manage large-scale and distributed data, we extend HR-BLS by introducing the distributed Huber-type robust broad learning system (DHR-BLS). Given the non-differentiability of the <span><math><msub><mrow><mi>ℓ</mi></mrow><mrow><mn>1</mn></mrow></msub></math></span>-norm, traditional gradient-based optimization methods are insufficient. Therefore, we adopt the alternating direction method of multipliers (ADMM) to train, ensuring convergence through the use of appropriate constraints. Experimental results on both synthetic and benchmark datasets show that HR-BLS outperforms traditional BLS and other state-of-the-art robust learning methods in terms of accuracy and robustness. Furthermore, DHR-BLS demonstrates exceptional scalability and effectiveness, making it suitable for distributed learning environments.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"314 ","pages":"Article 113184"},"PeriodicalIF":7.2,"publicationDate":"2025-02-21","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143471538","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-02-21DOI: 10.1016/j.knosys.2025.113178
Mohammad Arafah , Iain Phillips , Asma Adnane , Mohammad Alauthman , Nauman Aslam
Intrusion detection systems face significant challenges in handling high-dimensional, large-scale, and imbalanced network traffic data. This paper proposes a new architecture combining a denoising autoencoder (AE) and a Wasserstein Generative Adversarial Network (WGAN) to address these challenges. The AE-WGAN model extracts high-representative features and generates realistic synthetic attacks, effectively resolving data imbalance and enhancing anomaly-based intrusion detection. Our extensive experiments on NSL-KDD and CICIDS-2017 datasets demonstrate superior performance, achieving 98% accuracy and 99% F1-score in binary classification, surpassing recent approaches by 7%–15%. In multiclass cases, the model achieves 89% precision for DoS attacks and 84% for Probe attacks, while maintaining 79% precision for rare U2R attacks. Time complexity analysis reveals 23% reduced training time while maintaining high-quality synthetic attack generation, contributing a robust framework capable of handling modern network traffic complexities and evolving cyber threats.
{"title":"An enhanced BiGAN architecture for network intrusion detection","authors":"Mohammad Arafah , Iain Phillips , Asma Adnane , Mohammad Alauthman , Nauman Aslam","doi":"10.1016/j.knosys.2025.113178","DOIUrl":"10.1016/j.knosys.2025.113178","url":null,"abstract":"<div><div>Intrusion detection systems face significant challenges in handling high-dimensional, large-scale, and imbalanced network traffic data. This paper proposes a new architecture combining a denoising autoencoder (AE) and a Wasserstein Generative Adversarial Network (WGAN) to address these challenges. The AE-WGAN model extracts high-representative features and generates realistic synthetic attacks, effectively resolving data imbalance and enhancing anomaly-based intrusion detection. Our extensive experiments on NSL-KDD and CICIDS-2017 datasets demonstrate superior performance, achieving 98% accuracy and 99% F1-score in binary classification, surpassing recent approaches by 7%–15%. In multiclass cases, the model achieves 89% precision for DoS attacks and 84% for Probe attacks, while maintaining 79% precision for rare U2R attacks. Time complexity analysis reveals 23% reduced training time while maintaining high-quality synthetic attack generation, contributing a robust framework capable of handling modern network traffic complexities and evolving cyber threats.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"314 ","pages":"Article 113178"},"PeriodicalIF":7.2,"publicationDate":"2025-02-21","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143465240","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-02-19DOI: 10.1016/j.knosys.2025.113172
Cheng Ding , Zhi Zheng
Transfer learning has shown promising potentials in assisting multi-agent systems (MAS) to deal with complex collaborative tasks. In this work, we investigate MAS collaboration in 3D underwater environment. In response to the problem of high sampling cost in underwater operation when multi-agent without any prior knowledge, the multi-agent collaborative operation planning via cross-domain transfer learning (CDTL) is proposed. In CDTL, the training process of MAS is accelerated through learning the domain invariant knowledge from the samples of 2D ground collaborative tasks that easily obtained. First, the samples in ground tasks are divided into six state phases based on the semantic order of task execution, and a state transition graph is constructed accordingly. Then, a domain adaptation method with inter-class relationship (ICDA) is proposed, which focuses on the invariant semantic structure of the ground (source) and the underwater (target) task to capture prior knowledge. During the knowledge transferring, ICDA is used to correct decision of the agents’ policies that based on MAX-Q controller. Finally, the extensive experiments show that CDTL reduces the cost of physical time by 37.3% when the MAS completes the new task for the first time.
{"title":"Multi-agent collaborative operation planning via cross-domain transfer learning","authors":"Cheng Ding , Zhi Zheng","doi":"10.1016/j.knosys.2025.113172","DOIUrl":"10.1016/j.knosys.2025.113172","url":null,"abstract":"<div><div>Transfer learning has shown promising potentials in assisting multi-agent systems (MAS) to deal with complex collaborative tasks. In this work, we investigate MAS collaboration in 3D underwater environment. In response to the problem of high sampling cost in underwater operation when multi-agent without any prior knowledge, the multi-agent collaborative operation planning via cross-domain transfer learning (CDTL) is proposed. In CDTL, the training process of MAS is accelerated through learning the domain invariant knowledge from the samples of 2D ground collaborative tasks that easily obtained. First, the samples in ground tasks are divided into six state phases based on the semantic order of task execution, and a state transition graph is constructed accordingly. Then, a domain adaptation method with inter-class relationship (ICDA) is proposed, which focuses on the invariant semantic structure of the ground (source) and the underwater (target) task to capture prior knowledge. During the knowledge transferring, ICDA is used to correct decision of the agents’ policies that based on MAX-Q controller. Finally, the extensive experiments show that CDTL reduces the cost of physical time by 37.3% when the MAS completes the new task for the first time.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"314 ","pages":"Article 113172"},"PeriodicalIF":7.2,"publicationDate":"2025-02-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143474491","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-02-19DOI: 10.1016/j.knosys.2025.113209
Kui Hu , Qingbo He , Hao Xu , Changming Cheng , Zhike Peng
The cross-domain intelligent fault diagnosis (IFD) using unlabeled data has attracted more and more attention. However, most researchers focus on the improvement of single domain adaptive method (DAM). How to make full use of existing DAMs to improve the accuracy and generalization of the IFD model is a challenging problem. As a potential solution, a general dynamic domain adaptive ensemble (DDAE) framework is proposed. By introducing the optimal adaptation factor and combining with the proposed dynamic adaptive evaluating strategy, the DDAE can quantitatively evaluate the importance of different DAMs, and dynamically adjust the weight of DAMs during the training process. By this way, the ensemble strategy can be constructed adaptively within the model. We also design a feasible DDAE-based neural network model by integrating three different DAMs. Extensive experimental analysis indicates that the diagnostic performance of the model is superior to existing deep learning and transfer learning methods.
{"title":"Dynamic domain adaptive ensemble for intelligent fault diagnosis of machinery","authors":"Kui Hu , Qingbo He , Hao Xu , Changming Cheng , Zhike Peng","doi":"10.1016/j.knosys.2025.113209","DOIUrl":"10.1016/j.knosys.2025.113209","url":null,"abstract":"<div><div>The cross-domain intelligent fault diagnosis (IFD) using unlabeled data has attracted more and more attention. However, most researchers focus on the improvement of single domain adaptive method (DAM). How to make full use of existing DAMs to improve the accuracy and generalization of the IFD model is a challenging problem. As a potential solution, a general dynamic domain adaptive ensemble (DDAE) framework is proposed. By introducing the optimal adaptation factor and combining with the proposed dynamic adaptive evaluating strategy, the DDAE can quantitatively evaluate the importance of different DAMs, and dynamically adjust the weight of DAMs during the training process. By this way, the ensemble strategy can be constructed adaptively within the model. We also design a feasible DDAE-based neural network model by integrating three different DAMs. Extensive experimental analysis indicates that the diagnostic performance of the model is superior to existing deep learning and transfer learning methods.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"314 ","pages":"Article 113209"},"PeriodicalIF":7.2,"publicationDate":"2025-02-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143471539","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-02-18DOI: 10.1016/j.knosys.2025.113169
Ugur Yuzgec
This study considers the Single Candidate Optimizer (SCO) as an alternative to population-based heuristics, that is faster than them. Although the SCO algorithm is a fast single-candidate-based heuristic, it has certain limitations. To overcome these limitations and enhance the search performance of SCO, several solutions were proposed in this study. First, owing to the single-candidate nature of the SCO, the initial solution position can play a critical role. To compensate for this, an accelerated opposition-learning mechanism was integrated into the SCO. In addition, instead of the equation that is active when the number of unsuccessful improvement attempts is reached in the SCO structure, a mutation operator including chaotic functions (Levy, Gauss, and Cauchy) has been incorporated into the algorithm. Again, equations based on new approaches were added to the SCO algorithm to update the position of the candidate solution during the exploration and exploitation phases. Finally, the standard boundary value control mechanism is replaced with a more effective one. The algorithm developed in this study is named Accelerated Opposition Learning based Chaotic Single Candidate Optimizer (AccOppCSCO), inspired by the accelerated opposition learning mechanism and the mutation operator involving chaotic behaviors. The search capability of the proposed AccOppCSCO algorithm was first analyzed using four different methods: convergence, search history, trajectory, and computational complexity. The effectiveness of the mechanisms used in the AccOppCSCO algorithm for four different two-dimensional benchmark problems from the IEEE Congress on Evolutionary Computation 2014 (CEC2014) package was demonstrated. Subsequently, the performance of the proposed AccOppCSCO algorithm was evaluated on the CEC2014 and IEEE Congress on Evolutionary Computation 2020 (CEC2020) benchmark problems with different dimensions. The results show that the AccOppCSCO algorithm works effectively in the CEC2014 and CEC2020 test sets and offers better optimization results than SCO. The AccOppCSCO algorithm ranked first in the overall evaluation of the 30-dimensional CEC2014 comparison results with State of the Art (SOTA) heuristics from the literature. Finally, for ten different engineering design problems, the AccOppCSCO algorithm was analyzed and compared with the original SCO and other SOTA heuristics. The results show that AccOppCSCO is effective for engineering design problems. This emphasizes that the algorithm can work effectively on a wide range of problems and can be used in various applications. The source code of the AccOppCSCO algorithm for the CEC2014 benchmark suite is publicly available at https://github.com/uguryuzgec/AccOppCSCO.
{"title":"Accelerated opposition learning based chaotic single candidate optimization algorithm: A new alternative to population-based heuristics","authors":"Ugur Yuzgec","doi":"10.1016/j.knosys.2025.113169","DOIUrl":"10.1016/j.knosys.2025.113169","url":null,"abstract":"<div><div>This study considers the Single Candidate Optimizer (SCO) as an alternative to population-based heuristics, that is faster than them. Although the SCO algorithm is a fast single-candidate-based heuristic, it has certain limitations. To overcome these limitations and enhance the search performance of SCO, several solutions were proposed in this study. First, owing to the single-candidate nature of the SCO, the initial solution position can play a critical role. To compensate for this, an accelerated opposition-learning mechanism was integrated into the SCO. In addition, instead of the equation that is active when the number of unsuccessful improvement attempts is reached in the SCO structure, a mutation operator including chaotic functions (Levy, Gauss, and Cauchy) has been incorporated into the algorithm. Again, equations based on new approaches were added to the SCO algorithm to update the position of the candidate solution during the exploration and exploitation phases. Finally, the standard boundary value control mechanism is replaced with a more effective one. The algorithm developed in this study is named Accelerated Opposition Learning based Chaotic Single Candidate Optimizer (AccOppCSCO), inspired by the accelerated opposition learning mechanism and the mutation operator involving chaotic behaviors. The search capability of the proposed AccOppCSCO algorithm was first analyzed using four different methods: convergence, search history, trajectory, and computational complexity. The effectiveness of the mechanisms used in the AccOppCSCO algorithm for four different two-dimensional benchmark problems from the IEEE Congress on Evolutionary Computation 2014 (CEC2014) package was demonstrated. Subsequently, the performance of the proposed AccOppCSCO algorithm was evaluated on the CEC2014 and IEEE Congress on Evolutionary Computation 2020 (CEC2020) benchmark problems with different dimensions. The results show that the AccOppCSCO algorithm works effectively in the CEC2014 and CEC2020 test sets and offers better optimization results than SCO. The AccOppCSCO algorithm ranked first in the overall evaluation of the 30-dimensional CEC2014 comparison results with State of the Art (SOTA) heuristics from the literature. Finally, for ten different engineering design problems, the AccOppCSCO algorithm was analyzed and compared with the original SCO and other SOTA heuristics. The results show that AccOppCSCO is effective for engineering design problems. This emphasizes that the algorithm can work effectively on a wide range of problems and can be used in various applications. The source code of the AccOppCSCO algorithm for the CEC2014 benchmark suite is publicly available at <span><span>https://github.com/uguryuzgec/AccOppCSCO</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"314 ","pages":"Article 113169"},"PeriodicalIF":7.2,"publicationDate":"2025-02-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143454296","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-02-18DOI: 10.1016/j.knosys.2025.113162
Xinxi Xie , Quan Liu , Jun Yang , Hao Zhang , Zijun Zhou , Chuanjie Zhang , Junwei Yan
Mitigating the impact of adverse weather on images, such as rain, haze, snow and raindrops, poses a critical challenge in numerous computer vision tasks, particularly in outdoor scenarios like port security and traffic surveillance. Recent successful methods for restoring images affected by severe weather have predominantly embraced supervised learning, which heavily depend on the quality of collected image pairs. However, capturing ideal image pairs for adverse weather image restoration in real-world scenarios is nearly impossible. In practice, unpaired images are more commonly available. The absence of proper supervision among unpaired images can result in low-quality image restoration outcomes. Therefore, utilizing unpaired image data for adverse weather image restoration remains a significant challenge. In this paper, we propose an effective method for Unsupervised Adverse weather-degraded Image Restoration (UAIR). Our approach leverages contrastive learning to explore both the similarities and differences in deep feature space among images. We not only utilize the intrinsic similarities between restored image and original degraded image to guide the content of the restored image, but also take advantage of the categoricaly differences within unpaired image data, thereby strengthening the connections between the restored image and the clean image at category level. Extensive experiments conducted on benchmark datasets for various tasks, including image snow removal, combined image rain and haze removal and image raindrop removal demonstrate that our proposed method achieves state-of-the-art performance on both weather-specific and all-in-one weather image restoration.
{"title":"Unsupervised adverse weather-degraded image restoration via contrastive learning","authors":"Xinxi Xie , Quan Liu , Jun Yang , Hao Zhang , Zijun Zhou , Chuanjie Zhang , Junwei Yan","doi":"10.1016/j.knosys.2025.113162","DOIUrl":"10.1016/j.knosys.2025.113162","url":null,"abstract":"<div><div>Mitigating the impact of adverse weather on images, such as rain, haze, snow and raindrops, poses a critical challenge in numerous computer vision tasks, particularly in outdoor scenarios like port security and traffic surveillance. Recent successful methods for restoring images affected by severe weather have predominantly embraced supervised learning, which heavily depend on the quality of collected image pairs. However, capturing ideal image pairs for adverse weather image restoration in real-world scenarios is nearly impossible. In practice, unpaired images are more commonly available. The absence of proper supervision among unpaired images can result in low-quality image restoration outcomes. Therefore, utilizing unpaired image data for adverse weather image restoration remains a significant challenge. In this paper, we propose an effective method for Unsupervised Adverse weather-degraded Image Restoration (UAIR). Our approach leverages contrastive learning to explore both the similarities and differences in deep feature space among images. We not only utilize the intrinsic similarities between restored image and original degraded image to guide the content of the restored image, but also take advantage of the categoricaly differences within unpaired image data, thereby strengthening the connections between the restored image and the clean image at category level. Extensive experiments conducted on benchmark datasets for various tasks, including image snow removal, combined image rain and haze removal and image raindrop removal demonstrate that our proposed method achieves state-of-the-art performance on both weather-specific and all-in-one weather image restoration.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"314 ","pages":"Article 113162"},"PeriodicalIF":7.2,"publicationDate":"2025-02-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143454327","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Traffic prediction is a critical function of Intelligent Transportation Systems. Inspired by Graph/HyperGraph Neural Networks theory, researchers have proposed a series of effective methods for traffic prediction that have been proved as significant successes. Most methods construct an unchanging graph or hypergraph based on a fixed traffic network topology during prediction. These methods treat all traffic data (flow, speed, occupancy) equally, ignoring the different inherent attributes of traffic data. Other methods construct dynamic graph or hypergraph based on traffic data but ignore the topological structure of the road network itself. These methods will decrease the accuracy of prediction results. In this paper, we propose an innovative framework for traffic data prediction based on Sub-Hypergraph and Knowledge Distillation (SHKD), which effectively extracts traffic data features and combines them with road network topology. Specifically, we first cluster traffic data based on the inherent attributes and construct hypergraphs for data with similar attributes to represent their relationships, referred to as sub-hypergraphs. Then a teacher network is built from these sub-hypergraphs to extract the data features in traffic, while a student network is constructed based on the geographical connectivity among roads to extract global topological features. To integrate the two types of features, we apply a knowledge distillation method to transfer the data features learned by the teacher network into the training process of the student network, yielding the final prediction results. The proposed method has been assessed with several real-world datasets in predicting traffic status. The experimental results demonstrate the effectiveness of the proposed method.
{"title":"SHKD: A framework for traffic prediction based on Sub-Hypergraph and Knowledge Distillation","authors":"Xiangyu Yao, Xinglin Piao, Qitan Shao, Yongli Hu, Baocai Yin, Yong Zhang","doi":"10.1016/j.knosys.2025.113163","DOIUrl":"10.1016/j.knosys.2025.113163","url":null,"abstract":"<div><div>Traffic prediction is a critical function of Intelligent Transportation Systems. Inspired by Graph/HyperGraph Neural Networks theory, researchers have proposed a series of effective methods for traffic prediction that have been proved as significant successes. Most methods construct an unchanging graph or hypergraph based on a fixed traffic network topology during prediction. These methods treat all traffic data (flow, speed, occupancy) equally, ignoring the different inherent attributes of traffic data. Other methods construct dynamic graph or hypergraph based on traffic data but ignore the topological structure of the road network itself. These methods will decrease the accuracy of prediction results. In this paper, we propose an innovative framework for traffic data prediction based on Sub-Hypergraph and Knowledge Distillation (SHKD), which effectively extracts traffic data features and combines them with road network topology. Specifically, we first cluster traffic data based on the inherent attributes and construct hypergraphs for data with similar attributes to represent their relationships, referred to as sub-hypergraphs. Then a teacher network is built from these sub-hypergraphs to extract the data features in traffic, while a student network is constructed based on the geographical connectivity among roads to extract global topological features. To integrate the two types of features, we apply a knowledge distillation method to transfer the data features learned by the teacher network into the training process of the student network, yielding the final prediction results. The proposed method has been assessed with several real-world datasets in predicting traffic status. The experimental results demonstrate the effectiveness of the proposed method.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"312 ","pages":"Article 113163"},"PeriodicalIF":7.2,"publicationDate":"2025-02-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143436941","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-02-17DOI: 10.1016/j.knosys.2025.113158
Rui Zhao, Yuetong Li, Qing Zhang, Xinyi Zhao
Existing camouflaged object detection methods have made impressive achievements, however, the interference from highly similar backgrounds, as well as the indistinguishable object boundary, still hider the detection accuracy. In this paper, we propose a three-stage bilateral decoupling complementarity learning network (BDCL-Net) to explore how to utilize the specific advantages of multi-level encoded features for achieving high-quality inference. Specifically, all side-output features are decoupled into two branches to generate three complementary features. Different from previous methods that focus on obtaining the camouflaged object and body boundary, our body modeling stage, which includes a global positioning flow (GPF) module and a multi-scale body warping (MBW) module, is deployed to obtain a global contextual feature that provides coarse localization of potential camouflaged objects and a body feature that emphasizes learning the central areas of camouflaged objects. The detail preservation stage is designed to generate a detail feature that pays attention to the regions around the boundary. Consequently, the body prediction can avoid disturbances from the highly similar backgrounds, while the detail prediction can reduce errors caused by imbalanced boundary pixels. The complementary feature integration (CFI) module in the feature aggregation stage is designed to fuse these complementary features in an interactive learning manner. We conduct extensive experiments on four public datasets to demonstrate the effectiveness and superiority of our proposed network. The code is available at http://github.com/iuueong/BDCLNet.
{"title":"Bilateral decoupling complementarity learning network for camouflaged object detection","authors":"Rui Zhao, Yuetong Li, Qing Zhang, Xinyi Zhao","doi":"10.1016/j.knosys.2025.113158","DOIUrl":"10.1016/j.knosys.2025.113158","url":null,"abstract":"<div><div>Existing camouflaged object detection methods have made impressive achievements, however, the interference from highly similar backgrounds, as well as the indistinguishable object boundary, still hider the detection accuracy. In this paper, we propose a three-stage bilateral decoupling complementarity learning network (BDCL-Net) to explore how to utilize the specific advantages of multi-level encoded features for achieving high-quality inference. Specifically, all side-output features are decoupled into two branches to generate three complementary features. Different from previous methods that focus on obtaining the camouflaged object and body boundary, our body modeling stage, which includes a global positioning flow (GPF) module and a multi-scale body warping (MBW) module, is deployed to obtain a global contextual feature that provides coarse localization of potential camouflaged objects and a body feature that emphasizes learning the central areas of camouflaged objects. The detail preservation stage is designed to generate a detail feature that pays attention to the regions around the boundary. Consequently, the body prediction can avoid disturbances from the highly similar backgrounds, while the detail prediction can reduce errors caused by imbalanced boundary pixels. The complementary feature integration (CFI) module in the feature aggregation stage is designed to fuse these complementary features in an interactive learning manner. We conduct extensive experiments on four public datasets to demonstrate the effectiveness and superiority of our proposed network. The code is available at <span><span>http://github.com/iuueong/BDCLNet</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"314 ","pages":"Article 113158"},"PeriodicalIF":7.2,"publicationDate":"2025-02-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143453700","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-02-16DOI: 10.1016/j.knosys.2025.113188
Yuan Ding, Kaijun Wu, Bin Tian
In the digital conservation of Dunhuang murals, inpainting plays a pivotal role in the reconstructing of complete murals. Focusing on the semantic difference between masked and unmasked regions and the irrationality of the semantics of masked regions in the digitization of Dunhuang frescoes for inpainting, we applied the diffusion probability model based on image denoising to the inpainting of Dunhuang murals through official collaboration. We proposed a new frequency-domain guided diffusion model for Dunhuang mural painting to reconstruct the traditional structure-guided image painting process. First, we designed a module that generates frequency-domain information to decompose the frescoes into high- and low-frequency components. Then, we restored them to obtain the corresponding high- and low-frequency images. These images can guide the diffusion inpainting process of damaged murals, effectively resolving the semantic differences between masked and unmasked regions and yielding semantically consistent and reasonable inpainting results. In addition, we proposed a bilateral contrastive learning strategy to ensure that the inpainted images are closer to the positive samples and farther away from the negative samples in the representation space, thereby optimizing the inpainting performance. The effectiveness of the proposed method was verified by conducting large-scale inpainting experiments on real Dunhuang murals and the Place2 standard dataset. The proposed method outperformed other methods in terms of the subjective and objective evaluation metrics. The proposed method effectively recovered the details and content information of the damaged murals and provided more advanced technical support for the digitization and conservation of Dunhuang murals.
{"title":"Frequency-domain information guidance: Diffusion models for the Inpainting of Dunhuang murals","authors":"Yuan Ding, Kaijun Wu, Bin Tian","doi":"10.1016/j.knosys.2025.113188","DOIUrl":"10.1016/j.knosys.2025.113188","url":null,"abstract":"<div><div>In the digital conservation of Dunhuang murals, inpainting plays a pivotal role in the reconstructing of complete murals. Focusing on the semantic difference between masked and unmasked regions and the irrationality of the semantics of masked regions in the digitization of Dunhuang frescoes for inpainting, we applied the diffusion probability model based on image denoising to the inpainting of Dunhuang murals through official collaboration. We proposed a new frequency-domain guided diffusion model for Dunhuang mural painting to reconstruct the traditional structure-guided image painting process. First, we designed a module that generates frequency-domain information to decompose the frescoes into high- and low-frequency components. Then, we restored them to obtain the corresponding high- and low-frequency images. These images can guide the diffusion inpainting process of damaged murals, effectively resolving the semantic differences between masked and unmasked regions and yielding semantically consistent and reasonable inpainting results. In addition, we proposed a bilateral contrastive learning strategy to ensure that the inpainted images are closer to the positive samples and farther away from the negative samples in the representation space, thereby optimizing the inpainting performance. The effectiveness of the proposed method was verified by conducting large-scale inpainting experiments on real Dunhuang murals and the Place2 standard dataset. The proposed method outperformed other methods in terms of the subjective and objective evaluation metrics. The proposed method effectively recovered the details and content information of the damaged murals and provided more advanced technical support for the digitization and conservation of Dunhuang murals.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"314 ","pages":"Article 113188"},"PeriodicalIF":7.2,"publicationDate":"2025-02-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143454290","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
扫码关注我们
求助内容：
应助结果提醒方式：
京公网安备 11010802042870号