Pub Date : 2025-02-03DOI: 10.1016/j.knosys.2025.113094
Simi Job , Xiaohui Tao , Taotao Cai , Lin Li , Haoran Xie , Cai Xu , Jianming Yong
Classifying graph-structured data presents significant challenges due to the diverse features of nodes and edges and their complex relationships. While Graph Neural Networks (GNNs) are widely used for graph prediction tasks, their performance is often hindered by these intricate dependencies. Leveraging causality holds potential in overcoming these challenges by identifying causal links among features, thus enhancing GNN classification performance. However, depending solely on adjacency matrices or attention mechanisms, as commonly studied in causal prediction research, is insufficient for capturing the complex interactions among features. To address these challenges, we present HebCGNN, a Hebbian-enabled Causal GNN classification model that incorporates dynamic impact valuing. Our method creates a robust framework that prioritizes causal elements in prediction tasks. Extensive experiments on seven publicly available datasets across diverse domains demonstrate that HebCGNN outperforms state-of-the-art models.
{"title":"HebCGNN: Hebbian-enabled causal classification integrating dynamic impact valuing","authors":"Simi Job , Xiaohui Tao , Taotao Cai , Lin Li , Haoran Xie , Cai Xu , Jianming Yong","doi":"10.1016/j.knosys.2025.113094","DOIUrl":"10.1016/j.knosys.2025.113094","url":null,"abstract":"<div><div>Classifying graph-structured data presents significant challenges due to the diverse features of nodes and edges and their complex relationships. While Graph Neural Networks (GNNs) are widely used for graph prediction tasks, their performance is often hindered by these intricate dependencies. Leveraging causality holds potential in overcoming these challenges by identifying causal links among features, thus enhancing GNN classification performance. However, depending solely on adjacency matrices or attention mechanisms, as commonly studied in causal prediction research, is insufficient for capturing the complex interactions among features. To address these challenges, we present <em>HebCGNN</em>, a Hebbian-enabled Causal GNN classification model that incorporates <em>dynamic impact valuing</em>. Our method creates a robust framework that prioritizes causal elements in prediction tasks. Extensive experiments on seven publicly available datasets across diverse domains demonstrate that <em>HebCGNN</em> outperforms state-of-the-art models.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"311 ","pages":"Article 113094"},"PeriodicalIF":7.2,"publicationDate":"2025-02-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143340028","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-02-03DOI: 10.1016/j.knosys.2025.113034
Bo Peng , Huan Xu , Yupeng Sun , Quanle Liu , Xiangjiu Che
Graph convolutional networks (GCNs) (Zhang et al., 2018) are known for their exceptional graph learning performance. However, they have difficulties with complex topological structures, which can lead to over-smoothing and computational inefficiency. Existing hyperbolic GCN models are unable to efficiently capture the intricate features of such graphs, resulting in reduced accuracy. A hyperbolic tree-based graph convolutional neural network (HTGCN) is proposed that preserves the complexity of graph structures and mitigates the problem of over-smoothing by transforming graphs into tree structures and using hyperbolic models to aggregate key features. HTGCN improves the topological and spatial mapping of datasets. It also uses a parallel strategy for logarithmic and exponential approximation. Experimental results on six real-world datasets demonstrate that HTGCN improves accuracy in node classification and link prediction tasks compared to existing hyperbolic GCNs and topological structure models. This highlights its unique advantage in handling complex network structures. This paper extends the application of HGCNs and provides a new perspective and tools for deep learning models to handle complex networks.
{"title":"A construction of hyperbolic tree-based graph convolutional networks utilizing the Padé approximation","authors":"Bo Peng , Huan Xu , Yupeng Sun , Quanle Liu , Xiangjiu Che","doi":"10.1016/j.knosys.2025.113034","DOIUrl":"10.1016/j.knosys.2025.113034","url":null,"abstract":"<div><div>Graph convolutional networks (GCNs) (Zhang et al., 2018) are known for their exceptional graph learning performance. However, they have difficulties with complex topological structures, which can lead to over-smoothing and computational inefficiency. Existing hyperbolic GCN models are unable to efficiently capture the intricate features of such graphs, resulting in reduced accuracy. A hyperbolic tree-based graph convolutional neural network (HTGCN) is proposed that preserves the complexity of graph structures and mitigates the problem of over-smoothing by transforming graphs into tree structures and using hyperbolic models to aggregate key features. HTGCN improves the topological and spatial mapping of datasets. It also uses a parallel strategy for logarithmic and exponential approximation. Experimental results on six real-world datasets demonstrate that HTGCN improves accuracy in node classification and link prediction tasks compared to existing hyperbolic GCNs and topological structure models. This highlights its unique advantage in handling complex network structures. This paper extends the application of HGCNs and provides a new perspective and tools for deep learning models to handle complex networks.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"311 ","pages":"Article 113034"},"PeriodicalIF":7.2,"publicationDate":"2025-02-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143340036","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 : 2025-02-03DOI: 10.1016/j.knosys.2025.113074
Ziwei Wang, Siyang Li, Xiaoqing Chen, Dongrui Wu
Accurate decoding of electroencephalography (EEG) signals is crucial for brain–computer interfaces (BCIs); however, individual differences, non-stationarity of EEG signals, and limited training data make the decoding very challenging. Existing EEG data augmentation approaches usually operate in the temporal, frequency, or spatial domain only, which may not adequately capture the non-stationarity of EEGs. Moreover, these methods typically generate within-subject augmented trials, restricting their effectiveness in accommodating inter-subject variability. This paper proposes two time–frequency transform based EEG data augmentation approaches: Discrete Wavelet Transform Augmentation (DWTaug) and Hilbert–Huang Transform Augmentation (HHTaug). Both follow three steps: time–frequency domain decomposition, cross-subject sub-signal reassembling, and time domain reconstruction. Augmenting data expands the pool of labeled training samples, alleviating the data scarcity problem; time–frequency decomposition captures the non-stationary properties of EEG signals more effectively; finally, cross-subject reassembling of sub-signals handles individual differences. Experiments on 17 datasets from three different BCI paradigms demonstrated the superiority of DWTaug and HHTaug over nine existing EEG data augmentation approaches, improving 4% over baseline on average. By leveraging essential time–frequency information, DWTaug and HHTaug introduce new utility to traditional signal processing techniques, enhancing EEG data augmentation, thus effectively addressing key EEG decoding challenges. To our knowledge, this is the first work to simultaneously address individual variability, non-stationarity, and data scarcity in EEG decoding, significantly enhancing the real-world applicability of BCIs. Our code is publicized at https://github.com/wzwvv/CSDA.
{"title":"Time–frequency transform based EEG data augmentation for brain–computer interfaces","authors":"Ziwei Wang, Siyang Li, Xiaoqing Chen, Dongrui Wu","doi":"10.1016/j.knosys.2025.113074","DOIUrl":"10.1016/j.knosys.2025.113074","url":null,"abstract":"<div><div>Accurate decoding of electroencephalography (EEG) signals is crucial for brain–computer interfaces (BCIs); however, individual differences, non-stationarity of EEG signals, and limited training data make the decoding very challenging. Existing EEG data augmentation approaches usually operate in the temporal, frequency, or spatial domain only, which may not adequately capture the non-stationarity of EEGs. Moreover, these methods typically generate within-subject augmented trials, restricting their effectiveness in accommodating inter-subject variability. This paper proposes two time–frequency transform based EEG data augmentation approaches: Discrete Wavelet Transform Augmentation (DWTaug) and Hilbert–Huang Transform Augmentation (HHTaug). Both follow three steps: time–frequency domain decomposition, cross-subject sub-signal reassembling, and time domain reconstruction. Augmenting data expands the pool of labeled training samples, alleviating the data scarcity problem; time–frequency decomposition captures the non-stationary properties of EEG signals more effectively; finally, cross-subject reassembling of sub-signals handles individual differences. Experiments on 17 datasets from three different BCI paradigms demonstrated the superiority of DWTaug and HHTaug over nine existing EEG data augmentation approaches, improving 4% over baseline on average. By leveraging essential time–frequency information, DWTaug and HHTaug introduce new utility to traditional signal processing techniques, enhancing EEG data augmentation, thus effectively addressing key EEG decoding challenges. To our knowledge, this is the first work to simultaneously address individual variability, non-stationarity, and data scarcity in EEG decoding, significantly enhancing the real-world applicability of BCIs. Our code is publicized at <span><span>https://github.com/wzwvv/CSDA</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"311 ","pages":"Article 113074"},"PeriodicalIF":7.2,"publicationDate":"2025-02-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143339783","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 : 2025-02-03DOI: 10.1016/j.knosys.2025.113090
Yue Ge, Huaicheng Zhang, Jiguang Shi, Deyu Luo, Sheng Chang, Jin He, Qijun Huang, Hao Wang
Self-supervised learning (SSL) is a prevalent approach in the diagnosis of cardiovascular diseases. It leverages unlabeled ECG data for pre-training and thus alleviates the reliance on manual annotations. However, existing self-supervised methods tend to focus on either global or local features, hindering the integrated consideration of the global spatial relationships indicated by leads and the local cardiac activity information. To address this problem, a jigsaw-based autoencoder with masked contrastive learning (JAMC) is proposed. The model merges instance discrimination-based contrastive learning with jigsaw-based reconstruction task. This merge allows for simultaneous attention to the global invariant features and local detail features of the signals. In addition, lead, temporal, and mixed jigsaw transformations are applied, which rely on the contrast between different views to enable the model to synthesize the spatial position of each lead and the temporal features of ECG segments. The comparison of the three techniques highlights the impact of spatial lead structure and temporal dependencies on feature learning. To further emphasize local lead correlations, another lead mask transformation is introduced. It infers missing lead information from the remaining leads, promoting a focus on local features to certain regions of the heart. JAMC learns morphological, temporal, and spatial physiological features of ECG signals from both global and local perspectives. In contrast to existing methods, it demonstrates increases in F1-macro by 7.26% and 17.29% on the Chapman and the CPSC2018 dataset, respectively. The method explores a universal SSL framework that combines generative and contrastive tasks, demonstrating exceptional performance. This framework holds potential application value in the field of ECG signal classification and diagnosis.
{"title":"JAMC: A jigsaw-based autoencoder with masked contrastive learning for cardiovascular disease diagnosis","authors":"Yue Ge, Huaicheng Zhang, Jiguang Shi, Deyu Luo, Sheng Chang, Jin He, Qijun Huang, Hao Wang","doi":"10.1016/j.knosys.2025.113090","DOIUrl":"10.1016/j.knosys.2025.113090","url":null,"abstract":"<div><div>Self-supervised learning (SSL) is a prevalent approach in the diagnosis of cardiovascular diseases. It leverages unlabeled ECG data for pre-training and thus alleviates the reliance on manual annotations. However, existing self-supervised methods tend to focus on either global or local features, hindering the integrated consideration of the global spatial relationships indicated by leads and the local cardiac activity information. To address this problem, a jigsaw-based autoencoder with masked contrastive learning (JAMC) is proposed. The model merges instance discrimination-based contrastive learning with jigsaw-based reconstruction task. This merge allows for simultaneous attention to the global invariant features and local detail features of the signals. In addition, lead, temporal, and mixed jigsaw transformations are applied, which rely on the contrast between different views to enable the model to synthesize the spatial position of each lead and the temporal features of ECG segments. The comparison of the three techniques highlights the impact of spatial lead structure and temporal dependencies on feature learning. To further emphasize local lead correlations, another lead mask transformation is introduced. It infers missing lead information from the remaining leads, promoting a focus on local features to certain regions of the heart. JAMC learns morphological, temporal, and spatial physiological features of ECG signals from both global and local perspectives. In contrast to existing methods, it demonstrates increases in F1-macro by 7.26% and 17.29% on the Chapman and the CPSC2018 dataset, respectively. The method explores a universal SSL framework that combines generative and contrastive tasks, demonstrating exceptional performance. This framework holds potential application value in the field of ECG signal classification and diagnosis.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"311 ","pages":"Article 113090"},"PeriodicalIF":7.2,"publicationDate":"2025-02-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143360590","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 : 2025-02-02DOI: 10.1016/j.knosys.2025.113128
Fukai Zhang , Xiaobo Jin , Jie Jiang , Gang Lin , Mingzhi Wang , Shan An , Qiang Lyu
Fine-grained visual classification (FGVC) has numerous applications in various sectors, including industrial, service, and agricultural sectors. However, the existing FGVC methods do not allow simultaneous capturing of local and global image features, leading to unsatisfactory results. To solve FGVC problems, this study proposes a wavelet channel attention network (WCANet). A WCANet improves multiscale feature extraction within channel attention modules by combining global average pooling (GAP) which captures global features—with wavelet transform (WT), which captures local features. Intelligent identification of citrus varieties is urgently required for the differentiated management of different varieties of citrus trees planted in a smart citrus orchard. However, no public dataset is currently available for use in the fine-grained recognition of citrus varieties for intelligent management of citrus orchards. Therefore, we developed a dataset, named citrus variety dataset (CVD), based on the canopy images of eight common citrus varieties. Experimental results show that when a WCANet is added to a 50-layer residual network (ResNet-50) and a 101-layer residual network (ResNet-101), the citrus variety identification accuracies of these two models are 96.67 % and 96.83 %, respectively, which are better than the corresponding accuracies of other channel attention modules with the same settings. Finally, by adding a WCANet to ResNet-50 and pretraining on ImageNet, a citrus variety identification accuracy of 99.10 % was achieved. In this study, we provide a performance enhancement solution for the expert systems used in the identification of citrus varieties and agricultural products.
{"title":"Fine-grained recognition of citrus varieties via wavelet channel attention network","authors":"Fukai Zhang , Xiaobo Jin , Jie Jiang , Gang Lin , Mingzhi Wang , Shan An , Qiang Lyu","doi":"10.1016/j.knosys.2025.113128","DOIUrl":"10.1016/j.knosys.2025.113128","url":null,"abstract":"<div><div>Fine-grained visual classification (FGVC) has numerous applications in various sectors, including industrial, service, and agricultural sectors. However, the existing FGVC methods do not allow simultaneous capturing of local and global image features, leading to unsatisfactory results. To solve FGVC problems, this study proposes a wavelet channel attention network (WCANet). A WCANet improves multiscale feature extraction within channel attention modules by combining global average pooling (GAP) which captures global features—with wavelet transform (WT), which captures local features. Intelligent identification of citrus varieties is urgently required for the differentiated management of different varieties of citrus trees planted in a smart citrus orchard. However, no public dataset is currently available for use in the fine-grained recognition of citrus varieties for intelligent management of citrus orchards. Therefore, we developed a dataset, named citrus variety dataset (CVD), based on the canopy images of eight common citrus varieties. Experimental results show that when a WCANet is added to a 50-layer residual network (ResNet-50) and a 101-layer residual network (ResNet-101), the citrus variety identification accuracies of these two models are 96.67 % and 96.83 %, respectively, which are better than the corresponding accuracies of other channel attention modules with the same settings. Finally, by adding a WCANet to ResNet-50 and pretraining on ImageNet, a citrus variety identification accuracy of 99.10 % was achieved. In this study, we provide a performance enhancement solution for the expert systems used in the identification of citrus varieties and agricultural products.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"311 ","pages":"Article 113128"},"PeriodicalIF":7.2,"publicationDate":"2025-02-02","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143377685","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 : 2025-02-02DOI: 10.1016/j.knosys.2025.113038
Anima Pramanik , Soumick Sarker , Sobhan Sarkar , Sankar K. Pal
In this study, a new deep learning-based model, namely, transfer learning-based granulated Bi-LSTM (TLG-LSTM) is developed for fall detection on roads. The TLG-LSTM can handle the uncertainty issue arising between various ‘Fall’ and ‘No Fall’ events in complex scenarios concerning both indoor (i.e., home) and outdoor (i.e., road) areas. The TLG-LSTM consists of four phases: (i) object detection and tracking, (ii) MoveNet-Lightening and object-level feature(s) computation for all the detected objects, (iii) granule formation using these features, and (iv) temporal self attention mechanism-based Bi-LSTM for granule classification as ‘Fall’ or ‘No Fall’. Unlike state-of-the-art models, TLG-LSTM uses both MoveNet-Lightning and object-level features, enabling better modeling of both indoor and outdoor falls. For each detected object, two MoveNet-Lightning features, namely head-hip distance and hip-ankle distance are defined and used for obtaining a granule, namely pose granule. Whereas, three object-level features, namely change in aspect ratio, speed variation, and change in area are used for obtaining another granule, namely object granule. The commonality between these two granules represents the approximate regions concerning fall scenarios. Instead of the entire frame, these common granules are fed to the Bi-LSTM network for fall classification, thereby increasing speed as well as accuracy. Moreover, temporal self attention mechanism-based transfer learning is used to re-train the Bi-LSTM network, enhances the training speed and accuracy. Characteristics of TLG-LSTM are demonstrated over several real-time traffic videos acquired from ‘YouTube8M’. The superiority of the developed TLG-LSTM is also claimed over several state-of-the-art models.
{"title":"Real-time fall detection on roads using transfer learning-based granulated Bi-LSTM","authors":"Anima Pramanik , Soumick Sarker , Sobhan Sarkar , Sankar K. Pal","doi":"10.1016/j.knosys.2025.113038","DOIUrl":"10.1016/j.knosys.2025.113038","url":null,"abstract":"<div><div>In this study, a new deep learning-based model, namely, transfer learning-based granulated Bi-LSTM (TLG-LSTM) is developed for fall detection on roads. The TLG-LSTM can handle the uncertainty issue arising between various ‘Fall’ and ‘No Fall’ events in complex scenarios concerning both indoor (i.e., home) and outdoor (i.e., road) areas. The TLG-LSTM consists of four phases: (i) object detection and tracking, (ii) MoveNet-Lightening and object-level feature(s) computation for all the detected objects, (iii) granule formation using these features, and (iv) temporal self attention mechanism-based Bi-LSTM for granule classification as ‘Fall’ or ‘No Fall’. Unlike state-of-the-art models, TLG-LSTM uses both MoveNet-Lightning and object-level features, enabling better modeling of both indoor and outdoor falls. For each detected object, two MoveNet-Lightning features, namely head-hip distance and hip-ankle distance are defined and used for obtaining a granule, namely pose granule. Whereas, three object-level features, namely change in aspect ratio, speed variation, and change in area are used for obtaining another granule, namely object granule. The commonality between these two granules represents the approximate regions concerning fall scenarios. Instead of the entire frame, these common granules are fed to the Bi-LSTM network for fall classification, thereby increasing speed as well as accuracy. Moreover, temporal self attention mechanism-based transfer learning is used to re-train the Bi-LSTM network, enhances the training speed and accuracy. Characteristics of TLG-LSTM are demonstrated over several real-time traffic videos acquired from ‘YouTube8M’. The superiority of the developed TLG-LSTM is also claimed over several state-of-the-art models.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"311 ","pages":"Article 113038"},"PeriodicalIF":7.2,"publicationDate":"2025-02-02","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143340025","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 : 2025-02-01DOI: 10.1016/j.knosys.2025.113078
Wenqiang Hua , Nan Sun , Lin Liu , Chen Ding , YiZhuo Dong , Wei Sun
Deep learning methods have been widely applied in polarimetric synthetic aperture radar (PolSAR) image classification. However, these methods require abundant labeled data to achieve satisfactory performance, and it requires considerable effort to obtain a good deal of labeled data in practice, which requires huge human and resources. To tackle these issues, a new semi-supervised hybrid contrastive learning (SSHCL) is proposed for PolSAR image classification of finite labeled data. First, a hybrid contrastive learning model is constructed to extract feature information using only a handful of labeled data and unlabeled data. Second, combined with the idea of semi-supervised learning, a semi-supervised samples selection method is proposed. The pseudo-labeled samples are selected from unlabeled samples by combining superpixel and contrastive similarity, and are included in the label dataset. Then, the extended labeled samples are used to train the network for a new round. Finally, the optimized SSHCL network is used to classify the pixels of the whole image. The experimental outcomes conducted on three real PolSAR datasets demonstrate that the superior performance of the proposed compared to other existing methods under finite labeled data.
{"title":"Semi-supervised hybrid contrastive learning for PolSAR image classification","authors":"Wenqiang Hua , Nan Sun , Lin Liu , Chen Ding , YiZhuo Dong , Wei Sun","doi":"10.1016/j.knosys.2025.113078","DOIUrl":"10.1016/j.knosys.2025.113078","url":null,"abstract":"<div><div>Deep learning methods have been widely applied in polarimetric synthetic aperture radar (PolSAR) image classification. However, these methods require abundant labeled data to achieve satisfactory performance, and it requires considerable effort to obtain a good deal of labeled data in practice, which requires huge human and resources. To tackle these issues, a new semi-supervised hybrid contrastive learning (SSHCL) is proposed for PolSAR image classification of finite labeled data. First, a hybrid contrastive learning model is constructed to extract feature information using only a handful of labeled data and unlabeled data. Second, combined with the idea of semi-supervised learning, a semi-supervised samples selection method is proposed. The pseudo-labeled samples are selected from unlabeled samples by combining superpixel and contrastive similarity, and are included in the label dataset. Then, the extended labeled samples are used to train the network for a new round. Finally, the optimized SSHCL network is used to classify the pixels of the whole image. The experimental outcomes conducted on three real PolSAR datasets demonstrate that the superior performance of the proposed compared to other existing methods under finite labeled data.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"311 ","pages":"Article 113078"},"PeriodicalIF":7.2,"publicationDate":"2025-02-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143340027","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 : 2025-02-01DOI: 10.1016/j.knosys.2025.113089
Alimul Rajee , Md. Shahriare Satu , Mohammad Zoynul Abedin , K.M. Akkas Ali , Saad Aloteibi , Mohammad Ali Moni
Traffic accidents are unexpected incidents where one or multiple vehicles collide and damage properties, dying or injuring many individuals. It causes significant social burdens, including loss of life, serious injuries, and economic suppression from medical costs, property damages, and productivity losses. This kind of incident brings a miserable situation for the affected people. Many factors, including infrastructure, weather, vehicles, or driver-related issues, contribute to happening traffic accidents. This work explores an innovative approach by investigating contributing factors to ensure road safety. In this study, an ensemble machine learning model, namely Weighted Fusion-Based Feature Selection (WFFS), was proposed to identify different significant features to reduce the effects of traffic accidents. A large amount of traffic accident records from the United Kingdom (UK) were gathered and split into several folds, which were cleaned and balanced using different techniques such as removing percentages, Synthetic Minority Oversampling Technique (SMOTE), and random oversampling. Then, WFFS were employed in each fold and identified the most significant features to predict traffic accident severity more accurately. Different classifiers, such as tree-based, bagging, boosting, and voting classifiers, were implemented into WFFS-generated feature subsets and performed better than primary data and other feature subsets. In this case, the random tree-based bagging method provided the highest accuracy of 97.28% to predict accident severity for the WFFS subset, where its number of features is 18. However, different classifiers achieved better accuracies for 6 out of 11 times using WFFS. This method is highly recommended for policymakers and transportation engineers to identify potentially hazardous locations and take appropriate measures to diminish the effects of traffic accidents.
{"title":"WFFS—An ensemble feature selection algorithm for heterogeneous traffic accident data analysis","authors":"Alimul Rajee , Md. Shahriare Satu , Mohammad Zoynul Abedin , K.M. Akkas Ali , Saad Aloteibi , Mohammad Ali Moni","doi":"10.1016/j.knosys.2025.113089","DOIUrl":"10.1016/j.knosys.2025.113089","url":null,"abstract":"<div><div>Traffic accidents are unexpected incidents where one or multiple vehicles collide and damage properties, dying or injuring many individuals. It causes significant social burdens, including loss of life, serious injuries, and economic suppression from medical costs, property damages, and productivity losses. This kind of incident brings a miserable situation for the affected people. Many factors, including infrastructure, weather, vehicles, or driver-related issues, contribute to happening traffic accidents. This work explores an innovative approach by investigating contributing factors to ensure road safety. In this study, an ensemble machine learning model, namely Weighted Fusion-Based Feature Selection (WFFS), was proposed to identify different significant features to reduce the effects of traffic accidents. A large amount of traffic accident records from the United Kingdom (UK) were gathered and split into several folds, which were cleaned and balanced using different techniques such as removing percentages, Synthetic Minority Oversampling Technique (SMOTE), and random oversampling. Then, WFFS were employed in each fold and identified the most significant features to predict traffic accident severity more accurately. Different classifiers, such as tree-based, bagging, boosting, and voting classifiers, were implemented into WFFS-generated feature subsets and performed better than primary data and other feature subsets. In this case, the random tree-based bagging method provided the highest accuracy of 97.28% to predict accident severity for the WFFS subset, where its number of features is 18. However, different classifiers achieved better accuracies for 6 out of 11 times using WFFS. This method is highly recommended for policymakers and transportation engineers to identify potentially hazardous locations and take appropriate measures to diminish the effects of traffic accidents.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"311 ","pages":"Article 113089"},"PeriodicalIF":7.2,"publicationDate":"2025-02-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143340017","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-02-01DOI: 10.1016/j.knosys.2025.113120
Yuefei Wang , Yutong Zhang , Li Zhang , Yuxuan Wan , Zhixuan Chen , Yuquan Xu , Ruixin Cao , Liangyan Zhao , Yixi Yang , Xi Yu
Semantic segmentation is of great significance in the medical field, as it can help doctors intelligently, quickly, and accurately locate key lesion areas, providing crucial support for the diagnosis, treatment, and recovery of patients. The primary reason existing segmentation networks encounter accuracy bottlenecks is that the size of lesion images limits the network's attention to information, while also causing a lack of image information transmission between encoding and decoding. This study proposes a Multi-branch Partition Feature Enhancement Network (MPFE Net), which is essentially based on our proposed network system, the Multi-Branch Partition-Guided Decoding Network (MPD Net), using an image partition strategy. This approach divides the image at the encoding end and passes it to the decoding end, enabling parallel decoding with branches based on partitions to alleviate the problem of insufficient image information transmission. In terms of module design and optimization, under the processing idea of image partitioning, we have constructed the Bottleneck module Multi-semantic Progressive Interaction Guider (MPIG) for semantic progressive fusion. Furthermore, we have enhanced the ability to extract local information in the Vision Transformer, constructing the Multi-branched Feature Enhancement with Shared ViT (MFES ViT) to enhance control over image details. In the experiments, MPFE Net was compared with 20 models on 8 medical datasets for metric exploration, image comparison, process verification, and statistical analysis. We also discussed in detail 4 key issues. The results show that MPFE Net has better lesion universality and segmentation superiority.
{"title":"A feature enhancement network based on image partitioning in a multi-branch encoder-decoder architecture","authors":"Yuefei Wang , Yutong Zhang , Li Zhang , Yuxuan Wan , Zhixuan Chen , Yuquan Xu , Ruixin Cao , Liangyan Zhao , Yixi Yang , Xi Yu","doi":"10.1016/j.knosys.2025.113120","DOIUrl":"10.1016/j.knosys.2025.113120","url":null,"abstract":"<div><div>Semantic segmentation is of great significance in the medical field, as it can help doctors intelligently, quickly, and accurately locate key lesion areas, providing crucial support for the diagnosis, treatment, and recovery of patients. The primary reason existing segmentation networks encounter accuracy bottlenecks is that the size of lesion images limits the network's attention to information, while also causing a lack of image information transmission between encoding and decoding. This study proposes a Multi-branch Partition Feature Enhancement Network (MPFE Net), which is essentially based on our proposed network system, the Multi-Branch Partition-Guided Decoding Network (MPD Net), using an image partition strategy. This approach divides the image at the encoding end and passes it to the decoding end, enabling parallel decoding with branches based on partitions to alleviate the problem of insufficient image information transmission. In terms of module design and optimization, under the processing idea of image partitioning, we have constructed the Bottleneck module Multi-semantic Progressive Interaction Guider (MPIG) for semantic progressive fusion. Furthermore, we have enhanced the ability to extract local information in the Vision Transformer, constructing the Multi-branched Feature Enhancement with Shared ViT (MFES ViT) to enhance control over image details. In the experiments, MPFE Net was compared with 20 models on 8 medical datasets for metric exploration, image comparison, process verification, and statistical analysis. We also discussed in detail 4 key issues. The results show that MPFE Net has better lesion universality and segmentation superiority.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"311 ","pages":"Article 113120"},"PeriodicalIF":7.2,"publicationDate":"2025-02-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143349000","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 : 2025-02-01DOI: 10.1016/j.knosys.2025.113064
Damian Kusnik, Bogdan Smolka
Denoising remains one of the most crucial research areas within image processing given its effect on later analysis. During the different steps of image acquisition, transmission, and storage, noise considerably deteriorates image quality. Very poor image quality may prevent a vision system in some limiting situations from performing properly. In natural images, Gaussian and impulsive noise is probably the most frequent kind of noise and will be tackled in this paper. Various solutions to this problem are given in the contemporary literature, but all these techniques can be further improved for even better results, hence bringing better outcomes in further stages of image processing.
In this paper, we will present a Robust Adaptive Denoising technique (RAD) that makes use of elements of the Non-Local Means mechanism in combination with a new measure of the similarity of image patches, taking into account the impulsiveness component of image noise. The proposed methodology introduces an adaptive selection technique that relieves potential users from the complex process of parameter tuning to get the best results. We have also analyzed problems connected with the size of the patch and processing block, and also with some local weighting aspects.
Extensive experimentation demonstrates that the proposed approach outperforms the performance of existing filters and yields results superior to those obtained by recent neural network-based techniques. The proposed solution can be utilized without the cumbersome network training process, is independent of any a priori knowledge of the mixed noise level and image characteristics.
{"title":"Robust Adaptive Denoising of color images with mixed Gaussian and impulsive noise","authors":"Damian Kusnik, Bogdan Smolka","doi":"10.1016/j.knosys.2025.113064","DOIUrl":"10.1016/j.knosys.2025.113064","url":null,"abstract":"<div><div>Denoising remains one of the most crucial research areas within image processing given its effect on later analysis. During the different steps of image acquisition, transmission, and storage, noise considerably deteriorates image quality. Very poor image quality may prevent a vision system in some limiting situations from performing properly. In natural images, Gaussian and impulsive noise is probably the most frequent kind of noise and will be tackled in this paper. Various solutions to this problem are given in the contemporary literature, but all these techniques can be further improved for even better results, hence bringing better outcomes in further stages of image processing.</div><div>In this paper, we will present a Robust Adaptive Denoising technique (RAD) that makes use of elements of the Non-Local Means mechanism in combination with a new measure of the similarity of image patches, taking into account the impulsiveness component of image noise. The proposed methodology introduces an adaptive selection technique that relieves potential users from the complex process of parameter tuning to get the best results. We have also analyzed problems connected with the size of the patch and processing block, and also with some local weighting aspects.</div><div>Extensive experimentation demonstrates that the proposed approach outperforms the performance of existing filters and yields results superior to those obtained by recent neural network-based techniques. The proposed solution can be utilized without the cumbersome network training process, is independent of any a priori knowledge of the mixed noise level and image characteristics.</div></div>","PeriodicalId":49939,"journal":{"name":"Knowledge-Based Systems","volume":"311 ","pages":"Article 113064"},"PeriodicalIF":7.2,"publicationDate":"2025-02-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"143377618","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