Kun Zeng
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Kun Zeng has a total of 67 co-authors. Here are the top co-authors by number of shared publications:
- Zuoyong Li
Minjiang University
14 shared publications - Jinsheng Fang
School of Computer Science, Minnan Normal University, Zhangzhou, China
6 shared publications - Rong Hu
School of Computer Science and Mathematics, Fujian University of Technology, Fuzhou, China
4 shared publications - Haoyi Fan
School of Computer and Artificial Intelligence, Zhengzhou University, Zhengzhou 450001, China
3 shared publications - Qinghua Lin
College of Biomedical Engineering, Fudan University, Shanghai, 200433, China
3 shared publications - Zhiqiang Yan
Department of computer technology, Guilin University of Electronic Technology, Guilin, China
3 shared publications
Latest Publications
- 9 September 2026
- Published byĀ WileyĀ inĀ Concurrency and Computation: Practice and Experience
- Vol.Ā 38 Ā (18) , e70938
VGCAD: Visual GeometryāGuided Circulant Attention Distillation for Industrial Anomaly Detection
Industrial visual anomaly detection aims to identify and localize unknown defects using only normal samples for training. Efficient teacherāstudent methods reduce inference cost by avoiding largeāscale featureāmemory retrieval, but their compact reconstruction paths often provide limited modeling of spatial relations among visual tokens. This limitation is critical for industrial inspection, where defects commonly appear as subtle disruptions of regular textures, repeated layouts, or local geometric structures. To address this problem, this paper proposes Visual GeometryāGuided Circulant Attention Distillation (VGCAD), an efficient visual anomaly detection model. VGCAD introduces a sineāpositionāenhanced circulant attention module to strengthen global tokenācorrelation modeling with explicit twoādimensional spatial information, enabling the student decoder to reconstruct teacher features from structureāaware token representations. In addition, VGCAD employs a highālevel geoāsemantic manifold consistency constraint to overcome the limitation of pointwise cosine or mean squared error (MSE) distances in measuring abstract semantic features. Experiments on MVTec AD, VisA, and MPDD show that VGCAD improves imageālevel anomaly detection and pixelālevel localization while maintaining fast inference comparable to efficient baseline models. Extensive ablation and visualization results further verify the effectiveness of the proposed circulant attention module and geoāsemantic manifold constraint. The source code of the proposed method is publicly available at https://github.com/CreatedTRYNA/VGCAD .
- 9 September 2026
- Published byĀ WileyĀ inĀ Concurrency and Computation: Practice and Experience
- Vol.Ā 38 Ā (18) , e70938
VGCAD: Visual GeometryāGuided Circulant Attention Distillation for Industrial Anomaly Detection
Industrial visual anomaly detection aims to identify and localize unknown defects using only normal samples for training. Efficient teacherāstudent methods reduce inference cost by avoiding largeāscale featureāmemory retrieval, but their compact reconstruction paths often provide limited modeling of spatial relations among visual tokens. This limitation is critical for industrial inspection, where defects commonly appear as subtle disruptions of regular textures, repeated layouts, or local geometric structures. To address this problem, this paper proposes Visual GeometryāGuided Circulant Attention Distillation (VGCAD), an efficient visual anomaly detection model. VGCAD introduces a sineāpositionāenhanced circulant attention module to strengthen global tokenācorrelation modeling with explicit twoādimensional spatial information, enabling the student decoder to reconstruct teacher features from structureāaware token representations. In addition, VGCAD employs a highālevel geoāsemantic manifold consistency constraint to overcome the limitation of pointwise cosine or mean squared error (MSE) distances in measuring abstract semantic features. Experiments on MVTec AD, VisA, and MPDD show that VGCAD improves imageālevel anomaly detection and pixelālevel localization while maintaining fast inference comparable to efficient baseline models. Extensive ablation and visualization results further verify the effectiveness of the proposed circulant attention module and geoāsemantic manifold constraint. The source code of the proposed method is publicly available at https://github.com/CreatedTRYNA/VGCAD .
- 9 January 2026
- Published byĀ Springer NatureĀ inĀ Pattern Analysis and Applications
- Vol.Ā 29 Ā (1) , 24
A high-frequency information guiding attention network for super-lightweight image super-resolution
- 9 January 2026
- Published byĀ Springer NatureĀ inĀ Pattern Analysis and Applications
- Vol.Ā 29 Ā (1) , 24
A high-frequency information guiding attention network for super-lightweight image super-resolution
- 1 January 2026
- Published byĀ Institute of Electrical and Electronics Engineers (IEEE)Ā inĀ IEEE Transactions on Instrumentation and Measurement
- Vol.Ā 75, 4010111
Decontaminating Anomaly Detection: A Geometric Adversarial Approach With Stochastic Bottlenecks
- 1 January 2026
- Published byĀ Institute of Electrical and Electronics Engineers (IEEE)Ā inĀ IEEE Transactions on Instrumentation and Measurement
- Vol.Ā 75, 4010111
Decontaminating Anomaly Detection: A Geometric Adversarial Approach With Stochastic Bottlenecks
- 13 November 2025
- Published byĀ ElsevierĀ inĀ Biomedical Signal Processing and Control
- Vol.Ā 113
SwinGhostU-Net: A 3D U-shaped network combined with swin transformer and ghost convolution for brain tumor segmentation
- 13 November 2025
- Published byĀ ElsevierĀ inĀ Biomedical Signal Processing and Control
- Vol.Ā 113
SwinGhostU-Net: A 3D U-shaped network combined with swin transformer and ghost convolution for brain tumor segmentation
- 28 October 2025
- Published byĀ ElsevierĀ inĀ Pattern Recognition
- Vol.Ā 172
DSCIL: Dynamic selected contrastive instance learning for weakly supervised video anomaly detection
- 28 October 2025
- Published byĀ ElsevierĀ inĀ Pattern Recognition
- Vol.Ā 172
DSCIL: Dynamic selected contrastive instance learning for weakly supervised video anomaly detection
- Open Access
- 30 June 2025
- Published byĀ Springer NatureĀ inĀ Research Square
Reliable Attention Based Stereo Image Super-Resolution Network
Stereo image super-resolution (StereoSR) exploits complementary information between paired views to enhance image reconstruction quality. Existing attention-based methods for cross-view feature interaction overlook a fundamental principle: effective stereo reconstruction requires compensating insufficient features in one view with reliable features from its counterpart. This observation raises a critical question of quantifying and leveraging feature reliability. To address this, we propose a Reliable Attention Based Stereo Image Super-resolution Network (RASSR), which explicitly models and utilizes feature uncertainty for guided cross-view interaction. The core component of RASSR is the Reliable Stereo Cross Attention Module (RSCAM), which dynamically identifies high-uncertainty regions in one view and adaptively selects complementary features from low-uncertainty regions in the alternative view. To support this uncertainty-driven interaction, we propose two key components: a Monte Carlo Feature Extraction Block (MCFEBlock) that quantifies feature uncertainty through Monte Carlo sampling, and a Feature Compensation Module (FCM) that mitigates information loss during the sampling process. Extensive experiments on the KITTI 2012, KITTI 2015, Middlebury, and Flickr1024 datasets demonstrate that, by enhancing attention reliability, RASSR achieves state-of-the-art performance, outperforming existing methods in both PSNR and SSIM metrics.
- 30 June 2025
- Published byĀ Springer NatureĀ inĀ Research Square
Open AccessReliable Attention Based Stereo Image Super-Resolution Network
Stereo image super-resolution (StereoSR) exploits complementary information between paired views to enhance image reconstruction quality. Existing attention-based methods for cross-view feature interaction overlook a fundamental principle: effective stereo reconstruction requires compensating insufficient features in one view with reliable features from its counterpart. This observation raises a critical question of quantifying and leveraging feature reliability. To address this, we propose a Reliable Attention Based Stereo Image Super-resolution Network (RASSR), which explicitly models and utilizes feature uncertainty for guided cross-view interaction. The core component of RASSR is the Reliable Stereo Cross Attention Module (RSCAM), which dynamically identifies high-uncertainty regions in one view and adaptively selects complementary features from low-uncertainty regions in the alternative view. To support this uncertainty-driven interaction, we propose two key components: a Monte Carlo Feature Extraction Block (MCFEBlock) that quantifies feature uncertainty through Monte Carlo sampling, and a Feature Compensation Module (FCM) that mitigates information loss during the sampling process. Extensive experiments on the KITTI 2012, KITTI 2015, Middlebury, and Flickr1024 datasets demonstrate that, by enhancing attention reliability, RASSR achieves state-of-the-art performance, outperforming existing methods in both PSNR and SSIM metrics.
- Open Access
- 6 June 2025
- Published byĀ Association for Computing Machinery (ACM)
- p. 217-223
Combining Variable Convolution and Transformer Network for Image Super-Resolution
- 6 June 2025
- Published byĀ Association for Computing Machinery (ACM)
- p. 217-223
Open AccessCombining Variable Convolution and Transformer Network for Image Super-Resolution
- 2 May 2025
- Published byĀ WileyĀ inĀ Concurrency and Computation: Practice and Experience
- Vol.Ā 37 Ā (12-14)
MultiāModal Anomalous Driving Behavior Detection With Adaptive Masking
Intelligent Transportation Systems are tasked with enhancing road safety, a crucial challenge given that approximately 1.35 million fatalities occur globally each year, with 15%ā27% of these deaths attributed to Anomalous Driving Behaviors (ADBs). Detecting these behaviors in real time is vital for preventing accidents and improving traffic safety. However, the complexity of driving environments, characterized by diverse scenarios, drivers, and vehicle conditions, makes ADB detection a challenging task. This article proposes a novel approach for ADB detection, leveraging the advantages of multimodal data, adaptive masking, and multihead selfāattention mechanisms. The proposed method first employs an adaptive masking technique based on the Softmax function to sparsify input features, effectively reducing the influence of irrelevant information. By focusing on key features, the model becomes more resilient to noise, such as background clutter or irrelevant driver actions, which might otherwise interfere with the detection of abnormal behaviors. To further enhance feature integration across different data modalities (e.g., visual, infrared, and depth data), a multihead selfāattention mechanism is incorporated. This mechanism enables the model to prioritize important information from various sensor inputs, fostering more effective multimodal fusion and better decisionāmaking for behavior classification. In addition, a supervised contrastive learning strategy is utilized to mitigate memory usage, a common challenge in realātime systems where computational resources are limited. This approach ensures efficient learning by emphasizing the distinction between normal and abnormal behaviors while minimizing the memory footprint of the model. Extensive experiments on two benchmark datasets, 3MDAD and DAD, demonstrate the proposed method's superior performance in detecting ADBs. The results indicate a significant improvement in detection accuracy and robustness, highlighting the potential of this approach for deployment in realāworld Intelligent Transportation Systems aimed at enhancing road safety. This research provides a promising step forward in the development of more effective and scalable solutions for ADB detection, offering a foundation for future advancements in traffic safety technologies.
- 2 May 2025
- Published byĀ WileyĀ inĀ Concurrency and Computation: Practice and Experience
- Vol.Ā 37 Ā (12-14)
MultiāModal Anomalous Driving Behavior Detection With Adaptive Masking
Intelligent Transportation Systems are tasked with enhancing road safety, a crucial challenge given that approximately 1.35 million fatalities occur globally each year, with 15%ā27% of these deaths attributed to Anomalous Driving Behaviors (ADBs). Detecting these behaviors in real time is vital for preventing accidents and improving traffic safety. However, the complexity of driving environments, characterized by diverse scenarios, drivers, and vehicle conditions, makes ADB detection a challenging task. This article proposes a novel approach for ADB detection, leveraging the advantages of multimodal data, adaptive masking, and multihead selfāattention mechanisms. The proposed method first employs an adaptive masking technique based on the Softmax function to sparsify input features, effectively reducing the influence of irrelevant information. By focusing on key features, the model becomes more resilient to noise, such as background clutter or irrelevant driver actions, which might otherwise interfere with the detection of abnormal behaviors. To further enhance feature integration across different data modalities (e.g., visual, infrared, and depth data), a multihead selfāattention mechanism is incorporated. This mechanism enables the model to prioritize important information from various sensor inputs, fostering more effective multimodal fusion and better decisionāmaking for behavior classification. In addition, a supervised contrastive learning strategy is utilized to mitigate memory usage, a common challenge in realātime systems where computational resources are limited. This approach ensures efficient learning by emphasizing the distinction between normal and abnormal behaviors while minimizing the memory footprint of the model. Extensive experiments on two benchmark datasets, 3MDAD and DAD, demonstrate the proposed method's superior performance in detecting ADBs. The results indicate a significant improvement in detection accuracy and robustness, highlighting the potential of this approach for deployment in realāworld Intelligent Transportation Systems aimed at enhancing road safety. This research provides a promising step forward in the development of more effective and scalable solutions for ADB detection, offering a foundation for future advancements in traffic safety technologies.
- 25 February 2025
- Published byĀ Institute of Electrical and Electronics Engineers (IEEE)Ā inĀ IEEE Transactions on Instrumentation and Measurement
- Vol.Ā 74 Ā (00189456) , 1-13
KAC-Unet: A Medical Image Segmentation With the Adaptive Group Strategy and Kolmogorov-Arnold Network
In the field of deep learning-based medical image segmentation, convolutional neural networks extract image features by combining linear convolutional layers with nonlinear activation functions. However, excessive stacking of linear layers in the network limits the modelās ability to capture fine-grained details. In addition, the feature distribution imbalance caused by the traditional fixed grouping strategy can affect the deep modelās capacity to perceive the overall structure of the image. To address these challenges, we propose a medical image segmentation framework, called KAC-Unet. Firstly, we propose the Adaptive Group Strategy to balance the grouping of different input channels, alleviating the performance degradation caused by differences in group information. Then, we propose the Shift Tokenized Kolmogorov-Arnold Network Block to capture complex features in medical images through flexible nonlinear transformations and shift operations. Extensive experiments are conducted on three medical image segmentation datasets. The results demonstrate the effectiveness and superiority of our proposed method compared with state-of-the-art algorithms.
- 25 February 2025
- Published byĀ Institute of Electrical and Electronics Engineers (IEEE)Ā inĀ IEEE Transactions on Instrumentation and Measurement
- Vol.Ā 74 Ā (00189456) , 1-13
KAC-Unet: A Medical Image Segmentation With the Adaptive Group Strategy and Kolmogorov-Arnold Network
In the field of deep learning-based medical image segmentation, convolutional neural networks extract image features by combining linear convolutional layers with nonlinear activation functions. However, excessive stacking of linear layers in the network limits the modelās ability to capture fine-grained details. In addition, the feature distribution imbalance caused by the traditional fixed grouping strategy can affect the deep modelās capacity to perceive the overall structure of the image. To address these challenges, we propose a medical image segmentation framework, called KAC-Unet. Firstly, we propose the Adaptive Group Strategy to balance the grouping of different input channels, alleviating the performance degradation caused by differences in group information. Then, we propose the Shift Tokenized Kolmogorov-Arnold Network Block to capture complex features in medical images through flexible nonlinear transformations and shift operations. Extensive experiments are conducted on three medical image segmentation datasets. The results demonstrate the effectiveness and superiority of our proposed method compared with state-of-the-art algorithms.
- 15 February 2025
- Published byĀ ElsevierĀ inĀ Expert Systems with Applications
- Vol.Ā 273
Beyond-Skeleton: Zero-shot Skeleton Action Recognition enhanced by supplementary RGB visual information
- 15 February 2025
- Published byĀ ElsevierĀ inĀ Expert Systems with Applications
- Vol.Ā 273
Beyond-Skeleton: Zero-shot Skeleton Action Recognition enhanced by supplementary RGB visual information