KWKaijun Wu

Kaijun Wu

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Kaijun Wu has a total of 73 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Cluster Resource Load Prediction Method Based on Temporal-Feature Attention and Dynamic Stacking

    by Qiaoyan Zhang, Kaijun Wu, Chenshuai Bai

    Accurate prediction of computing-resource workloads is important for capacity planning, overload warning, and intelligent system management. Large-scale computing systems exhibit complex temporal fluctuations, sudden variations, and multi-resource coupling characteristics, making accurate workload prediction challenging. To address these challenges, this paper proposes a Hybrid Temporal-Feature Attention enhanced Multi-task Stacking model (HTAM-Stack) for multivariate cluster workload forecasting. First, a Temporal-Feature Hybrid Attention (TFHA) module is designed to jointly capture temporal dependencies and cross-resource feature interactions, enabling adaptive extraction of critical temporal patterns and important resource characteristics. Second, a Multi-Task Learning (MTL) framework is introduced to simultaneously predict CPU and Memory workloads by exploiting the correlations among heterogeneous resource variables. Furthermore, a Dynamic Stacking (DS) mechanism is developed to adaptively adjust the contributions of heterogeneous base learners through a weight generation network, and a Residual Corrector (RC) is incorporated to further enhance prediction robustness. Extensive experiments conducted on two widely used public cluster workload datasets, including Google Cluster Trace and Alibaba Cluster Trace, demonstrate that HTAM-Stack achieves competitive prediction performance under complex and dynamic workload conditions. The proposed model achieves MAE values of 0.0012 and 0.0010, RMSE values of 0.0031 and 0.0027, MAPE values of 0.20% and 0.16%, and R2 values of 0.9715 and 0.9782 on the two datasets, respectively. Moreover, HTAM-Stack requires only 3.20 ms inference time with 8.60 M parameters, achieving a favorable balance between prediction accuracy and computational efficiency. The results validate the general effectiveness of the proposed framework on public cluster workload benchmarks rather than its direct applicability to railway IT systems. Because no representative railway workload dataset was available, railway IT is discussed only as a potential application context that requires domain-specific validation.

  • Open Access

    Dynamic Neural Architecture-Driven Spatial Semantic Modeling and Multiscale Collaborative Sensing of Remote Sensing Objects

    by Chenshuai Bai, Kaijun Wu, Xiaofeng Bai, Xiaoqiang Wu

    Aiming at the severe challenges to the accuracy and robustness of the detection model caused by the complexity problems such as high object density, dramatic scale changes, occlusion, variable illumination and class imbalance in remote sensing image object detection, this study innovatively proposes a multi scale adaptive remote sensing network (MARS-Net). Firstly, the spatial context network (SCNet) is constructed. The adaptive feature extraction of complex object morphology is realized by the conditional convolution module, which significantly improves the adaptability of the model to diverse objects in remote sensing images. At the same time, SCNet integrates the shift window Transformer module to effectively capture the long-distance dependencies between objects and enhance the model 's semantic understanding of complex scenes. In addition, an enhanced multi scale feature pyramid network (EMFPN) is designed. The core of EMFPN is the global-offset and detail aggregation module, which integrates the low-stage feature alignment module and the high-stage feature alignment module of the Gold module. The simple aggregation module is used to efficiently fuse multi-scale information from different feature layers, thereby improving the feature utilization efficiency. The low-stage information fusion module and the high-stage information fusion module useomni dimensional dynamic convolution to more comprehensively model the spatial information of the object and accurately process the direction and deformation of the object. The high-stage information fusion module also incorporates the exponential moving average attention mechanism to dynamically adjust the attention weight of features, further enhancing the discriminative ability of key features and the robustness of the overall model. The experimental results show that MARS-Net improves mAP@50 by 3.38% and 1.21% on RSOD and DOTAv2.0 datasets, and mAP@50:95 by 14.61% and 0.55%, respectively, which fully verifies its superior performance and practical application potential in complex remote sensing scenarios. The implementation code is publicly available at: https://github.com/bcs1107/MARS-Net.git.

  • Enhancing DeepLabv3+ With Multiscale Sparse-Aware Dilated Convolution and Boundary-Interactive ASPP for River Ice Segmentation

    by Dingju Zhou, Kaijun Wu, Junmei Hu, Qi Sun

    Accurate segmentation of river ice in remote sensing imagery remains challenging due to heterogeneous ice textures, fragmented floes, and ambiguous ice–water boundaries. While DeepLabv3+ effectively captures multi-scale context, its standard atrous convolution suffers from uniform sampling and gridding effects, which limit boundary discrimination in complex river ice scenes. To address these issues, we propose a lightweight enhancement to DeepLabv3+ that integrates regionally weighted dilated convolution, a coordinated dilation configuration strategy, and a boundary-aware refinement branch. The proposed design improves central feature aggregation while preserving global context, without introducing additional computational overhead. Experiments on the NWPU_YRCC2 dataset demonstrate a 1.95% mIoU improvement over the baseline DeepLabv3+, validating the effectiveness of the proposed framework for precise and efficient river ice segmentation.


  • Open Access

    High-Resolution Remote Sensing Image Scene Classification Using Star Kernel Feature Integration

    by Chenshuai Bai, Kaijun Wu, Xiaofeng Bai, Xiaoqiang Wu

    In the classification of high-resolution remote sensing images, intricate scenes along with variations in lighting and viewpoint pose significant challenges for current convolutional neural networks to accurately capture essential features, leading to classification inaccuracies. The accurate categorization of these images is crucial for land management and environmental oversight. Thus, enhancing the model's resilience to guarantee dependable classification outcomes is an urgent issue that needs addressing. To this end, we propose a star-based multi-scale feature fusion network (StarFuNet), including StarFuNet-l, StarFuNet-b, StarFuNet-p, and StarFuNet-m. The model performs preliminary feature extraction and transformation operations on the input data by superimposing a four-layer EBlock module, which lays the foundation for subsequent processing. The feature fusion attention module is used to enhance the feature expression ability and improve the feature fusion effect, so that the model can better capture key information. The spatial pyramid pooling fast cross-stage partial connection module is used to transform the feature mapping of any size into a standardized feature vector to ensure that the model output has a high degree of adaptability and flexibility to adapt to different scales and types of data. By utilizing global average pooling, the normalized feature vectors are refined to produce highly precise classification outcomes, offering a dependable foundation for decision-making across numerous applications. We assessed the StarFuNet model using five publicly remote sensing datasets, and the findings indicate that its classification performance markedly surpasses that of current leading algorithms.

  • Multimodal VAD: Visual Anomaly Detection in Intelligent Monitoring System via Audio-Vision-Language

    by Dicong Wang, Qilong Wang, Qinghua Hu, Kaijun Wu

    Deep learning-based anomaly detection methods using visual sensors generally rely on a single modality or variants as raw signal inputs, which severely limits expressiveness and adaptability. The evolution of multimodal and visual-language pre-trained models is shaping new possibilities in video anomaly detection (VAD). So, how to efficiently leverage them to achieve reliable multimodal VAD presents a significant challenge worth investigating. In this work, we propose a novel dual-stream multimodal video anomaly detection network, which integrates coarse-grained and fine-grained streams combining video, audio, and text modalities. First, in the coarse-grained stream, we perform cross-modal fusion of audio features with temporally modeled visual features, utilizing contrastive optimization to achieve more accurate coarse-grained results. In the fine-grained stream, we constructed abnormal-aware context prompts by integrating visual information and prior knowledge related to anomalous events into the text modality. Through the “coarse-support-fine” strategy, we further enhanced the model’s ability to discriminate fine-grained anomalies. Our method achieved optimal performance in experiments on two large-scale anomaly datasets, demonstrating its effectiveness and superiority. It supports the development of highly robust intelligent monitoring systems and promotes the potential applications of multimodal video anomaly detection across industrial monitoring, public safety, smart cities, etc.


  • Temporal-Spatial Decoupled Self-Supervised Multi-Task Learning for Video Anomaly Detection and Localization in Intelligent Transportation Surveillance Systems

    by Dicong Wang, Qilong Wang, Qinghua Hu, Kaijun Wu

    Video anomaly detection (VAD) has become a critical research focus in the realms of computer vision and intelligent transportation systems, particularly with the rise of generative AI and advancements in self-supervised learning. These developments open up new frontiers in understanding complex scenarios within intelligent transportation systems. Since previous methods, confined to a single paradigm, inadequately captured intricate feature information, which led to a bottleneck that constrained the performance; individuals form unbiased and logical evaluations of matters through an in-depth review of a multitude of factors and rationalizations. Decoupled multi-task learning not only overcomes the limitations of single-task paradigms but also meets the requirements for comprehensiveness and objectivity in distinguishing things. Therefore, we propose a self-supervised multi-task learning network with object-level spatiotemporal decoupling in this paper. We decouple the video anomaly detection task in both spatial and temporal dimensions to construct multiple proxy tasks, optimizing overall performance through joint learning. Specifically, we first utilize a pre-trained object detector to construct spatiotemporal context cubes. Then, on the decoupled dimensions, we design several related proxy tasks, adjusting the context cube structure to accommodate each proxy task, and providing reliable information for anomaly detection through joint training. In the end, our method achieved performance that surpasses existing methods on three traffic road datasets, namely UCSD Ped2, Avenue, and ShanghaiTech, with AUCs of 99.6%, 95.6%, and 87.7%, respectively. It demonstrates the potential of generative AI-powered anomaly detection systems in advancing intelligent transportation applications.


  • Dunhuang Art Style Transfer via Hierarchical Vision Transformer and Color Consistency Constraints

    by Yan Cao, Yi Zhang, Yingjie Lin, Kaijun Wu

    As a significant part of China’s cultural heritage, Dunhuang murals hold profound value for digital preservation and artistic inheritance, particularly in enriching cultural content for modern consumer electronics. Image Style Transfer technology provides a promising solution for modernizing the presentation of Dunhuang mural art. However, existing deep learning-based style transfer methods often encounter challenges, including semantic information loss and color inconsistencies, when applied to the unique artistic style of Dunhuang murals. To address these issues, we propose Dunhuang Art Style Transfer Using Hierarchical Vision Transformer (Dh-StyHT), a framework designed to meet the demands of promoting Dunhuang mural art in modern contexts. Dh-StyHT employs a window-based attention mechanism to effectively capture long-range dependencies and non-local content-style correlations. Moreover, a specialized color consistency mechanism is introduced to mitigate the color fading caused by mural oxidation, enabling precise restoration of the murals’ distinctive color features and artistic style. Experimental results, both qualitative and quantitative, demonstrate the effectiveness of Dh-StyHT in preserving content and achieving balanced color distribution. This framework opens new avenues for applying style transfer techniques in traditional art preservation and modern consumer electronics. The code is available at https://github.com/Leaderzhangyi/Dh-StyHT.


  • An Intelligent Optimization Clustering Algorithm for Driver Risk Tendency Recognition in Intelligent Transportation Systems

    by Ning Ma, Kaijun Wu, Yubin Yuan, Shuoxun Liu, Jiawei Li

    Driving behavior assessment is of great significance in intelligent transportation systems to improve road safety and traffic efficiency. This paper proposes a novel weighted fuzzy C-means (FCM) clustering algorithm, called GSA-WFCM, which combines a genetic algorithm and a simulated annealing algorithm to optimize the evaluation process of driving behavior. GSA-WFCM first uses a genetic algorithm to optimize the clustering center to ensure the diversity and effectiveness of the initial selection; then, it uses a simulated annealing algorithm to avoid local optimal solutions, thereby improving the global nature of the clustering results. In order to enhance the robustness and accuracy of clustering, the algorithm introduces a point density weighting mechanism to improve the algorithm’s adaptability to data features of different dimensions. By integrating the advantages of genetic algorithms and simulated annealing algorithms, the algorithm achieves intelligent initialization, optimizes the initial clustering center, and significantly improves the convergence speed and stability of the algorithm. Experimental results on multiple data sets show that GSA-WFCM effectively overcomes the shortcomings of traditional FCM in initial cluster center selection, showing faster convergence and stronger robustness. Further verification on traffic data sets shows that the algorithm performs well in identifying different types of driving behaviors and can effectively classify cautious driving, normal driving and aggressive driving. The accuracy and stability of clustering are significantly improved, and the clustering performance reaches 0.81, providing a reliable tool for driving behavior analysis in intelligent transportation systems.