Rui Zhai
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Rui Zhai has a total of 64 co-authors. Here are the top co-authors by number of shared publications:
- Junyang Yu
School of Software, Henan University, Kaifeng, China
6 shared publications - Ke Lu
School of Software, Henan University, Kaifeng, China
5 shared publications - Haozhe Jin
School of Software, Henan University, Kaifeng, China
3 shared publications Ke Lu
3 shared publications- Shichong Liu
School of Software, Henan University, Kaifeng, China
3 shared publications Yalin Song
3 shared publications
Latest Publications
- 10 November 2026
- Published by Springer Nature in Multimedia Systems
- Vol. 32 (11) , 700
Sure-detr: semantically unified and resolution-enhanced DETR for UAV object detection
- 10 November 2026
- Published by Springer Nature in Multimedia Systems
- Vol. 32 (11) , 700
Sure-detr: semantically unified and resolution-enhanced DETR for UAV object detection
- 10 October 2026
- Published by Springer Nature in Multimedia Systems
- Vol. 32 (10) , 624
Multi-domain collaborative feature fusion network for underwater object detection
- 10 October 2026
- Published by Springer Nature in Multimedia Systems
- Vol. 32 (10) , 624
Multi-domain collaborative feature fusion network for underwater object detection
- 11 July 2026
- Published by Springer Nature in Multimedia Systems
- Vol. 32 (6) , 426
Fedud: a multi unmanned aerial vehicle collaborative detection framework
- 11 July 2026
- Published by Springer Nature in Multimedia Systems
- Vol. 32 (6) , 426
Fedud: a multi unmanned aerial vehicle collaborative detection framework
- 24 June 2026
- Published by Elsevier in Expert Systems with Applications
- Vol. 331
AHGT: Aspect-guided heterogeneous graph transformer for multimodal sentiment analysis
- 24 June 2026
- Published by Elsevier in Expert Systems with Applications
- Vol. 331
AHGT: Aspect-guided heterogeneous graph transformer for multimodal sentiment analysis
- Open Access
- 12 June 2026
- Published by Elsevier in SSRN Electronic Journal
A2Det: Adaptive Cross-Modal Aggregation and Alignment for RGB-IR Object Detection
Visible and infrared modalities offer complementary cues for robust object detection under challenging imaging conditions. Yet in RGB-IR detection, complementarity is not a guarantee of reliability: degraded modality responses may contaminate fusion, heterogeneous features may drift apart in representation space, and fused predictions may suffer from semantic–geometric inconsistency. To address these issues, we propose a unified RGB-IR detection framework that organizes RGB-IR perception around reliability-aware fusion, representation alignment, and prediction-stage coordination. First, we design an Adaptive Cross-modal Aggregation Fusion module to estimate modality reliability and suppress degraded responses before aggregation. Second, we introduce a Cross-Modal Alignment Constraint to regularize paired RGB-IR features in a shared semantic space, reducing modality-dependent representation drift. Third, we develop a Cross-Interaction Decoupled Head to coordinate classification confidence and localization quality through task-specific normalization and bidirectional branch calibration. Extensive experiments on M3FD and DroneVehicle show improved detection accuracy, feature response quality, and confidence–localization consistency in the evaluated settings. The full model improves the YOLOv11n dual-modal baseline and obtains 89.2%/57.9% mAP50/mAP50:95 on M3FD and 83.5%/62.2% on DroneVehicle.
- 12 June 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessA2Det: Adaptive Cross-Modal Aggregation and Alignment for RGB-IR Object Detection
Visible and infrared modalities offer complementary cues for robust object detection under challenging imaging conditions. Yet in RGB-IR detection, complementarity is not a guarantee of reliability: degraded modality responses may contaminate fusion, heterogeneous features may drift apart in representation space, and fused predictions may suffer from semantic–geometric inconsistency. To address these issues, we propose a unified RGB-IR detection framework that organizes RGB-IR perception around reliability-aware fusion, representation alignment, and prediction-stage coordination. First, we design an Adaptive Cross-modal Aggregation Fusion module to estimate modality reliability and suppress degraded responses before aggregation. Second, we introduce a Cross-Modal Alignment Constraint to regularize paired RGB-IR features in a shared semantic space, reducing modality-dependent representation drift. Third, we develop a Cross-Interaction Decoupled Head to coordinate classification confidence and localization quality through task-specific normalization and bidirectional branch calibration. Extensive experiments on M3FD and DroneVehicle show improved detection accuracy, feature response quality, and confidence–localization consistency in the evaluated settings. The full model improves the YOLOv11n dual-modal baseline and obtains 89.2%/57.9% mAP50/mAP50:95 on M3FD and 83.5%/62.2% on DroneVehicle.
- 19 February 2026
- Published by Elsevier in Neurocomputing
- Vol. 678
SAM2-driven dual-teacher framework using hierarchical cross-slice context for semi-supervised 3D medical image segmentation
- 19 February 2026
- Published by Elsevier in Neurocomputing
- Vol. 678
SAM2-driven dual-teacher framework using hierarchical cross-slice context for semi-supervised 3D medical image segmentation
- 1 January 2026
- Published by Aerospace Information Research Institute, Chinese Academy of Sciences in Journal of Image and Graphics
- Vol. null (null) , 1-17
面向医学图像公平诊断的潜空间条件扩散去偏
目的医学图像诊断模型易受年龄、性别和种族等人口统计学属性影响,造成子群间诊断性能差异。现有特征层去偏方法多通过对抗学习或相关性约束抑制人口统计学相关信息,但当其与病理特征耦合时,直接抑制可能削弱疾病判别信息,难以兼顾诊断性能与公平性。为此,本文提出一种潜空间条件扩散去偏方法,用于医学图像公平诊断中的特征学习。方法通过双空间正交约束降低潜在病理特征与人口统计学特征的相关性,对初步解耦的病理特征施加受控高斯扰动,以人口统计学属性和扩散时间步为条件完成去噪重构,并采用子群自适应阈值校准预测结果。结果在 CheXpert 的 4 项病理分类任务上,本文方法的准确率(accuracy,Acc)、受试者工作特征曲线下面积(area under the curve,AUC)、交叉子群AUC差异和等机会差异分别为 0.796、0.857、0.058 和 0.108。与 7 种对比方法在相应指标上的最佳平均结果相比,平均准确率和 AUC 分别由 0.781 和 0.841 提高至 0.796 和 0.857,等机会差异由 0.126 降至 0.108,交叉子群 AUC 差异为 0.058,与最佳对比结果持平。在 MIMIC-CXR 跨数据集分布外测试中,相较于公平性表现最佳的对比方法,交叉子群 AUC 差异和等机会差异分别由 0.121 和 0.287 降至 0.085 和 0.166。结论所提出的潜空间条件扩散去偏方法可在保持疾病判别能力的同时减小人口统计学子群间的诊断性能差异,为医学图像公平诊断提供了一种潜空间特征重构的去偏思路。
- 1 January 2026
- Published by Aerospace Information Research Institute, Chinese Academy of Sciences in Journal of Image and Graphics
- Vol. null (null) , 1-17
面向医学图像公平诊断的潜空间条件扩散去偏
目的医学图像诊断模型易受年龄、性别和种族等人口统计学属性影响,造成子群间诊断性能差异。现有特征层去偏方法多通过对抗学习或相关性约束抑制人口统计学相关信息,但当其与病理特征耦合时,直接抑制可能削弱疾病判别信息,难以兼顾诊断性能与公平性。为此,本文提出一种潜空间条件扩散去偏方法,用于医学图像公平诊断中的特征学习。方法通过双空间正交约束降低潜在病理特征与人口统计学特征的相关性,对初步解耦的病理特征施加受控高斯扰动,以人口统计学属性和扩散时间步为条件完成去噪重构,并采用子群自适应阈值校准预测结果。结果在 CheXpert 的 4 项病理分类任务上,本文方法的准确率(accuracy,Acc)、受试者工作特征曲线下面积(area under the curve,AUC)、交叉子群AUC差异和等机会差异分别为 0.796、0.857、0.058 和 0.108。与 7 种对比方法在相应指标上的最佳平均结果相比,平均准确率和 AUC 分别由 0.781 和 0.841 提高至 0.796 和 0.857,等机会差异由 0.126 降至 0.108,交叉子群 AUC 差异为 0.058,与最佳对比结果持平。在 MIMIC-CXR 跨数据集分布外测试中,相较于公平性表现最佳的对比方法,交叉子群 AUC 差异和等机会差异分别由 0.121 和 0.287 降至 0.085 和 0.166。结论所提出的潜空间条件扩散去偏方法可在保持疾病判别能力的同时减小人口统计学子群间的诊断性能差异,为医学图像公平诊断提供了一种潜空间特征重构的去偏思路。
- 26 June 2025
- Published by Springer Nature in Computing
- Vol. 107 (7) , 1-31
FedDIFFKD: integrating diffusion distillation and self-distillation for personalized federated learning
- 26 June 2025
- Published by Springer Nature in Computing
- Vol. 107 (7) , 1-31
FedDIFFKD: integrating diffusion distillation and self-distillation for personalized federated learning
- 28 April 2025
- Published by Springer Nature in Computing
- Vol. 107 (5) , 1-22
FedSplit: federated learning optimization algorithm for resource constrained clients
- 28 April 2025
- Published by Springer Nature in Computing
- Vol. 107 (5) , 1-22
FedSplit: federated learning optimization algorithm for resource constrained clients
- 12 March 2025
- Published by Springer Nature in Multimedia Systems
- Vol. 31 (3) , 1-16
An improved Multi-Scale Fusion and Small Object Enhancement method for efficient pedestrian detection in dense scenes
- 12 March 2025
- Published by Springer Nature in Multimedia Systems
- Vol. 31 (3) , 1-16
An improved Multi-Scale Fusion and Small Object Enhancement method for efficient pedestrian detection in dense scenes