Minjun Wang
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
Minjun Wang has a total of 24 co-authors. Here are the top co-authors by number of shared publications:
- Houjin Chen
School of Electronic and Information Engineering, Beijing Jiaotong University, Beijing 100044, China
8 shared publications - Yanfeng Li
Beijing Jiaotong University
6 shared publications - Luyifu Chen
School of Electronic and Information Engineering, Beijing Jiaotong University, Beijing, China
5 shared publications Jia Sun
3 shared publications- Jia Sun
School of Electronic and Information Engineering, Beijing Jiaotong University, Beijing, China
3 shared publications Yanfeng Li
3 shared publications
Latest Publications
- 1 November 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems for Video Technology
- Vol. 36 (11) , 16744-16757
Double Alignment for Generalizable Person Re-Identification
- 1 November 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems for Video Technology
- Vol. 36 (11) , 16744-16757
Double Alignment for Generalizable Person Re-Identification
- 7 July 2026
- Published by Elsevier in Neurocomputing
- Vol. 700
DCCD: Dual-contrastive channel disentanglement for single-source domain generalization in medical image segmentation
- 7 July 2026
- Published by Elsevier in Neurocomputing
- Vol. 700
DCCD: Dual-contrastive channel disentanglement for single-source domain generalization in medical image segmentation
- 6 January 2026
- Published by Elsevier in Biomedical Signal Processing and Control
- Vol. 116
DisFreAda: Distribution-driven multi-frequency adaptive network for generalizable medical image segmentation
- 6 January 2026
- Published by Elsevier in Biomedical Signal Processing and Control
- Vol. 116
DisFreAda: Distribution-driven multi-frequency adaptive network for generalizable medical image segmentation
- 20 November 2025
- Published by Elsevier in Computerized Medical Imaging and Graphics
- Vol. 126, 102670
Hallucinated domain generalization network with domain-aware dynamic representation for medical image segmentation
- 20 November 2025
- Published by Elsevier in Computerized Medical Imaging and Graphics
- Vol. 126, 102670
Hallucinated domain generalization network with domain-aware dynamic representation for medical image segmentation
- 12 May 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Instrumentation and Measurement
- Vol. 74 (00189456) , 1-13
Multilevel Feature Perturbation for Single-Domain Generalization Person Reidentification
Domain generalization person reidentification (DG ReID) is a critical research topic for intelligent video surveillance, aiming at building a robust model generalizing to the unseen domain. Particularly in single-source domain generalization (SDG) scenarios, the monotonous data style presents a significant challenge. To address this issue, existing domain generalization (DG) methods typically employ image- or feature-level data augmentation techniques. However, these approaches may have limitations when dealing with the complexities of real-world scenarios due to their reliance on a single augmentation strategy. To overcome these limitations, we propose a novel method called multilevel feature perturbation (MLFP) for SDG ReID. MLFP integrates two complementary augmentation modules: random background perturbation (RBP) at the image level and the uncertain sampling normalization module (USNM) at the feature level. Specifically, RBP is designed to perform a random mixture of image backgrounds at the image level, while preserving the integrity of the pedestrian subject. By focusing on the pedestrian objectives instead of the backgrounds, RBP encourages the model to learn domain-invariant, causal features and reduce the bias toward the source domain. USNM is designed to diversify the feature-level representations through manipulating feature statistics. To further enhance the learning process, we introduce a novel loss function called improved whitening loss (IW Loss). This loss function is designed to capture consistent representations and explore invariant patterns among diverse feature distributions. Extensive experimental results demonstrate that our MLFP outperforms state-of-the-art methods across several SDG ReID benchmarks, highlighting its effectiveness in addressing the challenges of domain generalization.
- 12 May 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Instrumentation and Measurement
- Vol. 74 (00189456) , 1-13
Multilevel Feature Perturbation for Single-Domain Generalization Person Reidentification
Domain generalization person reidentification (DG ReID) is a critical research topic for intelligent video surveillance, aiming at building a robust model generalizing to the unseen domain. Particularly in single-source domain generalization (SDG) scenarios, the monotonous data style presents a significant challenge. To address this issue, existing domain generalization (DG) methods typically employ image- or feature-level data augmentation techniques. However, these approaches may have limitations when dealing with the complexities of real-world scenarios due to their reliance on a single augmentation strategy. To overcome these limitations, we propose a novel method called multilevel feature perturbation (MLFP) for SDG ReID. MLFP integrates two complementary augmentation modules: random background perturbation (RBP) at the image level and the uncertain sampling normalization module (USNM) at the feature level. Specifically, RBP is designed to perform a random mixture of image backgrounds at the image level, while preserving the integrity of the pedestrian subject. By focusing on the pedestrian objectives instead of the backgrounds, RBP encourages the model to learn domain-invariant, causal features and reduce the bias toward the source domain. USNM is designed to diversify the feature-level representations through manipulating feature statistics. To further enhance the learning process, we introduce a novel loss function called improved whitening loss (IW Loss). This loss function is designed to capture consistent representations and explore invariant patterns among diverse feature distributions. Extensive experimental results demonstrate that our MLFP outperforms state-of-the-art methods across several SDG ReID benchmarks, highlighting its effectiveness in addressing the challenges of domain generalization.
- 12 December 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Information Forensics and Security
- Vol. 20 (15566013) , 1106-1118
Dualistic Disentangled Meta-Learning Model for Generalizable Person Re-Identification
Person re-identification (re-ID) is a research hotspot in the field of intelligent monitoring and security. Domain generalizable (DG) person re-identification transfers the trained model directly to the unseen target domain for testing, which is closer to the practical application than supervised or unsupervised person re-ID. Meta-learning strategy is an effective way to solve the DG problem, nevertheless, existing meta-learning-based DG re-ID methods mainly simulates the test process in a single aspect such as identity or style, while ignoring the completely different person identities and styles in the unseen target domain. As to this problem, we consider a double disentangling from two levels of training strategy and feature learning, and propose a novel dualistic disentangled meta-learning (D 2 ML) model. D 2 ML is composed of two disentangling stages, one is for learning strategy, which spreads one-stage meta-test into two-stage, including an identity meta-test stage and a style meta-test stage. The other is for feature representation, which decouples the shallow layer features into identity-related features and style-related features. Specifically, we first conduct identity meta-test stage on different person identities of the images, and then employ a feature-level style perturbation module (SPM) based on Fourier spectrum transformation to conduct the style meta-test stage on the image with diversified styles. With these two stages, abundant changes in the unseen domain can be simulated during the meta-test phase. Besides, to learn more identity-related features, a feature disentangling module (FDM) is inserted at each stage of meta-learning and a disentangled triplet loss is developed. Through constraining the relationship between identity-related features and style-related features, the generalization ability of the model can be further improved. Experimental results on four public datasets show that our D 2 ML model achieves superior generalization performance compared to the state-of-the-art methods.
- 12 December 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Information Forensics and Security
- Vol. 20 (15566013) , 1106-1118
Dualistic Disentangled Meta-Learning Model for Generalizable Person Re-Identification
Person re-identification (re-ID) is a research hotspot in the field of intelligent monitoring and security. Domain generalizable (DG) person re-identification transfers the trained model directly to the unseen target domain for testing, which is closer to the practical application than supervised or unsupervised person re-ID. Meta-learning strategy is an effective way to solve the DG problem, nevertheless, existing meta-learning-based DG re-ID methods mainly simulates the test process in a single aspect such as identity or style, while ignoring the completely different person identities and styles in the unseen target domain. As to this problem, we consider a double disentangling from two levels of training strategy and feature learning, and propose a novel dualistic disentangled meta-learning (D 2 ML) model. D 2 ML is composed of two disentangling stages, one is for learning strategy, which spreads one-stage meta-test into two-stage, including an identity meta-test stage and a style meta-test stage. The other is for feature representation, which decouples the shallow layer features into identity-related features and style-related features. Specifically, we first conduct identity meta-test stage on different person identities of the images, and then employ a feature-level style perturbation module (SPM) based on Fourier spectrum transformation to conduct the style meta-test stage on the image with diversified styles. With these two stages, abundant changes in the unseen domain can be simulated during the meta-test phase. Besides, to learn more identity-related features, a feature disentangling module (FDM) is inserted at each stage of meta-learning and a disentangled triplet loss is developed. Through constraining the relationship between identity-related features and style-related features, the generalization ability of the model can be further improved. Experimental results on four public datasets show that our D 2 ML model achieves superior generalization performance compared to the state-of-the-art methods.
- 19 September 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Sensors Journal
- Vol. 24 (22) , 37152-37168
WTFusion: Wavelet-Assisted Transformer Network for Multisensor Image Fusion
The objective of image fusion is to acquire a single image encompassing comprehensive information for better scene perception and visual comprehension with the benefit of multisensor images. However, existing fusion methods are often tailored to the specific fusion task, hindering the generalization of fusion networks across diverse tasks. To address this dilemma, an effective wavelet-assisted transformer network for multisensor image fusion is proposed in this study, named WTFusion. In WTFusion, the source images are treated as the degraded inputs, thus reframing image fusion as an image restoration process. This shift eliminates the necessity for complex fusion strategies, enhancing the versatility of WTFusion across diverse fusion tasks. Specifically, an encoder-decoder architecture incorporating multi-scale skip connections and transformer is designed for capturing both local and global representations. The encoder module is implemented for information extraction from the source images, while the decoder module with multiscale skip connections is utilized for the reconstruction of the fused image. In addition, the wavelet-based down-sampling layer is devised to replace the conventional down-sampling operation, thus averting the destruction of the basic structures and preserving details. To avoid hyperparameter dependency in loss functions, we design a wavelet domain loss to reconstruct low-frequency and high-frequency components regardless of trade-off hyperparameters, thus aligning with the intrinsic features of the source images and fostering universality across diverse fusion tasks. Extensive experiments conducted on multimodal image datasets reveal the superior performance of WTFusion over existing methods and the potential of WTFusion as a unified and effective method for image fusion tasks.
- 19 September 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Sensors Journal
- Vol. 24 (22) , 37152-37168
WTFusion: Wavelet-Assisted Transformer Network for Multisensor Image Fusion
The objective of image fusion is to acquire a single image encompassing comprehensive information for better scene perception and visual comprehension with the benefit of multisensor images. However, existing fusion methods are often tailored to the specific fusion task, hindering the generalization of fusion networks across diverse tasks. To address this dilemma, an effective wavelet-assisted transformer network for multisensor image fusion is proposed in this study, named WTFusion. In WTFusion, the source images are treated as the degraded inputs, thus reframing image fusion as an image restoration process. This shift eliminates the necessity for complex fusion strategies, enhancing the versatility of WTFusion across diverse fusion tasks. Specifically, an encoder-decoder architecture incorporating multi-scale skip connections and transformer is designed for capturing both local and global representations. The encoder module is implemented for information extraction from the source images, while the decoder module with multiscale skip connections is utilized for the reconstruction of the fused image. In addition, the wavelet-based down-sampling layer is devised to replace the conventional down-sampling operation, thus averting the destruction of the basic structures and preserving details. To avoid hyperparameter dependency in loss functions, we design a wavelet domain loss to reconstruct low-frequency and high-frequency components regardless of trade-off hyperparameters, thus aligning with the intrinsic features of the source images and fostering universality across diverse fusion tasks. Extensive experiments conducted on multimodal image datasets reveal the superior performance of WTFusion over existing methods and the potential of WTFusion as a unified and effective method for image fusion tasks.
- 14 December 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Instrumentation and Measurement
- Vol. 73 (00189456) , 1-16
RITFusion: Reinforced Interactive Transformer Network for Infrared and Visible Image Fusion
Infrared (IR) and visible (VIS) image fusion aims to obtain one comprehensive image, which can be consistent with the source images on details and thermal targets. However, the existing image fusion methods do not work well when faced with low-quality images in adverse conditions, such as overexposure, low illumination, smoke occlusion, and similar backgrounds. To address this dilemma, we propose an effective reinforced interactive transformer network for IR and VIS image fusion (RITFusion). To integrate different features from two modalities, we develop a novel fusion strategy composed of the intramodality self-attention (IMSA) block, the modal reinforcement (MR) block, and the intermodality interactive-attention (IMIA) block for implementing information association and interaction between modalities, while strengthening the weakened features caused by the adverse conditions. Moreover, the multiscale skip connection network is deployed to make full use of features from the source images. Furthermore, we design a morphology-based loss function and combine it with the intensity loss function, which guides network training to retain adequate details and ample thermal targets. Extensive experiments on public datasets demonstrate that our RITFusion outperforms other state-of-the-art fusion methods under both the common and the adverse conditions. The extended experiment on salient object detection reveals that the proposed method can boost the detection performance for subsequent high-level vision tasks.
- 14 December 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Instrumentation and Measurement
- Vol. 73 (00189456) , 1-16
RITFusion: Reinforced Interactive Transformer Network for Infrared and Visible Image Fusion
Infrared (IR) and visible (VIS) image fusion aims to obtain one comprehensive image, which can be consistent with the source images on details and thermal targets. However, the existing image fusion methods do not work well when faced with low-quality images in adverse conditions, such as overexposure, low illumination, smoke occlusion, and similar backgrounds. To address this dilemma, we propose an effective reinforced interactive transformer network for IR and VIS image fusion (RITFusion). To integrate different features from two modalities, we develop a novel fusion strategy composed of the intramodality self-attention (IMSA) block, the modal reinforcement (MR) block, and the intermodality interactive-attention (IMIA) block for implementing information association and interaction between modalities, while strengthening the weakened features caused by the adverse conditions. Moreover, the multiscale skip connection network is deployed to make full use of features from the source images. Furthermore, we design a morphology-based loss function and combine it with the intensity loss function, which guides network training to retain adequate details and ample thermal targets. Extensive experiments on public datasets demonstrate that our RITFusion outperforms other state-of-the-art fusion methods under both the common and the adverse conditions. The extended experiment on salient object detection reveals that the proposed method can boost the detection performance for subsequent high-level vision tasks.
- Open Access
- 2 May 2021
- Published by Institution of Engineering and Technology (IET) in IET Intelligent Transport Systems
- Vol. 15 (6) , 837-849
Multi‐scale pedestrian detection based on self‐attention and adaptively spatial feature fusion
Pedestrian detection is a classic problem in computer vision, which has an essential impact on the safety of urban autonomous driving. Although significant improvement has been made in pedestrian detection recently, small‐scale pedestrian detection is still challenging. To effectively tackle this issue, a multi‐scale pedestrian detector based on self‐attention mechanism and adaptive spatial feature fusion is proposed in this paper. In order to better extract global information, the spatial attention mechanism asymmetric pyramid non‐local block (APNB) module is applied. To achieve scale‐invariance detection, multiple detection branches are designed, which include a high‐resolution detection branch and a low‐resolution detection branch. In integrating multi‐scale features, the adaptively spatial feature fusion (ASFF) method is employed, which can solve the problem of feature inconsistency across different scales. Experimental results show that the proposed method obtains competitive performance on Caltech and CityPersons datasets.
- 2 May 2021
- Published by Institution of Engineering and Technology (IET) in IET Intelligent Transport Systems
- Vol. 15 (6) , 837-849
Open AccessMulti‐scale pedestrian detection based on self‐attention and adaptively spatial feature fusion
Pedestrian detection is a classic problem in computer vision, which has an essential impact on the safety of urban autonomous driving. Although significant improvement has been made in pedestrian detection recently, small‐scale pedestrian detection is still challenging. To effectively tackle this issue, a multi‐scale pedestrian detector based on self‐attention mechanism and adaptive spatial feature fusion is proposed in this paper. In order to better extract global information, the spatial attention mechanism asymmetric pyramid non‐local block (APNB) module is applied. To achieve scale‐invariance detection, multiple detection branches are designed, which include a high‐resolution detection branch and a low‐resolution detection branch. In integrating multi‐scale features, the adaptively spatial feature fusion (ASFF) method is employed, which can solve the problem of feature inconsistency across different scales. Experimental results show that the proposed method obtains competitive performance on Caltech and CityPersons datasets.
- 1 October 2018
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 1-9
Face Expression Recognition Based on Deep Convolution Network
A method based on depth volume and network for facial expression recognition was proposed. This method takes the facial expression image as the input of the CNN and trains the CNN network, and then uses the trained network to perform facial expression recognition. This paper uses jaffe and ck+ two face expression libraries to verify the algorithm, proves the effectiveness of the algorithm, and shows that its performance is better than the traditional method.
- 1 October 2018
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 1-9
Face Expression Recognition Based on Deep Convolution Network
A method based on depth volume and network for facial expression recognition was proposed. This method takes the facial expression image as the input of the CNN and trains the CNN network, and then uses the trained network to perform facial expression recognition. This paper uses jaffe and ck+ two face expression libraries to verify the algorithm, proves the effectiveness of the algorithm, and shows that its performance is better than the traditional method.