Hui Li
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
Recognitions
Hui Li has a total of 372 co-authors. Here are the top co-authors by number of shared publications:
- Xiao-Jun Wu
Jiangnan University
18 shared publications - Tianyang Xu
Jiangnan University
11 shared publications Xiao-Jun Wu
11 shared publications- Xiaoning Song
Jiangnan University
10 shared publications - Josef Kittler
University of Surrey
9 shared publications Chunyang Cheng
Jiangnan University
6 shared publications
Latest Publications
- Open Access
- 9 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 10258-10266
CRIF: Precision-Oriented Parallel Guidance for Controllable Referring Infrared and Visible Image Fusion
- 9 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 10258-10266
Open AccessCRIF: Precision-Oriented Parallel Guidance for Controllable Referring Infrared and Visible Image Fusion
- Open Access
- 20 September 2026
- Published by Elsevier in SSRN Electronic Journal
EMP-PIDNet: Enhanced Memory and PID-Controlled Dynamics for Stable Video Prediction
Video prediction remains a fundamental challenge due to non-rigid motion, complex spatial interactions, and long-range temporal dependencies. Existing approaches struggle to maintain both accurate motion modeling and stable memory evolution, often leading to motion blur and trend drift in dynamic scenes. To address these issues, a novel spatiotemporal prediction framework based on a closed-loop regulation mechanism is proposed, termed as EMP-PIDNet.Specifically, to avoid the static gating for global updates (such as conventional LSTM), an Adaptive Spatiotemporal Memory Unit (ASTM-ConvLSTM) is introduced which explicitly incorporates an importance-scoring mask within its memory path. This scheme enables spatially adaptive partitioning of input sequences, enhancing representation in critical motion regions while suppressing redundant background dynamics. Furthermore, to achieve self-regulated trend correction and transient disturbance filtering, a PID-Guided Motion Modulation Module (PGMM) is proposed which establishes a dynamic feedback pathway that models the deviation between expected and observed motion responses.Extensive experiments on Moving MNIST, TaxiBJ, and Taihu HAB datasets demonstrate that EMP-PIDNet achieves superior motion boundary sharpness, temporal consistency, and predictive stability, surpassing state-of-the-art baselines across diverse dynamic environments.
- 20 September 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessEMP-PIDNet: Enhanced Memory and PID-Controlled Dynamics for Stable Video Prediction
Video prediction remains a fundamental challenge due to non-rigid motion, complex spatial interactions, and long-range temporal dependencies. Existing approaches struggle to maintain both accurate motion modeling and stable memory evolution, often leading to motion blur and trend drift in dynamic scenes. To address these issues, a novel spatiotemporal prediction framework based on a closed-loop regulation mechanism is proposed, termed as EMP-PIDNet.Specifically, to avoid the static gating for global updates (such as conventional LSTM), an Adaptive Spatiotemporal Memory Unit (ASTM-ConvLSTM) is introduced which explicitly incorporates an importance-scoring mask within its memory path. This scheme enables spatially adaptive partitioning of input sequences, enhancing representation in critical motion regions while suppressing redundant background dynamics. Furthermore, to achieve self-regulated trend correction and transient disturbance filtering, a PID-Guided Motion Modulation Module (PGMM) is proposed which establishes a dynamic feedback pathway that models the deviation between expected and observed motion responses.Extensive experiments on Moving MNIST, TaxiBJ, and Taihu HAB datasets demonstrate that EMP-PIDNet achieves superior motion boundary sharpness, temporal consistency, and predictive stability, surpassing state-of-the-art baselines across diverse dynamic environments.
- Open Access
- 5 September 2026
- Published by Elsevier in SSRN Electronic Journal
SGCAFuse: A Unified Image Fusion Network with Spatial Gating and Channel Attention for Generalizable Multi-Task Fusion
Unified image fusion seeks to handle heterogeneous fusion tasks using a single model while preserving the task-specific complementary information required by different imaging scenarios. However, effectively exchanging complementary information across heterogeneous fusion tasks while preserving taskspecific characteristics remains a fundamental challenge. To address these issues, a unified image fusion network built upon spatially adaptive crosstask interaction and channel-aware reconstruction is proposed (SGCAFuse).The network first extracts shared low-level structural and texture representations from input image pairs through a CNN encoder and further models task-aware dependencies using dual Swin Transformer branches. A Spatial Gating Cross-Task Interaction Module is developed to estimate location-wise transfer weights, so that complementary auxiliary information can be routed to relevant regions while task-irrelevant responses are suppressed. A Channel Attention Fusion Decoder is then introduced to recalibrate latent responsechannels and strengthen informative details during reconstruction. Experiments on infrared-visible image fusion (IVIF) and multi-focus image fusion(MFIF) benchmarks demonstrate that SGCAFuse achieves competitive or superior performance compared with representative state-of-the-art methods. Additional evaluations on unseen multi-exposure image fusion (MEIF) and medical image fusion datasets further demonstrate its strong generalization capability to unseen fusion tasks.
- 5 September 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessSGCAFuse: A Unified Image Fusion Network with Spatial Gating and Channel Attention for Generalizable Multi-Task Fusion
Unified image fusion seeks to handle heterogeneous fusion tasks using a single model while preserving the task-specific complementary information required by different imaging scenarios. However, effectively exchanging complementary information across heterogeneous fusion tasks while preserving taskspecific characteristics remains a fundamental challenge. To address these issues, a unified image fusion network built upon spatially adaptive crosstask interaction and channel-aware reconstruction is proposed (SGCAFuse).The network first extracts shared low-level structural and texture representations from input image pairs through a CNN encoder and further models task-aware dependencies using dual Swin Transformer branches. A Spatial Gating Cross-Task Interaction Module is developed to estimate location-wise transfer weights, so that complementary auxiliary information can be routed to relevant regions while task-irrelevant responses are suppressed. A Channel Attention Fusion Decoder is then introduced to recalibrate latent responsechannels and strengthen informative details during reconstruction. Experiments on infrared-visible image fusion (IVIF) and multi-focus image fusion(MFIF) benchmarks demonstrate that SGCAFuse achieves competitive or superior performance compared with representative state-of-the-art methods. Additional evaluations on unseen multi-exposure image fusion (MEIF) and medical image fusion datasets further demonstrate its strong generalization capability to unseen fusion tasks.
- 1 August 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Pattern Analysis and Machine Intelligence
- Vol. 48 (8) , 9914-9929
EvaNet: Toward More Efficient and Consistent Infrared and Visible Image Fusion Assessment
Evaluation is essential in image fusion research, yet most existing metrics are directly borrowed from other vision tasks without proper adaptation. These traditional metrics, often based on complex image transformations, not only fail to capture the true quality of the fusion results but also are computationally demanding. To address these issues, we propose a unified evaluation framework specifically tailored for image fusion. At its core is a lightweight network designed efficiently to approximate widely used metrics, following a divide-and conquer strategy. Unlike conventional approaches that directly assess similarity between fused and source images, we first decompose the fusion result into infrared and visible components. The evaluation model is then used to measure the degree of information preservation in these separated components, effectively disentangling the fusion evaluation process. During training, we incorporate a contrastive learning strategy and inform our evaluation model by perceptual scene assessment provided by a large language model. Last, we propose the first consistency evaluation framework, which measures the alignment between image fusion metrics and human visual perception, using both independent no-reference scores and downstream tasks performance as objective references. Extensive experiments show that our learning-based evaluation paradigm delivers both superior efficiency (up to 1,000 times faster) and greater consistency across a range of standard image fusion benchmarks. Our code will be publicly available at https://github.com/AWCXV/EvaNet.
- 1 August 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Pattern Analysis and Machine Intelligence
- Vol. 48 (8) , 9914-9929
EvaNet: Toward More Efficient and Consistent Infrared and Visible Image Fusion Assessment
Evaluation is essential in image fusion research, yet most existing metrics are directly borrowed from other vision tasks without proper adaptation. These traditional metrics, often based on complex image transformations, not only fail to capture the true quality of the fusion results but also are computationally demanding. To address these issues, we propose a unified evaluation framework specifically tailored for image fusion. At its core is a lightweight network designed efficiently to approximate widely used metrics, following a divide-and conquer strategy. Unlike conventional approaches that directly assess similarity between fused and source images, we first decompose the fusion result into infrared and visible components. The evaluation model is then used to measure the degree of information preservation in these separated components, effectively disentangling the fusion evaluation process. During training, we incorporate a contrastive learning strategy and inform our evaluation model by perceptual scene assessment provided by a large language model. Last, we propose the first consistency evaluation framework, which measures the alignment between image fusion metrics and human visual perception, using both independent no-reference scores and downstream tasks performance as objective references. Extensive experiments show that our learning-based evaluation paradigm delivers both superior efficiency (up to 1,000 times faster) and greater consistency across a range of standard image fusion benchmarks. Our code will be publicly available at https://github.com/AWCXV/EvaNet.
- Open Access
- 28 July 2026
- Published by Elsevier in SSRN Electronic Journal
A Novel Few-Shot Learning Framework based on Self-Supervised Dual-Branch Covariance Model
The core challenge of few-shot classification lies in the limited discriminability and generalization ability of visual representations under data scarcity. Existing methods often fail to fully exploit the complementary value of self-supervised visual representations and multimodal prior knowledge. To address this issue, we propose a self-supervised dual-branch covariance framework, termed the Dual-Branch Deep Brownian Distance Covariance Network (D-BDCC). Within a meta-learning framework, the task-adaptive branch adopts a ResNet-12 backbone and the Deep Brownian Distance Covariance (DeepBDC) module to learn fine-grained discriminative covariance features. In parallel, a frozen DINOv2 branch provides general visual priors learned from large-scale self-supervised pretraining. The two visual representations are fused through normalization and feature concatenation to construct a more robust visual embedding. During inference, frozen CLIP textual priors are further incorporated to provide category-level semantic guidance for the construction of multimodal prototypes. Extensive experiments on miniImageNet, CUB, and FewMedical-XJAU demonstrate the effectiveness of the proposed method.
- 28 July 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessA Novel Few-Shot Learning Framework based on Self-Supervised Dual-Branch Covariance Model
The core challenge of few-shot classification lies in the limited discriminability and generalization ability of visual representations under data scarcity. Existing methods often fail to fully exploit the complementary value of self-supervised visual representations and multimodal prior knowledge. To address this issue, we propose a self-supervised dual-branch covariance framework, termed the Dual-Branch Deep Brownian Distance Covariance Network (D-BDCC). Within a meta-learning framework, the task-adaptive branch adopts a ResNet-12 backbone and the Deep Brownian Distance Covariance (DeepBDC) module to learn fine-grained discriminative covariance features. In parallel, a frozen DINOv2 branch provides general visual priors learned from large-scale self-supervised pretraining. The two visual representations are fused through normalization and feature concatenation to construct a more robust visual embedding. During inference, frozen CLIP textual priors are further incorporated to provide category-level semantic guidance for the construction of multimodal prototypes. Extensive experiments on miniImageNet, CUB, and FewMedical-XJAU demonstrate the effectiveness of the proposed method.
- Open Access
- 21 July 2026
- Published by MDPI AG in Big Data and Cognitive Computing
- Vol. 10 (7) , 244
Safety-Aware Event-Triggered Intervention for Motion Planning and Decision Making in Diffusion-Based Autonomous Driving
Diffusion-based trajectory planners achieve strong nominal performance in autonomous driving, but sparse safety intervention remains difficult to evaluate and realize effectively. This study addresses this problem by proposing a safety-aware event-triggered intervention framework on top of a fixed DiffusionDriveV2 planner. The method uses candidate-level risk signals and an auxiliary semantic risk trigger to decide when intervention should be activated, and realizes the intervention through conservative re-selection and mild action-space augmentation. To evaluate sparse interventions beyond global validation metrics, we further construct normal, conservative, and actual trajectories and introduce triggered-subset counterfactual evaluation. On NAVSIM navtest, global planner metrics remain nearly unchanged across sparse trigger policies, but the semantic trigger achieves better triggered-subset final score, TTC, and progress than matched-random and TTC-based triggers. Qualitative cases show that behaviorally distinct safety responses mainly arise from action-space augmentation rather than candidate reranking alone. These results show that the proposed framework can diagnose and partially alleviate the gap between risk recognition and action realization, while revealing that stronger semantic-conditioned action generation is needed to fully overcome the trigger-to-action bottleneck.
- 21 July 2026
- Published by MDPI AG in Big Data and Cognitive Computing
- Vol. 10 (7) , 244
Open AccessSafety-Aware Event-Triggered Intervention for Motion Planning and Decision Making in Diffusion-Based Autonomous Driving
Diffusion-based trajectory planners achieve strong nominal performance in autonomous driving, but sparse safety intervention remains difficult to evaluate and realize effectively. This study addresses this problem by proposing a safety-aware event-triggered intervention framework on top of a fixed DiffusionDriveV2 planner. The method uses candidate-level risk signals and an auxiliary semantic risk trigger to decide when intervention should be activated, and realizes the intervention through conservative re-selection and mild action-space augmentation. To evaluate sparse interventions beyond global validation metrics, we further construct normal, conservative, and actual trajectories and introduce triggered-subset counterfactual evaluation. On NAVSIM navtest, global planner metrics remain nearly unchanged across sparse trigger policies, but the semantic trigger achieves better triggered-subset final score, TTC, and progress than matched-random and TTC-based triggers. Qualitative cases show that behaviorally distinct safety responses mainly arise from action-space augmentation rather than candidate reranking alone. These results show that the proposed framework can diagnose and partially alleviate the gap between risk recognition and action realization, while revealing that stronger semantic-conditioned action generation is needed to fully overcome the trigger-to-action bottleneck.
- 19 June 2026
- Published by IOP Publishing in Medical Engineering & Physics
- Vol. 147 (7)
From diagnostic labels to radiology reports: a unified multi-modal framework for lesion detection and segmentation
Deep learning-based vision models are playing an increasingly pivotal role in clinical diagnosis and treatment.However, existing approaches predominantly rely on visual information, often neglecting the accompanying radiology reports. Even when textual data is considered, current methods typically restrict their input to discrete text labels, failing to exploit the comprehensive, fine-grained semantic information embedded in full-text real-world radiology reports. This limitation represents a significant underutilization of available training resources. Furthermore, effectively leveraging complementary information from multi-modal imaging and exploiting the synergy between detection and segmentation remain open challenges. To address these limitations, we propose TextDSNet, a unified multi-modal framework for joint medical detection and segmentation, incorporating radiology reports. Specifically, TextDSNet leverages reports as cross-modal guidance to orchestrate multi-modal image analysis. We introduce a novel paradigm termed "Lesion-Aware Segmentation," which enables the model to identify specific lesion targets, facilitating the joint optimization of coarse-grained localization and fine-grained segmentation. To overcome the scarcity of public datasets containing paired full-text reports, we curated radiology-style report annotations for the ISLES 2022 dataset under strict clinical supervision. Experimental results show that TextDSNet surpasses existing state-of-theart methods on this dataset. Our findings demonstrate that integrating full-text radiology reports is a viable path toward building more intelligent and clinically aligned medical AI systems. Code and dataset are available at https://github.com/wenxuan163/TextDSNet.git.
- 19 June 2026
- Published by IOP Publishing in Medical Engineering & Physics
- Vol. 147 (7)
From diagnostic labels to radiology reports: a unified multi-modal framework for lesion detection and segmentation
Deep learning-based vision models are playing an increasingly pivotal role in clinical diagnosis and treatment.However, existing approaches predominantly rely on visual information, often neglecting the accompanying radiology reports. Even when textual data is considered, current methods typically restrict their input to discrete text labels, failing to exploit the comprehensive, fine-grained semantic information embedded in full-text real-world radiology reports. This limitation represents a significant underutilization of available training resources. Furthermore, effectively leveraging complementary information from multi-modal imaging and exploiting the synergy between detection and segmentation remain open challenges. To address these limitations, we propose TextDSNet, a unified multi-modal framework for joint medical detection and segmentation, incorporating radiology reports. Specifically, TextDSNet leverages reports as cross-modal guidance to orchestrate multi-modal image analysis. We introduce a novel paradigm termed "Lesion-Aware Segmentation," which enables the model to identify specific lesion targets, facilitating the joint optimization of coarse-grained localization and fine-grained segmentation. To overcome the scarcity of public datasets containing paired full-text reports, we curated radiology-style report annotations for the ISLES 2022 dataset under strict clinical supervision. Experimental results show that TextDSNet surpasses existing state-of-theart methods on this dataset. Our findings demonstrate that integrating full-text radiology reports is a viable path toward building more intelligent and clinically aligned medical AI systems. Code and dataset are available at https://github.com/wenxuan163/TextDSNet.git.
- 11 March 2026
- Published by Elsevier in Signal Processing
- Vol. 245
Text-guided medical image fusion using unbalanced optimal transport: Semantic alignment and cross-modal interaction
- 11 March 2026
- Published by Elsevier in Signal Processing
- Vol. 245
Text-guided medical image fusion using unbalanced optimal transport: Semantic alignment and cross-modal interaction
- Open Access
- 12 January 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Multimedia Computing, Communications, and Applications
- Vol. 22 (1) , 1-19
BusReF: Infrared-Visible Images Registration and Fusion Focus on Reconstructible Area Using One Set of Features
In multi-modal imaging scenarios, the misalignment of images presents a persistent challenge. Conventional image fusion algorithms, aiming to enhance the performance of downstream vision tasks, presuppose strictly registered inputs to achieve satisfactory results. To relax this assumption, a common approach is to register the images first; however, existing multi-modal registration methods are often hindered by complex architectures and a heavy reliance on semantic information. This article proposes BusRef, a unified framework that jointly addresses image registration and fusion, with a specific focus on the Infrared-Visible Image Registration and Fusion (IVRF) task. Within this framework, unaligned image pairs are processed through three sequential stages: coarse registration, fine registration, and fusion. We demonstrate that this integrated approach enables more robust and accurate IVRF. Key to our framework is a novel training and evaluation strategy that employs masks to mitigate the influence of non-reconstructible regions on the loss function, thereby significantly improving the model’s accuracy and robustness. Furthermore, we introduce a gradient-aware fusion network designed to effectively preserve complementary information from both modalities. Comprehensive experiments demonstrate that BusRef achieves superior performance when compared against various state-of-the-art registration and fusion algorithms. Our code is available at https://github.com/Yukarizz/BusReF .
- 12 January 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Multimedia Computing, Communications, and Applications
- Vol. 22 (1) , 1-19
Open AccessBusReF: Infrared-Visible Images Registration and Fusion Focus on Reconstructible Area Using One Set of Features
In multi-modal imaging scenarios, the misalignment of images presents a persistent challenge. Conventional image fusion algorithms, aiming to enhance the performance of downstream vision tasks, presuppose strictly registered inputs to achieve satisfactory results. To relax this assumption, a common approach is to register the images first; however, existing multi-modal registration methods are often hindered by complex architectures and a heavy reliance on semantic information. This article proposes BusRef, a unified framework that jointly addresses image registration and fusion, with a specific focus on the Infrared-Visible Image Registration and Fusion (IVRF) task. Within this framework, unaligned image pairs are processed through three sequential stages: coarse registration, fine registration, and fusion. We demonstrate that this integrated approach enables more robust and accurate IVRF. Key to our framework is a novel training and evaluation strategy that employs masks to mitigate the influence of non-reconstructible regions on the loss function, thereby significantly improving the model’s accuracy and robustness. Furthermore, we introduce a gradient-aware fusion network designed to effectively preserve complementary information from both modalities. Comprehensive experiments demonstrate that BusRef achieves superior performance when compared against various state-of-the-art registration and fusion algorithms. Our code is available at https://github.com/Yukarizz/BusReF .