Selected work

Representative Papers

P2Fusion: Prompt-based Progressive Infrared-Visible Image Fusion via Dual-Prior Distillation

Aug 13, 2026

This work addresses the distortion in infrared and visible image fusion caused by inherent modality discrepancies and the limitations of existing methods that rely on static constraints or external priors while neglecting intrinsic modality characteristics. To this end, we propose a dual-intrinsic-prompt distillation framework that transforms thermal saliency and spatial quality priors into learnable dynamic prompts. Our approach enables decoupled and adaptive modality-specific feature fusion through a Teach-to-Fuse dual-granularity progressive guidance mechanism and a Gated Dynamic Expert Recalibration (GDER) module. The method achieves state-of-the-art performance across five benchmark datasets, outperforming competitors in 14 out of 20 evaluation metrics, and significantly enhances downstream object detection—improving mAP by 3.2%, 0.5%, and 0.9% on MSRS, M3FD, and DroneVehicle, respectively.

0 citationsRead paper

Beyond Reconstruction: Reconstruction-to-Vector Diffusion for Hyperspectral Anomaly Detection

Apr 13, 2026

This work addresses the limitations of conventional hyperspectral anomaly detection methods that rely on scalar reconstruction residuals, which often fail to preserve subpixel anomalies and suffer from confirmation bias due to anomaly-contaminated training. To overcome these issues, the authors propose a Reconstruction-to-Vector Diffusion (R2VD) framework that fundamentally shifts the detection paradigm from scalar reconstruction to vector diffusion. By leveraging a residual-guided generative dynamics mechanism within a four-stage pipeline, R2VD decouples targets from background through high-dimensional vector perturbation patterns. The method integrates physical priors, a full-context autoencoder, a diffusion Transformer, a physical spectral firewall, and vector dynamics inference to effectively retain subpixel structures and suppress spectral leakage. Evaluated across eight benchmark datasets, R2VD achieves state-of-the-art performance, significantly enhancing both anomaly detectability and background suppression.

0 citationsRead paper

ReIDMamba: Learning Discriminative Features with Visual State Space Model for Person Re-Identification

Nov 11, 2025

In person re-identification (ReID), existing Transformer-based methods suffer from quadratic computational and memory overhead with increasing sequence length. This work proposes ReIDMamba—the first pure-Mamba architecture for ReID—overcoming both the locality constraint of CNNs and the complexity bottleneck of Transformers. Methodologically, it introduces: (1) a global modeling backbone built upon the Vision State Space Model (VSSM); (2) a Multi-Granularity Feature Extractor (MGFE) to enhance fine-grained discriminability; and (3) Ranking-Aware Triplet Regularization (RATR) to improve feature robustness and diversity. ReIDMamba achieves state-of-the-art performance on five major benchmarks—including Market-1501 and DukeMTMC—while using only one-third the parameters of TransReID, requiring less GPU memory and enabling faster inference.

0 citationsRead paper

HistRetinex: Optimizing Retinex model in Histogram Domain for Efficient Low-Light Image Enhancement

Oct 23, 2025

To address the high computational cost of conventional Retinex methods for large-scale image enhancement, this paper pioneers the migration of the Retinex model into the histogram domain, establishing explicit histogram-level mappings between illumination and reflectance components. We propose a two-stage histogram-domain optimization framework: first, modeling via a position-count matrix; second, deriving closed-form iterative update rules for illumination and reflectance by integrating domain-specific priors, followed by efficient histogram matching for enhancement. Evaluated on a 1000×664 image, our method completes in only 1.86 seconds—accelerating over state-of-the-art approaches by at least 6.67 seconds—while achieving significant improvements in PSNR, SSIM, and perceptual quality. The core contribution lies in establishing the first histogram-domain Retinex modeling paradigm, uniquely balancing computational efficiency, reconstruction accuracy, and interpretability.

0 citationsRead paper

Recent publications

Latest Papers

P2Fusion: Prompt-based Progressive Infrared-Visible Image Fusion via Dual-Prior Distillation

Aug 13, 2026

This work addresses the distortion in infrared and visible image fusion caused by inherent modality discrepancies and the limitations of existing methods that rely on static constraints or external priors while neglecting intrinsic modality characteristics. To this end, we propose a dual-intrinsic-prompt distillation framework that transforms thermal saliency and spatial quality priors into learnable dynamic prompts. Our approach enables decoupled and adaptive modality-specific feature fusion through a Teach-to-Fuse dual-granularity progressive guidance mechanism and a Gated Dynamic Expert Recalibration (GDER) module. The method achieves state-of-the-art performance across five benchmark datasets, outperforming competitors in 14 out of 20 evaluation metrics, and significantly enhances downstream object detection—improving mAP by 3.2%, 0.5%, and 0.9% on MSRS, M3FD, and DroneVehicle, respectively.

0 citationsRead paper

Beyond Reconstruction: Reconstruction-to-Vector Diffusion for Hyperspectral Anomaly Detection

Apr 13, 2026

This work addresses the limitations of conventional hyperspectral anomaly detection methods that rely on scalar reconstruction residuals, which often fail to preserve subpixel anomalies and suffer from confirmation bias due to anomaly-contaminated training. To overcome these issues, the authors propose a Reconstruction-to-Vector Diffusion (R2VD) framework that fundamentally shifts the detection paradigm from scalar reconstruction to vector diffusion. By leveraging a residual-guided generative dynamics mechanism within a four-stage pipeline, R2VD decouples targets from background through high-dimensional vector perturbation patterns. The method integrates physical priors, a full-context autoencoder, a diffusion Transformer, a physical spectral firewall, and vector dynamics inference to effectively retain subpixel structures and suppress spectral leakage. Evaluated across eight benchmark datasets, R2VD achieves state-of-the-art performance, significantly enhancing both anomaly detectability and background suppression.

0 citationsRead paper

ReIDMamba: Learning Discriminative Features with Visual State Space Model for Person Re-Identification

Nov 11, 2025

In person re-identification (ReID), existing Transformer-based methods suffer from quadratic computational and memory overhead with increasing sequence length. This work proposes ReIDMamba—the first pure-Mamba architecture for ReID—overcoming both the locality constraint of CNNs and the complexity bottleneck of Transformers. Methodologically, it introduces: (1) a global modeling backbone built upon the Vision State Space Model (VSSM); (2) a Multi-Granularity Feature Extractor (MGFE) to enhance fine-grained discriminability; and (3) Ranking-Aware Triplet Regularization (RATR) to improve feature robustness and diversity. ReIDMamba achieves state-of-the-art performance on five major benchmarks—including Market-1501 and DukeMTMC—while using only one-third the parameters of TransReID, requiring less GPU memory and enabling faster inference.

0 citationsRead paper

HistRetinex: Optimizing Retinex model in Histogram Domain for Efficient Low-Light Image Enhancement

Oct 23, 2025

To address the high computational cost of conventional Retinex methods for large-scale image enhancement, this paper pioneers the migration of the Retinex model into the histogram domain, establishing explicit histogram-level mappings between illumination and reflectance components. We propose a two-stage histogram-domain optimization framework: first, modeling via a position-count matrix; second, deriving closed-form iterative update rules for illumination and reflectance by integrating domain-specific priors, followed by efficient histogram matching for enhancement. Evaluated on a 1000×664 image, our method completes in only 1.86 seconds—accelerating over state-of-the-art approaches by at least 6.67 seconds—while achieving significant improvements in PSNR, SSIM, and perceptual quality. The core contribution lies in establishing the first histogram-domain Retinex modeling paradigm, uniquely balancing computational efficiency, reconstruction accuracy, and interpretability.

0 citationsRead paper