Yulin Cao
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Yulin Cao has a total of 42 co-authors. Here are the top co-authors by number of shared publications:
- Le Sun
School of Computer Science, Nanjing University of Information Science and Technology, Nanjing, China
4 shared publications - Jiexing Li
College of Computer Qinghai Normal University Xining China
3 shared publications - Long Jin
Lanzhou University
3 shared publications - Wen Lu
College of Computer & State Key Laboratory of Tibetan Intelligent, Qinghai Normal University, Xining 810008, China
3 shared publications - Yuhui Zheng
Qinghai Normal University
3 shared publications - Zaobo He
College of Information Science and Technology, Jinan University, Guangzhou, China
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) , 17615-17627
PerturbMatch: Multi-Level Progressive Perturbation for Semi-Supervised Semantic Segmentation
- 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) , 17615-17627
PerturbMatch: Multi-Level Progressive Perturbation for Semi-Supervised Semantic Segmentation
- Open Access
- 2 October 2026
- Published by MDPI AG in Remote Sensing
- Vol. 18 (19) , 3381
From Prior to Frequency: A Progressive Structure Recovery Network for Remote Sensing Road Segmentation
Roads in remote sensing imagery often contain narrow branches, complex boundaries, and long-range connections. Background interference and repeated downsampling make these structures difficult to preserve. To address these challenges, we propose a Progressive Structure Recovery Network (PSRNet) for remote sensing road segmentation. The network integrates structural-prior guidance, frequency-domain enhancement, and multiscale decoding. A line-structure prior is generated from the input image and encoded alongside hierarchical image features. Local Detail Alignment integrates prior cues with shallow features to enhance local structural details. Structure–Semantic Alignment combines deep semantic features with structural cues to refine unreliable prior responses. At the bottleneck, a Frequency Detail-Aware Enhancement module adaptively modulates the amplitude spectrum of the aligned features while preserving phase information. The reconstructed features are combined with spatial representations through residual fusion. A deep-to-shallow progressive decoder then restores spatial details and generates multiscale side predictions, which are fused to produce the final segmentation map. Experiments on the Massachusetts Roads, DeepGlobe, and CHN6-CUG datasets demonstrate the effectiveness of PSRNet for road segmentation. Ablation studies further support the complementary contributions of structural-prior guidance, frequency-domain enhancement, and progressive decoding.
- 2 October 2026
- Published by MDPI AG in Remote Sensing
- Vol. 18 (19) , 3381
Open AccessFrom Prior to Frequency: A Progressive Structure Recovery Network for Remote Sensing Road Segmentation
Roads in remote sensing imagery often contain narrow branches, complex boundaries, and long-range connections. Background interference and repeated downsampling make these structures difficult to preserve. To address these challenges, we propose a Progressive Structure Recovery Network (PSRNet) for remote sensing road segmentation. The network integrates structural-prior guidance, frequency-domain enhancement, and multiscale decoding. A line-structure prior is generated from the input image and encoded alongside hierarchical image features. Local Detail Alignment integrates prior cues with shallow features to enhance local structural details. Structure–Semantic Alignment combines deep semantic features with structural cues to refine unreliable prior responses. At the bottleneck, a Frequency Detail-Aware Enhancement module adaptively modulates the amplitude spectrum of the aligned features while preserving phase information. The reconstructed features are combined with spatial representations through residual fusion. A deep-to-shallow progressive decoder then restores spatial details and generates multiscale side predictions, which are fused to produce the final segmentation map. Experiments on the Massachusetts Roads, DeepGlobe, and CHN6-CUG datasets demonstrate the effectiveness of PSRNet for road segmentation. Ablation studies further support the complementary contributions of structural-prior guidance, frequency-domain enhancement, and progressive decoding.
- 1 October 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. 13 (20) , 46359-46370
Fed-DAPT: Federated Dynamic Alignment and Prompt Tuning for Efficient Personalized Vision Transformers
- 1 October 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. 13 (20) , 46359-46370
Fed-DAPT: Federated Dynamic Alignment and Prompt Tuning for Efficient Personalized Vision Transformers
- 21 September 2026
- Published by Springer Nature in Nonlinear Dynamics
- Vol. 114 (18) , 1171
An efficient region-of-interest-driven simplified neural dynamics for large-scale inverse problems
- 21 September 2026
- Published by Springer Nature in Nonlinear Dynamics
- Vol. 114 (18) , 1171
An efficient region-of-interest-driven simplified neural dynamics for large-scale inverse problems
- Open Access
- 20 September 2026
- Published by MDPI AG in Remote Sensing
- Vol. 18 (18) , 3234
FCD-Mamba: A Frequency-Enhanced and Center-Pixel-Guided Dual-Branch Mamba Network for Hyperspectral Image Classification
Hyperspectral image classification is an important task in remote sensing image interpretation. However, existing Mamba-based methods remain limited in exploiting shallow frequency information, preserving neighborhood continuity during spatial serialization, and protecting center-pixel semantics. To address these limitations, we propose a frequency-enhanced and center-pixel-guided dual-branch Mamba network (FCD-Mamba). First, the Frequency-Enhanced Multiscale Convolution Module (FEMSCM) applies serial Fourier filtering to enhance multiscale spatial features and calibrates spectral features using low-, middle-, and high-frequency DCT energy descriptors derived from the original spectrum of the center pixel. Second, the Spatial Center-Adaptive Mamba (SCA-Mamba) employs two orthogonal Hilbert curves and their reverse traversals to preserve spatial neighborhood continuity, while adaptively aggregating multi-path features using center and global representations. The Spectral Center-Preserving Mamba (SCP-Mamba) constructs center-preserved and global spectral sequences and models their complementary spectral dependencies through bidirectional scanning. Finally, the Center-Pixel-Guided Residual Fusion Module (CGRFM) exploits center-to-global discrepancies to calibrate dual-branch features and introduces auxiliary supervision to maintain the discriminability of each branch. Experiments on multiple public hyperspectral datasets demonstrate that FCD-Mamba achieves stable and competitive classification performance, validating the effectiveness of its constituent modules.
- 20 September 2026
- Published by MDPI AG in Remote Sensing
- Vol. 18 (18) , 3234
Open AccessFCD-Mamba: A Frequency-Enhanced and Center-Pixel-Guided Dual-Branch Mamba Network for Hyperspectral Image Classification
Hyperspectral image classification is an important task in remote sensing image interpretation. However, existing Mamba-based methods remain limited in exploiting shallow frequency information, preserving neighborhood continuity during spatial serialization, and protecting center-pixel semantics. To address these limitations, we propose a frequency-enhanced and center-pixel-guided dual-branch Mamba network (FCD-Mamba). First, the Frequency-Enhanced Multiscale Convolution Module (FEMSCM) applies serial Fourier filtering to enhance multiscale spatial features and calibrates spectral features using low-, middle-, and high-frequency DCT energy descriptors derived from the original spectrum of the center pixel. Second, the Spatial Center-Adaptive Mamba (SCA-Mamba) employs two orthogonal Hilbert curves and their reverse traversals to preserve spatial neighborhood continuity, while adaptively aggregating multi-path features using center and global representations. The Spectral Center-Preserving Mamba (SCP-Mamba) constructs center-preserved and global spectral sequences and models their complementary spectral dependencies through bidirectional scanning. Finally, the Center-Pixel-Guided Residual Fusion Module (CGRFM) exploits center-to-global discrepancies to calibrate dual-branch features and introduces auxiliary supervision to maintain the discriminability of each branch. Experiments on multiple public hyperspectral datasets demonstrate that FCD-Mamba achieves stable and competitive classification performance, validating the effectiveness of its constituent modules.
- Open Access
- 21 August 2026
- Published by MDPI AG in Remote Sensing
- Vol. 18 (16) , 2843
Weakly Supervised Remote Sensing Segmentation via Decoupled Cross-Modal Distillation and Semantic-Guided Refinement
Pixel-level annotation of remote sensing imagery is costly, motivating weakly supervised semantic segmentation (WSSS) using only image-level labels. However, class activation maps (CAMs) often highlight only discriminative sub-regions and fail to separate adjacent land-cover regions, particularly in remote sensing scenes characterized by densely co-occurring land-cover classes and substantial variations in object scale. To address these limitations, we propose a three-stage framework that integrates complementary priors from Contrastive Language–Image Pre-training (CLIP), Self-Distillation with No Labels version 2 (DINOv2), and the Segment Anything Model (SAM). First, a lightweight CLIP adapter aligns vision–language priors with remote sensing imagery, while sigmoid-based multi-label decoupled distillation replaces class-competitive distillation with independent class-wise supervision, producing more complete CAMs. Second, DINOv2-guided feature clustering decomposes large merged regions before SAM prompt generation, while Spatial–Semantic Constraints are used to construct confidence-guided point-and-box prompts and reject excessively expanded or semantically inconsistent masks, thereby generating reliable pseudo-labels. Finally, a compact segmentation network is initialized with the weights learned in Stage 1 and retrained using the refined pseudo-labels generated in Stage 2, eliminating the need for foundation models during inference. Experiments on the Potsdam, LoveDA, and DeepGlobe datasets show that the proposed method achieves mean intersection over union (mIoU) scores of 53.16%, 52.66%, and 62.98%, respectively, outperforming state-of-the-art WSSS baselines by 6.55, 1.16, and 1.27 percentage points, respectively. These results demonstrate the effectiveness and generalizability of the proposed framework across diverse remote sensing scenarios under image-level supervision.
- 21 August 2026
- Published by MDPI AG in Remote Sensing
- Vol. 18 (16) , 2843
Open AccessWeakly Supervised Remote Sensing Segmentation via Decoupled Cross-Modal Distillation and Semantic-Guided Refinement
Pixel-level annotation of remote sensing imagery is costly, motivating weakly supervised semantic segmentation (WSSS) using only image-level labels. However, class activation maps (CAMs) often highlight only discriminative sub-regions and fail to separate adjacent land-cover regions, particularly in remote sensing scenes characterized by densely co-occurring land-cover classes and substantial variations in object scale. To address these limitations, we propose a three-stage framework that integrates complementary priors from Contrastive Language–Image Pre-training (CLIP), Self-Distillation with No Labels version 2 (DINOv2), and the Segment Anything Model (SAM). First, a lightweight CLIP adapter aligns vision–language priors with remote sensing imagery, while sigmoid-based multi-label decoupled distillation replaces class-competitive distillation with independent class-wise supervision, producing more complete CAMs. Second, DINOv2-guided feature clustering decomposes large merged regions before SAM prompt generation, while Spatial–Semantic Constraints are used to construct confidence-guided point-and-box prompts and reject excessively expanded or semantically inconsistent masks, thereby generating reliable pseudo-labels. Finally, a compact segmentation network is initialized with the weights learned in Stage 1 and retrained using the refined pseudo-labels generated in Stage 2, eliminating the need for foundation models during inference. Experiments on the Potsdam, LoveDA, and DeepGlobe datasets show that the proposed method achieves mean intersection over union (mIoU) scores of 53.16%, 52.66%, and 62.98%, respectively, outperforming state-of-the-art WSSS baselines by 6.55, 1.16, and 1.27 percentage points, respectively. These results demonstrate the effectiveness and generalizability of the proposed framework across diverse remote sensing scenarios under image-level supervision.
- Open Access
- 3 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 24-30
Large Models in Smart Water Systems
- 3 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 24-30
Open AccessLarge Models in Smart Water Systems
- 12 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Geoscience and Remote Sensing
- Vol. PP (01962892) , 1
SFIEET: Spectral Frequency-Induced Edge Enhancement Transformer for Hyperspectral Change Detection
Deep learning methods have achieved remarkable performance in hyperspectral image (HSI) change detection. However, they still face two primary limitations. First, most approaches treat all spectral channels equally, overlooking the varying importance of spectral information. Second, they often struggle to clearly delineate edges in low-resolution images, resulting in suboptimal feature extraction and increased false or missed detections. To address these challenges, we propose a novel Spectral Frequency-Induced Edge Enhancement Transformer (SFIEET) network to enhance the accuracy and robustness of HSI change detection. The proposed SFIEET framework comprises three key modules. The Spectral Frequency and Spatial Domain Fusion (SFSDF) module implements an adaptive frequency-domain attention mechanism based on grouped DCT, dynamically focusing on channels rich in weak semantic information. Simultaneously, it integrates multi-scale contextual spatial information to capture morphological changes such as texture fragmentation, effectively enhancing recognition of structural changes. The Sharpening and Laplace-Guided Edge Enhancement (SLGEE) module integrates multi-scale pooling with the Laplacian operator to fuse edge features across spectral bands, enabling precise boundary localization and more reliable detection in blurred regions. Finally, the Depthwise Convolution-Embedded Lightweight Transformer (DW-ELT) module incorporates depthwise separable convolutions into the Transformer architecture to enhance local feature modeling, while gated linear units (GLUs) strengthen nonlinear representations, making change-region features more discriminative. Extensive experiments on three benchmark datasets demonstrate that the proposed SFIEET consistently outperforms existing state-of-the-art methods. The source code is available at https://github.com/bcshi83/SFIEET.git for the sake of reproducibility.
- 12 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Geoscience and Remote Sensing
- Vol. PP (01962892) , 1
SFIEET: Spectral Frequency-Induced Edge Enhancement Transformer for Hyperspectral Change Detection
Deep learning methods have achieved remarkable performance in hyperspectral image (HSI) change detection. However, they still face two primary limitations. First, most approaches treat all spectral channels equally, overlooking the varying importance of spectral information. Second, they often struggle to clearly delineate edges in low-resolution images, resulting in suboptimal feature extraction and increased false or missed detections. To address these challenges, we propose a novel Spectral Frequency-Induced Edge Enhancement Transformer (SFIEET) network to enhance the accuracy and robustness of HSI change detection. The proposed SFIEET framework comprises three key modules. The Spectral Frequency and Spatial Domain Fusion (SFSDF) module implements an adaptive frequency-domain attention mechanism based on grouped DCT, dynamically focusing on channels rich in weak semantic information. Simultaneously, it integrates multi-scale contextual spatial information to capture morphological changes such as texture fragmentation, effectively enhancing recognition of structural changes. The Sharpening and Laplace-Guided Edge Enhancement (SLGEE) module integrates multi-scale pooling with the Laplacian operator to fuse edge features across spectral bands, enabling precise boundary localization and more reliable detection in blurred regions. Finally, the Depthwise Convolution-Embedded Lightweight Transformer (DW-ELT) module incorporates depthwise separable convolutions into the Transformer architecture to enhance local feature modeling, while gated linear units (GLUs) strengthen nonlinear representations, making change-region features more discriminative. Extensive experiments on three benchmark datasets demonstrate that the proposed SFIEET consistently outperforms existing state-of-the-art methods. The source code is available at https://github.com/bcshi83/SFIEET.git for the sake of reproducibility.
- 19 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Industrial Electronics
- Vol. 73 (4) , 5873-5883
Primal-Dual-Based Differential $k$-WTA Network for Competitive Coordination in Multi-Robot Systems
In multi-robot systems, the $\boldsymbol{k}$-winners-take-all ($\boldsymbol{k}$-WTA) problem serves as a competitive selection mechanism to identify the $\boldsymbol{k}$ most suitable robots from a group of $\boldsymbol{n}$ ($\boldsymbol{n} > \boldsymbol{k}$) ones for a given task. Existing $\boldsymbol{k}$-WTA networks either fail to eliminate lagging errors or involve a complex modeling process due to the introduction of additional Lagrangian variables to deal with bounded constraints. To address these challenges, a primal-dual-based differential $\boldsymbol{k}$-WTA (PDD-$\boldsymbol{k}$WTA) network is proposed in this article. This network incorporates time-derivative information to mitigate lagging errors and leverages linear variational inequalities (LVIs) to streamline the network structure. Theoretical analyses and comparative simulations validate the effectiveness and superiority of the PDD-$\boldsymbol{k}$WTA network. Furthermore, a multi-robot competitive coordination (MCC) model, which integrates the PDD-$\boldsymbol{k}$WTA network, a consensus estimator, and a robot control law, is developed to achieve the tracking of a mobile target.
- 19 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Industrial Electronics
- Vol. 73 (4) , 5873-5883
Primal-Dual-Based Differential $k$-WTA Network for Competitive Coordination in Multi-Robot Systems
In multi-robot systems, the $\boldsymbol{k}$-winners-take-all ($\boldsymbol{k}$-WTA) problem serves as a competitive selection mechanism to identify the $\boldsymbol{k}$ most suitable robots from a group of $\boldsymbol{n}$ ($\boldsymbol{n} > \boldsymbol{k}$) ones for a given task. Existing $\boldsymbol{k}$-WTA networks either fail to eliminate lagging errors or involve a complex modeling process due to the introduction of additional Lagrangian variables to deal with bounded constraints. To address these challenges, a primal-dual-based differential $\boldsymbol{k}$-WTA (PDD-$\boldsymbol{k}$WTA) network is proposed in this article. This network incorporates time-derivative information to mitigate lagging errors and leverages linear variational inequalities (LVIs) to streamline the network structure. Theoretical analyses and comparative simulations validate the effectiveness and superiority of the PDD-$\boldsymbol{k}$WTA network. Furthermore, a multi-robot competitive coordination (MCC) model, which integrates the PDD-$\boldsymbol{k}$WTA network, a consensus estimator, and a robot control law, is developed to achieve the tracking of a mobile target.