Yang Zhao
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Yang Zhao has a total of 138 co-authors. Here are the top co-authors by number of shared publications:
- Weijia Jia
Institute of Artificial Intelligence and Future Networks, Beijing Normal University, Zhuhai, China
38 shared publications - Ronggang Wang
Guangdong Provincial Key Laboratory of Ultra High Definition Immersive Media Technology, Shenzhen Graduate School, Peking University, Shenzhen, China
17 shared publications - Yuan Chen
School of Internet, Anhui University, Hefei, China
16 shared publications - Zhao Zhang
Hefei University of Technology, Hefei, China
12 shared publications - Xiaoping Liu
School of Computer and Information, Hefei University of Technology, Hefei, China
8 shared publications - Hai Min
School of Computer Science and Information Engineering, Hefei University of Technology, Hefei, China
7 shared publications
Latest Publications
- Open Access
- 9 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 5359-5368
Dual-Camera Video Demoiréing for Smartphones
- 9 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 5359-5368
Open AccessDual-Camera Video Demoiréing for Smartphones
- Open Access
- 9 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 15899-15908
Descriptive Video Degradation Assessment with Multimodal Large Language Models
- 9 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 15899-15908
Open AccessDescriptive Video Degradation Assessment with Multimodal Large Language Models
- Open Access
- 9 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 3570-3579
Uncertainty-Aware Consensus 3D Gaussian Splatting for High-Fidelity Lensless Novel View Synthesis
- 9 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 3570-3579
Open AccessUncertainty-Aware Consensus 3D Gaussian Splatting for High-Fidelity Lensless Novel View Synthesis
- 9 October 2026
- Published by Springer Nature in International Journal of Computer Vision
- Vol. 134 (10) , 469
Ultra-High-Definition Image Restoration via Prior-driven Patch Bilateral Compensation
- 9 October 2026
- Published by Springer Nature in International Journal of Computer Vision
- Vol. 134 (10) , 469
Ultra-High-Definition Image Restoration via Prior-driven Patch Bilateral Compensation
- Open Access
- 8 October 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Multimedia Computing, Communications, and Applications
- Vol. 22 (10) , 1-22
Boosting Representation Learning for High-Level Semantic Information in Facial Expression Recognition
Facial Expression Recognition (FER) requires high-level semantic cues to interpret abstract emotions. However, existing FER approaches commonly suffer from sparse semantic descriptions and unwanted noise from redundant facial pixels. To tackle these issues, we propose a boosting representation learning framework to build a robust high-level semantic projection space connecting textual semantics and facial images. To mitigate semantic scarcity for FER, we develop the Expression High-Level Semantic Encoding (EHSE) module. Guided by FACS-based physiological priors, this module constructs dense semantic prototypes to provide fine-grained supervised information for model learning. To remove irrelevant background noise mixed with redundant facial visual content, the Expression Key-Core Features Focusing (EKFF) module is built, which picks out discriminative facial cues, boosting the signal-to-noise ratio and semantic purity of extracted visual features. To eliminate ambiguous classification boundaries within the semantic projection space, the Expression High-Level Semantic Alignment (EHSA) module is designed, which is equipped with multi-dimensional constraints. Experiments on FER+, RAF-DB, and AffectNet verify the superiority of our approach. Our method obtains state-of-the-art results on FER+ and RAF-DB and achieves the optimal average accuracy over AffectNet’s 7-class and 8-class tasks. These results demonstrate that our framework can construct a balanced, discriminative, and semantically intact projection space for FER.
- 8 October 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Multimedia Computing, Communications, and Applications
- Vol. 22 (10) , 1-22
Open AccessBoosting Representation Learning for High-Level Semantic Information in Facial Expression Recognition
Facial Expression Recognition (FER) requires high-level semantic cues to interpret abstract emotions. However, existing FER approaches commonly suffer from sparse semantic descriptions and unwanted noise from redundant facial pixels. To tackle these issues, we propose a boosting representation learning framework to build a robust high-level semantic projection space connecting textual semantics and facial images. To mitigate semantic scarcity for FER, we develop the Expression High-Level Semantic Encoding (EHSE) module. Guided by FACS-based physiological priors, this module constructs dense semantic prototypes to provide fine-grained supervised information for model learning. To remove irrelevant background noise mixed with redundant facial visual content, the Expression Key-Core Features Focusing (EKFF) module is built, which picks out discriminative facial cues, boosting the signal-to-noise ratio and semantic purity of extracted visual features. To eliminate ambiguous classification boundaries within the semantic projection space, the Expression High-Level Semantic Alignment (EHSA) module is designed, which is equipped with multi-dimensional constraints. Experiments on FER+, RAF-DB, and AffectNet verify the superiority of our approach. Our method obtains state-of-the-art results on FER+ and RAF-DB and achieves the optimal average accuracy over AffectNet’s 7-class and 8-class tasks. These results demonstrate that our framework can construct a balanced, discriminative, and semantically intact projection space for FER.
- Open Access
- 3 October 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Multimedia Computing, Communications, and Applications
Controlling the Latent Diffusion Model for Generative Image Shadow Removal via Residual Generation
Large-scale generative models have achieved remarkable advancements in various visual tasks, yet their application to shadow removal in images remains challenging. These models often generate diverse, realistic details without adequate focus on fidelity, failing to meet the crucial requirements of shadow removal, which necessitates precise preservation of image content. In contrast to prior approaches that aimed to regenerate shadow-free images from scratch, this paper utilizes diffusion models to generate and refine image residuals. This strategy fully uses the inherent detailed information within shadowed images, resulting in a more efficient and faithful reconstruction of shadow-free content. Additionally, to prevent the accumulation of errors during the generation process, a cross-timestep self-enhancement training strategy is proposed. This strategy leverages the network itself to augment the training data, not only increasing the volume of data but also enabling the network to dynamically correct its generation trajectory, ensuring a more accurate and robust output. In addition, to address the loss of original details in the process of image encoding and decoding of large generative models, a content-preserved encoder-decoder structure is designed with a control mechanism and multi-scale skip connections to achieve high-fidelity shadow-free image reconstruction. Experimental results demonstrate that the proposed method can reproduce high-quality results based on a large latent diffusion prior and faithfully preserve the original contents in shadow regions.
- 3 October 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Multimedia Computing, Communications, and Applications
Open AccessControlling the Latent Diffusion Model for Generative Image Shadow Removal via Residual Generation
Large-scale generative models have achieved remarkable advancements in various visual tasks, yet their application to shadow removal in images remains challenging. These models often generate diverse, realistic details without adequate focus on fidelity, failing to meet the crucial requirements of shadow removal, which necessitates precise preservation of image content. In contrast to prior approaches that aimed to regenerate shadow-free images from scratch, this paper utilizes diffusion models to generate and refine image residuals. This strategy fully uses the inherent detailed information within shadowed images, resulting in a more efficient and faithful reconstruction of shadow-free content. Additionally, to prevent the accumulation of errors during the generation process, a cross-timestep self-enhancement training strategy is proposed. This strategy leverages the network itself to augment the training data, not only increasing the volume of data but also enabling the network to dynamically correct its generation trajectory, ensuring a more accurate and robust output. In addition, to address the loss of original details in the process of image encoding and decoding of large generative models, a content-preserved encoder-decoder structure is designed with a control mechanism and multi-scale skip connections to achieve high-fidelity shadow-free image reconstruction. Experimental results demonstrate that the proposed method can reproduce high-quality results based on a large latent diffusion prior and faithfully preserve the original contents in shadow regions.
- 23 May 2026
- Published by World Scientific Pub Co Pte Ltd in International Journal of Image and Graphics
A Performance Evaluation of Transformer and Mamba Models for Palmprint and Palm Vein Recognition
Palmprint and palm vein recognition are vital biometric technologies for secure authentication. The primary challenge lies in modeling the complex, unstructured textures of palmprints and the continuous vascular topology of palm veins, both of which require capturing long-range spatial dependencies. Traditional handcrafted and CNN-based methods partially address these challenges but are constrained by local receptive fields and manually designed features. In contrast, Transformers leverage global self-attention to model cross-scale correlations in unstructured palmprint patterns, while Mamba state space models (SSMs) offer linear computational complexity and hardware efficiency, making them well-suited for capturing continuous topological structures in palm veins. We present the first systematic evaluation of Transformer and Mamba architectures for palmprint and palm vein recognition, covering 13 public datasets (eight palmprint, five palm vein) and 17 deep models (nine Transformer variants, eight Mamba architectures), alongside traditional methods and CNN baselines. Key findings include: (i) XCiT achieves the best palmprint recognition (open-set ACC = 95.31%, EER = 1.75%) via cross-covariance attention; (ii) MambaVision achieves the best palm vein recognition (open-set ACC = 91.25%, TAR@FAR [ ] = 76.81%) via hybrid CNN–SSM design; (iii) synthetic data augmentation improves few-shot performance by up to 12.6% ACC on challenging datasets while exhibiting boundary effects on high-quality data. We further quantify the accuracy–efficiency trade-off, offering practical deployment insights and highlighting the complementary roles of Transformer and Mamba architectures.
- 23 May 2026
- Published by World Scientific Pub Co Pte Ltd in International Journal of Image and Graphics
A Performance Evaluation of Transformer and Mamba Models for Palmprint and Palm Vein Recognition
Palmprint and palm vein recognition are vital biometric technologies for secure authentication. The primary challenge lies in modeling the complex, unstructured textures of palmprints and the continuous vascular topology of palm veins, both of which require capturing long-range spatial dependencies. Traditional handcrafted and CNN-based methods partially address these challenges but are constrained by local receptive fields and manually designed features. In contrast, Transformers leverage global self-attention to model cross-scale correlations in unstructured palmprint patterns, while Mamba state space models (SSMs) offer linear computational complexity and hardware efficiency, making them well-suited for capturing continuous topological structures in palm veins. We present the first systematic evaluation of Transformer and Mamba architectures for palmprint and palm vein recognition, covering 13 public datasets (eight palmprint, five palm vein) and 17 deep models (nine Transformer variants, eight Mamba architectures), alongside traditional methods and CNN baselines. Key findings include: (i) XCiT achieves the best palmprint recognition (open-set ACC = 95.31%, EER = 1.75%) via cross-covariance attention; (ii) MambaVision achieves the best palm vein recognition (open-set ACC = 91.25%, TAR@FAR [ ] = 76.81%) via hybrid CNN–SSM design; (iii) synthetic data augmentation improves few-shot performance by up to 12.6% ACC on challenging datasets while exhibiting boundary effects on high-quality data. We further quantify the accuracy–efficiency trade-off, offering practical deployment insights and highlighting the complementary roles of Transformer and Mamba architectures.
- 6 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems for Video Technology
- Vol. 36 (7) , 10144-10157
Joint Resolution and Rendering Artifacts Removal for Cloud Gaming Image
With the rapid development of the cloud gaming industry, low-quality rendering and rescaling strategies are commonly employed to mitigate the high costs of cloud-based computation and bandwidth. As a result, client-side images often contain artifacts such as mixed aliasing and resolution distortions, which cannot be effectively handled by current super-resolution models. In response, this paper proposes a cloud gaming image enhancement (CGIE) model to tackle both rendering and resolution degradations. Initially, this paper builds a large dataset by rendering and rescaling paired data with different qualities from collected 3D game scenes. Subsequently, a lightweight dual-branch enhancement network is designed, which consists of a high-frequency branch primarily focused on detail enhancement and a sampling-space branch aimed at enlarging the receptive field and perceiving multi-scale aliasing artifacts. Experimental results demonstrate the superior anti-aliasing and image enhancement performance of the proposed method across various real-world cloud games and even mobile games. The dataset and codes are available at https://github.com/YCheno/CGIE/.
- 6 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems for Video Technology
- Vol. 36 (7) , 10144-10157
Joint Resolution and Rendering Artifacts Removal for Cloud Gaming Image
With the rapid development of the cloud gaming industry, low-quality rendering and rescaling strategies are commonly employed to mitigate the high costs of cloud-based computation and bandwidth. As a result, client-side images often contain artifacts such as mixed aliasing and resolution distortions, which cannot be effectively handled by current super-resolution models. In response, this paper proposes a cloud gaming image enhancement (CGIE) model to tackle both rendering and resolution degradations. Initially, this paper builds a large dataset by rendering and rescaling paired data with different qualities from collected 3D game scenes. Subsequently, a lightweight dual-branch enhancement network is designed, which consists of a high-frequency branch primarily focused on detail enhancement and a sampling-space branch aimed at enlarging the receptive field and perceiving multi-scale aliasing artifacts. Experimental results demonstrate the superior anti-aliasing and image enhancement performance of the proposed method across various real-world cloud games and even mobile games. The dataset and codes are available at https://github.com/YCheno/CGIE/.
- 17 February 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Consumer Electronics
- Vol. 72 (2) , 3479-3493
SIGNAL: Stereo Image Neural Learner for Continuous Super-Resolution
Stereo images are widely used in many scenarios such as autonomous driving and stereoscopic display. Stereo image super-resolution is a common technique to obtain high-definition stereo images. However, existing methods are mainly designed for fixed scales and cannot smoothly support arbitrary upscaling. To address these problems, this paper proposes a Stereo ImaGe NeurAl Learner (SIGNAL) based on implicit neural representation, which consists of three parts. First, a feature extraction module leveraging the iPASSR architecture to robustly extract hierarchical features from low-resolution stereo images. Second, a novel stereo implicit representation module that explicitly models the inherent geometric consistency and contextual relationships between the two views, enhancing the latent features through view-specific and cross-view interactions. Third, a cross-view decoding module designed to fully exploit the complementary information across views during the image reconstruction process. Extensive experiments across diverse datasets demonstrate that the proposed SIGNAL framework achieves superior performance. The results underscore the effectiveness of our approach in handling arbitrary scaling factors while preserving visual fidelity in stereo image restoration.
- 17 February 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Consumer Electronics
- Vol. 72 (2) , 3479-3493
SIGNAL: Stereo Image Neural Learner for Continuous Super-Resolution
Stereo images are widely used in many scenarios such as autonomous driving and stereoscopic display. Stereo image super-resolution is a common technique to obtain high-definition stereo images. However, existing methods are mainly designed for fixed scales and cannot smoothly support arbitrary upscaling. To address these problems, this paper proposes a Stereo ImaGe NeurAl Learner (SIGNAL) based on implicit neural representation, which consists of three parts. First, a feature extraction module leveraging the iPASSR architecture to robustly extract hierarchical features from low-resolution stereo images. Second, a novel stereo implicit representation module that explicitly models the inherent geometric consistency and contextual relationships between the two views, enhancing the latent features through view-specific and cross-view interactions. Third, a cross-view decoding module designed to fully exploit the complementary information across views during the image reconstruction process. Extensive experiments across diverse datasets demonstrate that the proposed SIGNAL framework achieves superior performance. The results underscore the effectiveness of our approach in handling arbitrary scaling factors while preserving visual fidelity in stereo image restoration.
- Top Cited Papers
- 16 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Multimedia
- Vol. 28 (15209210) , 3652-3666
Learning Dual Modality Interactions for Event-Based Motion Deblurring
Event cameras hold great potential for motion deblurring because they capture motion information with microsecond precision, offering robustness to motion blur. However, the limited interaction between RGB frames and event streams presents a significant challenge, preventing the full utilization of the event cameras' unique advantages. To address this, we propose Dual frame-event Interaction and introduce a multi-scale Network structure, DuInt-Net. DuInt-Net aims to tackle two key challenges: (1) enhancing the representational and interaction capabilities between RGB frames and event streams, and (2) adaptively selecting richer visual features for improved motion deblurring. We introduce an event-frame joint interaction module that consists of three branches: a base branch, a global awareness attention branch, and a local enhancement attention branch. The base branch processes essential pixel-level features that retain the original structural information. The global branch integrates event data to improve large-scale motion understanding, while the local branch uses large-kernel convolutions to refine fine-grained details in RGB frames. For superior reconstruction performance, we also propose the event-guided multi-scale fusion attention module, which effectively combines local visual information and global frame-event relationships. Extensive experiments demonstrate that DuInt-Net achieves superior performance, both quantitatively and qualitatively, showcasing its superior motion deblurring capabilities.
- 16 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Multimedia
- Vol. 28 (15209210) , 3652-3666
Top Cited PapersLearning Dual Modality Interactions for Event-Based Motion Deblurring
Event cameras hold great potential for motion deblurring because they capture motion information with microsecond precision, offering robustness to motion blur. However, the limited interaction between RGB frames and event streams presents a significant challenge, preventing the full utilization of the event cameras' unique advantages. To address this, we propose Dual frame-event Interaction and introduce a multi-scale Network structure, DuInt-Net. DuInt-Net aims to tackle two key challenges: (1) enhancing the representational and interaction capabilities between RGB frames and event streams, and (2) adaptively selecting richer visual features for improved motion deblurring. We introduce an event-frame joint interaction module that consists of three branches: a base branch, a global awareness attention branch, and a local enhancement attention branch. The base branch processes essential pixel-level features that retain the original structural information. The global branch integrates event data to improve large-scale motion understanding, while the local branch uses large-kernel convolutions to refine fine-grained details in RGB frames. For superior reconstruction performance, we also propose the event-guided multi-scale fusion attention module, which effectively combines local visual information and global frame-event relationships. Extensive experiments demonstrate that DuInt-Net achieves superior performance, both quantitatively and qualitatively, showcasing its superior motion deblurring capabilities.