Wan-Chi Siu
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Research Interests
Wan-Chi Siu has a total of 82 co-authors. Here are the top co-authors by number of shared publications:
- Yui-Lam Chan
Department of Electrical and Electronic Engineering, The Hong Kong Polytechnic University, Hong Kong, China
14 shared publications - Sik-Ho Tsang
Hong Kong Chu Hai College
9 shared publications - Zhi-Song Liu
Lappeenranta-Lahti University of Technology
7 shared publications - Ngai Fong Law
The Hong Kong Polytechnic University
6 shared publications Zhi-Song Liu
Center Multimedia Signal Processing, Department of Electronic and Information Engineering (EIE), Hong Kong Polytechnic University, Hung Hum, Hong Kong. 100077
5 shared publications- Wei Kuang
Department of Electronic and Information Engineering, The Hong Kong Polytechic University, Hong Kong Hong Kong
4 shared publications
Latest Publications
- Open Access
- 3 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Signal Processing Letters
- Vol. 33 (10709908) , 1386-1390
Quality-Assisted Domain Transfer for Fast Face Super-Resolution
Face super-resolution (FSR) is a challenging task, especially when the low-resolution (LR) image does not have sufficient information to build a high-resolution (HR) image. In this paper, we propose a quality-assisted domain transfer framework that uses a two-stage pipeline for efficient face SR. Instead of treating all LR images uniformly, we use a quality evaluator to classify images in different LR domains with distinct quality levels and determine the starting points for domain transfer adaptively. We further design a new residual domain learning strategy that integrates the LR domain with learned residual information. This approach allows us to avoid the correction formulation in diffusion models and the momentum correction in the originally proposed domain transfer approach and enables the model to learn the transformation from the LR domain to the HR domain gradually while preserving content information. The experimental results demonstrate competitive performance with the best inference speed among all approaches.
- 3 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Signal Processing Letters
- Vol. 33 (10709908) , 1386-1390
Open AccessQuality-Assisted Domain Transfer for Fast Face Super-Resolution
Face super-resolution (FSR) is a challenging task, especially when the low-resolution (LR) image does not have sufficient information to build a high-resolution (HR) image. In this paper, we propose a quality-assisted domain transfer framework that uses a two-stage pipeline for efficient face SR. Instead of treating all LR images uniformly, we use a quality evaluator to classify images in different LR domains with distinct quality levels and determine the starting points for domain transfer adaptively. We further design a new residual domain learning strategy that integrates the LR domain with learned residual information. This approach allows us to avoid the correction formulation in diffusion models and the momentum correction in the originally proposed domain transfer approach and enables the model to learn the transformation from the LR domain to the HR domain gradually while preserving content information. The experimental results demonstrate competitive performance with the best inference speed among all approaches.
- Open Access
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Image Processing
- p. 1
DTLS-Face-Inpaint: A Novel Inpainting Framework with Latent Space Domain Transfer
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Image Processing
- p. 1
Open AccessDTLS-Face-Inpaint: A Novel Inpainting Framework with Latent Space Domain Transfer
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Signal Processing Letters
- Vol. 33, 1-5
Domain Fusion in Latent Space for Face Video Super-Resolution
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Signal Processing Letters
- Vol. 33, 1-5
Domain Fusion in Latent Space for Face Video Super-Resolution
- Open Access
PUMPS: Skeleton-Agnostic Point-based Universal Motion Pre-Training for Synthesis in Human Motion Tasks
Motion skeletons drive 3D character animation by transforming bone hierarchies, but differences in proportions or structure make motion data hard to transfer across skeletons, posing challenges for data-driven motion synthesis. Temporal Point Clouds (TPCs) offer an unstructured, cross-compatible motion representation. Though reversible with skeletons, TPCs mainly serve for compatibility, not for direct motion task learning. Doing so would require data synthesis capabilities for the TPC format, which presents unexplored challenges regarding its unique temporal consistency and point identifiability. Therefore, we propose PUMPS, the primordial autoencoder architecture for TPC data. PUMPS independently reduces frame-wise point clouds into sampleable feature vectors, from which a decoder extracts distinct temporal points using latent Gaussian noise vectors as sampling identifiers. We introduce linear assignment-based point pairing to optimise the TPC reconstruction process, and negate the use of expensive point-wise attention mechanisms in the architecture. Using these latent features, we pre-train a motion synthesis model capable of performing motion prediction, transition generation, and keyframe interpolation. For these pre-training tasks, PUMPS performs remarkably well even without native dataset supervision, matching state-of-the-art performance. When fine-tuned for motion denoising or estimation, PUMPS outperforms many respective methods without deviating from its generalist architecture.
Open AccessPUMPS: Skeleton-Agnostic Point-based Universal Motion Pre-Training for Synthesis in Human Motion Tasks
Motion skeletons drive 3D character animation by transforming bone hierarchies, but differences in proportions or structure make motion data hard to transfer across skeletons, posing challenges for data-driven motion synthesis. Temporal Point Clouds (TPCs) offer an unstructured, cross-compatible motion representation. Though reversible with skeletons, TPCs mainly serve for compatibility, not for direct motion task learning. Doing so would require data synthesis capabilities for the TPC format, which presents unexplored challenges regarding its unique temporal consistency and point identifiability. Therefore, we propose PUMPS, the primordial autoencoder architecture for TPC data. PUMPS independently reduces frame-wise point clouds into sampleable feature vectors, from which a decoder extracts distinct temporal points using latent Gaussian noise vectors as sampling identifiers. We introduce linear assignment-based point pairing to optimise the TPC reconstruction process, and negate the use of expensive point-wise attention mechanisms in the architecture. Using these latent features, we pre-train a motion synthesis model capable of performing motion prediction, transition generation, and keyframe interpolation. For these pre-training tasks, PUMPS performs remarkably well even without native dataset supervision, matching state-of-the-art performance. When fine-tuned for motion denoising or estimation, PUMPS outperforms many respective methods without deviating from its generalist architecture.
- Open Access
- 11 December 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Consumer Electronics
- Vol. 71 (2) , 3575-3586
Back Projection Generative Strategy for Low and Normal Light Image Pairs With Enhanced Statistical Fidelity and Diversity
Low light image enhancement (LLIE) using supervised deep learning is limited by the scarcity of matched low/normal light image pairs. We propose Back Projection Normal-to-Low Diffusion Model (N2LDiff-BP), a novel diffusion-based generative model that realistically transforms normal-light images into diverse low-light counterparts. By injecting noise perturbations over multiple timesteps, our model synthesizes low-light images with authentic noise, blur, and color distortions. We introduce innovative architectural components -Back Projection Attention, BP Feedforward, and BP Transformer Blocks -that integrate back projection to model the narrow dynamic range and nuanced noise of real low-light images. Experiment and results show N2LDiff-BP significantly outperforms prior augmentation techniques, enabling effective data augmentation for robust LLIE. We also introduce LOL-Diff, a large-scale synthetic low-light dataset. Our novel framework, architectural innovations, and dataset advance deep learning for low-light vision tasks by addressing data scarcity. N2LDiff-BP establishes a new state-of-the-art in realistic low-light image synthesis for LLIE.
- 11 December 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Consumer Electronics
- Vol. 71 (2) , 3575-3586
Open AccessBack Projection Generative Strategy for Low and Normal Light Image Pairs With Enhanced Statistical Fidelity and Diversity
Low light image enhancement (LLIE) using supervised deep learning is limited by the scarcity of matched low/normal light image pairs. We propose Back Projection Normal-to-Low Diffusion Model (N2LDiff-BP), a novel diffusion-based generative model that realistically transforms normal-light images into diverse low-light counterparts. By injecting noise perturbations over multiple timesteps, our model synthesizes low-light images with authentic noise, blur, and color distortions. We introduce innovative architectural components -Back Projection Attention, BP Feedforward, and BP Transformer Blocks -that integrate back projection to model the narrow dynamic range and nuanced noise of real low-light images. Experiment and results show N2LDiff-BP significantly outperforms prior augmentation techniques, enabling effective data augmentation for robust LLIE. We also introduce LOL-Diff, a large-scale synthetic low-light dataset. Our novel framework, architectural innovations, and dataset advance deep learning for low-light vision tasks by addressing data scarcity. N2LDiff-BP establishes a new state-of-the-art in realistic low-light image synthesis for LLIE.
- Open Access
- 11 December 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Consumer Electronics
- Vol. 71 (1) , 379-387
Photo-Realistic Talking Face Generation Under Latent Space Manipulation
This paper focuses on generating photo-realistic talking face videos by leveraging on semantic facial attributes in a latent space and capturing the talking style from an old video of a speaker. We formulate a process to manipulate facial attributes in the latent space by identifying semantic facial directions. We develop a deep learning pipeline to learn the correlation between the audio and the corresponding video frames from a reference video of a speaker in an aligned latent space. This correlation is used to navigate a static face image into frames of a talking face video, which is moderated by three carefully constructed loss functions, for accurate lip synchronization and photo-realistic video reconstruction. By combining these techniques, we aim to generate high-quality talking face videos that are visually realistic and synchronized with the provided audio input. Our results were evaluated against some state-of-the-art techniques on talking face generation, and we have recorded significant improvements in the image quality of the generated talking face video.
- 11 December 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Consumer Electronics
- Vol. 71 (1) , 379-387
Open AccessPhoto-Realistic Talking Face Generation Under Latent Space Manipulation
This paper focuses on generating photo-realistic talking face videos by leveraging on semantic facial attributes in a latent space and capturing the talking style from an old video of a speaker. We formulate a process to manipulate facial attributes in the latent space by identifying semantic facial directions. We develop a deep learning pipeline to learn the correlation between the audio and the corresponding video frames from a reference video of a speaker in an aligned latent space. This correlation is used to navigate a static face image into frames of a talking face video, which is moderated by three carefully constructed loss functions, for accurate lip synchronization and photo-realistic video reconstruction. By combining these techniques, we aim to generate high-quality talking face videos that are visually realistic and synchronized with the provided audio input. Our results were evaluated against some state-of-the-art techniques on talking face generation, and we have recorded significant improvements in the image quality of the generated talking face video.
- Open Access
- 31 October 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Image Processing
- Vol. 33 (10577149) , 6324-6339
AnlightenDiff: Anchoring Diffusion Probabilistic Model on Low Light Image Enhancement
Low-light image enhancement aims to improve the visual quality of images captured under poor illumination. However, enhancing low-light images often introduces image artifacts, color bias, and low SNR. In this work, we propose AnlightenDiff, an anchoring diffusion model for low light image enhancement. Diffusion models can enhance the low light image to well-exposed image by iterative refinement, but require anchoring to ensure that enhanced results remain faithful to the input. We propose a Dynamical Regulated Diffusion Anchoring mechanism and Sampler to anchor the enhancement process. We also propose a Diffusion Feature Perceptual Loss tailored for diffusion based model to utilize different loss functions in image domain. AnlightenDiff demonstrates the effect of diffusion models for low-light enhancement and achieving high perceptual quality results. Our techniques show a promising future direction for applying diffusion models to image enhancement.
- 31 October 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Image Processing
- Vol. 33 (10577149) , 6324-6339
Open AccessAnlightenDiff: Anchoring Diffusion Probabilistic Model on Low Light Image Enhancement
Low-light image enhancement aims to improve the visual quality of images captured under poor illumination. However, enhancing low-light images often introduces image artifacts, color bias, and low SNR. In this work, we propose AnlightenDiff, an anchoring diffusion model for low light image enhancement. Diffusion models can enhance the low light image to well-exposed image by iterative refinement, but require anchoring to ensure that enhanced results remain faithful to the input. We propose a Dynamical Regulated Diffusion Anchoring mechanism and Sampler to anchor the enhancement process. We also propose a Diffusion Feature Perceptual Loss tailored for diffusion based model to utilize different loss functions in image domain. AnlightenDiff demonstrates the effect of diffusion models for low-light enhancement and achieving high perceptual quality results. Our techniques show a promising future direction for applying diffusion models to image enhancement.
- 30 October 2024
- Published by Elsevier in Expert Systems with Applications
- Vol. 262
Unveiling image source: Instance-level camera device linking via context-aware deep Siamese network
- 30 October 2024
- Published by Elsevier in Expert Systems with Applications
- Vol. 262
Unveiling image source: Instance-level camera device linking via context-aware deep Siamese network
- Open Access
- 9 April 2024
- Published by Elsevier in Journal of Visual Communication and Image Representation
- Vol. 100
Edge fusion back projection GAN for large scale face super resolution
- 9 April 2024
- Published by Elsevier in Journal of Visual Communication and Image Representation
- Vol. 100
Open AccessEdge fusion back projection GAN for large scale face super resolution
- 1 May 2022
- Published by SPIE-Intl Soc Optical Eng
- Vol. 12177, 132-430
Robust lane detection through automatic trajectory analysis with deep learning and big data environment
This paper gives focus on multi-lane detection from traffic cameras, which is based on automatic trajectory analysis and is promoted with advanced deep-learning technologies. Our proposed approach is based on big trajectory data that is robust to complex road scenes, which makes our approach particularly reliable and practical for Intelligent Transportation Systems. By using the deep learning object detection technology, it firstly generates big trajectory data on the road. Then, it detects the stop lines on the road and counts the number of lanes from the trajectories. Next, the trajectories are divided into different groups, where each group contains the trajectories of one lane. Finally, the lanes are fitted by the grouped trajectories. A large number of experiments have been done. The results show that the proposed approach can effectively detect the lanes on the road.
- 1 May 2022
- Published by SPIE-Intl Soc Optical Eng
- Vol. 12177, 132-430
Robust lane detection through automatic trajectory analysis with deep learning and big data environment
This paper gives focus on multi-lane detection from traffic cameras, which is based on automatic trajectory analysis and is promoted with advanced deep-learning technologies. Our proposed approach is based on big trajectory data that is robust to complex road scenes, which makes our approach particularly reliable and practical for Intelligent Transportation Systems. By using the deep learning object detection technology, it firstly generates big trajectory data on the road. Then, it detects the stop lines on the road and counts the number of lanes from the trajectories. Next, the trajectories are divided into different groups, where each group contains the trajectories of one lane. Finally, the lanes are fitted by the grouped trajectories. A large number of experiments have been done. The results show that the proposed approach can effectively detect the lanes on the road.