Huipeng Zhang
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Huipeng Zhang has a total of 7 co-authors. Here are the top co-authors by number of shared publications:
- Kuizhi Mei
State Key Laboratory of Human-Machine Hybrid Augmented Intelligence, Institute of Artificial Intelligence and Robotics, Xian Jiaotong University, Xian, China
5 shared publications - Xuyi Yu
State Key Laboratory of Human-Machine Hybrid Augmented Intelligence, Institute of Artificial Intelligence and Robotics, Xian Jiaotong University, Xian, China
5 shared publications - Yinghai Zhao
Beijing Huahang Institute of Radio Measurement, Beijing, China
3 shared publications Miaomiao Jin
Xi’an Jiaotong University
1 shared publication- Shang Lou
Research Institute of Jiangtong, Photovoltaic Technology Co. Ltd, Guangzhou, 510000, China
1 shared publication - Yaohua Liu
Guangdong Institute of Intelligence Science and Technology, Zhuhai, China
1 shared publication
Latest Publications
- 29 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems for Video Technology
- Vol. PP (10518215) , 1
Beyond Single Supervision: Multi-perspective Quantization-Aware Distillation for Semantic Segmentation
Efficient model compression is essential for deploying semantic segmentation models on edge devices. While quantization-aware training combined with knowledge distillation (KD) has emerged as a dominant paradigm, existing methods face a critical challenge: quantized teacher features provide deployment-compatible supervision but suffer from information loss, while full-precision features retain rich semantic details but exhibit representation mismatch with quantized students. Our analysis reveals that while single supervision sources each have inherent limitations, their combination produces superior results, suggesting the existence of more favorable intermediate representations in the feature space. To systematically explore this advantageous representation space, we propose Multi-perspective Feature Distillation (MFD). Our framework constructs multiple virtual quantized teachers by strategically perturbing the quantization parameters of a single full-precision teacher, generating diverse supervisory signals that bridge the gap between full-precision and quantized representations. This multi-perspective approach guides the student model toward learning robust features that maintain semantic richness while accommodating quantization constraints, all with minimal computational overhead. Comprehensive experiments show that MFD brings substantial improvements over existing distillation methods under 8-bit settings, while being efficient to train and readily integrable with existing KD frameworks. The source code will be available at https://github.com/Fateeeeee/MFD.
- 29 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems for Video Technology
- Vol. PP (10518215) , 1
Beyond Single Supervision: Multi-perspective Quantization-Aware Distillation for Semantic Segmentation
Efficient model compression is essential for deploying semantic segmentation models on edge devices. While quantization-aware training combined with knowledge distillation (KD) has emerged as a dominant paradigm, existing methods face a critical challenge: quantized teacher features provide deployment-compatible supervision but suffer from information loss, while full-precision features retain rich semantic details but exhibit representation mismatch with quantized students. Our analysis reveals that while single supervision sources each have inherent limitations, their combination produces superior results, suggesting the existence of more favorable intermediate representations in the feature space. To systematically explore this advantageous representation space, we propose Multi-perspective Feature Distillation (MFD). Our framework constructs multiple virtual quantized teachers by strategically perturbing the quantization parameters of a single full-precision teacher, generating diverse supervisory signals that bridge the gap between full-precision and quantized representations. This multi-perspective approach guides the student model toward learning robust features that maintain semantic richness while accommodating quantization constraints, all with minimal computational overhead. Comprehensive experiments show that MFD brings substantial improvements over existing distillation methods under 8-bit settings, while being efficient to train and readily integrable with existing KD frameworks. The source code will be available at https://github.com/Fateeeeee/MFD.
- 1 April 2026
- Published by Springer Nature in Machine Vision and Applications
- Vol. 37 (3) , 49
DKM-Net: self-supervised 3D occupancy prediction via multi-scale dynamic kernel method
- 1 April 2026
- Published by Springer Nature in Machine Vision and Applications
- Vol. 37 (3) , 49
DKM-Net: self-supervised 3D occupancy prediction via multi-scale dynamic kernel method
- 30 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. 13 (12) , 27181-27196
Toward Compact Representations: A Region-Based Logit Distillation Framework for Edge-Friendly Semantic Segmentation
Semantic segmentation enables fine-grained scene understanding for the Internet of Things (IoT) applications such as intelligent transportation and UAV-based monitoring. However, high-resolution segmentation models impose substantial memory and computational demands, making them impractical for resource-constrained IoT edge devices. Knowledge distillation (KD) transfers knowledge from a large teacher model to a compact student model. Yet, conventional logit distillation in segmentation suffers from two limitations: 1) dense pixel-wise alignment introduces spatial redundancy and 2) erroneous teacher predictions propagate to the student, degrading reliability. To address these issues, this article proposes a region-based logit distillation (RLD) framework. Instead of distilling all logits, RLD summarizes logit distributions within categories to reduce redundancy. To avoid interference from incorrect logits, we isolate their influence using label guidance and then summarize logits for a batch of samples via intraclass similarity. Teacher-guided singular value decomposition (SVD) facilitates logit dimensionality reduction. RLD retains only a small set of principal logits, for example, 19 for Cityscapes, achieving performance comparable to dense distillation. Combined with classic distillation baselines, RLD further achieves state-of-the-art results: 77.80% mIoU on Cityscapes and 69.90% on CamVid. Deployed on an NVIDIA Jetson Xavier NX, our distilled model achieves 46.78 FPS at 14.05 W in FP16 and 85.12 FPS at 7.73 W with INT8 quantization, demonstrating real-time capability and high energy efficiency for edge IoT applications.
- 30 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. 13 (12) , 27181-27196
Toward Compact Representations: A Region-Based Logit Distillation Framework for Edge-Friendly Semantic Segmentation
Semantic segmentation enables fine-grained scene understanding for the Internet of Things (IoT) applications such as intelligent transportation and UAV-based monitoring. However, high-resolution segmentation models impose substantial memory and computational demands, making them impractical for resource-constrained IoT edge devices. Knowledge distillation (KD) transfers knowledge from a large teacher model to a compact student model. Yet, conventional logit distillation in segmentation suffers from two limitations: 1) dense pixel-wise alignment introduces spatial redundancy and 2) erroneous teacher predictions propagate to the student, degrading reliability. To address these issues, this article proposes a region-based logit distillation (RLD) framework. Instead of distilling all logits, RLD summarizes logit distributions within categories to reduce redundancy. To avoid interference from incorrect logits, we isolate their influence using label guidance and then summarize logits for a batch of samples via intraclass similarity. Teacher-guided singular value decomposition (SVD) facilitates logit dimensionality reduction. RLD retains only a small set of principal logits, for example, 19 for Cityscapes, achieving performance comparable to dense distillation. Combined with classic distillation baselines, RLD further achieves state-of-the-art results: 77.80% mIoU on Cityscapes and 69.90% on CamVid. Deployed on an NVIDIA Jetson Xavier NX, our distilled model achieves 46.78 FPS at 14.05 W in FP16 and 85.12 FPS at 7.73 W with INT8 quantization, demonstrating real-time capability and high energy efficiency for edge IoT applications.
- 30 January 2026
- Published by Elsevier in Neural Networks
- Vol. 199, 108663
Efficient semantic segmentation via logit-guided feature distillation
- 30 January 2026
- Published by Elsevier in Neural Networks
- Vol. 199, 108663
Efficient semantic segmentation via logit-guided feature distillation