Bridging deep learning with real-world physical systems — from restoring degraded images to optimizing next-generation wireless networks. My work spans computer vision, graph neural networks, and communication systems.
- 🎓 Research Focus: Image Restoration · Low-Level Vision · Multimodal Learning · Wireless Communications
- 🧠 Interests: Efficient deep learning, knowledge distillation, implicit neural representations, graph neural networks
- 📫 Contact: Open an issue or PR on my repositories
A Unified Physical Model with INR & Knowledge Distillation
A parameter-efficient (0.10M params) framework tackling dehazing, low-light enhancement, and deraining in a single model:
- Unified Physical Model (UPM) — physically interpretable degradation factors
- Implicit Neural Representation (INR) — coordinate-based MLP for fine-grained recovery
- DINOv2 Distillation — semantic knowledge transfer from vision foundation models
- Frequency-Spatial Joint Loss — FFT-domain + spatial + distillation loss
PyTorch DINOv2 INR Knowledge Distillation
Low-Dose CT Image Denoising with Depth Guidance
A two-stage denoising framework for low-dose CT (LDCT) that jointly learns image restoration and depth estimation:
- Contrastive Prior Estimation — models noise via intensity & contrast analysis
- Depth-Guided Denoising — dual-task co-learning, exploiting depth-noise correlation
- Discrepancy-Aware Mechanism — weighted attention on high-discrepancy regions
- LEGM & SCAB — custom attention modules for multi-scale feature extraction
PyTorch Medical Imaging Contrastive Learning Depth Estimation
Node-Centric Feature Aggregation for IoT Scene Understanding
Optimizes IRS-aided THz MIMO resource allocation via heterogeneous graph neural networks:
- NCMG (Node-Centric Multimodal Graph) — fuses wireless channel features with ResNet visual features via Cross-Stitch units
- Joint Optimization — beamforming, IRS phase shifts, and bandwidth allocation in one end-to-end model
- Differentiable Channel Model — physics-based THz channel with path loss & molecular absorption
PyTorch GNN Heterogeneous Graph Cross-Modal Fusion THz Communication
Languages & Frameworks
CV & Vision
Tools & More
| ⭐ Stars |
📦 Repos |
🔧 Commits |
🐍 Top Lang |
Python · C++ · MATLAB · Jupyter