Chen Kang

Chen Kang

Peking University

← Back to Home

🧠 Neuromorphic Vision

Neuromorphic vision represents a paradigm shift in visual sensing, inspired by biological visual systems. My research focuses on leveraging spike cameras and event cameras to capture high-speed visual information with unprecedented temporal resolution.

First-Author Papers

SpikeReveal
### SpikeReveal: Unlocking Temporal Sequences from Real Blurry Inputs with Spike Streams **NeurIPS 2024 - Spotlight** We develop a self-supervised spike-guided image deblurring framework that addresses the synthetic-to-real domain gap. The core insight is that spike streams encode precise temporal information about scene motion, enabling sharp frame reconstruction without ground truth supervision. **Key Contributions:** - Self-supervised deblurring without paired training data - Theoretical analysis of spike-blur image fusion - State-of-the-art on real-world blurry datasets \[[Paper](https://arxiv.org/abs/2403.09486)\] \[[Code](https://github.com/chenkang455/S-SDM)\]
SpikeCLIP
### SpikeCLIP: Rethinking High-speed Image Reconstruction Framework with Spike Camera **AAAI 2025** We introduce CLIP-based supervision for spike-to-image reconstruction. Instead of relying solely on pixel-level supervision, we leverage CLIP's semantic understanding to guide reconstruction, achieving competitive performance with a lightweight network. **Key Contributions:** - First CLIP-based supervision for spike reconstruction - Lightweight network with strong performance - High-quality image generation pipeline \[[Paper](https://arxiv.org/abs/2501.04477)\] \[[Code](https://github.com/chenkang455/SpikeCLIP)\]
TRMD
### TRMD: Motion Deblur by Learning Residual from Events **IEEE TMM 2024** We propose a Two-Stage Residual-based Motion Deblurring framework for event cameras. Rather than directly predicting sharp images, we estimate residual sequences that bridge blurry inputs and sharp outputs. **Key Contributions:** - Two-stage residual-based deblurring framework - Residual sequences as intermediate supervision - Effective event-image fusion strategy \[[Paper](https://ieeexplore.ieee.org/document/10403964)\] \[[Code](https://github.com/chenkang455/TRMD)\]
Spike-Zoo
### Spike-Zoo: A Toolbox for Spike-to-Image Reconstruction An open-source benchmark providing state-of-the-art pretrained models for spike-to-image reconstruction tasks. \[[GitHub](https://github.com/chenkang455/Spike-Zoo)\] \[[Doc](https://spike-zoo.readthedocs.io/)\]