Reproduce “On-Device Super Resolution Imaging Using Low-Cost SPAD Array and Embedded Lightweight Deep Learning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zhenya Zang, Xingda Li, David Day Uei Li
Researchers at the University of Strathclyde developed LiteSR, a lightweight deep learning model that performs dual-branch super-resolution on low-resolution depth and intensity images from a 48x32 SPAD array. The system, deployed on an Arduino UNO Q, achieves real-time on-device inference at approximately 4.5-5.8 frames per second, enhancing spatial detail and accuracy for both modalities.
Key insights
- Lightweight deep learning architectures with residual learning and cross-modal guidance can effectively overcome spatial resolution limitations of consumer-grade SPAD arrays without external high-resolution sensors.
- Low-bit quantization (INT8) through QAT substantially reduces model size and computational cost (74% parameter reduction, >50% MAC reduction) with minimal performance degradation, making on-device deployment feasible.
- Real-time, on-device super-resolution of dual-modality SPAD data is achievable on resource-constrained microcontrollers, enabling immediate enhanced visual data at the edge for various applications.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.