On-Device Super Resolution Imaging Using Low-Cost SPAD Array and Embedded Lightweight Deep Learning

A team of researchers has unveiled a breakthrough in affordable, high-resolution imaging by combining a low-cost SPAD sensor with an embedded AI model. Their system, called LiteSR, uses a lightweight deep learning network to perform super-resolution on-the-fly. It takes the extremely low-resolution 48x32 pixel output from a consumer-grade Single-Photon Avalanche Diode (SPAD) array—a type of sensor that detects single photons for depth and intensity data—and reconstructs it into crisp 256x256 or even 512x512 images. This represents an 8x to 16x resolution boost, all processed locally on a simple Arduino UNO microcontroller to enable real-time video streaming without needing cloud computing.

The innovation lies in the co-design of the sensor hardware and the compressed AI software. The LiteSR model is specifically engineered to be small and efficient enough to run on embedded systems while maintaining high reconstruction fidelity, even with noisy input data. The researchers validated the system using both synthetic datasets and real-world indoor/outdoor measurements, confirming its robustness. This approach provides a scalable and cost-effective path to dramatically enhance the spatial resolution of current mass-market SPAD sensors, which are crucial for applications like LiDAR, low-light photography, and 3D sensing.