Deep-learning enhanced super-resolution imaging with low-cost single-photon avalanche diodes

This study presents a non-fusion super-resolution (SR) solution to enhance the performance of low-cost, consumer-grade single-photon avalanche diode (SPAD) arrays. We present a compact deep learning (DL) model that takes low-resolution (LR, 8×8) depth and intensity inputs and simultaneously reconstructs high-resolution (HR, 50×50) images. The model was evaluated on synthetic datasets spanning diverse scenes and real measurements from an STMicroelectronics VL53L8CX SPAD array. Results show high fidelity against ground truth images for synthetic datasets, and more precise structural details in real datasets. To facilitate hardware deployment, the model was further compressed using INT8 quantization, resulting in only a marginal accuracy loss. Both the original and quantized models achieve video-rate SR reconstruction on a mid-range GPU. Owing to its compact size, the DL model is well-suited to modern edge-computing platforms and offers strong potential for mobile and embedded applications.