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

Zhenya Zang; Xingda Li; David Day Uei Li

eess.IVMar 27, 2026

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

Authors: Zhenya Zang, Xingda Li, David Day Uei Li

Organizations: Department of Biomedical Engineering, University of Strathclyde, 16 Richmond Street, Glasgow, G1 1XQ, United Kingdom

Abstract

This work presents a lightweight super-resolution (LiteSR) neural network for depth and intensity images acquired from a consumer-grade single-photon avalanche diode (SPAD) array with a 48x32 spatial resolution. The proposed framework reconstructs high-resolution (HR) images of size 256x256. Both synthetic and real datasets are used for performance evaluation. Extensive quantitative metrics demonstrate high reconstruction fidelity on synthetic datasets, while experiments on real indoor and outdoor measurements further confirm the robustness of the proposed approach. Moreover, the SPAD sensor is interfaced with an Arduino UNO Q microcontroller, which receives low-resolution (LR) depth and intensity images and feeds them into a compressed, pre-trained deep learning (DL) model, enabling real-time SR video streaming. In addition to the 256x256 setting, a range of target HR resolutions is evaluated to determine the maximum achievable upscaling resolution (512x512) with LiteSR, including scenarios with noise-corrupted LR inputs. The proposed LiteSR-embedded system co-design provides a scalable, cost-effective solution to enhance the spatial resolution of current consumer-grade SPAD arrays to meet HR imaging requirements.

Figures & tables

Fig. 1: LiteSR architecture, composed of intensity-guided depth and intensity branches. The numbers of channels for the EARB, RLFB, Depth Head, and Intensity Head are denoted by CH_E, CH_R, CH_DH, and CH_IH.

Fig. 2: Comparison of super-resolution results on intensity test images for $4\times$ upscaling. From left to right: GT, LR, bicubic interpolation, LiteSR (FP32), and LiteSR (INT8). Quantitative metrics (MS-SSIM, PSNR, and GMSD) are reported for each reconstructed image. The input LR image has dimensions 32x48, resulting in different aspect ratios for the output HR images.

Fig. 3: Comparison of super-resolution results on depth test images for $4\times$ upscaling. From left to right: GT, LR, bicubic interpolation, LiteSR (FP32), and LiteSR (INT8). The Input LR image has dimensions 32x48, resulting in different aspect ratios for the output HR images.

Fig. 4: Hot pixel detection and compensation in photon counting mode using a white flat target. (a) Raw mean intensity map showing spatially distributed hot pixels with anomalously high DCRs. (b) Binary hot-pixel mask generated by thresholding at 250 photon counts. (c) Corrected intensity image obtained by replacing hot pixels with the local $3\times 3$ neighborhood average.

Fig. 5: SR results on real-world intensity captures. RGB views of the scenes are shown together with bicubic, LiteSR (FP32), and LiteSR (INT8) reconstructions and their LV scores.

Fig. 6: SR results on real-world depth maps. Bicubic, LiteSR (FP32), and LiteSR (INT8) outputs are compared using depth smoothness scores.

Fig. 7: System overview of the TMF8829-based depth sensing pipeline. (a) System architecture. (b) Software stack for configuration, acquisition, and sensor control. (c) Inference latency and model size for FP32 and INT8 LiteSR models across four channel configurations.

Fig. 8: Results on a stationary plush toy scene. (a) RGB reference photograph. (b) Raw LR depth map from the TMF8829 ToF sensor. (c) Upscaled depth map produced by LiteSR (INT8). (d) Raw LR intensity image. (e) Upscaled and normalized intensity image from LiteSR (INT8).

Fig. 9: Comparison of super-resolution results on depth test images for $8\times$ upscaling. From left to right: GT, LR, bicubic interpolation, LiteSR (FP32), and LiteSR (INT8).

Upscale ratio $4\times$ (256 $\times$ 256)
Model	Params	Param Size	MACs	OPs*
LiteSR FP32	355,949	1.36 MB	2.043 G	4.086 G
LiteSR INT8	93,133	363.80 KB	0.973 G	1.945 G
Upscale ratio $8\times$ (512 $\times$ 512)
LiteSR FP32	402,221	1.53 MB	5.148 G	10.296 G
LiteSR INT8	139,405	544.55 KB	4.077 G	8.155 G

TABLE I: Model Complexity Comparison of LiteSR Under FP32 and INT8 Precision for

4\times

and

8\times

Upscaling

Scene	SSIM $\uparrow$		PSNR (dB) $\uparrow$		RMSE $\downarrow$		$\delta<1.25$ $\uparrow$
Scene	w/	w/o	w/	w/o	w/	w/o	w/	w/o
Art	0.930	0.929	36.26	36.17	0.043	0.044	1.00	1.00
Books	0.987	0.987	45.12	44.89	0.016	0.016	1.00	1.00
Bowling	0.982	0.982	41.30	41.09	0.024	0.025	1.00	1.00
Dolls	0.988	0.988	46.53	46.16	0.013	0.014	1.00	1.00
Laundry	0.979	0.978	42.22	41.91	0.022	0.022	1.00	1.00

TABLE II: Ablation Study Evaluating the Contribution of EARB in the Depth Branch. Bold and underscored values indicate the better- and same-performing configurations for each metric.

Scene	MS-SSIM $\uparrow$		PSNR (dB) $\uparrow$		GMSD $\downarrow$		nGMSD $\downarrow$
Scene	w/	w/o	w/	w/o	w/	w/o	w/	w/o
Art	0.910	0.906	24.47	24.34	0.0014	0.0014	0.0044	0.0046
Books	0.912	0.910	23.55	23.43	0.0019	0.0019	0.0062	0.0064
Bowling	0.963	0.961	30.42	30.20	0.0006	0.0006	0.0036	0.0038
Dolls	0.906	0.902	23.47	23.28	0.0015	0.0016	0.0043	0.0045
Laundry	0.867	0.864	21.72	21.64	0.0026	0.0027	0.0073	0.0074

TABLE III: Ablation Study Evaluating the Contribution of RLFB in the Intensity Branch. Bold and underscored values indicate the better- and same-performing configurations for each metric.

Fig. 10: Comparison of super-resolution results on intensity test images for $8\times$ upscaling. From left to right: GT, LR, bicubic interpolation, LiteSR (FP32), and LiteSR (INT8).

Image Super-Resolution

Image super-resolution reconstructs high-resolution images from lower-resolution observations, balancing detail, fidelity, and computational cost.

See all 248 papers

Practical Noise Modeling for SPAD Intensity Imaging
Aug 1, 2026Wendi Liu, Yujie Lu, Zengxi Zhang +3Computational Imaging Image Denoising
The University of Tokyo
FreeTransformSR: Efficient Lightweight Image Super-Resolution via Free Low-Rank Learnable Transform
Sep 5, 2026Hongji Li, Yunhui LiHybrid CNN-Transformer Architectures Image Super-Resolution
Chinese Academy of Sciences Changchun Institute of Optics Fine Mechanics and Physics, Changchun, Jilin 130033, China · University of the Chinese Academy of Sciences, Beijing 100049, China
Physics-Informed Untrained Learning for RGB-Guided Superresolution Single-Pixel Hyperspectral Imaging
Apr 4, 2026Hao Zhang, Bilige Xu, Lichen Wei +2Hyperspectral Image Reconstruction Hyperspectral Imaging
College of Science, Northwest Agriculture & Forestry University, Yangling 712100, China · Key Laboratory of Photoelectronic Imaging Technology and System, School of Optics and Photonics, Beijing Institute of Technology, Beijing, China · State Key Laboratory of Digital Manufacturing Equipment and Technology, Huazhong University of Science and Technology, Wuhan 430074, China

Aug 1, 2026J/K move · Enter open · S save

Practical Noise Modeling for SPAD Intensity Imaging

Wendi Liu, Yujie Lu, Zengxi Zhang, Haiyang Jiang+2

The University of Tokyo

Computational Imaging Image Denoising

Open paper AskPDF arXiv

Single-photon avalanche diode (SPAD) cameras are promising for low-light and high-dynamic-range intensity imaging, but their practical use is limited by complex sensor-specific noise. Unlike time-correlated single-photon counting (TCSPC) systems, SPAD cameras record whether at least one detection occurred in each gate without photon timestamps in intensity imaging mode, making explicit noise decomposition difficult. We present a practical noise modeling and calibration framework for SPAD intensity denoising. Our forward model describes binary-frame accumulation with a Binomial observation process, models signal-independent dark noise as an exposure-dependent pure dark count term plus an exposure-independent dark-frame bias term, and incorporates pixel-wise response non-uniformity. We design a dedicated calibration procedure for the proposed model and use it to build a count-domain noise-synthesis pipeline for network training. For denoising, we further design a SPAD-specific dark-shading correction (SPAD-DSC) to remove most systematic noise before network training. We construct a real-world SPAD intensity dataset for testing. Experimental results demonstrate the superiority of the proposed noise model.

Abstract

Image Super-Resolution

Explore similar work

Practical Noise Modeling for SPAD Intensity Imaging

FreeTransformSR: Efficient Lightweight Image Super-Resolution via Free Low-Rank Learnable Transform

Physics-Informed Untrained Learning for RGB-Guided Superresolution Single-Pixel Hyperspectral Imaging