A model decomposition method for the real-time non-line-of-sight imaging

Real-time non-line-of-sight (NLOS) imaging faces a fundamental trade-off between acquisition efficiency and reconstruction quality. Although transient methods achieve high visual fidelity, they depend on extensive data collection. Regularization-based approaches allow undersampled reconstruction but often incur prohibitive computational costs. To overcome these limitations, we introduce MD-NLOS, a model decomposition method that formulates NLOS reconstruction as a least absolute shrinkage and selection operator (LASSO) problem improved by spectral filtering. By solving the optimization in the frequency domain, the method achieves notable computational efficiency. Furthermore, the reformulated problem can be effectively solved using simple gradient descent, avoiding the need for complex optimization schemes. The results show that our method reconstructs synthetic 256 × 256 and experimental 128 × 128 images, using only 36 and 64 sampling points, with reconstruction times of 3.1 and 4.6 s, respectively, yielding a structural similarity index (SSIM) of 0.7352, which is approximately 6-fold higher than that of FK.