Multispectral non-line-of-sight imaging via deep fusion photography

Passive non-line-of-sight (NLOS) imaging is a promising technique that extends visual perception to hidden objects around the corner, offering advantages such as low-cost, portability, and real-time. However, the low quality of current passive NLOS images remains a significant barrier to field application of NLOS targets imaging at long standoffs. This study introduces a multispectral NLOS imaging approach utilizing a deep fusion framework to reconstruct images from visible, short-wavelength infrared, and long-wavelength infrared raw data captured by portable devices. The nonlinear representation capabilities and learnable activation function of the Kolmogorov-Arnold network (KAN) are particularly suited to the inverse light field transmission model in NLOS imaging, enhancing the interpretability of the deep neural network. Experimental results demonstrate that this deep fusion photography method provides satisfied performance to image the occluded individuals despite the polynomial attenuation of effective signals with increasing distance between hidden objects and the relay wall. Notably, the passive NLOS experiments reveal successful imaging of hidden people at distance >5 m from the relay wall. Remarkably, even at distances three times greater than those in previous studies, quantitative metrics validate the superior performance of the proposed method in the task of passive NLOS imaging. © Science China Press 2025.