The light-cone transform (LCT) is a mathematical technique used in non-line-of-sight (NLOS) imaging; it is used to reconstruct hidden objects by analyzing the temporal evolution of scattered light paths. Wiener filtering is commonly employed to reduce noise that often distorts these three-dimensional reconstructions; however, its low-pass nature may attenuate high-frequency details, thereby compromising reconstruction accuracy and overall quality. In this work, we propose a learned light-cone transform (${\mathrm{L}}^2\mathrm{CT}$) method that rapidly converts transient data into three-dimensional hidden objects while preserving intricate physical details. We introduce a high-frequency-enhanced Wiener-filtering technique that retains low-frequency components while capturing high-frequency details through learnable parameterization, enabling more accurate reconstruction of hidden objects. Additionally, we integrate spatiotemporal feature-extraction and volume-projection modules to create an end-to-end NLOS image-reconstruction network. Extensive experiments demonstrate that our method significantly improves LCT reconstruction on real-world datasets captured by different NLOS imaging systems. The ${\mathrm{L}}^2\mathrm{CT}$ method surpasses existing approaches, achieving a 5-dB improvement in peak signal-to-noise ratio over the LCT and reaching state-of-the-art performance in NLOS reconstruction.

Close

PrevNext

View figure in article

• Classical optics
• Image processing
• Inverse problems
• Laser applications
• Confocal imaging
• Deep learning
• Machine learning
• Neural network simulations