Active Non-Line-Of-Sight Imaging Based on Fusion of Physical Prior and Deep Learning

Non-line-of-sight imaging technology allows for the reconstruction of hidden targets, with wide-ranging applications in fields such as autonomous driving, disaster response, and security. Traditional physics-based methods are often complex and prone to errors, while deep learning approaches commonly rely on 3D volume reconstruction, which can lead to the high computational costs and limited image quality. To overcome these challenges, this paper presents a novel imaging model that incorporates both channel and spatial attention mechanisms to selectively suppress irrelevant features and enhance important ones for better representation. The proposed dual-branch design improves feature quality by facilitating more efficient information extraction, while the multi-scale large-kernel attention mechanism with dilation enhances the model's robustness against variations in input data. Experimental results demonstrate that our network achieves high-quality reconstruction with significantly reduced computational complexity on both synthetic and real-world datasets.