A hybrid perceptron with cross-domain transferability towards active steady-state non-line-of-sight imaging

Abstract

Active steady-state non-line-of-sight (NLOS) imaging entails the acquisition and processing of continuous multi-bounce NLOS signals to facilitate the recovery of hidden scenes. Recently, learning-based methods have demonstrated competitive performance in NLOS imaging. However, most of them inadequately capture the underlying features inherent in acquired signals and fail to effectively exploit prior information from hidden scenes, thus constraining their ability to alleviate the ill-posedness. To address the above limitations, we propose a novel NLOS signal processing framework—hybrid perceptron with cross-domain transferability (HP-CDT). The HP enhances the utilization of primitive features through a hierarchical pooling and feature fusion (HPFI) mechanism while comprehensively capturing underlying correlations within the signals via local and global perception. Besides, it facilitates cross-level interactions and fusion of various features, thereby enriching the feature representation. The cross-domain transfer (CDT) strategy leverages line-of-sight (LOS) latent representations as priors to steer the NLOS feature extraction, facilitating the optimization of NLOS latent representations. Rendering experiments and practical assessment indicate that, compared with existing methods, our approach achieves superior imaging quality while maintaining a light-weight architecture for efficient deployment.