DFAR-Net: Dual-Input Three-Branch Attention Fusion Reconstruction Network for Polarized Non-Line-of-Sight Imaging

Polarized non-line-of-sight (NLOS) imaging is a promising visual perception technique for enhancing the visibility of occluded objects hidden behind walls. The main challenge of this task is that conventional single-angle relay wall projection polarization images provide limited effective information due to optical ill-posedness, resulting in poor imaging results. To address this problem, we designed a dual-input three-branch attention fusion reconstruction network, namely DFAR-Net, which utilizes our proposed channel-by-channel bilateral weighted fusion sub-network to fuse NLOS infrared polarization intensity and polarization degree image information and reconstruct hidden scenes. In the feature extraction part, we introduce a split kernel channel attention mechanism to emphasize or suppress features, aiming to improve the model’s generalization ability and robustness. Additionally, to enhance the reconstruction quality, we employ a combination of multi-scale loss functions to optimize the model’s expressive power. Experimental results on our self-collected full-optical polarization NLOS dataset, PI-ND, demonstrate the superior performance of DFAR-Net and its modules over current passive NLOS imaging methods.