Passive non-line-of-sight imaging with diffuse-aware attention-enhanced encoding.

Passive non-line-of-sight (NLOS) imaging reconstructs occluded targets using ambient light but suffers from low SNR and severe signal attenuation in deep networks. Unlike generic attentions, a diffuse-aware attention module (DAAM) is proposed that leverages two physical priors: the anisotropic angular distribution of diffuse reflections and the channel-wise SNR disparity. DAAM uses deformable convolution for anisotropic spatial attention and mean-std pooling for frequency-aware channel attention, fused by a learnable gate. Embedded in a residual-attention encoder, DAAM preserves weak signals while enhancing discriminative features. Experimental results show that DAAM offers significant performance improvements over traditional passive NLOS methods and generic attention mechanisms, achieving higher PSNR and better perceptual fidelity (LPIPS). These results demonstrate that incorporating physical priors into attention design is key to high-quality passive NLOS reconstruction.