Passive Non-Line-of-Sight imaging reconstruction based on dual input U-Net

In this paper, we propose a novel passive non-line-of-sight (NLOS) imaging method using a dual-input neural network algorithm. Our method employs a dual-input U-Net architecture that takes two RGB blurred images of different sizes as inputs. To enhance the extraction of shallow features from the smaller image, we introduce a new input convolutional attention encoder (ICAE) module, which works in tandem with attention modules to weigh shallow features with deep semantic information and improve scene reconstruction accuracy. We evaluate our method on a public NLOS dataset and demonstrate its superior performance in terms of SSIM and PSNR compared to existing methods. Additionally, our results show that the dual-input network outperforms the single-input network in terms of reconstruction accuracy.