Lightweight multi-scale feature fusion with attention guidance for passive non-line-of-sight imaging

Passive non-line-of-sight (NLOS) imaging is an emerging technology to enhance the perception of hidden objects. Existing methods often suffer from issues of imaging quality and model complexity. To address these challenges, we propose a lightweight multi-scale NLOS imaging network (LMS-NLOS). Our network combines a multi-scale encoder-decoder structure with a secondary detail-enhanced transformer module to capture fine-grained details. An asymmetric cross-fusion module is introduced to fuse shallow and deep features, reducing noise and redundancy. Additionally, a multi-scale loss function (MSLoss) is designed to enhance contour features and guide model training. To ensure model efficiency, the LMS-NLOS employs a perception-enhanced feed-forward network with a spatial shift operation (PEFNs). Experimental results on the VIS dataset collected by our laboratory and the public NLOS-Passive dataset demonstrate the effectiveness of our proposed methods. Compared with existing approaches, our multi-scale NLOS imaging network (MS-NLOS) achieves higher imaging quality, with a PSNR of 24.80 dB and an SSIM of 0.9234 on the VIS dataset. Meanwhile, LMS-NLOS offers a more compact model, reducing the model size by nearly half while maintaining good imaging performance. It achieves a PSNR of 24.26 dB and an SSIM of 0.8697 on the VIS dataset. Our work highlights the potential of attention-guided multi-scale feature fusion for lightweight passive NLOS imaging. The code is available at https://github.com/CS-wpf/LMS-NLOS.