增强反射的被动非视域成像 U-NET 重构

Non-line-of-sight（NLOS）imaging reconstructs hidden scenes by capturing indirect optical signals reflected from targets occluded by obstacles via intermediate surfaces，and has broad applications in security monitoring，autonomous driving，and related fields. Existing methods often suffer from weak effective signals，strong noise，and limited reconstruction accuracy due to insufficient exploitation of the nonLambertian reflection characteristics of intermediate surfaces. To address these limitations，a passive NLOS imaging experimental system was constructed using tile intermediate surfaces of different materials，and the NLOS-Passive-TILES-stl10 dataset was established，comprising eight groups and more than 800 000 projected images. An enhanced-reflection improved U-Net architecture（ER-UNET）was proposed for passive NLOS imaging，incorporating a reflection feature extractor，differentiable CLAHE，adaptive instance normalization，and other modules. RGB images are transformed into Y，R-Y，and B-Y channels，with an increased weight assigned to the Y channel. Multi-scale features are extracted through an encoder-decoder structure，and SVD-based low-rank decomposition is performed to enable gradient aware RGB reconstruction. Experimental results demonstrate that ER-UNET achieves an average PSNR of 15. 14 dB and an SSIM of 0. 50 on the NLOS-Passive test set，outperforming NLOS-OT and C-GAN. The proposed method also exhibits strong generalization ability on test data such as faces and icons，providing a promising approach for the practical application of passive NLOS imaging.