ing at the problem of non-line-of-sight imaging under incoherent illumination, we propose a solution based on deep learning. Combining the classical semantic segmentation and residual model in the field of computer vision, a URNet network structure is constructed and the classical bottleneck layer structure is improved. The experimental results show that the improved model has more details of recovery images and generalization ability. Compared with speckle autocorrelation imaging method under incoherent illumination, the recovery performance of this method is greatly improved.