| Non-line-of-sight imaging is an emerging technology in the field of computational imaging,which has important research value in automatic driving,medical imaging,and anti-terrorism rescue.Compared with traditional imaging technology,non-line-of-sight imaging technology can collect information returned from target objects utilizing scattering bodies such as walls and floors,to realizing imaging of targets outside the field of view,such as behind the corners of walls,indoor objects behind windows,etc.This effectively solves the problem that traditional imaging solutions cannot image targets hidden outside the blind spot of the field of view and expands the imaging range.In the non-line-of-sight imaging scenario,Poisson noise has a large impact on the quality of non-line-of-sight imaging due to the presence of occlusion between the detection target and the imaging device resulting in a large reduction in effective echo photons.Traditional image Poisson noise reduction algorithms have long iteration times,fixed modes and manual parameter settings.Deep learning as an efficient image recovery algorithm has been studied and applied to the field of Poisson noise reduction with good results.Base on the building of a non-line-of-sight imaging system based on the photon time-of-flight synchronization measurement,this paper carries out the research on Poisson noise reduction technology for non-line-of-sight imaging based on deep learning to further improve the quality of non-line-of-sight imaging.The main research contents and results are as follows:(1)A simulation model of the non-line-of-sight domain imaging system based on the bi-directional scattering distribution function is designed.Through the approximate modeling of the optical properties of random rough surfaces,the scattering angle distribution of photons after a collision is modeled according to the bidirectional scattering distribution function,finally,the motion trajectory of photons in the non-line-of-sight domain scenario is tracked and simulated by using Monte Carlo method.The influence of photon number,distance,and reflection characteristics of the middle interface on the imaging results are analyzed by comparing the simulation with the actual experiment,and the correctness of the model is verified.(2)The Attention Enhanced Feature(AEF-Net)network based on the attention mechanism is designed.The images in the BSD200 and SET11 datasets were used as imaging targets,respectively,and the datasets for training the network were produced using a non-line-of-sight imaging simulation model.The simulated data are used to optimize the training of the network,and experiments were designed to compare the noise reduction effect of the proposed network AEF-Net with that of the conventional noise reduction algorithms,and to compare the effect of the network trained by different noise models on removing Poisson noise in the non-line-of-sight domain scene.The experimental results show that AEF-Net has a better noise reduction effect than the traditional Poisson image noise reduction algorithm,and the data set generated by the designed non-line-of-sight domain simulation model can better improve the Poisson noise reduction performance of the network.(3)A multi-scale feature fusion image denoising network based on the multi-label network M-Net is designed.To achieve cross-channel-level pixel fusion,the nonlinear-activation function-free image fusion module(Nonlinear-activation Feature Fusion,NFF-Block)and the multi-deconvolution head transposition attention cascade fusion module(Multi-scale Feature Fusion,MFF-Block)are designed respectively.Experiments were designed to compare the noise reduction performance of M-Net,NFF-Net and MFF-Net in non-line-of-sight domain simulations and actual experimental results,and the experimental results showed that the network MFF-Net composed of the multi-deconvolution head transposition attention cascade fusion module has better noise reduction performance. |
