基于深度学习的单光子非视域成像泊松降噪  

Poisson Noise Suppression for Single-Photon Non-Line-of-Sight Imaging Based on Deep Learning

作　　者：涂敏[1] 鄢秋荣[1] 郑勇健 熊贤成 邹荃 戴钱玲 卢孝强[2] Tu Min;Yan Qiurong;Zheng Yongjian;Xiong Xiancheng;Zou Quan;Dai Qianling;Lu Xiaoqiang(School of Information Engineering,Nanchang University,Nanchang 330031,Jiangxi,China;Xi’an Institute of Optics and Precision Mechanics,Chinese Academy of Sciences,Xi’an 710119,Shaanxi,China)

机构地区：[1]南昌大学信息工程学院,江西南昌330031 [2]中国科学院西安光学精密机械研究所,陕西西安710119

出　　处：《激光与光电子学进展》2023年第20期86-93,共8页Laser & Optoelectronics Progress

基　　金：国家自然科学基金(61865010,62165009);江西省研究生创新专项基金(YC2021-S146)。

摘　　要：在非视域成像场景中,有效的回波光子大量减少,泊松噪声对非视域成像的质量影响较大。传统图像泊松降噪算法存在迭代时间长、模式固定和手动设置参数等问题。为提高非视域成像质量,设计一种基于深度学习的单光子非视域成像泊松降噪方法。为解决训练样本不足的问题,利用几何光学近似和蒙特卡罗方法对非视域场景下的光子运动轨迹进行追踪建模,对非视域成像过程进行仿真,利用仿真数据重建的泊松噪声图像制作数据集。设计基于注意力机制的特征增强降噪网络(AEF-Net),利用仿真数据对网络进行优化训练。最后,搭建一套非视域成像系统对网络的泊松降噪性能进行验证。实验结果表明所提AEF-Net去除非视域场景下的泊松噪声效果优于传统降噪算法。In non-line-of-sight imaging scenes,while effective echo photons are reduced to a great extent,Poisson noise largely impacts the non-line-of-sight imaging quality.Moreover,issues such as long iteration time,fixed mode,and manual parameter setting have been identified with traditional image Poisson noise suppression algorithms.Therefore,to improve the quality of non-line-of-sight imaging,this study designed a deep learning-based Poisson noise suppression method for single-photon non-line-of-sight imaging.First,geometrical optics approximation and Monte Carlo methods were implemented to track and model the photon motion trajectory in the non-line-of-sight scene,simulate the non-line-of-sight imaging process,produce a dataset using the Poisson noise images reconstructed from the simulation data,and address the problem of insufficient training samples.Subsequently,we designed an attention-based feature-enhanced noise suppression network(AEF-Net),followed by optimization and training of the network using simulation data.Furthermore,we built a non-line-of-sight imaging system to verify the Poisson noise suppression performance of the network.The experimental results show an outperformance of our designed noise suppression attention-based feature-enhanced noise suppression network than the conventional noise suppression algorithms for removing Poisson noise from non-line-of-sight scenes.

关 键 词：非视域成像 仿真分析 深度学习 泊松降噪 

分 类 号：O438[机械工程—光学工程]