基于远场焦斑测量数学模型改进的CNN去噪方法研究  

Research on CNN Denoising Algorithm Based on an Improved Mathematical Model for the Measurement of Far-field Focal Spot

作　　者：王拯洲[1] 王力 谭萌 段亚轩[1] 王伟[1] 田新锋[1] 魏际同 WANG Zheng-zhou;WANG Li;TAN Meng;DUAN Ya-xuan;WANG Wei;TIAN Xin-feng;WEI Ji-tong(Xi'an Institute of Optics and Precision Mechanics,Chinese Academy of Science,Xi'an 710119,China)

机构地区：[1]中国科学院西安光学精密机械研究所,西安710119

出　　处：《光子学报》2020年第12期118-137,共20页Acta Photonica Sinica

基　　金：国家自然科学基金(Nos.61705254,U1930118);陕西省重点研发计划(No.2020GY‐114)。

摘　　要：针对基于纹影的高动态范围远场焦斑测量数学模型没有考虑噪声对测量结果影响的缺点,本文对基于纹影的远场焦斑测量方法从三个方面进行优化.首先,改进基于纹影的远场焦斑测量数学模型,将噪声作为影响实验结果的重要因素引入数学模型中,使该数学模型和真实的实验环境相匹配,提高了该数学模型的实用性和理论支撑作用;其次,将基于卷积神经网络的去噪算法(DnCNN)引入主瓣和旁瓣CCD图像去噪处理中,并改进该去噪算法存在的不足,使得能够有效去除主瓣和旁瓣12位图像、不同级别(0~75 dB)的噪声;最后,完整仿真了远场焦斑测量实验的全过程,包括分光、衰减、加噪声、纹影小球遮挡、去噪、衰减倍率放大、焦斑重构等,获得了有效的焦斑重构实验结果,其中重构焦斑图像和理论焦斑图像的相关系数为0.9989,重构焦斑动态范围与理论焦斑动态范围之间误差为3.22%.仿真实验结果表明,通过该数学模型和DnCNN去噪算法的改进措施,验证了改进的数学模型必要性和DnCNN去噪算法在提高重构焦斑二维分布和动态范围精度方面的优越性能,提高了基于纹影的高动态范围远场焦斑测量的可信度,满足了高动态范围远场焦斑测量对于精度和效率的要求.Aim at the shortcomings that the mathematical model for the measurement of far-field focal spot with high dynamic range does not consider the influence of noise on the measurement results,this paper optimizes the measurement method of far-field focal spot based on schlieren from three aspects.Firstly,the mathematical model for the measurement of far-field focal spot based on schlieren is improved,and the noise is added to the mathematical model,which makes the mathematical model match with the real experimental environment,and improves the practicability and theoretical support of the mathematical model;Secondly,the denoising algorithm based on Convolution Neural Network(DnCNN)is used in the de-noise processing of the main lobe and side lobe CCD image,and the original denoising algorithm is improved effectively in this paper,which can remove the noise of different levels(0~75 dB)of the mainlobe and sidelobe 12-bit images;Finally,the whole experimental process of far-field focal spotmeasurement is simulated,including light splitting,attenuation,adding noise,schlieren sphere occlusion,denoising,attenuation magnification,focal spot reconstruction,etc.,and the effective experimental resultsof reconstructed focal spot is obtained,which the correlation coefficient between the reconstructed andtheoretical focal spot images is 0.998 9,and the error of dynamic range between the reconstructed andtheoretical focal spot is 3.22%. The simulation results show that through the improvement of themathematical model and the DnCNN denoising algorithm,the necessity of the improved mathematicalmodel and the superior performance of the DnCNN denoising algorithm in improving the accuracy of thetwo-dimensional distribution and dynamic range of reconstructed focal spot are verified. The reliability ofthe measurement of far-field focal spot with high dynamic range based on schlieren is improved,and theaccuracy and efficiency of the measurement of far-field focal spot in high dynamic range is met in the end.

关 键 词：远场焦斑测量 纹影法 焦斑重构 DnCNN 去噪方法 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]