基于深度降噪卷积神经网络的宽波段共相检测研究

Broad-band co-phase detection based on denoising convolutional neural network

拼接镜的共相误差检测是当前科学研究的热点问题之一,基于宽波段光源的共相检测技术解决了夏克哈特曼法由于目标流量低引起的测量时间长的问题,从而提升了piston误差的检测精度和量程。然而,当前宽波段算法在实际应用中,由于复杂的环境以及相机扰动等干扰因素的存在导致获取的圆形孔径衍射图像含有一定量的噪声,从而导致相关系数值低于设定阈值,最终使该方法精度降低,甚至失效。针对这一问题,本文提出将基于深度降噪卷积神经网络(DnCNN)的算法集成到宽波段算法中,以实现对噪声干扰的控制,并保留远场图像的相位信息。首先,将使用MATLAB获得的圆孔衍射图像作为DnCNN的训练数据,然后,将不同噪声水平的图像导入到训练好的降噪模型中,即可得到降噪后的图像以及降噪前、后圆孔衍射图像的峰值信噪比和二者与清晰无噪声图像间的结构相似度。结果表明:降噪处理后的图像与理想清晰图像之间的平均结构相似度较处理之前有了明显提升,获得了理想的降噪效果,有效增强了宽波段算法在高噪声条件下的应对能力。该研究对于探索用于实际共相检测环境宽波段光源算法具有较强的理论意义和应用价值。

The co-phase error detection of segmented mirrors is currently a critical focus of scientific research.Co-phase detection technology based on a broad-band light source solves the problem of long measurement times caused by the Shackle-Hartmann method’s low target flow rates,thereby improving the accuracy and range of piston error detection.However,in the application of the current broad-band algorithm,the complex environment and the presence of disturbing factors such as camera perturbations cause the acquired circular aperture diffraction images to contain a certain amount of noise,which leads to a correlation coefficient value below the set threshold,reduces the accuracy of the method,and even makes it ineffective.To solve the problem,we propose a method by integrating an algorithm based on Denoising Convolutional Neural Network(DnCNN)into the broad-band algorithm in order to control the noise interference and retain the phase information of the far-field image.First,the circular hole diffraction image obtained by using MATLAB is used as the training data for DnCNN.After the training,the images with different noise levels are imported into the trained noise reduction model to obtain the denoised image as well as the peak signalto-noise ratios of the circular hole diffraction images before and after denoising.The structural similarity between the two images and the clear and noiseless image are also obtained.The results indicate that the average structural similarity between the denoised image and the ideal clear image has significantly improved compared to the image before processing,and this achieves an ideal denoising effect,which effectively increases the ability of broad-band algorithms to cope with the effects of high noise conditions.This study has strong theoretical significance and application value for exploring the broad-band light source algorithm for applications in practical co-phase detection environments.