Research On Multi-Information Extraction Method Of X-Ray Grating Imaging Based On CNN

The grating imaging method based on Talbot-Lau effect is widely used as the X-ray grating imaging technology.This will easily obtain the absorption signal,scattering signal and phase signal inside the object.Many studies have shown that the phase signal can effectively improve the contrast of soft tissue detection,while the scattered signal can greatly improve the detection sensitivity of particles or void structures inside objects.Therefore,the X-ray grating imaging technology based on the Talbot-Lau effect has received widespread attention from researchers,and it is hoped that this technology will be applied to clinical detection.At present,the extraction of absorption,scattering,and phase signals is based on multiple phase step projection images,which has has the certain defects.The inaccuracy of the step displacement operation and the instability of the optomechanical output will cause the existence of noise and estimation bias,and the mathematical analysis method will amplify these errors.In addition,for the disadvantages of large time cost and low radiation dose utilization rate of the phase stepping method,although there are multiple signal extraction methods to solve the time cost problem,all of them have lost space cost to a certain extent(i.e.the image quality has been different).The degree of limitation cannot meet the requirements of practical application.The paper proposes a new X-ray grating imaging multi-signal extraction method based on the convolutional neural network,which aims to achieve the function of the original method and further improve its defects.Firstly,the XPNET network model which completely replace the phase stepping method through the end-to-end training structured.We use synchrotron radiation data to explore the effect of different step numbers on the extraction effect.The main problem is to reduce the signal-to-noise ratio to solve the time cost and radiation dose problem.Then,by adding a noise reduction module to optimize on the basis of XPNET,we propose to the XPNr NET network model with noise reduction function,which can improve the signal-to-noise ratio of the image with a small number of steps,the image quality of the picture and indirectly reduce the time cost.Then,on the issue of estimation bias,the XPDb NET network model is added a cascade structure based on the XPNET model.Instead of reducing noise through traditional noise reduction,the image resolution and signal extraction accuracy are reduce the estimation deviation of the phase signal by 15% to the maximum extent,thereby improving the accuracy of signal extraction.All network models in this paper are using an end-to-end training method trained through natural picture data sets.All the trained models have been verified by simulation data?synchrotron radiation data and common laboratory light source data,then have achieved good results.The network model has wide applicability.