Scattering Imaging Based On Non-experimental Training Neural Networks

The light scattering caused by inhomogeneous media(such as biological tissues,clouds,etc.)reduces the signal-to-noise ratio of the imaging system,and even makes the image information of interest submerged in noise.How to overcome the scattering or extract effective object information from the scattered light field has always been a research hotspot for scientific researchers.In recent years,deep learning-based scattering imaging has developed rapidly,solving a series of bottlenecks encountered in traditional imaging technology,such as scattering imaging of complex objects,object imaging in dynamic scattering environments,and deep positioning scattering imaging.However,the current deep learning scattering imaging research requires a large amount of experimental data to train the neural network to reconstruct the target to be detected,which brings a certain burden to researchers in terms of time and hardware cost,and even some experimental data is difficult to obtain.It’s not enough to support neural network training.By analyzing the optical memory effect of inhomogeneous media(scattering media)and the connection between speckle and objects,this thesis innovatively proposes to use simulation data to train neural networks in order to reconstruct the image of the target from experimental speckle.method.The specific research contents are as follows:(1)Due to the optical memory effect of the scattering medium,the autocorrelation of the speckle and the autocorrelation of the object correspond one-to-one under the condition of spatially incoherent illumination.Based on this physical prior information,we use the simulated autocorrelation of objects to train the neural network model,and then reconstruct the unknown objects from experimental speckle image.In order to compare the imaging performance of the proposed scheme with that of previous research schemes,a model trained with neural network using experimental data is constructed and also used to reconstruct unknown targets.Experimental results show that the method of using simulation data to train the network can effectively reconstruct unknown objects in reality.However,compared with the experimental data training network,the mean square error,Pearson correlation coefficient and peak signal-to-noise ratio are slightly inferior.(2)In order to further improve the imaging performance of the simulation training neural network method,the speckle autocorrelation is considered to contain some noise.The author simulates the autocorrelation of point spread function of the scattering imaging system and calculates the convolution between target object and autocorrelation of point spread function to achieve the purpose of correcting the simulation data.The modified simulation model can achieve more similar effect to the real scattering imaging scene,and the trained reconstructed network model can restore the reconstructed image with higher quality from the experimental speckle data.Comparing the performance of neural networks trained by simulation and experimental data,the quality index of target reconstruction is similar,but the generalization performance of the former on unknown objects with different magnifications is much higher than that of the latter.