The Research Of Focusing And Imaging Through Scattering Media

Imaging through scattering media such as biological tissue,ground glass,and the fog has been a widely used but challenging subject in the optical field.The commonly used method for imaging through scattering materials utilize ballistic photon filtered from the scattered background.The ballistic photon travels in the scattering medium at most one transport mean free path,which limits the imaging distance of this method.As scattered light goes further in a scattering medium than ballistic photons,using scattered light for imaging can obtain a longer imaging distance.However,the scattering process of light is full of randomness and complexity;how to "decode" the original object from a random speckle pattern is a difficult problem.In recent years,the wavefront shaping technique(WFS)has brought a breakthrough in the field of imaging through scattering media using scattered light.WFS precisely modulates the phase or amplitude of the incident light using the spatial light modulator to realize the focusing of light through the scattering medium.WFS experimentally proved that by accurately modulating the incident wavefront,a scattering medium could be used as an optical element like a lens,opening up a new field of focusing and imaging through scattering media using scattered light.This paper focuses on focusing and imaging through scattering media.Starting from the two aspects of focusing through scattering media based on WFS and computational imaging based on deep learning algorithms,the research content is divided into the following three parts:(1)Focusing through scattering media using the harmony search algorithmThe WFS method focuses through scattering media by using iterative algorithms to find the optimized phase or amplitude mask to compensate for the disordered wavefront.Thus identifying a fast and robust iterative algorithm is a critical task in this field.In this paper,we introduce the harmony search(HS)algorithm for use in wavefront shaping.To verify the effectiveness of the HS algorithm,we built a simulation model for focusing light through the scattering medium based on the random matrix theory.The measurement noise is simulated by adding Gaussian white noise,and the random disturbance is added to the elements of the random matrix of scattering medium to simulate the vibration environment.The numerical simulation results show that the HS algorithm has higher convergence speed,higher overall intensity enhancement,and better robustness in noise and vibration environments than other modulation algorithms.We establish a scattering optical imaging system and a WFS software system to modulate 1024 input channels of input wavefront with the HS algorithm,and the focused light intensity is 180 times higher than the average intensity of the speckle pattern before modulation.The experimental results agree well with the simulation results and show that the HS algorithm can be effectively applied to focus through a scattering medium,and has faster-focusing speed and higher intensity enhancement.(2)Object recognition through scattering media using convolutional neural networkAs the scattering process of light is highly complex and has many possible scattering paths,the relationship between the original object and its speckle pattern can not be expressed directly by expressions.While the data-driven method can automatically learn the hidden features from speckle-object image pairs,especially with the rapid development of deep learning algorithms,computational imaging from speckle patterns not only has the advantages of simple imaging setup and robustness to speckle decoherence but also has excellent imaging performance.We established an optical imaging system and a software system to collect speckle patterns dataset for training our convolutional neural networks(CNN).A classification CNN architecture is established to recognize objects through a scattering medium.CNN can extract the non-linear relationship from the speckle pattern,and the accuracy of identifying the same face dataset is 4.2%higher than that of the linear classifier support vector machine.(3)Imaging through scattering media using fully convolutional neural networkTo imaging through scattering media,we extend the CNN classifier to the fully convolutional neural network.To improve the reconstruction quality of non-sparse objects,we use the U-Net architecture.The downsampling and upsampling module of the U-Net are improved to make it more suitable for the task of speckle reconstruction.Aiming at the problem that the large parameters of U-Net architecture limit the spatial bandwidth product of reconstructed images,the network architecture is lightweight with depth separable convolution layers to reduce the model parameters to 11%of the original.By analyzing the influence of different training data sets,data quantity and loss function on the reconstruction ability,we find that use datasets composed of speckles of mixed sparse objects and non-sparse objects to train the neural network can improve the generalization performance of the network architecture,and use structural similarity(SSIM)as a loss function to train neural network can reconstruct images with more detailed information and clear edge than the neural networks trained by mean square error and mean absolute error.