Research On Physical Prior And Data Driven Scattering Imaging And Its Application

Scattering is a common phenomenon in optical imaging.In the imaging path,the inhomogeneity of the scattering medium interferes with the propagation of photons carrying target information,which results in a lower detection signal-to-noise ratio(SNR),which makes direct imaging impossible.Therefore,how to achieve the high-resolution scattering imaging is a critical problem in the field of computational optical imaging.To achieve imaging using scattered optical signals,scattering imaging methods based on wavefront shaping,transmission matrix,scattering correlation,and time-of-flight techniques have been proposed and improved one after another.However,most of the current methods are limited to the recovery of sparse targets while requiring the determinism of the scattering process,which limits their application in practice.With the continuous development of photoelectric detection equipment as well as computer technology,deep learning plays a great role in the field of scattering imaging with its capabilities of powerful data fitting and information extraction.Deep learning-based scattering imaging methods show significant advantages in terms of imaging speed,quality,and information dimensionality.However,the problems such as the acquisition of datasets and the generalization ability of models constrain the development of the method.Therefore,the physical prior information can be introduced into the training of neural network to guide the parameter optimization,effectively reducing the difficulty of data acquisition and the requirements of the experimental environment while ensuring the imaging quality.This physical prior-based learning strategy greatly improves the generalization ability of the neural network imaging model to key nodes such as medium and target type.Based on deep learning algorithms and derived from the scattering imaging mechanism,this thesis develops training optimization approaches from supervised to semi-supervised and finally to unsupervised.The proposed different network models and supervision strategies not only effectively reduce the hardware requirement and time cost of the imaging system,but also are beneficial for the application of scattering imaging methods in fields such as biomedicine and unmanned vehicles.The main research contents and innovation points of this thesis are as follows:(a)Two imaging schemes through dynamic scattering medium are proposed.When using array detection for imaging through dynamic scattering medium,the low SNR measurements hinder high-resolution optical imaging.To address this problem,we characterize the low SNR measurement data by power spectrum estimation and obtain a noise model for scattering imaging.Then,the proposed noise model is used to synthesize the simulation dataset for network training,and then we can solve the problems related to phase recovery of low SNR in the case of unknown experimental scenarios.We experimentally validate the robustness of the proposed method to noise.The spatial resolution and total acquisition time both outperform the Fourier-domain shower-curtain effect-based imaging system.For the point detection scheme,a dynamic scattering medium correlation imaging method based on high compression ratio single-pixel detection is proposed,which uses the orthogonal compression property of the onedimensional discrete cosine transform to reconstruct high-quality target images.In the case of low sampling,the method obtains spectral coefficients with large weights and generates the target image by performing an inverse discrete cosine transform on them.Simulation results and optical experiments show that the method still achieves high-quality imaging even at a sampling ratio of 0.03.(b)An optical reconstruction scheme for non-sparse objects through scattering medium is proposed.Unlike previous imaging methods that use the optical angular memory effect,this scheme can achieve simultaneous retrieval of the distribution functions of the original target and the scattering medium by measuring multiple complex target objects under the same scattering medium,provided that the Nyquist sampling theorem is satisfied.The scheme iterates by setting a constraint of amplitude range for each measurement,and finds the decoupling path that best matches the phase distribution of the scattering medium during the reconstruction process by using a global optimization approach.We explored scattering media with different statistical properties to prove the feasibility of the scheme.It has the advantages of high resolution,non-intrusive,high robustness,and low cost at the same time,and thus has higher practical value and broader development prospects.(c)A non-line-of-sight imaging scheme based on iterative optimization of the neural network is proposed.The scheme exploits the interaction between the neural network and the physical model of non-line-of-sight imaging,which can automatically update the network weight parameters without training data.The optimized network model is able to reconstruct hidden scenes from irradiance patterns on relay surfaces,with high quality and under strong ambient light.Simulation and experimental results show that the method has good performance in terms of detail processing and reconstruction robustness.Our method will further facilitate the practical application of non-line-of-sight domain imaging in real scenes.(d)We propose an optical cryptanalysis scheme based on a physical prior.Benefited from the full understanding of the randomness of light scattering,we obtain the statistical invariance properties between ciphertexts through a message preprocessing step and then use neural networks to learn the potential mapping relationships between plaintexts and ciphertexts,which can finally achieve high fidelity prediction of complex plaintext messages.In addition,by increasing the diversity of random phase encoding,the generalization ability of the network model for ciphertext prediction can be significantly improved,and the feasibility and effectiveness of the proposed method are verified using optical experiments.