Study Of Speckle Correlation Imaging Algorithm In The Low-signal-to-noise Ratio Low-resolution Condition Based On Physical-sensing

Lights form disordered speckle pattern when they pass through difusers such as clouds,glass,and biological tissue,making it diffcult to directly detect targets.In the feld of optical imaging,imaging through difusers is an important and widespread research element with wide applications in astronomical observation and biomedical applications.Speckle correlation imaging(SCI)algorithms extract the speckle pattern’s energy spectrum and bispectrum by calculating autocorrelation and triple-correlation to retrieve the target’s Fourier amplitude and phase information within the range of optical memory efects(OME).Existing SCI algorithms usually require the high signal-to-noise ratio(SNR)high-resolution(HR)speckle pattern images to suppress statistical noise,and have limited reconstruction capability in low SNR low-resolution(LR)environments.In recent years,deep learning has shown great vitality and potential in solving imaging through difusers.Therefore,this doctoral thesis will design scattered image reconstruction algorithms based on physical prior knowledge guidance by combining the principles of SCI algorithms and deep learning for the low-SNR LR problems in visible and near-infrared imaging systems,and the main research work and innovations are as follows:SNR models for the energy spectrum and bispectrum of the speckle pattern are developed,and an adaptive fltering algorithm is proposed to improve the robustness of the speckle triple-correlation imaging(STCI)algorithm.The SNR model of the energy spectrum is used to estimate the high quality autocovariance of the speckle pattern,which can be the physical priori knowledge to sense the efective scattering region under diferent targets.Thus the adaptive reconstruction without manual parameters is realized.Theoretical and experimental results show that the method is not only robust to low-SNR data caused by inappropriate optical element parameters,but also has good immunity to simulated Gaussian noise,simulated Gaussian white noise,and real noise caused by ambient interfering light in the optical system.A two-stage(de-noising and reconstruction)learning algorithm with autocorrelation as a physical constraint is proposed for the low-SNR problem caused by ambient interference light under visible illumination.This physical constraint is enable by the consistency of autocorrelations between the speckle pattern and the target within the OME range.Thus,the autocorrelation is used as an intermediate optimization parameter to optimize the learning strategy and improve the convergence effciency of the network model training.The two-stage learning algorithm can improve the robustness to noise and the generalization capability to various difusers.Under the imaging conditions of unknown difusers and unknown noise,it is experimentally verifed that the de-noising stage can signifcantly improve the quality of the autocorrelation.The reconstruction stage can successfully recover the target hidden behind difusers by using the enhanced autocorrelation as input.The reconstruction phase with successfully recovers the target information hidden behind unknown difusers by using the enhanced autocorrelation as input.A super-resolution reconstruction learning algorithm is proposed based on the physical mechanism of resolution degradations for LR problems caused by the difraction limitation of the detector or the optical imaging system under near-infrared illumination.Using the principle of the speckle redundancy and resolution degradations when the detector observes the signal,the two inverse problems(the super-resolution reconstruction and imaging through difusers)are modeled simultaneously to build a network that can sense features of the LR speckle pattern.Then the learning is driven by a small amount of speckle pattern to fully exploit the feature of low-resolution data.It is experimentally verifed that the algorithm can efectively compensate the low-resolution loss information and can successfully reconstruct the HR target from a single-frame LR speckle pattern.In conclusion,this doctoral thesis proposes a physical mechanism-aware method to enhance SNR and compensate resolution in the low-SNR LR imaging environment,which can signifcantly reduce the dependence of imaging algorithms on high-performance detectors and HR data.The proposed method has important research value and application prospects in astronomical observation,biomedicine,and military reconnaissance.