Research On Coding Technique And Image Reconstruction Algorithm Based On Single Photon Detection

LIDAR(Light Detection And Ranging)has been widely used in various fields such as photoelectric reconnaissance and remote sensing thanks to its excellent range resolution and well-orientation properties,now it is playing an indispensable role in some cutting-edge technique(e.g.,autonomous vehicles).However,the attenuation property of laser raises the concern for its effective range,thus becoming one of the bottlenecks in LIDAR research.Single photon detection technique is capable of detecting echo signal at single photon sensitivity by adopting single photon detectors,leading to significant improvement on its sensitivity.Thus enable the users to acquire target information at long range with low uncertainty.Geiger mode Avalanche Photon Diode(Gm APD)is one of the most popular single photon detectors thanks to its low cost and binary response.Nevertheless,the single photon detection system has to accumulate enough photons to recover the echo signal as the output of Gm APD is either 0 or 1.This lead to a sparse data cube under some applications(e.g.,long range imaging),which is challenging to reconstruct 3D profile of the target.On the other side,laser with high repetition rate is used to acquire photon-time data in limited time,resulting in range ambiguity.This dissertation is centered on the way to solve range ambiguity and how to exploit limited echo information efficiently,which are the cores of acquiring accurate target information at long range.The main contributions of this dissertation are as follows:(1)The dissertation proposed 2 pseudorandom code generating algorithms to solve range ambiguity,namely fast multi-phase pesudorandom code generating algorithm and best multi-phase pesudorandom code generating algorithm.Meanwhile,Fast Fourier Transform(FFT)algorithm is used to achieve fast decoding.This dissertation increased the effective range by adopting pseudorandom code.As m sequence meets drawbacks(e.g.,can not be modulated with high speed),new algorithms are proposed.Fast multi-phase pesudorandom code generating algorithm modulates pulse intervals based on the parameters of the system and the random number from Mersenne-Twister algorithm,then the modulated bits are assigned with an intensity value by the random number and predetermined intensity projection.Best multi-phase pesudorandom code generating algorithm considers the code as a permutation of bits and solved it by Partical Swarm Optimization (PSO) algorithm,where the initialization was done by linear congruential method with low complexity.By performing on simulated data,the pseudorandom code generated by either fast multi-phase pesudorandom code generating algorithm or best multiphase pesudorandom code generating algorithm showed better self-correlation performance compared to dual detector pesudorandom code generating algorithm and m sequence.Furthermore,both fast multi-phase pesudorandom code generating algorithm and best multi-phase pesudorandom code generating algorithm is easy to implemented at a low cost,their nice properties enables the use of FFT based decoding algorithm which showed best decoding efficiency.(2)Since imaging at long range could lead to a sparse data cube,this dissertation proposed a Non-Local framework named NL3DR(Non-Local 3D Restoration)for the reconstruction of 3D images under photon sparse case.This dissertation exploits the correlations between all the pixels by building an affinity graph,then a Laplacian regularizer is applied as a constraint during the reconstruction.Finally,the spatial correlations and reconstruction results shall be updated alternatively.To avoid the empty pixels and noisy pixels that corrupt the learned Non-Local correlations,MSA(Multi-Scale Analysis)algorithm is proposed as the initialization step of NL3 DR to improve the robustness by building super-pixels and analyzing multi-scale information.According to state-of-the-art methods,the echo signal is modeled as Poisson distribution,and the cost function is formed by negative log likelihood and Laplacian regularization term.The cost function is separable w.r.t parameters of interest,therefore the minimization of cost function can be efficiently solved by ADMM(Alternating Direction Method of Multipliers)algorithm.Furthermore,considering the computational complexity of NL3 DR,this dissertation proposed non-uniform sampling and various optimization strategies to reduce the complexity,where NL3 DR is optimized as OPN3DR(Optimized based Nonlocal 3D Restoration).Finally,reconstruction of 3D images with photon per pixel(PPP)less than 1 was achieved on both simulated data and real data,where OPN3 DR showed robust results compared to competitive algorithms(e.g.,RDI-TV)under limited sampling points and even better results are shown if they are performed under the same sampling points.(3)To solve the range ambiguity by using the true randomness of quantum devices,this dissertation studies the three-dimensional imaging system based on single photon source,and the reconstruction is achieved by adopting a variant of OPN3 DR algorithm named F3DRSI(Fast 3D Restoration for Single photon source Imaging).As single photon pair source emit photons at true random intervals,any range ambiguity is avoided,thus the system is free from using extra modulation module and pseudorandom codes.The system split the emitted pair of photons using polarizing beam splitter,and both photons were detected by detectors.Then the timing difference between detectors refers to the time-of-flight(TOF)and hence target distance.The data acquired by single photon source were even sparse although the ultra low energy has a lot of advantages.To make full use of the received photons,F3 DRSI algorithm is applied to reconstruct the 3D image of the target with PPP less than 1,while the uncertainty remains to be millimeter scale under the extreme environment.(4)Considering the different properties of spectral bands in different media,this dissertation studies the multispectral single photon imaging system and its reconstruction algorithm MS3DR(Multispectral Sparse 3D Restoration).Multispectral single photon imaging system is capable of acquiring 3D information and spectral signature,where the histograms from multiple bands could improve the robustness of estimation under photon starved cases.To learn from the sparse 4D hypercube,this paper generalized OPN3 DR as MS3 DR algorithm,where both spatial correlations and spectral correlations were used as prior for the reconstruction.In the experiments,MS3 DR achieved multispectral 3D imaging under photon per pixel per spectral less than 1;the algorithm was further validated in complex scenes under photon per pixel per spectral only 0.57 and SBR(Signal-to-Background Ratio)of 8.72.