Research On Multimode Fiber Imaging Based On Phase Modulation And Deep Learning

Endoscopy has emerged as a practical and efficient tool,which has wide applications in both industrial non-destructive testing,monitoring and medical minimally invasive diagnosis and treatment.Optical fiber,having many advantages,is an ideal invasive light guide medium for endoscopes.Current optical fiber endoscopes are mainly applied in the form of fiber bundles or single-mode fiber(SMF)combining with hybrid mechanical scanning actuators,which inevitably increase the volume of the probe and limit the resolution and structure miniaturization of the endoscope,and thus are usually unsuitable for the small biological tissue cavity imaging.The application of multimode fibers(MMFs),which have the characteristics of small core diameter and parallel transmission of multiple guided modes,shows an attractive solution to the defects of large size and low resolution of fiber endoscope.Therefore,it has both significant theoretical and practical meaning for the progress of endoscopy technology to study approaches of imaging through MMF.Among several approaches,wavefront modulation method dispenses with measuring the transmission matrix,employs simple optical path and shows good system stability.However,it usually requires multiple iterations to form a batch of focused spots.While deep learning-based method,as a data-driven approach,no longer relies on the forward mathematical model and image inverse reconstruction algorithm,its imaging quality is dependent on the characteristics and quantity of training samples.In this thesis,researches on two technique routes of multimode fiber imaging,including wavefront phase modulation-based MMF focused spot scanning endoscopic imaging and deep learning-based MMF speckle image reconstruction,are carried out.The main research work and innovative achievements of this thesis are summarized as follows:(1)Theoretically,a phase compensation scheme using two-step phase shifting based on LC-SLM is proposed,which can be used to quickly calculate optimal compensation phase information to be loaded on LC-SLM in order to modulate the phase of the MMF input wavefront to realize spot focusing at the output;According to the incoherent imaging theory and scattering medium transmission matrix theory,the theoretical model of multimode fiber focused spot scanning imaging is established,and efficient solution of the theoretical model is established by firstly calibrating the impulse response function of the imaging system and then scanning the real object for imaging to reconstruct the surface reflectance of the two-dimensional object.(2)Since MMF focused spot scanning imaging based on wavefront shaping requires to solve the problem of quickly generating multi focused spots at the output of MMF for spot scanning imaging,a two-step phase-shifting phase-compensating technique based on LC-SLM and a parallel algorithm are proposed.The method dispenses with multiple iterations and is capable of efficiently acquiring several spots to realize fast and parallel focusing.An MMF with the length of 5 m and the core diameter of 50 μm is used in experiment and fast light focusing is realized through it.a focused efficiency up to 30.31% is achieved with the help of background subtraction,and minimum diameter of the focused spots is 1.6158 μm;Feasibility of effectively spot focusing employing two-step phase-shifting method combined with parallel algorithm is experimentally demonstrated and it takes only 65 seconds to calculate the compensated phases 100 focused spots at contiguously different locations of 10 rows and 10 columns at the MMF output facet.(3)Since spot scanning endoscopic imaging system utilizing single MMF has a complicated optical structure,an endoscopic imaging scheme based on multimode fiber coupler is proposed.Besides,the focused spot scanning endoscopic imaging system based on the 1×2(three-port)multimode fiber coupler is designed and built.The phase modulation technique is utilized to quickly acquire 900 focused spots,which are located in array locations of 30 rows and 30 columns at the output of MMF coupler,for subsequent scanning focusing.The average focused efficiency and the average diameter of these experimentally obtained 900 focused spots is 19.35% and 2.0423 μm.The total reflective area of the USAF1951 resolution test board is scanned to calibrate the impulse response function of imaging system and with that,the reflected light intensity and impulse response function information is combined to solve the reflectance information of the object surface to reconstruct the surface image of the twodimensional(2D)object.An MMF with core diameter of 105 μm is used to realize scanning imaging of the object within 60 μm at MMF output facet,realizing imaging the Group.5 Element.6(line width of 8.77 μm)in the USAF 1951 resolution test board.(4)Technique of image reconstruction through the MMF based on a deep learning method is exploited.A deep learning model combining the convolutional Autoencoder(AE)with the self-normalizing neural networks(SNNs)is proposed.Simulations on speckle image reconstruction through the MMF are conducted using the proposed AESNNs combined deep learning model to validate both its feasibility and its effectiveness at different noise levels.Simulation results show that under ideal conditions,the image reconstruction accuracy of handwritten test digits is 97.84%;Experimental setup of MMF speckle acquisition based on digital micromirror device is built to acquire 4000 train data pairs of speckle patterns with corresponding handwritten digits from the train database;300 epochs of training is carried out for the AE-SNNs deep learning model using the obtained train data pairs.And then both handwritten digits and capital letters are used as test samples to realize speckle image reconstruction with the accuracy of92.18% and 91.62%,the structural similarity of two kinds of reconstructed images with corresponding ground-truth objects is 0.6996 and 0.6293,experimentally demonstrating that proposed AE-SNNs deep learning model realizes high-quality speckle image reconstruction and shows generalization and feature transfer capability to some extent.