Deep Learning Based Adaptive Optics In Flow Measurement

Imaging based flow measurement method,like particle image velocimetry(PIV),is a type of non-contact,real-time,quantitively and global flow measurement technique.It has been widely used in fluid mechanics research and industrial application,has great significance for the study of complex flow.However,the measurement accuracy of these optical imaging-based measurement techniques can be seriously impaired when the refractive index changes in the fluid medium,or when an open gas-liquid phase boundary with random fluctuation occurred in the measurement optical path.The wavefront in imaging optical path is distorted due to the refractive index change,which causes distortion on the PIV image.These random and time-varying wavefront distortions in the optical path cause significant errors in terms of the particle position distribution on the PIV particle image,i.e.,geometric distortion,and seriously degrade the image quality.The traditional PIV technology estimate the fluid motion field based on the particle displacements on the continuous multi-frame PIV particle images,and the fluid motion field is estimated by the cross-correlation algorithm or the optical flow algorithm.Therefore,such geometric distortions on the particle image and the degradation of the image quality can seriously affect the measurement accuracy of the PIV,result in unreliable measurement results.Adaptive Optics(AO)technology is an advanced technique to correct time-varying wavefront distortion for an optical system.The traditional AO system measures wavefront distortion in real-time by wavefront sensors,and then the control signal is calculated by the electronic system,controls the wavefront corrector to perform realtime wavefront correction.In this way the optical system can remain good working condition when it is disturbed by wavefront distortion.In order to correct the wavefront distortion for flow measurement in real time,which is also for the reduction of the flow measurement uncertainty,adaptive optics techniques can be applied in such optical systems.However,it is shown that the wavefront distortion caused by fluctuating airwater interface in optical flow field measurement has the features of high frequency,large dynamic range,and high spatial resolution.The correction performance for a traditional adaptive optical system based on hardware wavefront correctors is limited by their technical specifications in terms of dynamic range,available spatial frequency and correction bandwidth.In this dissertation,in order to eliminate limitations of a traditional actuator-based adaptive optics system,we propose a new concept of adaptive optics,which is actuatorfree adaptive optics based on deep leaning methods.Proposed new approach is applied in wavefront distortion correction in flow measurements.The main research in this dissertation is organized into four parts as follow:1.Distortion model in flow measurement is analyzed first,and a measurement method for such distorted wavefront is proposed,which is based on Hartmann-Shack wavefront sensor and a spatial distributed laser guide star.The measurement device,which is PIV camera of the particle image velocimetry system,and the measurement object are merged in different media,and the imaging system needs to pass through the fluctuating air-liquid phase boundary.Such procedure induces a random wavefront aberration before the imaging system for PIV,i.e.,the imaging system needs to image through a randomly fluctuating phase boundary.The relationship between wavefront distortion in the optical path and the actual degradation model on the image is discussed.Then a wavefront measurement method for such wavefront distortion is proposed,where the concept of a spatial distributed laser guide star is applied first,and combined with the utilizing of Hartmann-Shack wavefront sensor.2.For the measurement of aberrated wavefront caused by the fluctuating phase boundary in flow measurement,traditional Hartmann-Shack wavefront sensor is limited by its dynamic range.A Hartmann-Shack wavefront measurement algorithm with large dynamic range is proposed in this paper.A centroid estimation algorithm based on image segmentation is proposed,that is,the spots centroid estimation is performed from segmented spot area on a Hartmannogram,eliminated the limitation from the sub-aperture in traditional centroid algorithm,obtained the centroid of all spots from the whole Hartmannogram,and eliminated the effect of noise during the process of centroid estimation.Then a neighboring searching-matching algorithm is proposed to expand the dynamic range,the spot centroid and the calibration coordinate position are matched correspondingly.The performance of the proposed Hartmann-Shack wavefront measurement algorithm with large dynamic range is evaluated in terms of centroid estimation error,linearity,quantitatively evaluation of dynamic range expansion and algorithm is tested on an experiment setup.3.This part is the core of the entire research.Based on the deep learning technique,a new concept of actuator-free adaptive optics system is proposed.The AOPIV-MIUN algorithm,which is a multiple-input convolutional neural network architecture is proposed.This is the first time that a multi-input convolutional neural network is applied to an image regression problem.Based on proposed network architecture,the wavefront distortion information obtained by the Hartmann-Shack wavefront sensor is used as an additional input,and the distorted image in the particle image velocimetry system is corrected by the proposed network.The data sets for neural network training and test are generated from a designed experiment setup.The distortion correction performance of the proposed neural network is evaluated by taking the correction effect of the corrected particle image and the measurement uncertainty reduction of the flow as the evaluation criterion.The correction performance in terms of the measurement uncertainty reduction for proposed actuator-free adaptive optics method reaches 82%,and such result is better than the traditional closed-loop adaptive optical system,which reaches 77% under exactly same experimental condition and same performance evaluation criterion.4.In this part,the flow motion estimation and distortion correction are combined by a deep learning model,where a fluid motion estimation algorithm can also conduct distortion correction is proposed.Two different multi-input convolutional neural network structures are proposed,where the input are two frames of distorted PIV particle images and measured wavefront distortion information,and directly output the corrected flow motion field.Deep-learning based dense estimation and sparse estimation are proposed for flow field estimation under different resolution.The flow motion estimation and distortion correction are simultaneously integrated into the deep convolutional neural network.Based on the PIV particle image generation model,the Hartmann-Shack wavefront sensor simulation model,and the image distortion model,a synthetic dataset for training is generated.And the performance of the proposed algorithm is evaluated from different perspective.The dissertation focuses on an actuator-free adaptive optics technique applied in distortion correction in flow measurement.Proposed method eliminated limitations of a traditional close-loop adaptive optics system in such application.It is the first domestic research on adaptive optics applied in flow measurement,and for the first time the concept of deep-learning based actuator-free adaptive optics is proposed,furthermore,the first application of the concept of multi-input convolutional neural networks for an image regression or image translation problem.As a perspective,proposed approach can be used to measure the liquid flow inside droplets or gas flows within Tayler bubbles,which are characterized by an all-side open surface that hindered undistorted optical measurements so far.