Research On Image Location And Blur Recovery Technology Of Background Schlieren Calibration Based On Deep Learning

Background Oriented Schlieren(BOS)is a technique for flow visualization by measuring flow field density.The BOS technology only needs the camera and the background image stand on the both sides of the flow field,and can realize the non-contact measurement of the flow field to be measured.It has the advantages of simple structure,large field of view,and strong anti-interference ability.If multiple cameras are installed outside the measured area from different angles and the BOS technique is combined with the CT technique,the transient 3D CT reconstruction of the flow field to be measured can be realized,it provides an effective technology for quantitative measurement of density parameters in wind tunnel experiment.CT-BOS technology is becoming one of the popular technologies for flow visualization and quantitative measurement of complex flow field.Before the multi-direction BOS offset projection is used to reconstruct the flow field by CT algorithm,it is necessary to calibrate the multi-cameras in the wind tunnel experiment environment and obtain the pose parameters of each camera,which is the key step to determine the reliability of CT-BOS reconstruction.However,when BOS technology is used in background offset extraction,the camera needs to be focused on the background,which causes the defocusing of the reconstruction area and leads to the blur of calibration image,hence,the reliability of calibration results of multi-camera is seriously affected.To address the issues mentioned above,the following researches are performed in this article:Firstly,in order to improve the efficiency of calibration image processing,the adaptive localization algorithm of calibration image is studied.The feature extraction layer,the feature discrimination layer and the fine-tuning frame structure are designed based on the features of the nested ring in the calibration image and the deep learning method.In this paper,a new algorithm of nested ring recognition for calibration plate images with different angles and different degrees of blur is proposed based on the Convolutional Neural Networks(CNN).The experimental results show that the proposed algorithm is better than the existing one,the recognition accuracy of the algorithm is improved from 37.14% to 88.57%,which effectively improves the recognition accuracy of the modulus and inclination calibration images.Secondly,after determining the calibration area by the above algorithm,the research of the image restoration of the checkerboard calibration image is carried out.On the basis of analyzing the existing fuzzy restoration algorithms and according to the characteristics of BOS calibration image,a self-adaptive knife-edge method is proposed to improve the accuracy of cutting edge region selection and the efficiency of processing the cutting edge calibration image.At the same time,this paper analyzes the problem of blurred image restoration from the perspective of deep learning,and recovers the calibration image by using CNN and VDSR(Very-Deep Super-Resolution)networks.On the other hand,by combining the generic adversary network(GAN)with the VDSR Network,the VDSR-GAN network is designed.Experimental results show that compared with other algorithms,VDSR-GAN network has the most obvious feature of reconstructed calibration image.Finally,the calibration experiments of single-camera,multi-camera and wind tunnel test section are carried out,and the calibration results before and after image restoration are compared respectively.The experimental data show that the calibration efficiency and accuracy of the BOS calibration image can be improved by the algorithm.