Research On Method Of Image Distortion Correction And Restoration In Scattering Imaging

Laser active detection imaging is an active detection imaging technology that can accurately and quickly obtain the three-dimensional spatial information of the object.It has high resolution,measurement accuracy and anti-interference ability,so it is widely used in military reconnaissance,remote sensing mapping and automatic driving.However,when the laser active detection passes through the scattering medium,the object image obtained by scattering imaging will be seriously distorted,and the accurate object information cannot be obtained.The intensity distribution of the distorted image obtained by scattering imaging fluctuates randomly,and it is difficult for traditional image processing methods to restore a clear original object image.In recent years,the image distortion correction and restoration method of scattering imaging has become a research hotspot at home and abroad.At present,the existing methods to restore distorted images through scattering medium include wavefront modulation,transmission matrix,speckle correlation and deep learning.Among them,the wavefront modulation method and the transmission matrix method need complex modulation and measurement processes,and have poor real-time performance,which are severely limited in practical applications.The speckle correlation method uses the optical memory effect of scattering medium to suppress the scattering effect of the medium,obtains the Fourier amplitude information of the object,and then uses the iterative phase recovery algorithm to restore the distorted image.Although the time resolution of this method is very high,the convergence speed and operation time of the phase recovery algorithm are limited,and the accurate object image cannot be recovered quickly.At the same time,the size of the object is also limited by the optical memory effect range of scattering medium.The deep learning method can obtain the mapping relationship between the original image and the distorted image through training,so as to realize the rapid restoration of the distorted image to the original image.However,the training process of the current deep learning methods is usually supervised,and the training data need to be aligned and labeled in advance,which limits the development of scattering distortion images restoration in practical applications.For the problem that the convergence speed and operation time of the iterative phase recovery algorithm are not fast enough,this thesis first studies the distortion mechanism of images passing through scattering medium,and analyzes the working principle of speckle correlation based scattering distortion image restoration.A laser active imaging system through scattering medium is designed.The influencing factors of scattering distorted image restoration based on iterative phase recovery algorithm are analyzed.An improved method of iterative phase recovery algorithm is proposed.The absolute value is used as the output condition,which improves the information utilization efficiency and accelerates the convergence speed of the algorithm.Through numerical simulation,the restoration quality of several typical phase recovery algorithms is compared and analyzed.Moreover,an imaging system through scattering medium is built.The experimental results show that the improved method can restore the distorted image through scattering medium in only 30 iterations,and the operation time is less than that of hybrid input-output/error reduction algorithm.For the problem that the distorted image restoration through scattering medium is limited by the range of optical memory effect,a laser active imaging system through scattering medium is designed in this thesis.This thesis proposes an image restoration method based on reflectivity difference through scattering medium beyond memory effect range,according to the difference of object reflectivity and the difference of object autocorrelation intensity distribution.The autocorrelation of the highest part of the object reflectivity is obtained by peak searching,and this part of the object image is restored by iterative phase recovery algorithm.Then,based on the correlation between the restored object and its adjacent objects,the deconvolution recursive algorithm is used to restore the adjacent objects.The numerical simulation and experimental results show that this method can still restore the distorted image through scattering medium beyond three times optical memory effect range without the help of prior information,and the peak signal-to-noise ratio is more than 20 d B.For the problem that the deep learning method depends on the supervised training and labeled data,this thesis proposes a method to restore the distorted image through scattering medium based on the unsupervised learning by combining the speckle correlation theory and the unsupervised learning method,which avoids the labels of the distorted image and the original image in the data preprocessing process.According to the principle of image transformation consistency,a network architecture for restoration of distorted images through scattering medium based on unsupervised learning is designed.The image transformation between the autocorrelation of distorted images and the autocorrelation of original object images is realized through unlabeled data training.Then,a pre trained convolutional neural network is used to restore the object image autocorrelation to a clear object.The experimental results show that this method can accurately restore the unknown object images with different complexity through scattering medium with different statistical characteristics,and the restoration results are more stable than those of the supervised learning method,which is not affected by the unlabeled training data.This shows that the unsupervised learning scheme in this thesis is very suitable for processing the unlabeled data with disordered order obtained in practical applications.The research results in this thesis improve the restoration speed of the distorted image obtained by scattering imaging,overcome the limitation of the optical memory effect range on the restoration of the distorted image of scattering imaging,and realize the training and restoration prediction of unlabeled scattering distorted image dataset.It provides support for the development of image distortion correction and restoration methods of scattering imaging,and provides a powerful reference for the application of scattering image in biomedical and military security fields.