Research On Scattering Imaging Technology Based On Speckle Autocorrelation Perception Learning

Optical imaging is an important way for people to record information.However,there are various scattering media in nature,and scattering phenomena are inevitable in the imaging process.When the light wave is scattered by the medium,the phase distortion of the wavefront will be caused,and the distorted speckle image will be received at the image plane,making it impossible to directly and clearly image the target object.Traditional scattering imaging schemes physically model the causes of speckle.Due to the complexity of the scattering system,this method is insufficient in imaging speed and imaging quality.Therefore,it is of great significance to study how to achieve faster and clearer imaging through scattering media.This paper mainly uses speckle autocorrelation and deep learning technology,relying on the huge advantages of deep learning technology in imaging speed and quality,combined with the physical process of speckle autocorrelation and deep learning data-driven method to carry out joint modeling optimization of physical perception learning,through Train the network model to achieve speckle image reconstruction based on physical prior information.The specific research contents are:Firstly,the imaging method of speckle autocorrelation behind scattering medium is studied according to the speckle formation process and the prior characteristics of image statistics.The optical speckle size and imaging field of view are theoretically deduced,and the influence of background items on imaging quality is analyzed.It solves the failure problem of Fienup-type phase recovery algorithm under low-power incoherent light illumination,and realizes the reconstruction of target object information from low-contrast and low signal-to-noise ratio images.Secondly,in order to further improve the imaging speed and quality,the powerful optimization ability of deep learning technology is introduced into the anti-scattering reconstruction.By improving the Unet network model,a deep residual shrinkage segmentation network DRS-Unet is proposed.The proposed network can Remove the redundant noise in the speckle image,and use the attention mechanism to achieve more efficient feature extraction,so as to improve the low quality of network model reconstruction and poor generalization ability.Finally,by combining speckle autocorrelation technology and deep learning technology,the physics-driven and data-driven joint modeling of scattering problems is realized,and the autocorrelation prior information is used to guide the neural network optimization process.Compared with the pure data-driven approach,the physical-data combined drive method greatly reduces the dependence on the amount of data,and has been verified by simulation and experiments.The results show that the method can effectively suppress image noise and nonlinear distortion,improve the accuracy and stability of reconstruction,and is suitable for measurement and imaging of complex objects.