Application Research Of Deep Learning In Optical Information Security And Computational Ghost Imaging

With the advent of the era of big data and the vigorous development of computer science,deep learning technology has solved the long-term problems in different fields by virtue of the powerful model fitting ability of neural network,and set off a wave of cross-integration between industries and deep learning technology.As a data-driven optimization algorithm,deep learning can significantly improve computational efficiency and image quality in the optical field,and break through the physical limitations of traditional imaging methods to a certain extent.In recent years,deep learning technology has been widely used in optical information security and computational ghost imaging.The main applications in optical information security are the optimization of information encoding and decoding process and the cryptanalysis of encryption system.The main applications in computational ghost imaging are to improve imaging quality,shorten imaging time and optimize the design of sampling base.In traditional methods,these applications are difficult to be compatible with or require expensive imaging equipment.However,the introduction of deep learning technology has paved a new way for the development of optical information security and computational ghost imaging.A multi-image encryption method based on frequency multiplexing and deep learning is proposed to solve the problems of low number of encrypted images and long decryption time in current multi-image encryption methods.Before encryption,a number of plaintext images are grouped.The plaintext images in each group are coded by random matrix and modulated by sinusoidal fringe successively.Specially designed sinusoidal fringe can stagger the main frequency information of each group of plaintext images in the Fourier frequency domain,so as to achieve frequency domain reuse.Finally,multiple groups of images are superimposed and scrambled to obtain the ciphertext image.In the decryption process,the deep neural network is used to improve the quality and speed of decrypted images.After the scrambled ciphertext is recovered,it is downsampled in the frequency domain,and then sent to the trained neural network,so that each group of corresponding plaintext images can be directly reconstructed.Numerical simulation experiments show that when 32 images are encrypted,the correlation coefficient between decryption results and real plaintext images can still reach more than 0.99.In addition,through histogram analysis,correlation analysis of adjacent pixels and anti-noise attack analysis,it is proved that the encryption method has the advantages of large encryption capacity,high robustness and fast decryption speed.Aiming at the problems such as strong data dependence and easy overfitting in the supervised learn-based computational ghost imaging method,an equalvariant computational ghost imaging method based on unsupervised learning is proposed,which can realize highquality image reconstruction without the original image as a label and without additional image prior constraints.The method mainly uses the consistency of measured values and group invariance of natural images to constrain the solution space.Firstly,the inherent priori conditions of deep neural networks can be used to reduce the uncertainty in the process of image reconstruction.Secondly,the image equalvariant properties derived from the group invariance of natural images can capture effective information outside the range space of the original measurement matrix,so as to obtain higher imaging quality and better generalization ability.In addition,after the method is trained on the simulated data,the trained deep neural network can directly reconstruct the image from the light intensity signal collected in the experiment without additional training.Compared with other traditional methods,the feasibility and advantages of this imaging method are demonstrated in numerical simulation experiments and optical experiments.This new method can provide new insights into computational ghost imaging technology and further promote its practical application in daily life.