Research On Object Identification In Computational Ghost Imaging And Optimization Of Recovery Algorithm Based On Deep Learning

Ghost imaging is a new imaging method that uses the second-order correlation characteristics of the light field to calculate the target image.Due to this special imaging method,compared with traditional imaging methods,it can achieve anti-scatter imaging in poor optical environments such as rain,snow,fog and haze,and has strong anti-noise.Therefore,in some specific scenes,ghost imaging has more advantages than traditional imaging method.However,the common problem of ghost imaging is that its imaging quality is positively related to the number of samples.High-quality imaging results mean a lot of time cost.This article starts from the principle and uses deep learning methods to optimize the performance of the computational ghost imaging system from two perspectives.First of all,for the scene of using computational ghost imaging to classify and recognize objects,this article jumps out of the mindset that recognizing objects is to recognize their images,and proposes a method that can use deep network models to directly recognize light intensity sequences before imaging.The experiment uses the data sampling method based on the hadamard matrix to verify the effectiveness of the proposed two-way network model(Ghost Imaging Hadamard Neural Network,GIHNN)to directly identify the light intensity sequence,in this way,it can achieve fast end-to-end recognition,and greatly improve the recognition efficiency.Through comparative experiments of different sampling rates,this method can effectively identify objects at a sampling rate as low as 9.77%.The image reconstructed by the basic correlation function at this sampling rate is full of noise and difficult to recognize.Moreover,the processing time of this method is shorter than identify an image after the imaging process.Then we considered the problem of basic correlation function that requires multiple sampling to image effectively.Based on the idea of feature fusion,we took advantage of its ability to effectively fuse the semantic information of different levels of images,and proposed a deep learning model to reconstruct the image directly from the light intensity sequence instead of the basic correlation function.By comparing the results of the two loss functions,this method can reconstruct higher quality images at a low sampling rate compared with the similar method.At the same time,we analyze the common defects of the existing methods and our method caused by the supervised learning training method,and points out the direction for future research.In this paper,this article organically combines deep learning and computational ghost imaging to optimize system performance from two perspectives.The identification method before imaging jumped out of the mindset that the identifying an object means identifying its image,using a two-way network model to achieve rapid recognition at low sampling rates,and it also proved that one-dimensional light intensity information has basic image semantic information;The image reconstruction algorithm in the imaging process is replaced with a network based on feature fusion,which realizes that only the target image is reconstructed from the light intensity sequence,and the original image can be effectively reconstructed at a low sampling rate.From two perspectives,this paper uses deep learning to solve the contradiction between the sampling time and imaging quality in computing ghost imaging,which is conducive to promoting the practical process of computing ghost imaging and popularizing deep learning to more application fields.