Research On Deep Neural Network Method Based On Computational Ghost Imaging

The method of Computational Ghost Imaging(CGI)is based on correlation calculation by speckle patterns and the corresponding bucket values.The information of correlation calculation comes from the reference arm which contains the spatial information of projection pattern and the object arm which contains the information of the object to be imaged.CGI reduces the loss of traditional optical imaging in extreme environment to some extent because of its unique imaging mode.However,the disadvantages of the method still exist.The first problem is that the ghost imaging restoration effect is not as good as expected and even cannot be imaged at low sampling rate.This thesis discusses the depth neural network method based on CGI,and discusses the reconstruction of CGI from multiple perspectives using a variety of depth neural network mechanisms.In this thesis,we have completed the reconstruction of CGI based on multi-scale fusion of Dense Net,CGAN(Conditional Generative Adversarial Nets)and convGRU-U-Net,and analyzed the imaging results.Specifically,the following research work has been completed:1.A reconstruction algorithm of CGI based on multiscale fusion of density net is proposed.Firstly,reconstruction results at low sampling rate are recovered by traditional CGI.Then,feature extraction is carried out according to the Dense Net network,and the extracted features at different scales are fused through the pixel level spatial attention mechanism.Finally,results are obtained by reconstructing the network.The algorithm is trained by simulation data.In the simulation test set,this method can recover the reconstruction results with 10%sampling rate.At the same time,the analysis and verification in real environment demonstrate the practicability of this method.2.A reconstruction algorithm of CGI based on CGAN is proposed.Firstly,reconstruction results at low sampling rate are recovered by traditional CGI.According to the principle of game theory,the generator and discriminator are calculated.Different from the loss function of other networks,the generator is trained according to the results calculated by the discriminator.At the same time,in order to ensure the training effect of CGAN network,the critic mechanism is introduced to control the weight of parameters.Compared with the general GAN network,this method also introduces an additional loss function to the generator to ensure that the generation effect is gradually close to the label image.Finally,the traditional computational ghost image restoration method is used to obtain the reconstruction results of the test set respectively through the simulation data and the real data,which are used as the input of the network to verify the feasibility and effectiveness of the method,and a good reconstruction result is obtained under the sampling rate of 5%.3.A reconstruction algorithm of CGI based on convGRU-U-Net is proposed.Firstly,based on the convGRU(convolutional Gated Recurrent Unit)network,the projection pattern sequence and the corresponding bucket detector value sequence are used as the inputs of the network to reconstruct the target object.The number of convGRU cycles,that is,the number of convGRU neurons in the convgru network,is the sampling number of CGI.Then,the result of convGRU network is input into U-net network to enhance the reconstructed image and get the final reconstruction result.When the sampling rate of this algorithm is 3.12%,good reconstruction results are obtained through simulation data and real data,and the non-handwritten content is used as the test image to prove the generalization of the algorithm.