Correlation Imaging Was Reconstructed Based On Residual Convolutional Neural Networks

Correlation imaging,also called ghost imaging or two-photon imaging,is an indirect imaging technique based on the quantum entanglement or intensity correlation.Different from traditional optical imaging technology,the imaging process of ghost imaging does not require an imaging lens,only two beams,one of which is called a signal beam and the other is called a reference beam.The former passes through a random mask(such as rotating ground glass)and irradiates the speckle field on the object,and then collects the reflected or transmitted signal from the object through a barrel detector that does not have spatial resolution capabilities,while the latter is used by a spatially resolved barrel detector.It is collected by a capable detector,and we then correlate the signals detected by the two detectors to calculate the image of the object.Therefore,ghost imaging has extremely strong penetrating power and ability to resist atmospheric turbulence,and has very broad application prospects in military,detection,search and rescue,medical and other fields.However,compared with traditional optical imaging,ghost imaging requires a lot of sampling to get high-quality images.Therefore,how to obtain a clear image when the number of samples is limited has become a difficult problem in real life applications of ghost imaging.Although the compressed sensing algorithm can greatly reduce the sampling rate,thereby reducing the imaging time,it still requires a long signal processing time,so the final imaging speed is still slow.Therefore,in this manuscript,we propose a method for reconstructing high-quality images using deep learning algorithms at low sampling rates.This experiment is based on the traditional ghost imaging experimental device through residual convolutional neural network to improve the speed and quality of imaging.First,use the correlation algorithm of traditional ghost imaging to perform correlation calculation on the data collected by the traditional ghost imaging experimental device,and then train the processed data.In the process of repeated training,we can obtain a model with all the image characteristics of the database,and then reconstruct the untrained image,so that the network can reconstruct a low-noise image from the new test image.A series of derivation hypotheses and experimental data show that this scheme greatly improves the efficiency of data collection and processing.When the number of samples is insufficient,especially at low sampling rates,the imaging speed of this scheme is much faster than that of the traditional scheme,Is about forty times that of traditional correlation imaging and about one hundred times that of pseudo-inverse algorithm,and the quality of the generated image is far superior to traditional correlation imaging algorithms and compressed sensing algorithms while ensuring the imaging speed.In addition,we also found that the scheme has strong universality.When the number of samples is 100,the reconstructed picture accuracy can reach more than 90%,and in the reconstruction of some cluttered numbers and not including It also has a good effect when training the handwritten letter images in the model,which makes the scheme have important potential in practical applications.