Research On Photon Counting Single Pixel Imaging Technology Based On Attention Mechanism

Photon counting single pixel imaging combines photon counting technology with single pixel imaging based on compressed sensing,with the advantages of high sensitivity and low cost,so it has important application in optical remote sensing imaging,spectral imaging,biomedical imaging and other fields.Traditional reconstruction algorithms have the problems of long sampling time and long reconstruction time,major breakthroughs have been made in recent years by using deep learning for compressed sensing reconstruction.The compressed sampling and reconstruction network which is based on deep learning,using the fully connected layer without the offset and activation function as the measurement matrix,achieves faster and higher quality reconstructed images by avoiding the huge amount of calculation caused by traditional iterative algorithms.However,when the fully connected layer is used for block-compressed sensing of high-resolution images,the reconstructed image will produce block artifact.At the same time,most deep reconstruction networks treat the extracted image features equally,which leads to the network cannot make full use of the important information in the image.Centering on these two problems,this dissertation focuses on studying the deblock sampling network and attention mechanism reconstruction network for the photon counting single pixel imaging,the main research content and results are as follows:(1)This dissertation investigates the causes of block artifact,and proposed the overlapping block sampling network(Os＿net),nested sampling network(Ns＿net),convolutional sampling network(Cs＿net)to replace the fully connected layer sampling.The simulation experiment is designed to compare the reconstruction performance and deblocking ability of different algorithms.The experimental result shows that Cs＿net has a very good reconstruction performance compared with other algorithms and can completely remove the block artifact.Cs＿net is applied to a photon counting single pixel imaging system by binarizing the floating-point weights in the sampling network of Cs＿net.This dissertation compares the performance of Cs＿net with the traditional iterative algorithm.The experimental result shows that the proposed network model achieves a better imaging quality and can completely remove block artifact.(2)The reconstructed network based on the attention mechanism is designed.In order to explore the effect of different attention mechanisms for compressive sensing reconstruction,this dissertation proposes channel attention mechanism network(ACS＿net＿V1),spatial attention mechanism network(ACS＿net＿V2)and hybrid attention mechanism network(ACS＿net).The simulation experiment is designed to compare the reconstruction performance of three network,and the result shows that all the attention mechanism modules can improve the deep reconstruction network,among which ACS＿net is the most obvious.The binary simulation experiment compares the performance of ACS＿net with the traditional iterative algorithm,and the result shows that the proposed scheme can achieve higher imaging accuracy.(3)Based on spatial attention mechanism network,we designed the differentiation-weighted loss function ALoss,weighted separately for the mean variance of the different regions.This loss function gives larger weight to the important regions and smaller weight to the non-essential regions.The simulation experiment is designed to compare the influence of designed loss function and ordinary function on the network,and the result shows that the loss function we proposed improves the network with spatial attention mechanism.