Research On Single Pixel Imaging Algorithm Based On Deep Learning

Single pixel imaging technology is a new kind of imaging technology,through the light field coding,using only a single pixel detector at a relatively low spatial resolution without sampling rate can rebuild the target image,the technology in high sensitive imaging,multi-spectral imaging,THZ imaging,infrared imaging,and many other fields has received the widespread attention.The single pixel imaging technique relies on the principle of compressed sensing to detect and reconstruct the target.In this paper,a single pixel imaging algorithm based on deep learning is studied.This paper introduces the research background and significance,and summarizes the research status of single pixel imaging technology at home and abroad.This paper introduces the basic principle of single pixel imaging,presents the typical D-AMP compressed sensing reconstruction algorithm,studies the two kinds of compressed sensing reconstruction algorithms based on ReconNet and DR2-Net,and retrains the two kinds of deep neural networks,and presents the simulation results of each algorithm.In this paper,a single pixel imaging network based on least squares generative adversarial networks is designed,and the corresponding generative networks and adversarial networks are constructed.Generative networks is composed of a sampling network and a reconstruction network,including sampling network using fully connected network to simulate compression perception of compression,sampling process,and at the same time in order to be able to in the digital micro mirror array and spatial modulator used in attachments,designed the binary function is used to sample network training,implementation of {0,1} matrix or {1,1} measurement matrix optimization.Deep expansion convolutional neural network is introduced into the reconstruction network to realize the reconstruction process of nonlinear signals,and the end-to-end mapping relationship between measured values and reconstructed images is learned.The network consists of a fully connected layer and seven convolutional layers.The discriminant network uses the deep convolutional neural network,which consists of nine convolutional layers,two pooling layers and two fully connected layers.The loss function is designed,and the perception loss function is used to obtain the loss function.The perception loss is obtained by mean square error loss,VGG network loss and counter loss.The influence of network parameter setting on network is analyzed.The training and testing of a single pixel imaging network based on the least squares generative adversarial networks are completed.For 128*128 resolution,the sampling rate is 0.1,0.05,0.01 respectively three conditions for training and obtained the sample matrix and the reconstruction of the corresponding network model,and the D-AMP algorithm,ReconNet,DR2-net and single pixel imaging network based on generative network without adversarial composition carried on the comparative analysis of the simulation study results show that the method in this paper,both subjective visual effect and objective image quality assessment have improved significantly,and the method in this paper is fast.According to the principle of single pixel imaging experiment,the experiment scheme was designed,the platform of single pixel imaging experiment was built,and the components of the experiment device were introduced.A single pixel imaging experiment is carried out by using the experimental platform,the experimental results verify the effectiveness of the proposed method,and the reconstruction quality is obviously better than that of the existing algorithm.