Research On Single-Pixel Imaging Based On The Physical Models And Data-Driven Method

Optical imaging technology is the basis of human understanding and perception of the natural world.Traditional point-to-point imaging methods mostly rely on high-resolution array detectors,which directly detect target objects through fixed optical paths and can only obtain two-dimensional spatial intensity information of the scene.As a new computational imaging technology,single-pixel imaging uses structured light time series coding to compress the spatial distribution features of the target to be measured into one-dimensional detection light intensity.With high sensitivity,wide spectrum,low-cost single-pixel detector and time-space,spectrum,and other multiplexing technologies,single-pixel imaging has an excellent performance in high dimension,hyperspectral,harsh environment imaging,and other fields,but there is also a tradeoff between imaging speed and reconstruction quality.Starting from single-pixel imaging technology,the first two chapters systematically comb through the generation and development process of single-pixel imaging and deep learning.Aiming at structural coding and algorithm reconstruction,imaging schemes under different sampling bases are introduced respectively from three aspects:intensity correlation operation,compressive sensing technology,and deep learning algorithm.In Chapter 3,a single-pixel imaging scheme of an untrained neural network driven by physics is proposed to aim at the common problems of dataset acquisition difficulty,generalization,and poor network interpretation in the single-pixel imaging method based on data-driven deep learning.The neural network input is only the single one-dimensional light intensity sequence collected by the single-pixel detector.The forward physical model of the single-pixel imaging is integrated with the network structure.Combined with the deep image prior characteristics of the randomly initialized network,the weight parameters of the network are automatically optimized and updated under the supervision of light intensity.When the loss function converges to a certain area,the network model can establish a mapping relationship from the measured value to the high signal-to-noise ratio image.Simulation and optical experiments verify the effectiveness of the proposed method.High-quality reconstruction results can be obtained under a high compression ratio,and the imaging efficiency is significantly improved.In Chapter 4,an imaging method based on a deep prior generalized alternate projection algorithm is proposed to aim at the low imaging efficiency and low color fidelity in single-pixel color imaging.After being modulated by the Bayer array,the object information is color-coded.In the computational decoding stage,the plug-and-play generalized alternate projection algorithm solves the optimization problem under under-sampling,and the pre-trained deep denoising network provides regularization constraints for parameter updating.At the same time,the image demosaicing algorithm is integrated into the iterative process,and the pre-trained deep demosaicing network is used to provide the prior information of object color for alternate projection.Under the low sampling ratio of 0.0625,the proposed method can obtain a high signal-to-noise ratio and high color fidelity of the color single-pixel reconstruction image.In Chapter 5,aiming at the problems that the efficiency of single-pixel imaging is limited by the encoding resolution and the paired data needs to be obtained in network training,a singlepixel super-coding resolution imaging method based on unpaired data-driven deep learning is proposed.In terms of measurement,the zig-zag spectrum sampling path is used to achieve Hadamard single-pixel imaging under low sampling.In terms of reconstruction,twodimensional images with low pixel resolution are reconstructed preferentially,and high-quality super-resolution images are reconstructed through trained generators,which can reduce the number of measurements exponentially.The cycle generative adversarial network is designed and used to realize the mutual mapping from low-resolution degraded images to high-resolution clear images by unpaired data-driven training.The network trained by the simulation data can be used for experimental reconstruction,and the super-coding resolution image recovery is realized at the low sampling rate(one-quarter of 3.125%).