The Application Of Deep Learning Technology In Terahertz Singlepixel Imaging

Terahertz imaging refers to the two-dimensional imaging of a target using electromagnetic radiation between 0.1 and 10 THz frequencies.Because terahertz wave has the advantages of low photon energy,penetration and spectral resolution,terahertz imaging technology is widely used in security detection,biomedicine,product quality inspection and other fields.Different from the visible light field,mature commercial terahertz cameras or array detectors are difficult to obtain,which greatly restricts the practical promotion of terahertz imaging technology.Thanks to the development of single-pixel imaging technology,terahertz single-pixel imaging can realize twodimensional imaging using terahertz single-point detector and imaging algorithm,which reduces the cost of the system.However,THz single-pixel imaging still has the problems of long imaging time and poor imaging quality,and simply shortening the imaging time of the encoding mask will certainly lead to the deterioration of the image.At present,deep learning is on the rise and widely used in image processing,speech recognition,target detection and other fields.There are few researches on deep learning network in terahertz single-pixel imaging,which means that how to use deep learning to solve the above problems is of great significance.In view of the above problems,this thesis designed two neural networks for terahertz single-pixel imaging to improve the imaging quality and shorten the imaging time.The main innovative research work includes:1.Aiming at the contradiction between terahertz single-pixel imaging quality and imaging time,a residual dual-path dense network（RDPDN）was proposed to raise the efficiency of terahertz single-pixel imaging.The original image is simulated by adding different levels of noise to the optical image,and the existing denoising network and RDPDN are used to train the image to verify the robustness of network denoising.Simulation results show that RDPDN has better image quality than existing denoising networks.The THz experiment results show that the sampling time of each mask can be reduced to 1/20 of that of the conventional system,the number of Hadamard patterns can be reduced to 10% of the pixels,while maintaining high image quality,the acceptable SNR is higher than 20 d B,and the structural similarity is more than 0.85.2.In addition,an end-to-end convolutional neural network is also proposed,which aims to use Hadamard encoding coefficient to directly image,eliminating the intermediate step of conventional image algorithm resolution.Optical images are encoded by Hadamard mask and reconstructed by a small number of coding coefficients.The results show that the end-to-end network proposed in this paper only needs 5%Hadamard coefficient to reconstruct THz image,which greatly reduces the time required for coding acquisition and reconstruction,and further reduces the imaging time without guaranteeing the imaging quality.