Research On High-resolution Reconstruction Algorithm For Terahertz Image

In recent years,frequent international terrorist incidents have put forward stricter requirements for security check system.Terahertz waves are popular because of their good penetrability and non-destructive detection.Therefore,security inspection systems using Terahertz waves are developing rapidly.However,due to the diffraction limit of optical imaging and the technological level of the sensor itself,the Terahertz image obtained by the imaging system is of low resolution and severe edge blur.The most direct way to solve this problem is to upgrade the hardware equipment of the imaging system.However,the current technology makes the cost of upgrading the hardware expensive.Therefore,it is particularly necessary to study software methods to improve the resolution of the Terahertz image to make up the lack of existing hardware equipment.In this paper,based on the practical Terahertz scanning imaging system,the super-resolution reconstruction of Terahertz images is studied,focusing on the improved convex set projection and convolutional neural network superresolution reconstruction algorithm.Firstly,the mechanism of Terahertz detection and imaging is analyzed,an imaging platform based on the existing equipment in the laboratory is built to provide image sources for subsequent algorithm verification.According to the imaging platform,the terahertz image quality reduction model is discussed,and the commonly used reconstruction and learningbased super-resolution reconstruction theories are analyzed and compared.The advantages and disadvantages of each theory are clarified,which provides a theoretical basis for the improvement of the algorithm.Secondly,the theory and realization process of convex set projection algorithm are studied,and the classical algorithm is improved to solve the problems of less high-frequency information and blurred edges of Terahertz image.A new edge-guided interpolation is used instead of bilinear interpolation to obtain the initial image.A method of weighting the templet factor of the central pixel at the verge is used to optimize the PSF,where the weighting value can be adjusted adaptively with the change of the angle between the edge and the horizontal direction.It is verified by related experiments that the improved convex set projection algorithm maintains better edge details of the reconstructed Terahertz image.Finally,on the basis of the fast super resolution convolutional neural network,the network model is optimized for the characteristics of Terahertz image edge information blurring and the defects of the model itself.The number of feature extraction layers in the network model is optimize to extract richer feature information.The size of the deconvolution layer to is optimized to save model parameters.A residual structure is introduced to deepen the depth of the network model.An edge-related information to the loss function and the comparison between detection and structure similarity is added to improve the edge retention of the reconstructed image.The performance of the algorithm is verified by using the actual tool Terahertz image,which shows that the edge strength and average gradient are improved compared with the original algorithm.