The Application Research On Image Super-Resolution Algorithm Of Terahertz Continuous Wave Image

Terahertz imaging technology has a broad application prospect in the field of nondestructive testing and security inspection.However,due to the limitation of hardware and diffraction limit,terahertz images have low resolution,lack of detail information and poor quality,so it is of great significance to improve the resolution and quality of terahertz images.In this paper,the terahertz frequency-modulated continuous wave imaging system is used as the experimental object,and the image super-resolution algorithm based on deep convolutional neural network is used to improve the resolution and image quality of terahertz images.The degradation model of terahertz image is studied,which is modeled as the original high-definition image and blur kernel convolution and then down-sampled plus noise.The experimental method is used to measure the point of the terahertz frequency-modulated continuous wave imaging system at the focal plane.The diffusion function is given and its approximate mathematical expression is given.A deep convolutional neural network model with excellent performance is designed.The multi-level residual structure and long-short jump connection are used to increase the network depth without “gradient disappearance”.The 1×1 convolution is used properly to maintain performance and reduce the number of parameters.The batch normalization is used to speed up the network training.Training the network based on the degradation model of terahertz images.Adding different widths of blur kernels and different levels of noise during training,the network can simultaneously improve the pixel resolution and spatial resolution of the image,and its generalization ability is greatly enhanced,and multiple degradation conditions can be handled by only one network.The super-resolution method and other classical algorithms of this paper are compared and tested on the simulated terahertz image and the real terahertz image.The experimental results show that the super-resolution method can significantly improve the resolution and image quality of terahertz images,and its effect.Significantly better than the traditional method.The deep learning algorithm based on deep learning is extended to the image restoration algorithm.Denoising and deblurring algorithms based on deep learning are implemented by the same network model and training method.Experiments show that these algorithms are superior to traditional methods.Based on the MATALB GUI,an image restoration software was developed,which integrates image denoising,deblurring and super-resolution algorithms based on deep learning.The software interface is friendly and easy to operate,so ordinary users can easily apply these algorithms to process various images.