Mathematical Degradation Model Learning for Terahertz Image Super-Resolution

This study proposes a super-resolution (SR) method for terahertz time-domain spectroscopy (THz-TDS) images, combining convolutional neural network (CNN) and mathematical degradation model. The model considers three possible factors affecting the quality of THz images: blur kernel, noise, down-sampler. Specifically, kernel characterizes continual change image extent with imaging distance. designed CNN learns from then copes distance dependent restoration problem based on learned mapping between low high-resolution pairs. two-stage comparative experiment shows that proposed significantly improved images. To be specific, our enhanced resolution by factor 1.95 to 0.61 mm respect diffraction limit. In addition, achieved greatest improvement in terms quality, an increase 4.35 PSNR 0.10 SSIM. We believe could offer satisfactory solution THz-TDs SR applications.