Mathematical Degradation Model Learning for Terahertz Image Super-Resolution

• Published in: IEEE access 2021, Vol.9, p.128988-128995
• Main Authors: Lu, Yao, Mao, Qi, Liu, Jingbo
• Format: Article
• Language: English
• Subjects: Artificial neural networks; Blurring; CNN; Convolution; Degradation; Image enhancement; Image quality; Image resolution; Image restoration; image super-resolution; Imaging; Kernel; Kernels; Mathematical analysis; mathematical degradation model; Mathematical model; Mathematical models; Superresolution; THz-TDS imaging; Time domain analysis
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2021.3113258

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