Terahertz Single Image Super-Resolution reconstruction algorithm based on the deep residual network

Abstract:
Glass fiber reinforced plastics (GFRP) are widely used in aviation, electrical insulating materials and chemical equipment. But its production quality is not stable. Through fast scanning imaging by Terahertz (THz) domain spectrometer to identify small defects in the image, usually obtaining only a single original image with low resolution (LR). Therefore, improving the resolution of terahertz image has become a research hotspot. In this paper, we propose a Terahertz Single Image Super-Resolution (SISR) method based on residual networks for non-destructive testing (NDT) of composite materials, using prefabricated defects as samples. In order to improve the resolution of terahertz images, a very deep convolutional neural network (VDCNN) based on residual learning which is based on the Super-Resolution Convolutional Neural Network (SRCNN) is constructed. This network can be used for Super-Resolution reconstruction of original LR images at terahertz. The effect of network layer number on Super-Resolution performance under different scaling time is studied. Over time, Super-Resolution performance improves. In addition, the peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) are calculated, and the results show that the high resolution (HR) image based on the deep residual network is superior to the traditional method.