Terahertz Image Super-Resolution Reconstruction Using Unpaired Real-World Images

Terahertz imaging technology has been widely used in applications of security detection, defect detection and biomedicine. The in- sufficient resolution is still the main defect of this technology so far. The use of deep neural network for super-resolution reconstruction of image is a mainstream enhancement of resolution methods. Aiming at solving the shortcomings of existing methods to obtaining training sets and the design of networks, this paper proposes to use real- world terahertz images as low-resolution images and real-world optical images as ground truth to organize unpaired training sets, also a cycle adversarial super-resolution network is designed for the characteristics of unpaired training sets and different picture styles. In this paper we trained the network using training sets organized with the terahertz images which was collected through the Internet, and the images suitable for training in the Office-Home Datasets. The obtained results with real terahertz images shown that through the proposed network the resolution of terahertz images improved, which proves the feasibility of this network.