Super-Resolution Reconstruction Algorithm for Terahertz Images

Terahertz imaging technology has great prospect in applications of nondestructive testing, biomedicine and security inspection. However, terahertz imaging has not received a satisfactory image quality so far due to problems such as complex noise, artifacts and low resolution. In this work, an image preprocessing algorithm by combining an improved wavelet thresholding function and a generative adversarial network image super-resolution algorithm was proposed. Specifically, the constant deviation between the original wavelet coefficients and the estimated wavelet coefficients could be reduced by improving the threshold function and setting appropriate adjustment parameters to dynamically select a fixed threshold. As such the approximation between of the reconstructed image and the original image could be improved. In addition, the generative network combined with the improved loss function was designed to obtain the detailed information of the images at multiple scales. Moreover, an image database containing 51,300 images was built by degrading and data expanding the original 3650 images collected from the DF2K dataset so as to evaluate the proposed super-resolution reconstruction algorithm of terahertz images. The obtained results show that there were more than 2 dB and at less 1 dB improvement compared with previous median filtering and the hard/soft threshold wavelet, respectively. Consequently, the preliminarily reconstructed terahertz images suggested that the improved threshold wavelet function together with the generative adversarial network may offer the possibility of provide a high quality image especially for nondestructive testing.