Classification of concealed objects using terahertz imaging and artificial neural networks

Abstract

Imaging in the terahertz frequency band is applied in a number of fields, such as security, medical or quality control. However, a low resolution or distortions of the images hinder the identification or recognition of the objects. To cope with the processing of visual information, artificial neural networks are broadly employed. In this work, the monochromatic radiation of 253 GHz was used to collect the image set of the investigated objects either in the air or covered with a packing material. Such a set was later used to train convolutional and generative adversarial neural networks poised for three tasks: (i) the classification of objects; (ii) the enhancement of image resolution; (iii) the identification of cover material. The obtained results demonstrated that the packaging materials were identified with an accuracy of 83.33%, while the investigated objects were classified with an accuracy of 89.42%. The PSNR metric of images with improved resolution reached up to 22.44 dB. The optical properties such as refractive indices and absorption coefficients of the packaging materials were also defined using terahertz time-domain spectroscopy, and it was found that the accuracy of object and material classification in general does not depend on the physical properties and type of a package.