Infrared thermography as a non-invasive scanner for concealed weapon detection

Non-invasive scanning techniques are vital for threat detection in areas of heavy human traffic to ensure civilian safety. Longer waves in the electromagnetic spectrum, such as millimetre waves and terahertz, have been successfully deployed in commercial personnel scanning systems. However, these waves suffer from lower image resolution due to their longer wavelengths.

Infrared has a shorter wavelength compared to millimetre waves and terahertz. Infrared has a lower penetration potential compared to its counterparts but boosts higher image resolution due to its shorter wavelength. Machine learning techniques, i.e., principal component analysis, active contour, and Fuzzy-c, were applied to the infrared images to improve the visualization of concealed objects.

Convolutional neural networks, i.e., ResNet-50, were explored as an automatic classifier for the presence of concealed objects. A transfer learning approach was applied to an ImageNet pre-trained ResNet-50 model. After preprocessing the IR images using Fuzzy-c, two models were trained, using 900 and 3082 images, respectively. Evaluating the models using a confusion matrix and receiver operating characteristic curve, an area-under-curve of 0.869 and 0.922 was obtained. An optimization procedure was used to determine the model threshold, resulting in a prediction error of 19.9% and 14.9%, respectively.