Efficient Multimodal Data Augmentation Method for Training Deep Neural Networks in Millimeter-Wave Imaging Systems

Monitoring the security of important and busy places such as airports and train stations is very important. Millimeter-wave (MMW) security scanning systems are among the most prominent technologies used for this purpose. One of the key advantages of these systems is the automatic detection of objects hidden under clothing from millimeter wave images using deep neural networks, which increases the speed and accuracy of object identification and preserves people’s privacy. In this article, three different deep neural networks are investigated for automatic object detection in millimeter wave images. Still, due to the variety of humans’ body structures, objects and materials, a huge dataset gathered in a variety of conditions is a must. Furthermore, materials with low reflection coefficients have a very low contrast on human body, which makes their detection with any automatic detection algorithm a real challenge. To overcome these challenges, in this article two new data augmentation methods; i.e., the spectral data enhancement and geometrical data augmentation; are proposed, tailored based on the characteristics of the millimeter wave imaging system. The proposed geometrical data augmentation method increases the number and diversity of the training dataset required for training deep neural networks in the targeted subject. While, the proposed spectral data enhancement adds an extra dimension to the input data to overcome the low contrast of some materials on the body. The presented results in Section V show that the use of these data augmentation methods increases the accuracy of deep neural networks significantly, which ultimately help to detect objects of variant reflectivity level millimeter wave images with a limited dataset.