A Balanced Multi-Modal Thermal and Passive Millimeter-Wave Dataset for Deep Learning-Based Concealed Weapon Detection

In the field of smart city security surveillance, automated concealed weapon detection using non-invasive imaging remains a challenge. However, no publicly available dataset provides balanced, paired, and pre-processed thermal infrared (TIR) and passive millimeter-wave (PMMW) imagery specifically designed for training deep learning models and facilitating fair cross-modality comparison. The present study aimed to curate and standardize a balanced, multimodal dataset that enables the systematic evaluation of detection algorithms under controlled conditions. A dataset was designed that comprised 686 images from each of the secondary Thermal Infrared (TIR) and Passive Millimeter Wave (PMMW) modalities, using the FLIR Breach PTQ136 camera and Ka-band radiometer of 34 GHz frequency, respectively. The images were subjected to image flipping, rotation, scaling to a size of 224 × 224 pixels, normalizing to [0, 1] scale, and splitting into stratified partitions (70:20:10). This led to balanced binary classes with an augmentation of 1.72× for the TIR modality. Thus, the study curated a balanced dataset of 1,372 thermal and millimeter-wave images. Results show successful class parity and standardized preprocessing. Recommendations include collecting synchronized image pairs and expanding to outdoor environments.