Comparative Study of Deep Learning Architectures for Thermal vs. Millimeter-Wave Concealed Weapon Detection

Abstract

CWD is becoming more and more essential for smart city security, while TIR and PMMW are emerging as promising sensing techniques that can be used in a privacy-respectful way to achieve the CWD objective. However, current studies focus on evaluating a single deep learning architecture with respect to a single sensor modality, raising the question of which architecture achieves the best performance on each sensor modality. This work aims to fill this gap by carrying out a systematic comparison among eight different deep learning architectures, a custom CNN, and seven transfer-learning architectures (namely, ResNet-18, ResNet-50, ResNet-152, DenseNet-121, DenseNet-201, and GoogleNet) in terms of their performance on balanced and curated sets of TIR and PMMW images, comprising respectively 686 samples for each category. The models were trained using the same training settings (Adam optimisation function, cross-entropy loss, learning rate 0.001, and batch size 32), while being evaluated with respect to accuracy, precision, recall, F1-score, inference latency, and size. The custom CNN was the only architecture to generalise well to both modalities, achieving 97.6% accuracy. On the other hand, TIR outperforms its nearest competitor on the same modality (ResNet-50, 78.5%) by 21.5 percentage points and also happens to be the most compact and computationally efficient network for this modality. This suggests that the choice of architecture for hidden weapon recognition should be modality-specific and not generic, and that specialized domain-trained architectures are worth considering for sensing modalities, which deviate significantly from natural images on which transfer learning architectures are pre-trained conventionally.