Transformer-Based Anchor-Free Detection of Concealed Objects in Passive Millimeter Wave Images

Passive millimeter-wave (PMMW) imaging has merits of nonradiation and good penetrability to most clothes, hence it has been a reliable security technique for the detection of concealed objects. At present, deep learning-based approaches have shown a great advantage in automatic detection. However, the low resolution and high background noise of PMMW images make the task tricky, especially for small objects. In this article, we propose a transformer-based anchor-free detector with the integration of local/global information and adaptive label assignment to address the aforementioned issues. Compared with the existing anchor-based methods adopted for PMMW image detection, our detector can further improve the efficiency and remove the handcraft anchor boxes. To be specific, we first employ hierarchical transformer architecture as the backbone, which has the capacity to model long-range dependencies of the feature at different scales. We propose a new strategy that calculates self-attention within the local region/global region in turn, providing detailed and global features of small objects. Second, we design a learnable position encoding module to obtain positional information between pixels. We propose an attention weighting module that enables the network to adaptively refine the features and distinguish positive and negative samples. Finally, we propose an adaptive label assignment strategy to dynamically optimize the number of positive samples used for detections. The proposed method is validated on our self-developed PMMW imager. The experimental results show that our method achieves better performance on accuracy and competitive speed compared with the state-of-the-art methods.