A Boundary-Enhanced Multibranch Network for Improved Object Detection in Low-Resolution Millimeter-Wave Images

With the rapid development of millimeter-wave technology in security inspection, its application in detecting contraband has demonstrated significant advantages. However, the low resolution, low pixel density, and unclear boundaries of terahertz images generated by millimeterwave imaging significantly limit the accurate detection of contraband. This paper proposes a Boundary-Enhanced Multi-Branch Network (BEMBNet) model based on a three-branch architecture. By designing a dedicated boundary branch and dynamic fusion modules, the model optimizes the extraction of target boundary features and the representation of global semantic information. The model efficiently integrates multi-scale features using Dynamic Reinforced Feature Pyramid Pooling (DRFPPM) and Boundary- Aware Multi-scale Fusion (BAMF) modules. Experiments demonstrate that BEMBNet outperforms existing advanced models across multiple metrics, particularly in low-resolution and boundary-blurred scenarios, exhibiting exceptional segmentation accuracy and computational efficiency. This research provides an innovative solution for millimeter-wave image processing and offers new directions for advancing security inspection technologies.