FA-UNet: Semantic Segmentation of Passive Millimeter–Wave Images for Concealed Object Detection

Abstract

Passive millimeter–wave (PMMW) scanners are widely used for personal security screening in public places due to their nonradiation and high real-time capabilities. However, the images obtained by these scanners frequently exhibit low signal-to-noise ratios and contrast, presenting challenges for automated detection systems. To address this issue, we propose an efficient semantic segmentation approach, FA-UNet, that employs a UNet architecture with a fusion attention mechanism to conduct binary classification (human body vs. background, including objects) for PMMW images. This approach incorporates a spatial attention mechanism into the lateral connections between the encoder and decoder and introduces a channel attention mechanism during the feature fusion process in the decoder. By combining these attention mechanisms, FA-UNet leads to more precise segmentation outcomes. The segmented image is then fused with the original image using our multistage fusion method, in which, first, the two images are blended in a 1:1 ratio for object detection. Then, a new fused image is obtained by adjusting the ratio within a certain range (0.3–0.5). Finally, the object detection results are overlaid on this fused image to generate a directly displayable image. We evaluate our method using a self-made dataset. Experimental results demonstrate that FA-UNet can accurately segment the human body region and preserve object shapes effectively. Using the fused image for object detection helps reduce false detections caused by background noise interference while improving the detection rate of weak targets. Additionally, the fused image aids in manual image interpretation in locations with higher security inspection levels and contributes to protect the privacy of individuals undergoing inspection to the greatest extent possible.