A dangerous item segmentation algorithm for passive terahertz imaging security inspection is proposed in response to the difficulty and low precision of dangerous item recognition in the passive terahertz imaging. First of all, the hypothesis of the dangerous item local structural difference and the hypothesis of the local luminance difference are made to locate the Region Of Interest(ROI) where dangerous items might exist in terahertz images. Meanwhile, the shallow convolutional network containing a few feature channels and nerve cells is chosen for super-resolution processing of images in ROI regions. The images are input into the U-net to obtain high-quality and clearly-outlined partitioned images of dangerous items. Finally, an experiment is conducted to verify the improvement of the detection accuracy of the proposed algorithm in comparison with traditional partitioning algorithms. This is conducive to improving the recognition rate of dangerous items by the passive terahertz imaging security inspection system.