The data volume of millimeter-wave security imaging is relatively small, and the data volume for uncommon types of dangerous objects is even smaller, with the data exhibiting a long-tail distribution. Security inspection requires fast detection speed, and existing deep learning methods are not yet fully applicable to millimeter-wave security imaging datasets. To meet the needs of daily security inspection, a method based on the YOLOv5 algorithm is proposed for feature extraction of human hidden targets in millimeter-wave security radar imaging. Firstly, the YOLOv5 algorithm framework is studied, and a Focal CIoU loss function is proposed to re-weight the samples and reduce the long-tail effect. Next, the dataset is processed by cropping the targets and randomly pasting them into existing images to expand and balance the data volume among different categories, achieving the purpose of resampling. Finally, the Squeeze-and- Excitation Network(SENet) attention mechanism is introduced to improve the accuracy of target recognition. The validation results show that the mean Average Precision (mAP) of human hidden targets using the proposed method reaches 85.4%, which is a 4.7% improvement compared to the original YOLOv5 algorithm. This performance meets the detection requirements for daily usage scenarios.