Abstract: Conventional security screening methods face difficulties in effectively identifying non-metallic prohibited items (such as ceramic knives and liquid explosives), and some technologies carry risks of ionizing radiation exposure. In contrast, millimeter-wave security screening technology can efficiently detect such hazardous materials while maintaining a significant safety profile for humans. This paper proposes HM-YOLO, an improved millimeter-wave radar image object detection method based on the YOLOv8s network, designed to address critical challenges in millimeter-wave images, including low signal-to-noise ratio, blurred boundaries, and complex backgrounds. By developing a Haar wavelet transform attention downsampling (HWC) module that integrates Haar wavelet transform with a convolutional block attention module (CBAM) to replace conventional downsampling convolution modules, target recognition performance is effectively optimized. Simultaneously, the introduction of the multi-scale fusion module (MSFM) and multi-scale pyramid pooling fusion (MPPF) module significantly enhances the discernment capability for target details and boundaries. Experimental results demonstrate that the proposed method substantially improves target recognition accuracy and robustness, Compared to the baseline model YOLOv8s, our method achieves improvements of 2.2% in mAP0.5 on the human prohibited items WM-SAR dataset and 2.7% on the remote sensing dataset LS-SSDD, while outperforming existing models including YOLOv8s and YOLOv9s across multiple evaluation metrics.