An Efficient Advanced-YOLOv8 Framework for THz Object Detection

With the merits of low radiation energy and high sensitivity, terahertz (THz) imaging has become a promising candidate in security screening. However, objects with small size and low spatial resolution, substantial background noise, and inadequate training samples are still issues that affect the detection accuracy in THz images. To address these issues, an advanced YOLOv8 (A-YOLOv8) is proposed, in which an improved path aggregation network (I-PANet) is designed to excel in detecting objects with small size and low spatial resolution in THz images. Then, an effective image preprocessing strategy based on sliding partition and small area noise filling (SPSANF) is presented to eliminate background noise and enhance image contrast. Furthermore, a THz data augmentation (TDA) module specialized in THz images is introduced in A-YOLOv8 to generate rich and diverse samples. Numerical results demonstrate that A-YOLOv8 outperforms YOLOv8 by 5.0% and 3.9% on THz secure-detecting datasets, respectively. These results underscore the potential of the proposed approach in THz image object detection.