MDT-YOLOv8s: Enhanced YOLOv8s model for X-ray contraband detection

To address the challenges in X-ray security inspection, such as complex backgrounds, high rates of missed and false detections, severe occlusions, and scale variations, this paper proposes an improved detection model based on YOLOv8s, named MDT-YOLOv8s, to enhance detection accuracy. First, a multi-head self-attention mechanism is added to the final layer of the backbone network to improve feature extraction. Second, a Dynamic Head module is introduced to enhance the network's scale, spatial, and multi-task awareness, significantly improving detection performance. Finally, a Trapezoid loss function is designed to increase the relative loss of hard classified samples by assigning higher weights near the threshold μ, thus focusing attention on the optimization of difficult classified and error-prone samples. Experimental results on the SIXray dataset show that MDT-YOLOv8s improves the mean average precision (mAP) by 2.3% compared to the original YOLOv8s. The proposed model enhances detection capability and provides effective support for real-time security inspection applications.