Attention-Guided YOLOv11 and Transformer Fusion for Enhanced Detection of Prohibited Items in X-Ray Security Images

This repository contains the implementation code, model architecture configuration, training scripts, and evaluation utilities for YOLOv11-TFNet, a Transformer-enhanced YOLOv11-based detector for prohibited-item detection in X-ray security images. The proposed model integrates C3TR Transformer blocks into the YOLOv11 multi-scale feature aggregation neck, enabling global contextual reasoning over pseudo-color X-ray composites. Experiments are conducted on the Balanced X-Ray Contraband Detection Dataset (Zhang et al., 2024), comprising 13,728 images across 12 contraband categories. Key results (seed=42, 50 epochs, early stopping):- mAP@50: 0.886- mAP@50-95: 0.712- Precision: 0.885- Recall: 0.823- F1-Score: 0.853 Multi-seed validation (seeds 0, 42, 2026): YOLOv11-TFNet 0.887 ± 0.004 mAP@50; ΔmAP = +0.046 over YOLOv11s baseline (95% CI: [+0.037, +0.049]). Files included:- yolov11_tfnet_train.py: Training, evaluation, multi-seed validation, and plot generation- yolov11_tfnet.yaml: C3TR Transformer neck architecture definition- data.yaml: Dataset configuration (12 classes, 70/15/15 split)- requirements.txt: Python dependencies- README_zenodo.md: Setup and usage instructions Associated paper: Scientific Reports (under review, 2026).Dataset: Balanced X-Ray Contraband Detection Dataset (Zhang et al., 2024, Scientific Data).License: CC BY 4.0