XraySafe-YOLO: Compact Scale-Aware Fusion and Contour-Motivated Attention for Recall-Oriented Detection of Occluded Prohibited Items in X-ray Images

Occlusion, weak texture, and scale variation make prohibited-item detection difficult in X-ray security images. This study presents XraySafe-YOLO, a compact detector built on YOLO11n that combines Contour-Guided Occlusion Attention with Lightweight Scale-Aware Feature Fusion. On OPIXray, at a fixed confidence threshold of 0.25 and intersection-over-union threshold of 0.50, the full model reduced false negatives from 201 to 163 in the seed-42 experiment. Across three split-and-training trials, the full model reduced false negatives and improved fixed recall over YOLO11n in every trial, with a mean false-negative count of 172.7 and a mean fixed recall of 0.9030. A supplemental single-seed evaluation on the 12-class PIDray benchmark increased mean average precision at an intersection-over-union threshold of 0.50 from 0.7938 to 0.8035 and reduced false negatives from 10,148 to 9,318. LSAFF produced the more stable individual improvement, whereas CGOA alone was sensitive to the OPIXray trial seed. Results on both datasets show the same main behavior: the proposed combination improves recall and reduces missed detections, while some false-positive and precision costs remain. The method is therefore positioned as a compact recall-oriented option for human-assisted X-ray security inspection rather than as a detector that improves every metric.

@article{Liu2026,
  title     = {XraySafe-YOLO: Compact Scale-Aware Fusion and Contour-Motivated Attention for Recall-Oriented Detection of Occluded Prohibited Items in X-ray Images},
  journal   = {International Journal of Advanced Computer Science and Applications},
  volume    = {17},
  number    = {9},
  year      = {2026},
  publisher = {The Science and Information Organization},
  author    = {Liu Liu and Wei Huang},
  doi       = {10.14569/IJACSA.2026.0170984},
  url       = {https://doi.org/10.14569/IJACSA.2026.0170984}
}