# YOLOv11-XRBS: Enhanced Identification of Small and Low-Detail Explosives in X-Ray Backscatter Images

## TL;DR

This paper introduces YOLOv11-XRBS, a new framework for detecting small explosives in X-ray backscatter images, achieving high accuracy with a custom dataset and improved detection strategies.

## Contribution

The novel framework YOLOv11-XRBS introduces adaptive architecture, Size-Aware Focal Loss, and a recomposed loss function for enhanced explosive detection in XRBS images.

## Key findings

- YOLOv11-XRBS achieved a mean average precision (mAP) of 94.8% on the SBCXray dataset.
- The framework outperformed existing YOLO variants and classical models like Faster R-CNN and SSD512.
- The proposed methods improved detection of small and low-detail explosives in cluttered backgrounds.

## Abstract

Identifying concealed explosives in X-ray backscatter (XRBS) imagery remains a critical challenge, primarily due to low image contrasts, cluttered backgrounds, small object sizes, and limited structural details. To address these limitations, we propose YOLOv11-XRBS, an enhanced detection framework tailored to the characteristics of XRBS images. A dedicated dataset (SBCXray) comprising over 10,000 annotated images of simulated explosive scenarios under varied concealment conditions was constructed to support training and evaluation. The proposed framework introduces three targeted improvements: (1) adaptive architectural refinement to enhance multi-scale feature representation and suppress background interference, (2) a Size-Aware Focal Loss (SaFL) strategy to improve the detection of small and weak-feature objects, and (3) a recomposed loss function with scale-adaptive weighting to achieve more accurate bounding box localization. The experiments demonstrated that YOLOv11-XRBS achieves better performance compared to both existing YOLO variants and classical detection models such as Faster R-CNN, SSD512, RetinaNet, DETR, and VGGNet, achieving a mean average precision (mAP) of 94.8%. These results confirm the robustness and practicality of the proposed framework, highlighting its potential deployment in XRBS-based security inspection systems.