Luggage Security Scanning Optimization to Enhance Efficiency and Safety at Airports Using Deep Learning

Abstract

One of the most difficult and important challenges in aviation security is detecting concealed threats in the luggage at airports. Manual inspection is slow, operator dependent and is becoming impractical due to the volumes of passengers at modern airports. This paper presents a deep learning-based luggage security scanning system built on the CLCXray dataset, targeting twelve threat categories that span both sharp objects such as blades, scissors, knives, daggers, and Swiss army knives, as well as liquid-container threats including plastic bottles, cans, vacuum cups, glass bottles, carton drinks, tins, and spray cans. Detecting liquid threats is particularly challenging because their X-ray appearance varies significantly with content, container shape, and overlap with surrounding items. Our proposed system adopts YOLOv8n as the primary detection model and uses a structured three-phase transfer learning strategy — backbone freezing for domain warm-up, partial unfreezing for mid-level feature adaptation, and full fine-tuning for precision refinement — all optimised for CPU-only execution on commodity hardware. A class-aware oversampling scheme and lightweight Albumentations augmentation pipeline specifically address the imbalance between liquid and sharp-object classes. On the held-out test split of the CLCXray dataset, the proposed YOLOv8n model achieves a mean average precision at IoU 0.5 ([email protected]) of 0.835, a precision of 0.856, a recall of 0.770, an F1-score of 0.811, and a [email protected]:0.95 of 0.654. These results represent an improvement of approximately 9.6% in [email protected] and 20.9% in [email protected]:0.95 over the best published baseline. A comparative evaluation against YOLOv10n confirms that YOLOv8n delivers superior accuracy for this safety-critical domain, making it the recommended architecture for real-world airport deployment.