X-ray security inspection images often suffer from low texture, small object scale, and severe occlusion, which lead to missed detections of prohibited items such as metal knives. Existing detectors, including YOLO-based methods, still have difficulty transferring the strong performance achieved on natural-light images to complex X-ray screening scenarios. This paper introduces a one-stage framework, X-Safe, to achieve systematic improvement through a three-level collaborative optimization strategy. First, a Large-Kernel Efficient Layer Aggregation Network with Attention (LKNA) is designed for backbone feature extraction. By progressively enlarging the convolutional kernels and integrating coordinate attention with global attention mechanisms, LKNA expands the effective receptive field and strengthens global shape representation. Second, a Lightweight Integrated Feature Fusion from Backbone into Neck (IFFBN) mechanism is proposed. Through multi-level feature alignment, adaptive gating fusion, and cross-scale feature injection, IFFBN compensates for small-target details that are weakened during multi-scale feature transmission. Third, a Task Alignment for Simplified Optimal Transport Assignment (TATA) strategy is incorporated. It combines optimal-transport-based dynamic matching with task-aligned classification-regression supervision, improving positive-sample assignment while maintaining linear complexity O(NM). X-Safe achieves 93.1% mAP50 on OPIXray, improving the YOLOv7(SimOTA) baseline by 9.4%. On CLCXray, X-Safe attains 85.1% mAP50, outperforming YOLOv7(SimOTA), YOLOv9-m, and YOLOv11-m by 1.5, 0.4, and 0.5%, respectively. Under a unified ablation protocol on OPIXray, removing LKNA, IFFBN, or TATA from the complete model reduces mAP50 by 2.7, 0.8, and 1.3%, respectively; in addition, within the complete framework, the progressive <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(5\times 5\)</EquationSource> </InlineEquation>/<InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(7\times 7\)</EquationSource> </InlineEquation>/<InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(9\times 9\)</EquationSource> </InlineEquation> kernel configuration outperforms an all-<InlineEquation ID="IEq4"> <EquationSource Format="TEX">\(3\times 3\)</EquationSource> </InlineEquation> variant by 4.6% mAP50, and TATA outperforms the SimOTA and TAL assignment strategies by 1.8% and 0.4% mAP50. In addition, X-Safe has 39.4 M parameters with a model size of 79.8 MB, requires 87.9 GFLOPs, and reaches 69.7 FPS on an RTX 3090 under a batch size of 1, showing a favorable balance between accuracy and real-time inference for X-ray security inspection.

X-safe: an X-ray security detection method based on incremental Kernel aggregation, hierarchical co-optimization and task-aligned labeling

期刊论文
发表日期：

2026年7月27日

查看全文

Jiajie Wu,
Xianghua Xu

摘要

X-ray security inspection images often suffer from low texture, small object scale, and severe occlusion, which lead to missed detections of prohibited items such as metal knives. Existing detectors, including YOLO-based methods, still have difficulty transferring the strong performance achieved on natural-light images to complex X-ray screening scenarios. This paper introduces a one-stage framework, X-Safe, to achieve systematic improvement through a three-level collaborative optimization strategy. First, a Large-Kernel Efficient Layer Aggregation Network with Attention (LKNA) is designed for backbone feature extraction. By progressively enlarging the convolutional kernels and integrating coordinate attention with global attention mechanisms, LKNA expands the effective receptive field and strengthens global shape representation. Second, a Lightweight Integrated Feature Fusion from Backbone into Neck (IFFBN) mechanism is proposed. Through multi-level feature alignment, adaptive gating fusion, and cross-scale feature injection, IFFBN compensates for small-target details that are weakened during multi-scale feature transmission. Third, a Task Alignment for Simplified Optimal Transport Assignment (TATA) strategy is incorporated. It combines optimal-transport-based dynamic matching with task-aligned classification-regression supervision, improving positive-sample assignment while maintaining linear complexity O(NM). X-Safe achieves 93.1% mAP50 on OPIXray, improving the YOLOv7(SimOTA) baseline by 9.4%. On CLCXray, X-Safe attains 85.1% mAP50, outperforming YOLOv7(SimOTA), YOLOv9-m, and YOLOv11-m by 1.5, 0.4, and 0.5%, respectively. Under a unified ablation protocol on OPIXray, removing LKNA, IFFBN, or TATA from the complete model reduces mAP50 by 2.7, 0.8, and 1.3%, respectively; in addition, within the complete framework, the progressive \(5\times 5\) / \(7\times 7\) / \(9\times 9\) kernel configuration outperforms an all- \(3\times 3\) variant by 4.6% mAP50, and TATA outperforms the SimOTA and TAL assignment strategies by 1.8% and 0.4% mAP50. In addition, X-Safe has 39.4 M parameters with a model size of 79.8 MB, requires 87.9 GFLOPs, and reaches 69.7 FPS on an RTX 3090 under a batch size of 1, showing a favorable balance between accuracy and real-time inference for X-ray security inspection.