YOLO-XRAY：跨尺度注意力与深度可分离卷积在实时X光违禁品检测中的应用

YOLO-XRAY: Cross-Scale Attention and Depthwise Separable Convolutions for Real-Time X-Ray Contraband Detection

摘要: X射线安检系统在公共安全领域发挥着至关重要的作用，其核心任务是实现对潜在违禁品的实时检测，以预防安全风险和突发事件。然而，现有的传统X射线违禁品检测模型普遍存在结构复杂、参数量大、计算开销高等问题，不仅难以在轻量化的小型安检设备上高效部署，而且在检测精度和推理速度方面也存在明显局限，难以兼顾实际应用对高效性与实时性的双重要求。针对上述问题，本文提出了一种基于改进YOLOv8的轻量化检测模型——YOLO-XRAY。该模型在设计过程中充分兼顾了复杂场景下违禁品检测的实际需求。首先，引入跨尺度注意力机制，有效融合高、低分辨率特征层的信息，实现全局上下文与局部细节的协同建模，从而显著提升了小目标在复杂背景与遮挡条件下的显著性与检测精度。其次，采用深度可分离卷积替代标准卷积，有效降低冗余计算与参数规模，从而大幅减少计算成本并提升模型的部署效率。在SIXray数据集上的实验结果表明，改进后的YOLO-XRAY模型在保持轻量化结构的同时，mAP50提升了5.8%，参数量减少了5.5%。这一结果不仅验证了模型在违禁品识别任务中的有效性与优越性，也充分证明了其在实际安检场景中实现高精度、低延迟检测的应用潜力。该研究为智能安检系统的轻量化设计与实用化落地提供了一种切实可行的解决方案，对推动公共安全领域智能化发展具有重要意义。

Abstract: X-ray security inspection systems play a crucial role in the field of public safety, with their core mission being the real-time detection of potential contraband items to prevent security risks and emergencies. However, existing conventional X-ray contraband detection models often suffer from excessive structural complexity, large parameter sizes, and high computational costs. These limitations hinder their efficient deployment on lightweight inspection devices and restrict their performance in terms of detection accuracy and inference speed, making it difficult to simultaneously meet the dual requirements of efficiency and real-time applicability in practical scenarios. To address these challenges, this paper proposes a lightweight detection model based on an improved YOLOv8, termed YOLO-XRAY. The model is designed with full consideration of the practical demands of contraband detection in complex scenarios. First, a cross-scale attention mechanism is introduced to effectively integrate information from both high- and low-resolution feature layers, achieving collaborative modeling of global context and local details. This significantly enhances the saliency and detection accuracy of small objects under complex backgrounds and occlusion conditions. Second, depthwise separable convolutions are employed to replace standard convolutions, effectively reducing redundant computation and parameter scale, thereby substantially lowering computational cost and improving deployment efficiency. Experimental results on the SIXray dataset demonstrate that the improved YOLO-XRAY model achieves a 5.8% increase in mAP@50 while reducing the number of parameters by 5.5%, all while maintaining a lightweight structure. These findings not only validate the effectiveness and superiority of the proposed model for contraband recognition tasks but also confirm its potential for achieving high-accuracy, low-latency detection in real-world security inspection scenarios. Overall, this study provides a practical and feasible solution for the lightweight design and deployment of intelligent security inspection systems and holds significant implications for advancing the intelligent development of public safety technologies.