首页|反向加权融合多尺度特征的X射线图像违禁品检测

反向加权融合多尺度特征的X射线图像违禁品检测

扫码查看
针对现阶段违禁品检测方法存在的混叠效应以及在类内变化显著的场景下检测精度较低等问题,提出一种反向加权融合多尺度特征的X射线图像违禁品检测算法,通过反向自适应地引导融合多尺度上下文特征来实现准确的违禁品目标检测.首先,使用多尺度场景感知模块获取从局部到全局的目标表征信息,帮助处理显著的类内变化.其次,利用反向加权融合结构采用特征引导加权的方式,高效融合蕴含丰富上下文特征的多级特征,缓解融合过程中易出现的混叠效应.最后,设计了一种Focal-SIOU损失函数,用于平衡不同质量违禁品目标预测框之间的贡献差异,并结合角度和边长损失进一步提升预测框的收敛速度和回归精度.本文方法在SIXray、OPIXray、PIDray等3个非常具有挑战性的基准数据集上进行了广泛的评测实验,平均精度均值(mAP)分别达到93.2%、90.7%和85.1%.实验结果充分表明,方法相比于最新方法性能更优,并且能够满足实时目标检测的实际应用需求.
Reverse weighted fusion of multi-scale features for prohibited object detection in X-ray images
Aiming at the problems of aliasing effects and low detection accuracy in scenes with significant intra-class variations generally found in the existing prohibited object detection methods,this paper proposes a prohibited object detection algorithm for X-ray images with reverse weighted fusion of multi-scale features,so as to accurately detect the prohibited object by reverse adaptively guiding the fusion of multi-scale context features.First,a multi-scale scene perception module is used to obtain the object representation information from local and global,which helps to deal with significant intra-class variations.Second,by utilizing the reverse weighted fusion structure,the feature-guided weighting is employed to efficiently fuse multi-level features with rich context features,so as to alleviate the aliasing effects during the fusion process.Finally,a Focal-SIOU loss function is designed to balance the contribution differences between the predicted box of different quality for prohibited objects,and the convergence speed and regression accuracy of predicted box are further improved by combining the angle and side length losses.Extensive experiments were carried out on three very challenging benchmark datasets of SIXray,OPIXray,and PIDray by using the method proposed in this paper,and the mAP reached 93.2%,90.7%,and 85.1%on the three datasets,respectively.The experimental results have fully demonstrated that our method is not only better than the state-of-the-art methods,but also can meet the practical application requirements of real-time object detection.

prohibited object detectionmulti-scale fusioncontext featureloss function

马昌嵩、裴晓芳、周磊、周进、杨继海

展开 >

南京信息工程大学电子与信息工程学院 南京 210044

无锡学院电子信息工程学院 无锡 214105

违禁品检测 多尺度融合 上下文特征 损失函数

国家自然科学青年基金高校哲学社会科学研究课题苏高教会专项江苏职业教育研究课题无锡学院教改课题

422050782022SJYB09792022JDKT138XHYBLX2023282JGZD202107

2024

国外电子测量技术
北京方略信息科技有限公司

国外电子测量技术

CSTPCD
影响因子:1.414
ISSN:1002-8978
年,卷(期):2024.43(4)
  • 31