基于SuNet的公共交通安检违禁品的检测  

Prohibited items detection in public transportation security inspection based on SuNet

作　　者：张缓缓[1] 刘鹏程 姜萌 王雨欣 ZHANG Huanhuan;LIU Pengcheng;JIANG Meng;WANG Yuxin(School of Electronics and Information,Xi’an Polytechnic University,Xi’an 710048,China)

机构地区：[1]西安工程大学电子信息学院,陕西西安710048

出　　处：《西安工程大学学报》2025年第2期47-56,共10页Journal of Xi’an Polytechnic University

基　　金：国家自然科学基金青年项目(61902302);陕西省科技厅重点研发计划项目(2024GX-YBXM-231);浙江省博士后科研项目择优资助(ZJ2022154)。

摘　　要：在公共交通安检场景中,违禁品与非违禁品相互重叠,导致现有模型难以有效地识别被遮挡违禁品类别问题。针对这一问题,文中提出基于SuNet的违禁品检测模型。首先,设计了强化注意力定位特征金字塔网络(augmented attention localization feature pyramid network,AALFPN),以强化违禁品的语义信息,并将违禁品定位信息和语义信息融合,引导模型准确定位被遮挡的违禁品位置,增强违禁品的特征轮廓。其次,引入了密集注意力机制(dense attention mechanism,DAM),以有效地识别和提取被遮挡违禁品。最后,引入了SmoothL1 Loss损失函数解决在回归过程中违禁品类别信息丢失的问题。该实验在PIDray数据集上对SuNet能够有效识别被遮挡违禁品类别进行验证,在CLCXray数据集上对SuNet在其他违禁品数据集上具有泛化性进行验证。结果表明:在PIDray数据集上,相较于RoIAttn模型,SuNet在AP@0.5∶0.95、AP@0.5和AP@0.75指标上分别提升了2.9%、4.4%和3.3%;在CLCXray数据集上,相较于RoIAttn模型,SuNet在AP@0.5∶0.95、AP@0.5和AP@0.75指标上分别提升了1.4%、1.4%和0.4%。说明SuNet不仅能有效识别被遮挡违禁品类别,还在其他违禁品数据集上具有良好的泛化性能,可为公共交通安检场景提供一种有效的违禁品检测解决方案。In the scenario of security inspection in public transportation,the overlapping of prohibited and non-prohibited items made it difficult for existing models to effectively identify obscured prohibited item categories.To address this issue,a prohibited item detection model based on SuNet was proposed in this paper.Firstly,an augmented attention localization feature pyramid network(AALFPN)was designed to enhance the semantic information of prohibited items and fuse the localization information and semantic information of prohibited items to guide the model in accurately locating obscured prohibited items,enhancing the feature contour of prohibited items.Secondly,a dense attention mechanism(DAM)was introduced to effectively identify and extract obscured prohibited items.Finally,the SmoothL1 Loss loss function was introduced to address the problem of loss of prohibited item category information during regression.To verify SuNet′s ability to effectively identify obscured prohibited item categories,this study conducted experiments on the PIDray dataset.To assess SuNet′s generalization on other prohibited item datasets,this study conducted experiments on the CLCXray dataset.Experimental results show that on the PIDray dataset,compared to the RoIAttn model,the SuNet improves by 2.9%,4.4%and 3.3%on the AP@0.5∶0.95,AP@0.5 and AP@0.75 metrics,respectively.On the CLCXray data set,compared to the RoIAttn model,the SuNet improves by 1.4%,1.4%and 0.4%on the AP@0.5∶0.95,AP@0.5 and AP@0.75 metrics,respectively.The experimental results demonstrate that SuNet not only effectively identifies obscured prohibited item categories but also exhibits good generalization performance on other prohibited item datasets,providing an effective solution for prohibited item detection in the public transportation security inspection scenario.

关 键 词：违禁品检测 SuNet 强化注意力定位特征金字塔网络 密集注意力机制 SmoothL1 Loss 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]