基于递归特征金字塔和可切换空洞卷积的轨道交通智慧安检违禁品识别技术研究

Research on Contraband Identification Technology For Rail Transit Intelligent Security Inspection Based on Recursive Feature Pyramid and Switchable Hole Convolution

为保证城市轨道交通的安全性,X光行包安检系统已经被广泛运用。但现阶段行包安检主要依赖人工判定,存在高人工、低效率、欠客观等问题。本文以实现智能化、自动化行包安检为目标,研究并设计了一种基于ResNet50的目标探测模型。该模型使用递归特征金字塔结构,将附加的反馈链接与自下而上的骨架层结合起来。在特征提取时使用可切换空洞卷积,提取不同尺度的特征。通过将两种方法集成到检测器中,自主学习违禁品的关键性特征,实现了X光行包违禁品自动识别。使用大规模公开数据集PIDray对所提算法进行测试,并设计消融实验验证各模块的有效性。实验结果表明本文算法对12大类违禁品的平均识别准确率可达到91%以上。证明了使用人工智能算法进行自动安检的可行性,为城市轨道交通行包安检自动监测的落地应用提供有力支撑。

In order to ensure the safety of urban rail transit,X-ray luggage security inspection system has been widely used.However,currently luggage security inspection mainly depends on manual judgment,which has the problems of high labor,low efficiency and less objectivity.To realize intelligent and automatic luggage security inspection,this paper proposes a ResNet50-based object detection model.The model uses the recursive feature pyramid structure to combine the additional feedback links with the bottom-up skeleton layer.For feature extraction,it uses the switchable hole convolution to extract the features of different scales.By integrating the two methods into the detector,the proposed method can learn the key characteristics of contraband and realize the automatic identification of X-ray luggage contraband.The proposed algorithm is tested using the large-scale public data set PIDray,and ablation experiments are designed to verify the effectiveness of each module.Experimental results show that the average recognition accuracy of the proposed algorithm for 12 categories of contraband achieves more than 91%.It proves the feasibility of using artificial intelligence algorithm for automatic security inspection,and provides a strong support for the landing of automatic luggage security inspection in urban rail transit.