摘要
X光扫描是对安检过程中通过的包裹进行违禁品检查的重要成像手段,而传统的人工识图方式具有效率低、易受主观因素影响等缺陷.为推动违禁品自动识别技术的发展,从数据、任务和方法 3个层面对X光图像识别特性递进分析.在数据层面,通过对X光与可见光的数据和数据集分别进行对比,揭示了X光安检图像在数据上的独特性.在任务层面,从数据特性、行业监管和实际业务要求等多个维度深入分析X光安检图像目标识别任务的复杂性.在方法层面,基于数据与任务特性,对现有X光安检图像目标识别的特定方法策略进行归类和简述.研究结果表明:X光安检图像目标识别技术需要应对数据特性导致的困境,适应行业监管的变化,处理安检对象特点的差异,以及满足细粒度的监管要求;在应对部分特性带来的挑战时,数据预处理、数据扩充、重叠遮挡处理和多视角融合等有效策略被提出,并存在可能的提升空间和拓展方向.研究结果能够为本领域研究人员提供参考和启发,以更好地满足持续变化的安检任务要求.
X-ray scanning is a crucial imaging technique for contraband inspection of packages during security checks.However,traditional manual recognition methods suffer from low efficiency and sus-ceptibility to subjective factors.To promote the development of automatic contraband identification technology,a progressive analysis of X-ray image recognition characteristics is conducted from three levels:data,task,and method.At the data level,the data and datasets of X-ray and visible light are compared separately,revealing the uniqueness of X-ray security images.At the task level,the com-plexity of the object recognition task of X-ray security images is deeply analyzed from multiple as-pects,such as data characteristics,industry supervision,and actual business requirements.At the method level,based on data and task characteristics,the specific methods and strategies for existing X-ray security image object recognition are classified and briefly described.The results indicate that X-ray image object recognition technology needs to cope with the dilemma caused by data characteris-tics,adapt to changes in industry supervision,deal with differences in the characteristics of inspected objects,and meet fine-grained regulatory requirements.In response to the challenges brought by cer-tain characteristics,effective strategies such as data preprocessing,data enrichment,occlusion pro-cessing,and multi-view fusion have emerged,presenting potential areas for improvement and expan-sion.This study’s findings can offer targeted references and inspiration to researchers in the field,aid-ing them in better meeting the evolving requirements of security screening tasks.
作者
孙运达
孙嘉龙
SUN Yunda;SUN Jiaong(Nuctech Company Limited,Beijing 100083,China;School of Computer Science and Technology,Beijing Jiaotong University,Beijing 100044,China)
出处
《北京交通大学学报》
CAS
CSCD
北大核心
2024年第2期47-56,共10页
JOURNAL OF BEIJING JIAOTONG UNIVERSITY
基金
国家重点研发计划(2022YFC3310200)。
关键词
X光安检
图像处理
目标识别
违禁品查验
X-ray security inspection
image processing
object recognition
contraband inspection