To address the shortage of public datasets for customs X-ray images of contraband and the difficulties in deploying trained models in engineering applications,a method has been proposed that employs the Extract-Transf...To address the shortage of public datasets for customs X-ray images of contraband and the difficulties in deploying trained models in engineering applications,a method has been proposed that employs the Extract-Transform-Load(ETL)approach to create an X-ray dataset of contraband items.Initially,X-ray scatter image data is collected and cleaned.Using Kafka message queues and the Elasticsearch(ES)distributed search engine,the data is transmitted in real-time to cloud servers.Subsequently,contraband data is annotated using a combination of neural networks and manual methods to improve annotation efficiency and implemented mean hash algorithm for quick image retrieval.The method of integrating targets with backgrounds has enhanced the X-ray contraband image data,increasing the number of positive samples.Finally,an Airport Customs X-ray dataset(ACXray)compatible with customs business scenarios has been constructed,featuring an increased number of positive contraband samples.Experimental tests using three datasets to train the Mask Region-based Convolutional Neural Network(Mask R-CNN)algorithm and tested on 400 real customs images revealed that the recognition accuracy of algorithms trained with Security Inspection X-ray(SIXray)and Occluded Prohibited Items X-ray(OPIXray)decreased by 16.3%and 15.1%,respectively,while the ACXray dataset trained algorithm’s accuracy was almost unaffected.This indicates that the ACXray dataset-trained algorithm possesses strong generalization capabilities and is more suitable for customs detection scenarios.展开更多
针对X射线安检场景中违禁品目标检测精度低,检测模型过于复杂的问题,在YOLOv7-Tiny模型的基础上,提出了一种新的轻量化检测方法。首先在骨干网络中融合改进的轻量化模块GhostNetV2,在减少模型参数的同时,提高训练效率;其次在YOLOv7-Tin...针对X射线安检场景中违禁品目标检测精度低,检测模型过于复杂的问题,在YOLOv7-Tiny模型的基础上,提出了一种新的轻量化检测方法。首先在骨干网络中融合改进的轻量化模块GhostNetV2,在减少模型参数的同时,提高训练效率;其次在YOLOv7-Tiny的颈部网络部分加入金字塔拆分注意力机制,有效解决参数减少导致的提取特征不足问题,提高背景复杂以及多尺度目标回归的准确性;最后,通过使用归一化Wasserstein距离方法来度量损失,替代了原有的Intersection over Union度量,降低了小目标位置偏差的敏感性,增强了小目标的回归准确性。实验结果表明,改进模型在SIXray、CLCXray和OPIXray数据集上平均检测精度达到92.9%、76.2%和91.2%,相比原始算法分别提升了6.5%、2%和1.8%;所提出模型在轻量化的同时能够进一步提高检测能力,可以满足实时检测要求,具有较好的应用价值。展开更多
基金supported by the National Natural Science Foundation of China(Grant No.51605069).
文摘To address the shortage of public datasets for customs X-ray images of contraband and the difficulties in deploying trained models in engineering applications,a method has been proposed that employs the Extract-Transform-Load(ETL)approach to create an X-ray dataset of contraband items.Initially,X-ray scatter image data is collected and cleaned.Using Kafka message queues and the Elasticsearch(ES)distributed search engine,the data is transmitted in real-time to cloud servers.Subsequently,contraband data is annotated using a combination of neural networks and manual methods to improve annotation efficiency and implemented mean hash algorithm for quick image retrieval.The method of integrating targets with backgrounds has enhanced the X-ray contraband image data,increasing the number of positive samples.Finally,an Airport Customs X-ray dataset(ACXray)compatible with customs business scenarios has been constructed,featuring an increased number of positive contraband samples.Experimental tests using three datasets to train the Mask Region-based Convolutional Neural Network(Mask R-CNN)algorithm and tested on 400 real customs images revealed that the recognition accuracy of algorithms trained with Security Inspection X-ray(SIXray)and Occluded Prohibited Items X-ray(OPIXray)decreased by 16.3%and 15.1%,respectively,while the ACXray dataset trained algorithm’s accuracy was almost unaffected.This indicates that the ACXray dataset-trained algorithm possesses strong generalization capabilities and is more suitable for customs detection scenarios.
文摘针对X射线安检场景中违禁品目标检测精度低,检测模型过于复杂的问题,在YOLOv7-Tiny模型的基础上,提出了一种新的轻量化检测方法。首先在骨干网络中融合改进的轻量化模块GhostNetV2,在减少模型参数的同时,提高训练效率;其次在YOLOv7-Tiny的颈部网络部分加入金字塔拆分注意力机制,有效解决参数减少导致的提取特征不足问题,提高背景复杂以及多尺度目标回归的准确性;最后,通过使用归一化Wasserstein距离方法来度量损失,替代了原有的Intersection over Union度量,降低了小目标位置偏差的敏感性,增强了小目标的回归准确性。实验结果表明,改进模型在SIXray、CLCXray和OPIXray数据集上平均检测精度达到92.9%、76.2%和91.2%,相比原始算法分别提升了6.5%、2%和1.8%;所提出模型在轻量化的同时能够进一步提高检测能力,可以满足实时检测要求,具有较好的应用价值。
