X射线安检图像高精度实时目标检测模型与基准数据集

支洪平; 孙立峰; 王旭

doi:10.13700/j.bh.1001-5965.2024.0459

X射线安检图像高精度实时目标检测模型与基准数据集

doi: 10.13700/j.bh.1001-5965.2024.0459

1.
清华大学计算机科学与技术系，北京 100084
2.
科大讯飞(苏州)科技有限公司，苏州 215000

详细信息

通讯作者:
E-mail：sunlf@tsinghua.edu.cn

中图分类号: TP391.41；TP73
计量
- 文章访问数: 323
- HTML全文浏览量: 57
- PDF下载量: 12
- 被引次数: 0
出版历程
- 收稿日期: 2024-06-21
- 录用日期: 2024-08-16
- 网络出版日期: 2024-09-09
- 整期出版日期: 2026-02-28

High-precision real-time object detection model and benchmark for X-ray security inspection images

1.
Department of Computer Science and Technology，Tsinghua University，Beijing 100084，China
2.
Iflytek (Suzhou) Technology Co.，Ltd.，Suzhou 215000，China

More Information

Corresponding author: E-mail：sunlf@tsinghua.edu.cn

摘要

摘要:
图像目标检测技术辅助提高了安检工作效率，进一步保障了公共安全。然而，不同型号安检机成像的差异性、X射线安检图像的复杂性及昂贵的数据标注成本制约了X射线安检图像目标检测技术的深入研究。为此，针对不同安检机厂商相同物质X射线成像颜色的差异，基于风格迁移算法进行数据集扩充，提高目标检测算法的泛化性；针对X射线安检图像中同类待识别物品尺寸的明显差异，提出一种细化的特征金字塔网络结构提取更加丰富的不同层次语义信息；为进一步提高检测精度，提出一个易于集成的细粒度分类模块，该模块能很好地适配主流目标检测模型。同时，构造一个大规模的基准数据集，包含56659张X射线安检图像，37种违禁品，每张图像均进行高质量标注。该公开X射线安检图像数据集包含的违禁品种类和图像数量较多。基于该X射线违禁品数据集进行对比实验，结果显示，所提模型结构较基线模型YOLOX-L的均值平均精度(mAP)提高约0.056。
- X射线安检图像 /
- 风格迁移 /
- 目标检测 /
- 细粒度分类模块 /
- 基准数据集
Abstract:
Image object detection technology has greatly improved the work efficiency of the security inspection and further guaranteed public security. However, the differences in imaging standards among different types of security inspection machines, the complexity of X-ray images, and the expensive cost of data annotation have constrained further research of object detection technology based on X-ray security inspection images. To improve the universality of our item detection system, we extend the dataset using a style transfer approach to account for variations in X-ray imaging hues of the same substance across various security equipment manufacturers. A refined feature pyramid network structure is proposed to extract richer semantic information from different levels in response to the significant differences in the size of similar objects to be recognized in X-ray images. A fine-grained classification module, which is simple to plug into the general object detectors, is what we suggest in order to increase detection accuracy even more. Meanwhile, this dataset contains 56659 X-ray images, featuring 37 types of contraband, with each image being high-quality annotated. This is a larger publicly available X-ray image dataset in terms of both the variety of contraband types and the number of images. Based on comparative experiments conducted on this X-ray contraband dataset, the model structure proposed in this article achieved an approximate 0.056 improvement in mean average precision (mAP) compared to the baseline model.
- X-ray security inspection images /
- style transfer /
- object detection /
- fine-grained classification module /
- benchmark

HTML全文

图 1 部分违禁品示例X射线安检图像

Figure 1. Example X-ray security inspection images of some prohibited items

下载: 全尺寸图片幻灯片

图 2 条件生成对抗模型结构

Figure 2. Structure of conditional generative adversarial model

下载: 全尺寸图片幻灯片

图 3 部分域之间风格迁移效果

Figure 3. Style transfer effect between some domains

下载: 全尺寸图片幻灯片

图 4 检测流程及检测模型结构

Figure 4. Detection process and structure of detection model

下载: 全尺寸图片幻灯片

图 5 细化的特征金字塔网络结构

Figure 5. Refined feature pyramid network structure

下载: 全尺寸图片幻灯片

图 6 细粒度分类模块结构

Figure 6. Fine-grained classification module structure

下载: 全尺寸图片幻灯片

表 1 数据集样本信息

Table 1. Sample information of the dataset

大类标签	细类标签	样本数
刀具	刀片	2935
	美工刀	2448
	特殊刀片	844
	全金属折叠刀	2539
	非全金属折叠刀	383
	全金属柄刀	1561
	非全金属柄刀	785
	线形刀	52
	厨刀	462
	刮胡刀	309
玻璃容器	玻璃瓶	58307
玻璃容器	酒瓶	1417
有机物容器	饮料瓶	19057
有机物容器	易拉罐	1086
电脑		3816
充电宝		5344
小电子设备		16328
枪		104
压力罐		3418
工具	钳子	6568
	扳手	3416
	锤子	578
	抹泥板	1607
	斧头	36
	锯条	71
	螺丝刀	6796
	铅坠	281
	凿子	192
指虎		85
手铐		56
甩棍		38
雨伞		8010
剪刀		4683
打火机		8403
金属杯		7553
打火机油罐		63
弹弓		26

下载: 导出CSV

表 2 各数据集对比

Table 2. Comparison of various datasets

数据集	类别数	图像数量
GDXray^[16]	3	8150
OPIXray^[17]	5	8885
PIDXray^[14]	12	47677
本文	37	56659

下载: 导出CSV

表 3 各检测模型实验结果

Table 3. Experimental results of each detection model

大类标签	样本数	平均精度
大类标签	样本数	Cascade R-CNN^[4]	YOLOX-L^[18]	YOLOX-L-FG
刀具	2453	0.326	0.696	0.744
工具	3776	0.478	0.730	0.773
玻璃容器	11867	0.747	0.887	0.870
有机物容器	3975	0.633	0.828	0.826
电脑	784	0.937	0.971	0.898
充电宝	1139	0.731	0.863	0.850
小电子设备	3378	0.790	0.904	0.867
枪	23	0.351	0.333	0.517
压力罐	669	0.410	0.574	0.731
指虎	19	0.506	0.657	0.789
手铐	9	0.722	0.749	0.495
雨伞	1664	0.922	0.968	0.897
剪刀	891	0.488	0.619	0.731
打火机	1657	0.583	0.741	0.773
金属杯	1507	0.947	0.974	0.899
打火机油罐	4	0.250	0.250	0.545
弹弓	8	0.042	0.198	0.432
甩棍	7	0.071	0.429	0.732
注：mAP为表中所有18个类别平均精度的算术平均值；Cascade R-CNN、YOLOX-L、YOLOX-L-FG的mAP分别为0.552、0.687、0.743。

下载: 导出CSV

表 4 FG模块集成前后均值平均精度对比

Table 4. Comparison of mean average precision before and after FG module integration

基线模型	是否集成FG	mAP	增益
Cascade R-CNN^[4]	否	0.552	0.047
Cascade R-CNN^[4]	是	0.599	0.047
YOLOX-S^[18]	否	0.577	0.082
YOLOX-S^[18]	是	0.659	0.082
YOLOX-M^[18]	否	0.647	0.043
YOLOX-M^[18]	是	0.690	0.043
YOLOX-L^[18]	否	0.687	0.056
YOLOX-L^[18]	是	0.743	0.056

下载: 导出CSV

表 5 各模型推理延迟

Table 5. Inference delay of each model

检测模型	推理芯片	图像尺寸/(像素×像素)	数据类型	延迟/ms	mAP
Cascade R-CNN^[4]	NVIDIA GeForce GTX 1660	1024×1280	32位浮点数	162	0.552
Cascade R-CNN-FG				210	0.599
YOLOX-L^[18]				116	0.687
YOLOX-L-FG				123	0.743

下载: 导出CSV

参考文献(18)

[1]	GAUS Y F A, BHOWMIK N, AKCAY S, et al. Evaluating the transferability and adversarial discrimination of convolutional neural networks for threat object detection and classification within X-ray security imagery[C]//Proceedings of the 2019 18th IEEE International Conference on Machine Learning and Applications. Piscataway: IEEE Press, 2019: 420-425.
[2]	张顺, 龚怡宏, 王进军. 深度卷积神经网络的发展及其在计算机视觉领域的应用[J]. 计算机学报, 2019, 42(3): 453-482. doi: 10.11897/SP.J.1016.2019.00453 ZHANG S, GONG Y H, WANG J J. The development of deep convolution neural network and its applications on computer vision[J]. Chinese Journal of Computers, 2019, 42(3): 453-482(in Chinese). doi: 10.11897/SP.J.1016.2019.00453
[3]	LIU W, ANGUELOV D, ERHAN D, et al. SSD: single shot MultiBox detector[C]//Proceedings of the Computer Vision-ECCV 2016. Berlin: Springer, 2016: 21-37.
[4]	CAI Z W, VASCONCELOS N. Cascade R-CNN: delving into high quality object detection[C]//Proceedings of the 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE Press, 2018: 6154-6162.
[5]	王坤, 冯康威. 基于改进YOLOv5的交通场景小目标检测算法[J/OL]. 北京航空航天大学学报, 2024(2024-03-12)[2024-06-19]. https://link.cnki.net/doi/10.13700/j.bh.1001-5965.2024.0003. WANG K, FENG K W. Small target detection algorithm for traffic scene based on improved YOLOv5[J/OL]. Journal of Beijing University of Aeronautics and Astronautics, 2024 (2024-03-12)[2024-06-19]. https://link.cnki.net/doi/10.13700/j.bh.1001-5965.2024.0003(in Chinese).
[6]	JIAO L C, ZHANG F, LIU F, et al. A survey of deep learning-based object detection[J]. IEEE Access, 2019, 7: 128837-128868. doi: 10.1109/ACCESS.2019.2939201
[7]	GIRSHICK R, DONAHUE J, DARRELL T, et al. Rich feature hierarchies for accurate object detection and semantic segmentation[C]//Proceedings of the 2014 IEEE Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE Press, 2014: 580-587.
[8]	HE K M, ZHANG X Y, REN S Q, et al. Spatial pyramid pooling in deep convolutional networks for visual recognition[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2015, 37(9): 1904-1916. doi: 10.1109/TPAMI.2015.2389824
[9]	GIRSHICK R. Fast R-CNN[C]//Proceedings of the 2015 IEEE International Conference on Computer Vision. Piscataway: IEEE Press, 2016: 1440-1448.
[10]	REDMON J, DIVVALA S, GIRSHICK R, et al. You only look once: unified, real-time object detection[C]//Proceedings of the 2016 IEEE Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE Press, 2016: 779-788.
[11]	ARKIN E, YADIKAR N, XU X B, et al. A survey: object detection methods from CNN to Transformer[J]. Multimedia Tools and Applications, 2023, 82(14): 21353-21383. doi: 10.1007/s11042-022-13801-3
[12]	VASANTHI P, MOHAN L. Multi-Head-Self-Attention based YOLOv5X-Transformer for multi-scale object detection[J]. Multimedia Tools and Applications, 2024, 83(12): 36491-36517.
[13]	LIU J Y, LENG X X, LIU Y. Deep convolutional neural network based object detector for X-ray baggage security imagery[C]//Proceedings of the IEEE 31st International Conference on Tools with Artificial Intelligence. Piscataway: IEEE Press, 2020: 1757-1761.
[14]	WANG B Y, ZHANG L B, WEN L Y, et al. Towards real-world prohibited item detection: a large-scale X-ray benchmark[C]//Proceedings of the 2021 IEEE/CVF International Conference on Computer Vision. Piscataway: IEEE Press, 2022: 5392-5401.
[15]	LIANG K J, SIGMAN J B, SPELL G P, et al. Toward automatic threat recognition for airport X-ray baggage screening with deep convolutional object detection[EB/OL]. (2019-12-13)[2024-06-19]. https://arxiv.org/abs/1912.06329.
[16]	MERY D, RIFFO V, ZSCHERPEL U, et al. GDXray: the database of X-ray images for nondestructive testing[J]. Journal of Nondestructive Evaluation, 2015, 34(4): 42. doi: 10.1007/s10921-015-0315-7
[17]	WEI Y L, TAO R S, WU Z J, et al. Occluded prohibited items detection: an X-ray security inspection benchmark and de-occlusion attention module[C]//Proceedings of the 28th ACM International Conference on Multimedia. New York: ACM, 2020: 138-146.
[18]	GE Z, LIU S T, WANG F, et al. YOLOX: exceeding YOLO series in 2021[EB/OL]. (2021-08-06)[2024-06-19]. https://arxiv.org/abs/2107.08430.