融合多尺度深度特征的残缺鞋印匹配算法

doi:10.11871/jfdc.issn.2096-742X.2026.04.005

数据与计算发展前沿 ›› 2026, Vol. 8 ›› Issue (4): 72-86.

CSTR: 32002.14.jfdc.CN10-1649/TP.2026.04.005

doi: 10.11871/jfdc.issn.2096-742X.2026.04.005

融合多尺度深度特征的残缺鞋印匹配算法

宋久常¹(),唐云祁^1,^*(),郭威¹,张嘉楠²

¹ 中国人民公安大学侦查学院，北京 100038
² 北京市公安局朝阳分局刑事侦查支队，北京 100123

收稿日期:2026-01-12 出版日期:2026-08-20 发布日期:2026-08-21
通讯作者: *唐云祁（E-mail: tangyunqi@ppsuc.edu.cn）
作者简介:宋久常，中国人民公安大学刑事科学技术专业，硕士研究生，主要研究方向为计算机视觉、鞋印识别。
本文中主要负责算法设计、实验实施与论文撰写。
SONG Jiuchang is a master’s student in Criminal Science and Technology at the People’s Public Security University of China. His reseach interests include computer vision and footwear impression recognition.
In this paper, he is mainly responsible for algorithm design, conducting experiment, and manuscript writing.
E-mail: 1725162048@qq.com|唐云祁，中国人民公安大学教授，博士研究生导师，主要研究方向为电子数据检验、深度伪造鉴别等。
本文中负责模型设计、模型指导优化、论文修改。
TANG Yunqi is a professor and doctoral supervisor at the School of Investigation, People’s Public Security University of China. His research interests include electronic data forensics and deepfake detection.
In this paper, he is mainly responsible for model design,guidance on model optimization, and manuscript revision.
E-mail: tangyunqi@ppsuc.edu.cn
基金资助:
中国人民公安大学刑事科学技术双一流创新研究专项(2023SYL06)

Deficient Shoeprint Matching via Fusion of Multi-Scale Deep Features

SONG Jiuchang¹(),TANG Yunqi^1,^*(),GUO Wei¹,ZHANG Jianan²

¹ School of Investigation, People’s Public Security University of China, Beijing 100038, China
² Criminal Investigation Detachment, Chaoyang Sub-bureau, Beijing Municipal Public Security Bureau, Beijing 100123, China

Received:2026-01-12 Online:2026-08-20 Published:2026-08-21

摘要/Abstract

摘要：

【背景】 鞋印作为犯罪现场最常见的物证之一，在刑事侦查中发挥着重要作用。“鞋印+监控”侦查技术通过将现场鞋印与视频监控相结合，已成为锁定犯罪嫌疑人的有效手段。【目的】 针对现有鞋印匹配算法存在的针对性不强、鲁棒性差、匹配精度不高以及计算复杂度大等问题，本文提出一种基于改进EfficientNetV2-S的融合多尺度深度特征的残缺鞋印匹配算法。【方法】 构建了包含1,342类共7,455张鞋印样本的DP-database数据集，其中残缺鞋印占比约70%。采用轻量化的EfficientNetV2-S为骨干网络，引入ECA注意力机制优化Fused-MBConv模块，并通过多层级特征融合策略整合Stage2至Stage6的深度特征，实现不同尺度鞋印信息的有效提取。【结果】 在CSS-200数据集上取得top1准确率79%、top10准确率89.5%、top200查全率100%。在CS-Database数据集上，高质量鞋印和血迹鞋印的top1检出率均达100%，灰尘鞋印的top5检出率达100%，优于现有主流算法。【结论】 本文方法在保持高效率的同时显著提升了匹配精度和鲁棒性，对残缺、模糊等低质量鞋印具有较强识别能力，可有效满足公安实战“快侦快破”需求。

关键词: 匹配算法, 特征融合, 残缺鞋印, EfficientNetV2-S

Abstract:

[Background] Shoeprints, common physical evidence at crime scenes, play a key role in criminal investigations. The “shoeprint+surveillance” technology narrows down suspect identification by combining on-site shoeprints with video surveillance. [Objective] To address limitations of existing shoeprint matching algorithms (limited task specificity, poor robustness, low accuracy, high computational complexity), this paper proposes a deficient shoeprint matching algorithm based on improved EfficientNetV2-S with multi-scale deep feature fusion. [Methods] We constructed the DP-database with 7,455 shoeprint samples (1,342 categories, 70% deficient). Using lightweight EfficientNetV2-S as backbone, we optimized Fused-MBConv via ECA attention and integrated Stage2-Stage6 deep features through multi-level fusion for multi-scale shoeprint information extraction. [Results] On CSS-200, it achieved 79% top-1 accuracy, 89.5% top-10 accuracy, and 100% top-200 recall. On CS-Database, top-1 detection rate reached 100% for high-quality and bloodstained shoeprints, 100% top-5 for dust ones, outperforming mainstream algorithms. [Conclusions] The proposed method improves accuracy and robustness while maintaining efficiency, effectively recognizes low-quality shoeprints (deficient, blurred), and meets public security's rapid investigation requirements.

Key words: matching algorithm, feature fusion, partial shoe print, EfficientNetV2-S

宋久常, 唐云祁, 郭威, 张嘉楠. 融合多尺度深度特征的残缺鞋印匹配算法[J]. 数据与计算发展前沿, 2026, 8(4): 72-86.

SONG Jiuchang, TANG Yunqi, GUO Wei, ZHANG Jianan. Deficient Shoeprint Matching via Fusion of Multi-Scale Deep Features[J]. Frontiers of Data and Computing, 2026, 8(4): 72-86, https://cstr.cn/32002.14.jfdc.CN10-1649/TP.2026.04.005.

图/表 21

图1

图2

表1

表2

图3

图4

表3

图5

图6

图7

图8

图9

表4

图10

图11

表5

表6

表7

图12

图13

图14

参考文献 44

[1]	金益锋, 白艳平, 刘寰. 全国16个省份足迹自动识别系统应用情况分析[J]. 刑事技术, 2017, 42(6): 504-507. doi: 10.16467/j.1008-3650.2017.06.016
[2]	张嘉楠, 孟昕, 唐云祁. 浅谈视频侦查与现场足迹联系方法[J]. 法制博览, 2016(14): 172-173.
[3]	张涛, 朱振东, 王慧, 等. 基于鞋印图像的性别预测方法[J]. 计算机工程, 2022, 48(11): 306-313. doi: 10.19678/j.issn.1000-3428.0062940
[4]	袁小萍. 对快侦快破现行案件的几点思考[J]. 公安大学学报(社会科学版), 1998(4): 34-37.
[5]	刘伟, 姜乐怡, 朱婷鸽, 等. 鞋印图像匹配研究现状与发展趋势[J]. 西安邮电大学学报, 2021, 26(2): 70-76.
[6]	BELSER C, INEICHEN M, PFEFFERLI P. Evaluation of the ISAS system after two years of practical experience in forensic police work[J]. Forensic Science International, 1996, 82(1): 53-58. doi: 10.1016/0379-0738(96)01966-4
[7]	MIKKONEN S, ASTIKAINEN T. Databased classification system for shoe sole patterns-identification of partial footwear impression found at a scene of crime[J]. Journal of Forensic Sciences, 1994, 39(5): 1227-1236. doi: 10.1520/JFS13708J
[8]	ALEXANDRE G. Computerized classification of the shoeprints of burglar soles[J]. Forensic Science International, 1996, 82(1): 59-65. doi: 10.1016/0379-0738(96)01967-6
[9]	ASHLEY W. What shoe was that? The use of computerized image database to assist in identification[J]. Forensic Science International, 1996, 82(1): 7-20. doi: 10.1016/0379-0738(96)01962-7
[10]	ALGARNI G, AMIANE M. A novel technique for automatic shoeprint image retrieval[J]. Forensic Science International, 2008, 181(1): 10-14. doi: 10.1016/j.forsciint.2008.07.004
[11]	TEAGUE M R. Image analysis via the general theory of moments[J]. Journal of the Optical Society of America, 1980, 70(8): 920-930. doi: 10.1364/JOSA.70.000920
[12]	HUYNH C, DE CHAZAL P, MCERLEAN D, et al. Automatic classification of shoeprints for use in forensic science based on the Fourier transform[C]// International Conference on Image Processing, IEEE, 2003: 569-572.
[13]	DE CHAZAL P, FLYNN J, REILLY R B. Automated processing of shoeprint images based on the Fourier transform for use in forensic science[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2005, 27(3): 341-350. doi: 10.1109/TPAMI.2005.48 pmid: 15747790
[14]	CERVELLI F, DARDI F, CARRATO S. A translational and rotational invariant descriptor for automatic shoe marks[C]// European Signal Processing Conference. Aalborg, Denmark: IEEE, 2010: 1665-1669.
[15]	OJALA T, PIETIKAINEN M, HARWOOD D. Performance evaluation of texture measures with classification based on Kullback discrimination of distributions[C]// Pattern Recognition, 1994. Vol. 1 - Conference A: Computer Vision & Image Processing. Proceedings of the 12th IAPR International Conference on IEEE, 1994: 582-583.
[16]	LOWE D G. Distinctive image feature from scale-invariant key points[J]. International Journal of Computer Vision, 2004, 60(2): 91-110. doi: 10.1023/B:VISI.0000029664.99615.94
[17]	DALAL N, TRIGGS B. Histograms of oriented gradients for human detection[C]// IEEE Computer Society Conference on Computer Vision & Pattern Recognition, IEEE, 2005: 886-893.
[18]	WANG X, SUN H, YU Q, et al. Automatic shoeprint retrieval algorithm for real crime scenes[C]// Asian Conference on Computer Vision. Lecture Notes in Computer Science, Springer, 2014: 399-413.
[19]	彭飞. 基于局部语义块和流行排序的鞋印图像匹配[D]. 大连: 大连海事大学, 2019.
[20]	周思越. 基于局部语义滤波器组的低质量鞋印图像匹配算法[D]. 大连: 大连海事大学, 2020.
[21]	焦扬, 杨传颖, 石宝. 基于SVM相关反馈的鞋印图像匹配算法[J]. 计算机科学, 2020, 47(S2): 244-247.
[22]	CUI J, ZHAO X, LI D. An automatic shoeprint retrieval method using neural codes for commercial shoeprint scanners[C]// Chinese Conference on Computer Vision. Lecture Notes in Computer Science, Springer, 2017: 158-169.
[23]	KONG B, SUPANCIC J, RAMANAN D, et al. Cross-domain image matching with deep feature maps[J]. International Journal of Computer Vision, 2018, 127(3): 1750-1768.
[24]	CUI J, ZHAO X, LIU N, et al. Robust shoeprint retrieval method based on local-to-global feature matching for real crime scenes[J]. Journal of Forensic Sciences, 2018, 64(1): 422-430. doi: 10.1111/jfo.2019.64.issue-2
[25]	MA Z, DING Y, WEN S, et al. Shoeprint image retrieval with multi-part weighted CNN[J]. IEEE Access, 2019, 7: 59728-59736. doi: 10.1109/Access.6287639
[26]	史文韬, 唐云祁. 基于微调VGG-16的现场鞋印匹配算法[J]. 中国人民公安大学学报(自然科学版), 2020, 26(3): 22-29.
[27]	史文韬, 唐云祁. 基于选择性卷积特征描述子融合的鞋印匹配算法[J]. 科学技术与工程, 2021, 21(16): 6772-6779.
[28]	韩雨彤. 基于深度学习网络的鞋印匹配算法[D]. 北京: 中国人民公安大学, 2022.
[29]	赵梦影. 基于深度特征的跨模态鞋印图像匹配[D]. 北京: 北京邮电大学, 2020.
[30]	韩雨彤, 郭威, 唐云祁, 等. 基于混合域注意力机制的鞋印匹配算法[J]. 中国人民公安大学学报(自然科学版), 2022, 28(2): 22-30.
[31]	辛一冉, 唐云祁, 蔡能斌. 融合全局与局部特征的鞋印特征提取网络设计[J]. 激光与光电子学进展, 2023, 60(6): 162-171.
[32]	WEN Z, CURRAN J M, WEVERS G. Shoeprint image retrieval and crime scene shoeprint image linking by using convolutional neural network and normalized cross correlation[J]. Science & Justice, 2023, 63(4): 439-450.
[33]	TAN M, LE Q V. EfficientNetV2: Smaller models and faster training[C]// Proceedings of the 38th International Conference on Machine Learning (ICML), PMLR, 2021: 10096-10106.
[34]	HOWARD A, SANDLER M, CHU G, et al. Searching for MobileNetV3[C]// Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), IEEE, 2019: 1314-1324.
[35]	WANG Q, WU B, ZHU P, et al. ECA-Net: Efficient channel attention for deep convolutional neural networks[C]// Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), IEEE, 2020: 11534-11542.
[36]	ZHOU Z, SIDDIQUEE M M R, TAJBAKHSH N, et al. UNet++: Redesigning skip connections to exploit multiscale features in image segmentation[J]. IEEE Transactions on Medical Imaging, 2020, 39(6): 1856-1867. doi: 10.1109/TMI.2019.2959609 pmid: 31841402
[37]	ZHANG P, WANG D, LU H, et al. Deep hierarchical feature concatenation for visual tracking[J]. IEEE Signal Processing Letters, 2017, 24(8): 1228-1232. doi: 10.1109/LSP.2017.2717083
[38]	HE K, ZHANG X, REN S, et al. Deep residual learning for image recognition[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), IEEE, 2016: 770-778.
[39]	史文韬. 基于卷积神经网络的鞋印匹配算法研究[D]. 北京: 中国人民公安大学, 2021: 13-17.
[40]	RICHETELLI N, LEE M C, LASKY C A, et al. Classification of footwear outsole patterns using Fourier transform and local interest points[J]. Forensic Science International, 2017, 275: 102-109. doi: S0379-0738(17)30084-1 pmid: 28343023
[41]	段硕古. 细粒度鞋印图像检索算法研究[D]. 大连: 大连海事大学, 2022.
[42]	KORTYLEWSKI A, ALBRECHT T, VETTER T. Unsupervised footwear impression analysis and retrieval from crime scene data[C]// Asian Conference on Computer Vision. Lecture Notes in Computer Science, Springer, 2015: 644-658.
[43]	WANG X, ZHANG C, WU Y, et al. A manifold ranking based method using hybrid features for crime scene shoeprint retrieval[J]. Multimedia Tools and Applications, 2017, 76(9): 21629-21649. doi: 10.1007/s11042-016-4029-3
[44]	LUOSTARINEN T, LEHMUSSOLA A. Measuring the accuracy of automatic shoeprint recognition methods[J]. Journal of Forensic Sciences, 2014, 59(6): 1627-1634. doi: 10.1111/1556-4029.12474 pmid: 24846765

模型	参数量/h	推理时间/h	训练时间/h	跨域测试mAP/%
ResNet-50	23.5	11.2	5	54.2
EfficientNetV2-S	86.1	12.3	12	66.4
VisionTransformer-Base	144.2	14.9	16	52.3
SwinTransformer-Small	74.8	13.2	33	44.7

层级	模块	尺寸	通道	模块数
1	Conv3×3	300×300	24	1
2	FusedMBConv	150×150	48	2
3	FusedMBConv	75×75	64	4
4	FusedMBConv	38×38	128	4
5	MBConv	19×19	160	6
6	MBConv	19×19	256	9
7	MBConv	10×10	518	15
8	FC	10×10	1,280	1

符号	符号说明
Q	查询图像的集合
q	单个查询图像
\|Q\|	查询集合的大小（查询图像总数）
k	排名位置或指定的排名阈值
top(q)	查询q的匹配在匹配结果中的排名
M	查询q的相关（正确匹配）样本总数
P(k)	前k个结果中正确匹配的数量/k
rel(k)	指示函数，正确匹配为1，否则为0
Σ	求和符号
AP(q)	查询q的Average Precision值

Stage	accuracy/%
Stage	Top1	Top5	Top1%	Top5%	Top10%
Original	62.5	71	91	97	99
P2346	70	80	94	97.5	99.5
P2356	67	77	96	99.5	100
P2345	66	76.5	96	98	99.5
P2456	64.5	72	95.5	99	100
P3456	67.5	77.5	95	98.5	100
P23456	77.5	81	96	99.5	100

Methods	accuracy/%
Methods	Top1	Top10	Top1%	Top10%
微调VGG-16[26]	52	75.5	83	87
VGG-19＋SCDA[27]	62.5	85.5	92.5	94.5
Resnet34＋混合域[30]	56	79	89.5	94.5
Efficientnet-B3＋PCA[31]	69	75	95.5	99.5
本文方法	79	89.5	98	100

融合多尺度深度特征的残缺鞋印匹配算法

Deficient Shoeprint Matching via Fusion of Multi-Scale Deep Features

RichHTML

PDF

可视化

摘要/Abstract

引用本文

使用本文

图/表 21

参考文献 44

相关文章 9

编辑推荐

Metrics

本文评价

Methods		accuracy/%
		Top1	Top5
高质量	POC[40]	99	100
	MCNCC[23]	99	100
	局部语义块和流形排序[19]	99	100
	混合特征和邻域图像[43]	99	100
	局部语义滤波器组[20]	99	100
	细粒度特征和流行排序[41]	100	100
	VGG-19＋SCDA[27]	100	100
	Efficientnet-B3＋PCA[31]	100	100
	本文方法	100	100
血足迹	POC[40]	45.3	66
	MCNCC[23]	92.5	100
	局部语义块和流形排序[19]	79.2	96.2
	混合特征和邻域图像[43]	92.5	100
	局部语义滤波器组[20]	94.3	100
	细粒度特征和流行排序[41]	94.3	100
	VGG-19＋SCDA[27]	100	100
	Efficientnet-B3＋PCA[31]	100	100
	本文方法	100	100
灰尘足迹	POC[40]	47	54.5
	MCNCC[23]	86.4	95.5
	局部语义块和流形排序[19]	83.3	90.9
	混合特征和邻域图像[43]	89.4	95.5
	局部语义滤波器组[20]	95.45	98.48
	细粒度特征和流行排序[41]	86.4	94.45
	VGG-19＋SCDA[27]	87.87	90.9
	Efficientnet-B3＋PCA[31]	95.45	98.48
	本文方法	93.94	100

Methods	Top1%	Top10%	方法分类
局部语义块和流形排序[19]	73	93.7	基于传统人工设计特征的方法
混合特征和邻域图像[44]	61.3	88
局部语义滤波器组[20]	71.8	87.3
MCNCC[23]	79	89
CABM[43]	58	79
细粒度特征和流行排序[42]	82	93.3
Conv-25088[26]	47.16	76.59	基于深度学习的方法
VGG-19＋SCDA[27]	50.84	79.93
Efficientnet-B3＋PCA[31]	77.73	91.28
Proposed method	59	84

[1]	安真仟, 刘为军. 基于多层次情感与语义特征融合的虚假新闻检测方法[J]. 数据与计算发展前沿, 2026, 8(2): 111-122.
[2]	琚子政, 陈鹏, 隋晋光, 朱隆昇. 基于多尺度时空图融合的警情预测模型[J]. 数据与计算发展前沿, 2026, 8(2): 154-170.
[3]	潘语泉,袁得嵛,贾源,王安然. 基于融合特征的VGAT-VGAN跨社交网络身份关联算法[J]. 数据与计算发展前沿, 2026, 8(1): 103-118.
[4]	令狐荣微, 张瑜, 石元泉, 杨玉军. 基于多特征融合的PE恶意代码检测与分类研究[J]. 数据与计算发展前沿, 2025, 7(6): 77-91.
[5]	李勇,任勇毛,殷卓然,周旭. 一种基于深度学习的轻量化流量识别模型[J]. 数据与计算发展前沿, 2025, 7(2): 3-11.
[6]	卢成浩,陈秀宏. 基于隐式分区学习深度特征融合重建曲面网络[J]. 数据与计算发展前沿, 2024, 6(6): 19-31.
[7]	刘琦玮,李俊,顾蓓蓓,赵泽方. TSAIE：图像增强文本的多模态情感分析模型[J]. 数据与计算发展前沿, 2022, 4(3): 131-140.
[8]	陈涛,安俊秀. 基于特征融合的微博短文本情感分类研究[J]. 数据与计算发展前沿, 2020, 2(6): 21-29.
[9]	冷佳旭,刘莹. 基于深度学习的小目标检测与识别[J]. 数据与计算发展前沿, 2020, 2(2): 120-135.