主管：中华人民共和国司法部
主办：司法鉴定科学研究院
ISSN 1671-2072  CN 31-1863/N
中国司法鉴定

中国司法鉴定 ›› 2026 ›› Issue (3): 55-64.DOI: 10.3969/j.issn.1671-2072.2026.03.006

• 专题研究：文书物证量化分析与智能鉴定技术前沿 • 上一篇    下一篇

基于小样本量的印章印文种类识别研究

陈    琦1，李    冰1，2，张    磊3，黄    旭4   

  1. 1.中国政法大学 证据科学研究院； 2.中国政法大学 证据科学教育部重点实验室；
    3.最高人民检察院检察技术信息研究中心； 4.广东南天司法鉴定所
  • 出版日期:2026-05-15 发布日期:2026-05-19

Research on Recognition of Stamp Impression Types Based on Small Sample Quantity

CHEN Qi1, LI Bing1,2, ZHANG Lei3, HUANG Xu4   

  1. 1. Institute of Evidence Law and Forensic Science, China University of Political Science and Law;
    2. Key Laboratory of Evidence Law and Forensic Science, Ministry of Education, China University of Political Science and Law; 3. Procuratorial Technology and Information Research Center of the Supreme People’s Procuratorate; 4. Guangdong Nantian Institute of Forensic Science
  • Published:2026-05-15 Online:2026-05-19

PDF

摘要：  目的 探究计算机视觉领域经典模型在小样本量印章印文种类识别中的可行性，以及在不同盖印条件下对识别结果的影响。方法 利用VGG16、ResNet50、视觉Transformer三种网络模型，对光敏印章、激光雕刻渗透印章、回墨印章、铜章、木章和橡胶印章共六种常见印章所盖印的印文进行种类识别。结果 在实验阶段，三种网络模型对六种印章印文的识别效果均表现优异，对光敏印章印文、回墨印章印文、木章印文和橡胶印章印文的识别准确率基本达到100%，仅在激光雕刻渗透印章印文和铜章印文的识别上精度略有下降；在盲测阶段，三种模型对六种印章印文的识别准确率与实验阶段相比普遍下降了3个百分点到35个百分点，证实了网络模型在真实复杂场景中的局限性。结论 计算机视觉领域经典模型能够辅助印章印文的种类识别，但在识别精度上仍需改进。

关键词: 文件检验；印章印文检验；种类识别；卷积神经网络；视觉Transformer 

Abstract: Objective To investigate the feasibility of classical computer vision models in recognizing stamp impression types with small sample quantity, as well as the impact of different stamping conditions. Methods Three network models, VGG16, ResNet50, and visual Transformer, were utilized to identify the kinds of stamp impressions imprinted by six common types of stamps, including photosensitive stamp, laser-engraved penetration stamp, self-inking stamp, copper stamp, wooden stamp, and rubber stamp. Results In the experiment, the three network models performed well in recognizing all six types of stamp impressions. The recognition accuracy for photosensitive stamp, self-inking stamp, wooden stamp and rubber stamp all basically reached 100%, while only a slight decrease was observed in the recognition accuracy for laser-engraved penetration stamp and copper stamp. In the following blind test, the recognition accuracy of the three models for the six types of stamp impressions generally dropped by 3 to 35 percentage points, confirming the limitations of the network models in real complex scenarios. Conclusion Classical models in the field of computer vision can assist in recognizing the types of stamp impressions, but their recognition accuracy needs to be improved.

Key words: document examination, examination of stamp impression, type recognition, convolutional neural network, visual Transformer

中图分类号: