基于孪生网络的小样本人脸识别研究

打开文本图片集

中图分类号：TP391.41 文献标志码：A 文章编号:1003-5168(2025)14-0023-05

DOI:10.19968/j.cnki.hnkj.1003-5168.2025.14.004

Research on Few-Shot Face Recognition Based on Siamese Networks

XU Qinan XIA ChunguanLIN Yuqing (ZhouKou Normal University, Zhoukou 466oo0, China)

Abstract: [Purpose] This study aims to explore few-shot face recognition technology based on Siamese networks to address the performance degradation of face recognition systems caused by data scarcity. [Methods]An enhanced Siamese network model is constructed and integrated with the MTCNN face detection algorithm to achieve precise facial region extraction.A mutual information metric is introduced to measure feature similarity,where the mutual information value between the joint distribution and marginal distribution of dual-branch feature vectors serves as the similarity criterion.By combining a contrastive loss function，the feature embedding space is optimized to maximize mutual information forintra-class samples while minimizing it for inter-class samples,thereby enhancing feature discriminabilityunder few-shot conditions.Additionaly,mutual information is normalized to the [O,1]interval to improve the robustness of similarity quantification，strengthening model stability and generalization.[Findings] Experimental results on the AT&T dataset demonstrate that the proposed method achieves a recognition accuracy of 98.64% under few shot conditions,outperforming traditional approaches such as PCA + SVM (63.87%), DeepFace (96. 11% ),and FaceNet ( 97.41% ),whichvalidates its effectiveness and innovation.[Conclusions] The enhanced Siamese network facilitates high-precision face recognition with limited data,providing novel insights and methodologies to the field of few-shot learning.

Keywords: siamese network; face recognition; few-shot learning; mutual information metric

0 引言

近年来，随着人工智能技术的迅速发展，人脸识别已成为计算机视觉领域中一个备受关注的研究方向，在安防监控、金融支付、社交媒体和智能家居等多种应用中展现出巨大的潜力。（剩余8444字）