基於自監督預訓練與孿生網路架構的指紋辨識方法;Research on Fingerprint Recognition Method Based on Self-supervised Pre-training and Siamese Network Architecture

NCUIR > College of Electrical Engineering & Computer Science > Graduate Institute of Computer Science and Information Engineering > Electronic Thesis & Dissertation > Item 987654321/96305

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/96305

Title:	基於自監督預訓練與孿生網路架構的指紋辨識方法;Research on Fingerprint Recognition Method Based on Self-supervised Pre-training and Siamese Network Architecture
Authors:	壽柏安;Shou, Po-An
Contributors:	資訊工程學系
Keywords:	自監督學習;孿生網路架構;Self-Supervised Learning;Siamese Neural Networks
Date:	2025-01-22
Issue Date:	2025-04-09 17:39:00 (UTC+8)
Publisher:	國立中央大學
Abstract:	本研究針對指紋辨識領域中資料稀缺性與應用場景多樣化的挑戰，提出一套基於自監督預訓練與孿生網路架構的指紋辨識方法，以提升模型的準確性與穩健性。指紋辨識作為身份認證的重要技術，已廣泛應用於支付系統、安全管理及個人裝置等領域。然而，由於公開指紋資料集的數量有限且多為標準化樣本，難以涵蓋實際應用中的複雜場景，如指紋部分缺損、噪聲干擾與跨感測器變化等問題，使得模型在應用中的效能受限。私有資料集的隱私性與保密性更進一步限制了訓練數據的獲取，成為當前指紋辨識研究面臨的重要瓶頸。為了解決上述問題，本研究首先採用自監督學習技術，通過對比學習與預訓練，充分挖掘未標記指紋資料中的潛在特徵，為模型提供強大的初始化，減少對標記資料的依賴。接著，設計孿生網路架構，利用兩個共享參數的預訓練子網路對指紋影像進行特徵提取，並通過歐幾里得距離計算影像間的相似性。此架構能有效應對部分缺損、噪聲干擾及跨感測器變化等挑戰場景，提升辨識效能。本研究結果顯示，提出的方法在指紋資料集上的辨識效能達到良好效果，預訓練模型使下游微調可以更快速收斂，節省訓練資源與時間，且在處理複雜場景的指紋時展現出高度的穩健性與泛化能力。;This study addresses the challenges of data scarcity and the diverse application scenarios in the field of fingerprint recognition by proposing a fingerprint recognition method based on self-supervised pretraining and a Siamese network architecture. This method aims to improve the accuracy and robustness of the model. As an essential technology for identity authentication, fingerprint recognition has been widely applied in payment systems, security management, and personal devices. However, the limited availability of public fingerprint datasets, which are predominantly standardized samples, fails to adequately cover the complex scenarios encountered in real-world applications, such as partial fingerprint degradation, noise interference, and cross-sensor variations. Consequently, the performance of models in practical applications is restricted. Moreover, the privacy and confidentiality of proprietary datasets further limit access to training data, representing a critical bottleneck in current fingerprint recognition research. To address these challenges, this study first employs self-supervised learning techniques to fully explore the latent features in unlabeled fingerprint data through contrastive learning and pretraining. This approach provides a robust initialization for the model and reduces dependency on labeled data. Subsequently, a Siamese network architecture is designed, where two pretrained subnetworks with shared parameters are used to extract features from fingerprint images. The similarity between images is calculated using the Euclidean distance. This architecture effectively tackles challenging scenarios such as partial degradation, noise interference, and cross-sensor variations, thereby enhancing recognition performance. The experimental results demonstrate that the proposed method achieves satisfactory recognition performance on fingerprint datasets. The pretrained model enables faster convergence during downstream fine-tuning, saving training resources and time. Furthermore, it exhibits high robustness and generalization capabilities when processing fingerprints in complex scenarios.
Appears in Collections:	[Graduate Institute of Computer Science and Information Engineering] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	26	View/Open

社群 sharing

Loading...