發展基於深度學習之掌紋及掌靜脈辨識系統;Development of Deep Learning-based Palmprint and Palm Vein Recognition System

NCU Institutional Repository > 資訊電機學院 > 電機工程研究所 > 博碩士論文 > Item 987654321/98412

請使用永久網址來引用或連結此文件: https://ir.lib.ncu.edu.tw/handle/987654321/98412

題名:	發展基於深度學習之掌紋及掌靜脈辨識系統;Development of Deep Learning-based Palmprint and Palm Vein Recognition System
作者:	楊皓翔;Yang, Hao-Hsiang
貢獻者:	電機工程學系
關鍵詞:	深度學習;掌紋辨識;deep learning;palm recognition
日期:	2025-07-24
上傳時間:	2025-10-17 12:45:31 (UTC+8)
出版者:	國立中央大學
摘要:	本研究利用深度學習神經網路實現一個掌紋及掌靜脈辨識系統，並將其應用於行動裝置上進行身份驗證。掌紋及掌靜脈辨識技術因其高安全性與穩定性，在生物辨識領域具有重要應用價值。本研究設計了使用MediaPipe手部關鍵點檢測功能建立穩固的幾何座標系，由此開發一種自適應精準定位掌紋和掌靜脈影像ROI區域的方法，並結合Gabor濾波器整合灰階化、多尺度多方向 Gabor 濾波器與直方圖均衡化強化等圖像預處理增強紋理特徵，確保深度特徵能完整捕捉掌紋與靜脈細節。接著採用在ImageNet-21k 預訓練之EfficientNetV2-S進行遷移學習，輸入掌紋及掌靜脈影像特徵提取與分類，再透過TensorFlow Lite將深度學習模型輕量化後部署於Android平台。同時，在Android平台上設計應用程式，提供完整的身分驗證功能。實驗以 MPD、TCD 及 CASIA (850 nm + 940 nm) 三個公開資料集驗證模型之辨識性能與泛化能力，於 MPD 與 TCD 分別獲得 99.88 % 與 99.94 % 的最高準確率，而經波段融合後之 CASIA 資料集亦提升至 90.75%。為實踐行動端應用，本研究完成 PyTorch→ONNX→TensorFlow→TFLite 的轉換與動態範圍量化，將原始 94.2 MB 模型壓縮至 21.7 MB，成功於 Samsung Galaxy Tab A T590 達成掌紋及掌靜脈辨識，實測平均推論延遲僅 373.62 ms，兼具高準確率與即時性。綜合而言，本研究主要貢獻為提供了從註冊到辨識完整的手掌辨識流程，實現一種兼顧安全性與實用性的移動端生物辨識系統，證實了以行動裝置整合掌紋與掌靜脈生物特徵的可行性與實用價值，亦為日後多模態身分驗證與移動端應用部署提供實證基礎。 ;This research implements a palmprint and palm vein recognition system using deep learning neural networks and deploys it on mobile devices for identity authentication. Palmprint and palm vein recognition technologies possess significant application value in the biometric field due to their high security and stability. This study designs a robust geometric coordinate system utilizing MediaPipe hand keypoint detection functionality, thereby developing an adaptive method for precisely localizing regions of interest (ROI) in palmprint and palm vein images. The approach integrates Gabor filters with image preprocessing techniques including grayscale conversion, multi-scale multi-directional Gabor filtering, and histogram equalization enhancement to strengthen texture features, ensuring that deep features can comprehensively capture palmprint and vein details. Subsequently, transfer learning is employed using EfficientNetV2-S pre-trained on ImageNet-21k for feature extraction and classification of palmprint and palm vein images. The deep learning model is then optimized through TensorFlow Lite for lightweight deployment on the Android platform. Concurrently, an application is developed on the Android platform to provide complete identity authentication functionality. The experiments validated the model′s recognition performance and generalization capability using three public datasets: MPD, TCD, and CASIA (850 nm + 940 nm). The proposed method achieved the highest accuracy rates of 99.88% and 99.94% on the MPD and TCD datasets, respectively. For the CASIA dataset, the accuracy was enhanced to 90.75% after spectral band fusion. To enable mobile deployment, this study accomplished the complete conversion pipeline from PyTorch → ONNX → TensorFlow → TensorFlow Lite with dynamic range quantization. The original 94.2 MB model was successfully compressed to 21.7 MB, enabling palmprint and palm vein recognition on the Samsung Galaxy Tab A T590. The experimental results demonstrated an average inference latency of only 373.62 ms, achieving both high accuracy and real-time performance. In summary, the proposed method provides a comprehensive palm recognition pipeline from enrollment to recognition, implementing a mobile biometric system that balances security and practicality. This work demonstrates the feasibility and practical value of integrating palmprint and palm vein biometric features on mobile devices, while providing an empirical foundation for future multimodal identity authentication and mobile deployment applications.
顯示於類別:	[電機工程研究所] 博碩士論文

文件中的檔案:

檔案	描述	大小	格式	瀏覽次數
index.html		0Kb	HTML	5	檢視/開啟

在NCUIR中所有的資料項目都受到原著作權保護.

社群 sharing

資料載入中.....