植基於倒傳遞神經網路與 GPU 加速技術之訪客管理系統 設計與實作

Design and Implementation of a Visitor Management System by Using Graphics Processing Unit-Accelerated Back-Propagation Neural Networks

本研究中，吾人設計並實作出一款名為 Bio-IdGate 的自動化訪客管理系統。為了改善傳 統作業有關訪客登錄以及訪客資料管理等處理方式的缺失，Bio-IdGate 運用了倒傳遞神經網路 技術與 GPU 平行加速運算方法實現其核心功能。在 Bio-IdGate 系統中，根據擷取所得之生 物特徵，倒傳遞神經網路被用於實現訪客身份識別之服務。有鑑於倒傳遞神經網路於訓練過程 中會隨著資料樣本數量的增加而耗用大量的運算資源，Bio-IdGate 採用自行研發的 GPU 平行加速運算方法，俾使系統得以提供更為即時的服務。實驗結果顯示，運用 GPU 加速倒傳遞神經網路的執行確實有其成效。整體而言，Bio-IdGate 為傳統訪客管理作業方式的不足提供了有效的解決方案。此一系統預期將可成為企業機構施行安全管制措施所不可或缺的利器。

This paper presents Bio-IdGate as an advanced visitor management system to improve visitor registration and information management activities. In the Bio-IdGate system, back-propagation neural networks realize personnel recognition schemes to confirm the identity of an individual requesting services. Moreover, the implementation of neural networks by using graphics processing units(GPUs) provides remarkable performance compared with central processing units (CPUs) for computationally-intensive applications, such as training of back-propagation neural networks for large data sets. Experimental results revealed that GPU-accelerated implementation reduced computational costs compared with the standalone CPU version. Bio-IdGate is capable of offering an organized overview of visitor records and reducing time spent managing visitor information.