隨著AI時代的到來，許多繁雜且重複性的事物逐漸能由AI協助處理。傳統警察在追緝犯人或通緝犯時，主要依靠人眼仔細監看監視器畫面，然而這種方式面臨人力成本高與容易疲勞失誤之雙重限制，因此本論文提出了一套即時且準確的人臉辨識系統。 本論文採用YOLO與MobileFaceNet搭配遷移學習的各種策略建構即時人臉辨識系統，解決目標人臉資料量不足的問題，並且設計一套去識別化模組提升資訊安全性，最後在辨識到目標時透過Telegram通知相關人員做二次確認，期望未來有機會能以這種有效率的方式實現智慧的安全城市。 ;With the advent of the AI era, many complex and repetitive tasks are increasingly managed through AI-assisted systems. Traditionally, police officers track suspects or wanted criminals primarily by manually reviewing surveillance footage, a method that is both labor-intensive and prone to human error due to fatigue. To address these issues, this thesis proposes a real-time and accurate facial recognition system. This thesis utilizes YOLO and MobileFaceNet, combined with various Transfer Learning strategies, to construct a real-time facial recognition system. The proposed system effectively tackles the issue of insufficient facial data for target individuals, and further integrates De-identification technique to enhance data security. Upon identifying the target, the system sends notifications to relevant personnel via Telegram for secondary verification. It is anticipated that this efficient method will contribute towards achieving a smart and secure urban environment in the future.
