利用Dlib深度殘差學習實作人臉辨識－以google meet線上會議點名為例

本文利用現今人工智慧的技術，實作自動執行人臉辨識與比對程式，並將此程式實際應用於疫情時代大量使用的線上教學系統中，以快速得到線上教學參與人數之數量及參與對象的身分識別。經過一連串程式設計的過程，實作成果如下：一、利用OpenCV的CascadeClassifier分類器來搜尋影像中的人臉時，使用的是Haar特徵產生的人臉分類器，針對影像中過於歪斜的人臉，這種方法還是偵測不到的。二、Dlib的人臉搜尋主要是利用HOG特徵實作，這種方法準確度較高，當找出了人臉位置後，Dlib進而使用Ensemble of Regression Trees理論來預測每一個人臉上的指定特徵位置。當有了所有人臉的特徵位置資料，經過人臉對正程序，最後使用深度殘差學習模型計算每張人臉的128D向量資訊而能確實使用歐式距離的計算進行人臉比對。在我們的實作中，人臉辨識的準確度可以百分之百的達成。

關鍵字

哈爾特徵；方向梯度直方圖；集成回歸樹；深度殘差學習

並列摘要

This article uses today's artificial intelligence technology to implement an automatic face recognition and comparison program, and practically applies this program to the online teaching system that is widely used in the epidemic era, so as to quickly obtain the number of online teaching participants and the number of participants. After a series of programming process, the implementation results are as follows: 1. When using OpenCV's CascadeClassifier to search for faces in the image, the face classifier generated by Haar features is used. For faces that are too skewed in the image, this method is still unable to detect. 2. Dlib's face search is mainly implemented using HOG features. This method has high accuracy. When the face position is found, Dlib then uses the Ensemble of Regression Trees theory to predict the specified feature position on each face. When the feature position data of all faces is available, after the face alignment procedure, the deep residual learning model is used to calculate the 128D vector information of each face, and the Euclidean distance calculation can be used for face comparison. In our implementation, the accuracy of face recognition can be achieved 100%.