陳美勇Chen, Mei-Yung陳勁凱Chen, Chin-Kai2024-12-172024-08-122024https://etds.lib.ntnu.edu.tw/thesis/detail/baf4fa8d8615f5645d96f6221d111431/http://rportal.lib.ntnu.edu.tw/handle/20.500.12235/123066臉部辨識是一種重要的生物識別技術，在多種應用中得到廣泛使用。然而，化妝以及年齡變化會使人臉發生變化，進而影響人臉上的特徵，從而降低臉部辨識的準確性。為了解決化妝以及年齡變化造成的臉部辨識問題，本論文提出了一種基于MediaPipe的FaceMesh和類神經網路的臉部辨識方法，以解決化妝以及年齡變化造成的臉部辨識問題，該方法將在Python內部逐步構成。MediaPipe FaceMesh模型的人臉偵測是以 BlazeFace 人臉偵測器為基礎，該偵測器會對圖像進行操作並計算人臉位置。偵測到人臉後，FaceMesh模型會使用一個自定義殘差神經網絡提取名為landmark的臉部特徵，並利用歐式距離和landmark蘊含的座標資料計算指定的landmark之間的距離以及比值，作為訓練用的臉部特徵。主成分分析用於提高準確率，降低過擬合現象。類神經網路用於訓練模型。實驗結果表明，該方法在化妝以及年齡變化下的臉部辨識有一定的準確性，具有一定的應用價值。Face recognition is an important biometric technology widely used in various applications. However, makeup and age changes can alter facial features, reducing the accuracy of face recognition. To address the issues caused by makeup and age changes, this paper proposes a face recognition method based on MediaPipe's FaceMesh and neural networks. This method aims to tackle the problems posed by makeup and age changes, and will be implemented step by step in Python.The face detection in the MediaPipe FaceMesh model is based on the BlazeFace face detector, which processes images and calculates the position of faces. After detecting a face, the FaceMesh model uses a custom residual neural network to extract facial features called landmarks. Euclidean distances and the coordinates embedded in these landmarksare used to calculate distances and ratios between specified landmarks as facial features for training. Principal Component Analysis (PCA) is employed to improve accuracy and reduce overfitting. Neural networks are then used to train the model.Experimental results demonstrate that this method achieves a certain level of accuracy in face recognition under makeup and age changes, showing potential for practical applications.MediapipeBlazefaceFacemesh跨年龄化妆人臉識別主成分分析類神經網路MediapipeBlazefaceFacemeshCross-ageMakeupFace RecognitionPrincipal Component AnalysisNeural Networks學術論文