基於CNN對於多人環境進行人臉辨識之研究
Abstract
人臉辨識於現今社會為熱門的議題,每個人皆有獨一的臉部特徵,相較於密碼或是個人證件等傳統的識別方式,人臉辨識既不需要隨時攜帶實體證件也不用擔心忘記密碼。當經由辨識而取得臉部影像後,就能夠藉由不同的臉部特徵與人臉資料庫進行比對來驗證身分。
本研究以設置於教室上方的攝影機拍攝課堂環境,取得之臉部影像解析度較低,因此人臉特徵較不突出,且亦有光線亮度不均勻以及臉部偏移等問題,導致傳統人臉辨識效果不佳。本研究運用YOLOv3結合深度學習的人臉偵測技術取得個人的臉部影像,並搭配卷積神經網路 (Convolutional Neural Network)訓練合適的模型進行人臉辨識,對於20 × 20以上之低解析度且包含不同角度的臉部影像,皆能達到97%以上的辨識準確率。由於人臉長時間下來會有些許的變化,根據實驗結果,經由四個月後之臉部影像仍能維持94%的辨識準確率。
Face recognition has become an important research topic. Compared with more traditional identification methods, such as passwords or personal ID card, automatic face recognition does not require physical ID cards nor need to worry about forgetting passwords. After obtaining the facial image, it is possible to verify the identity by comparing each person's different facial features with the face database. In this study, cameras are placed above the classroom for capturing the classroom setting. Therefore, the resolution of facial images was relatively low, which resulted in less obvious facial features, and faced problems such as brightness and face angle at the same time. This study uses YOLOv3 combined with deep learning face detection method to obtain each person's face image, and use the convolutional neural network (CNN) to train a suitable face recognition model. For low-resolution facial images with a resolution as low as 20×20 and at different angles, the trained model can achieve an accuracy of 97%. From experimental results, the trained model can still maintain a recognition rate of 94%, even with faces taken four months after the initial training faces.
Face recognition has become an important research topic. Compared with more traditional identification methods, such as passwords or personal ID card, automatic face recognition does not require physical ID cards nor need to worry about forgetting passwords. After obtaining the facial image, it is possible to verify the identity by comparing each person's different facial features with the face database. In this study, cameras are placed above the classroom for capturing the classroom setting. Therefore, the resolution of facial images was relatively low, which resulted in less obvious facial features, and faced problems such as brightness and face angle at the same time. This study uses YOLOv3 combined with deep learning face detection method to obtain each person's face image, and use the convolutional neural network (CNN) to train a suitable face recognition model. For low-resolution facial images with a resolution as low as 20×20 and at different angles, the trained model can achieve an accuracy of 97%. From experimental results, the trained model can still maintain a recognition rate of 94%, even with faces taken four months after the initial training faces.
Description
Keywords
人臉辨識, 深度學習, 卷積神經網路, face recognition, deep learning, convolutional neural network