以階層式深度卷積網路實現少樣本的人臉辨識系統;Low-shot Face Recognition using Hierarchical Deep Convolutional Neural Networks

NCUIR > College of Electrical Engineering & Computer Science > Graduate Institute of Computer Science and Information Engineering > Electronic Thesis & Dissertation > Item 987654321/74730

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/74730

Title:	以階層式深度卷積網路實現少樣本的人臉辨識系統;Low-shot Face Recognition using Hierarchical Deep Convolutional Neural Networks
Authors:	沈明訢;Shen, Ming-Hsin
Contributors:	資訊工程學系
Keywords:	深度學習;人臉辨識;即時系統;少樣本
Date:	2017-08-16
Issue Date:	2017-10-27 14:37:41 (UTC+8)
Publisher:	國立中央大學
Abstract:	近年來機器學習蓬勃發展，人臉偵測 (face detection) 與人臉辨識 (face recognition) 技術被廣泛地應用在各種實務、商務、娛樂系統上，像是門禁系統、監控系統、身分認證的登入系統、智慧型裝置等。然而，人臉偵測與人臉辨識常常受到許多因素影響，包含光照環境的不同、表情的不同、臉部旋轉、遮蔽物、及樣本數少的情形等。因此，我們利用卷積神經網路克服以上問題，並提高人臉辨識系統的辨識率。本論文分為兩個部分：第一個部分使用更快速的區域卷積神經網路 (Faster Region Convolutional Neural Network, Faster R-CNN)，在樣本充足的條件下克服光影變化、模糊雜訊、臉部旋轉等因素，並且能夠及時偵測與辨識人臉；第二個部分使用雙胞胎神經網路 (Siamese neural network)，在樣本不充足的條件下提升小樣本類別的辨識率。用我們自己蒐集的多種光源、角度、清晰度變化的人臉資料庫，透過階層式卷積神經網路架構來訓練學習人臉特徵。在實驗分析中，我們以自己拍攝的影片做測試 (包含不同光線變化，不同角度的人臉影像)。偵測方面，依不同參數的調整，偵測率可以達到 96.84%，誤判率 0%。本實驗設計的更快速區域卷積網路的辨識率為 99.65%；在 1920×1080 解析度的影片測試下平均速度為每秒 12.76 張影像；在 960×540解析度的影片測試下平均速度為每秒 24.03張影像。雙胞胎網路，以特徵差異當作是少樣本的分類網路，最後得到整體辨識率98.17%，少樣本類別辨識率92.4%，我們改善了樣本不足的問題，鑄造較好的分類器。 ;In recent years, machine learning has flourishingly developed in face detection and face recognition which are widely used in variety applications, such as access control, monitoring, identity authentication, smart devices, etc. However, face detection and face recognition are always encountered difficult factors, such as different lighting conditions, different facial expressions, facial rotation, occlusion, and small number of samples. Based on the traditional methods, the detected and recognized result are not accepted. Thus, in this study, we use convolutional neural networks to overcome the problems, and improve the recognition rate in face recognition system. The proposed system consists of two parts. In the first part, we use the faster R-CNN (Faster Region Convolutional Neural Network) with sufficient samples to recognize faces with overcoming the various lighting conditions, blurred, and various views of faces. In the second part, we use the Siamese neural network to recognize faces in the minor classes with a few samples. In the experiments, we use our own videos to test the face detection and recognition in various environments such as different lighting conditions, face sizes, and face directions. In the detection stage, the detection rate can reach 96.84%, false positive rate (Misjudgment Ratio) is almost 0%. In the case of face recognition of 1920×1080 images, the recognition rate is 99.65% with 12.76 frames per second (FPS). In the other case of 960×540 images, the FPS is 24.03. With the Siamese network, we distinguish two face images to achieve the recognition rate being 92.4%.
Appears in Collections:	[Graduate Institute of Computer Science and Information Engineering] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	366	View/Open

社群 sharing

Loading...