標題: 基於微雲端架構之即時人臉辨識系統
Real-time Face Recognition System Based on Cloudlet Architecture
作者: 王翌倫
林寶樹
Wang, Yi-Lun
Lin, Bao-Shuh
多媒體工程研究所
關鍵字: 即時人臉辨識;微雲端;高維度區域二元特徵;人臉辨識;real-time face recognition;cloudlet;high dimension local binary feature;face recognition
公開日期: 2015
摘要: 人臉辨識已經被廣為研究並且在準確率上面具有非常好的成果,然而,有研究指出,人臉偵測正確率和所使用的特徵描述子大小有顯著關係,即用越大量的特徵將會得到越好的精準度,若要從人臉中,產生大量的特徵值,必然會需要進行大量的運算進而導致速度的下降。傳統人臉辨識架構包含人臉偵測、人臉輪廓偵測、人臉特徵擷取、特徵值比對等構件,本研究針對這些構件,利用微雲端框架以及高維度的區域二元特徵值為基底進行人臉辨識,對人臉辨識系統進行負載平衡。人臉偵測為傳統人臉辨識架構之主要瓶頸之一,本研究提出一方法結合多執行緒及降圖形金字塔將其延遲時間降低至數毫秒,對於整體效能有顯著提升,此外,高維度之人臉區域二元特徵值擷取亦為人臉辨識系統中的主要瓶頸,本研究利用區域二元特徵值擷取多尺度之特性,並對其擷取時間做評估測試,將高維度之區域二元特徵值擷取平均地分給微雲端架構中的處理器進行運算,實驗結果指出其對於特徵值擷取構件的延遲時間有顯著降低之效果,然而,高維度的特徵擷取產生的高維度特徵值必然對網路傳輸造成非常大的負擔,本研究亦提出分解式主成分分析使其高維度之特徵能被各別降維後,再進行傳輸,本研究對於效能的顯著提升顯示微雲端架構之可行性。最後,對於現行具有高辨識率的人臉辨識演算法進行測試,發現其應用在真實生活中會有低精準度之現象,本研究亦提出多重尺度之支持向量機比對使其在真實生活中能被應用。
Traditional face recognition architecture consists of several components, including face detection, face landmark detection, face feature extraction and feature matching. In this research, we focus on these components and build up a real-time face recognition system based on cloudlet architecture and high dimension local binary pattern features, and we also put emphasis on load balance among processors in cloudlet. Face detection is one of the main bottleneck, and we propose a hybrid method to achieve several milliseconds face detection. Furthermore, high dimension feature extraction is also one of the main bottleneck, we utilize the property of high dimension local binary pattern feature to offload the task to other processors in cloudlet. The result of experiment indicates that the performance increases significantly as expected. However, high dimension features would cause heavy loading for network, so we also propose split PCA to reduce the dimension individually in each processor. Finally, high dimension local binary pattern feature achieves high rate of accuracy in face recognition for some preprocess image, but it doesn’t perform well in the wild. We also propose a multi-scale SVM for such problem. To sum up, all experiments in this study show that it is feasible to achieve high-accuracy and low-latency face recognition in a cloudlet architecture.
URI: http://etd.lib.nctu.edu.tw/cdrfb3/record/nctu/#GT070256641
http://hdl.handle.net/11536/139239
Appears in Collections:Thesis