以區域二元圖樣與部分比對為基礎之人臉辨識

隨著數位相機的普及，人們出門遊玩時，總會拍攝許多照片。我們認為辨認出誰在這些照片中是件有意義且有趣的事。因此與傳統的人臉辨識不同，我們著重的地方在於我們處理的是一般大眾出遊拍的照片。這些照片可能會導致人臉上有不同的光影變化，人臉可能不是正對攝影機，或是臉上有頭髮墨鏡等遮掩五官的物品。這使得我們想要解決的問題無法單純使用傳統方式解決。因此，在這篇論文中，我們使用區域二元圖樣(Local Binary Patterns)及自行提出的部份比對(Partial Matching)的方式來實做人臉辨識。在結果的評估上，我們使用「AR」及「FERET」的正臉資料當作評估基礎，另外加入兩組實驗室成員出遊所拍攝的照片當作最後的評估結果。在資料集一(Dataset I)的309張照片中，我們的系統在100 個群組時，可以達到99.46%的正確率，且Google線上版(Online Version)有94個群組正確率在99.92%的，而單機版有99個群組，正確率可達到100%。而在資料集二(Dataset II)的838張照片中，若將照片分成253個群組，我們可以達到99.57%的正確率，而Google線上版有195個群組正確率為99.49%，單機版有253個群組，正確率為100%。為了執行效能，我們將系統實做在四核心(quad-core)系統上，並將部分工作平行化處理。在我們309張生活照的實驗中，使用單一執行緒需要73分鐘，但使用四個執行緒只需24分鐘。因此我們約有3倍的效能加速。

關鍵字

人臉辨識；區域二元圖樣；部分比對；平行計算

並列摘要

Due to the popularity of digital cameras, when people go on vacation, they will take many pictures. We think it is very meaningful and interesting to identify who are in these pictures. Therefore, different from traditional face recognition problem, we focus on those pictures taken by everyday people. These pictures may have different illumination, different poses, or partially occlusion, which will lead to significant performance dropping using traditional face recognition algorithm. Therefore, in this paper, we present a novel algorithm based on Local Binary Patterns and then combined with Partial Matching. In result evaluation, we will use the AR dataset, FERET dataset, and two home-photo datasets. In addition, we will compare with Google Picasa, which is almost the industry standard, and our performance is no worse than the performance of Google Picasa is using two home-photo datasets. In our system, we get the precision 99.46% in the home photo dataset I (309 images) with 100 clusters, and Picasa will get 99.92% precision with 94 clusters in web version and 100% precision with 99 clusters in download version. In addition, we will get the precision 99.59% in the home photo dataset II (838 images) from 253 images, and Picasa will get 99.49% precision with 190 clusters in web version and 100% precision with 253 clusters in download version. Moreover, we implement the system in a quad-core system, and also implement certain parts of our system in parallel. In our experiment, if we use only a single thread in our system, the executing time of 309 images is 73 minutes. However, if we use four threads in our quad-core PC, we can finish the same job in 24 minutes. It is almost three times faster than single-thread.

並列關鍵字

Face recognition ； Local Binary Patterns ； Partial Matching ； Multithreads

參考文獻

A.M. Martinez, R. Benavente. The AR Face Database. CVC Technical Report #24, June 1998.

Andrew Wagner, Hohn Wright, Arvind Ganesh, Zihan Zhou, and Yi Ma. Towards a Practical Face Recognition System: Robust registration and Illumination by Sparse Representation. In IEEE International Conference on Computer Vision and Pattern Recognition, 2009.

Baochang Zhang, Yongsheng Gao. Local Derivative Pattern Versus Local Binary Pattern: Face Recognition With High-Order Local Pattern Descriptor. In IEEE Transactions on Image Processing, VOL. 19, No. 2, February 2010.

Carlos D. Castillo, David W. Jacobs. Using Stereo Matching with General Epipolar Geometry for 2D face Recognition across Pose. In IEEE Transaction on Pattern Analysis and Machine Intelligence, VOL. 31, No. 12, December 2009.

Che-Hua Yeh, Pei-Ruu Shih, Kuan-Ting Liu, Yin-Tzu Lin, Huang-Ming Chang, Ming Ouhyoung. A Comparison of Three Methods of Face Recognition for Home Photos. ACM SIGGRAPH Poster, August 2009.

國際替代計量

以區域二元圖樣與部分比對為基礎之人臉辨識

全文下載

主題瀏覽