以影像描述子為基礎之人臉辨識

Abstract

人臉辨識一直是電腦視覺領域中最重要的議題之一。經過數十年的研究，在光線以及臉部表情受到控制的情形下，對於人的正臉，其辨識率超過99%。但在光線、臉部表情、人臉的姿勢有所變化，或是人臉受到遮蔽時，其辨識率就會大幅下降。為了解決這些問題，在本文中以影像描述子(Local Image Descriptor)描述人臉，並且以部分比對(Partial Matching)的方式比對兩張人臉照片。在本文中，我們使用了各種不同的影像描述子，並且試著找出最好的一個。為了提升計算效率，我們將部分的運算平行化，並且實作在多核心的系統上。我們將重點放在非監督式學習(Unsupervised Learning)上，在第一組測試資料的309張人臉中，將其分類為100組時，準確率為99.25%。在第二組測試資料的838張人臉中，將其分類為253個群組時，準確率為99.82%。此結果與Google Picasa單機版本的結果相當相似。但我們的速度仍慢了近8倍以上，故在加速計算上仍有進步的空間。

Face recognition is an important topic in computer vision in the past decades. The rec-ognition rate of frontal faces now is higher than 99% if lighting and facial expressions are controlled. However, if the lighting, facial expression, and pose are various, or the face is under partial occlusion, the recognition rate becomes much lower. In this paper, following Hua and Akbarzadeh 2009’s approach, we implement face representation us-ing local image descriptor, and compare two faces by partial matching. We try kinds of local image descriptors to find the best one. To improve our performance, we parallelize some parts of our computation, and implement it in a quad-core system. In our first da-taset with 309 faces of 5 subjects, we get a recognition precision of 99.25% with 100 clusters; and in our second dataset with 838 faces of 8 subjects, we get a recognition precision of 99.82% with 253 clusters. These results are similar to that of Google Picasa PC version, however, ours is currently at least 8 times slower. Further speedup is ex-pected in the future work.