基於不均勻度特徵及K-L 轉換之人臉辨識

Face Recognition Base on Gini Features and K-L Transform

劉忠榮(LIU Zhong-rong)

指導教授 ： 李正宇

亞洲大學/資訊學院/生物資訊學系碩士在職專班/碩士(2009年)

摘要

近年來， 人臉辨識受到各界的矚目， 以人臉進行識別的應用
也越來越廣泛；這些應用大概可分為人臉偵測和人臉辨識兩部分。
人臉偵測的重點是找出人臉的部分， 並加以標示或擷取。人臉辨識
則是把擷取到的人臉圖像和資料庫中的人臉圖像做比對， 並且分辨
其身份。本文是屬於人臉辨識部分。
在1991 年利用主成分分析做人臉辨識的方法被提出來， 但
是這個方式計算的是整張圖像的變異量，計算量大，計算速度低。
因此本論文以圖像的不均勻度（ Gini） 的值來取特徵， 然後再用
這個特徵當作模版進行辨識， 希望能提升辨識速度和辨識率。同
時， 為了在判斷辨識正確與否時能有一個客觀的標準， 我們使用
Otsu 分群法， 對K-L 轉換時求出的歐幾里德距離做分群。
本文所使用的資料庫是ORL 人臉資料庫，裡面共有四十個人，
每人十張影像， 全部有400 張112×92 的影像。我們取用每個人
的第一張作為樣版（ Template）， 在只用K-L 轉換時， 辨識時間
平均約為4.12 秒， 平均辨識率是88.65％ ， 而用Gini 值0.79～
0.81 取樣後， 再使用K-L 轉換， 辨識時間平均約為1.81 秒， 平
均辨識率則變為89.25％ 。如果把Gini 值設為0.77～ 0.81， 則
平均辨識率可提高到90.91％，平均時間則為2.06 秒。從這裡可
以發現， 在辨識率方面沒有顯著的改善， 但是在辨識的速度上則
有非常好的效果。

關鍵字

人臉辨識 ； K-L 轉換 ； 不均勻度 ； Otsu

並列摘要

In recent years, face recognition was attractive by the public's
attention, and the application of face recognition has become
increasingly widespread; about these applications can be divided into
two parts ---human face detection and human face recognition .The
point of face detection is to identify the key parts of face, and then
mark or capture it. Face recognition is to capture images of human
faces and do the matching with the face images ins ide the
database and then distinguish the identity. This article is part of
face recognition.
In 1991 the method of human face recognition that use the
principal component analysis have been proposed, However, this
approach is to calculate the variance of the whole image. The compute
capacity is too large and the pace of the calculation is too slow.
Therefore, the research of this essay is to check the characteristics by
the unevenness of the image (Gini), and then use this as a template to
identify Hope to enhance the recognition speed and rate. At the same
time, in order to have an objective standard to determine whether the
recognition is correct or not, we use the Otsu group method to
divide groups by the European several Reed distances which were
got by the K-L transform.
In this essay, we used the ORL face database, which have 40
people, and each person has 10 images. There are totally 400 images
V
of 112 × 92 in it. We used everyone's first image as a template, when
only using K-L transform. The average time of recognition was about
4.12 seconds, the average recognition rate was 88.65%, and then we
sampled Gini value--- 0.79 ~ 0.81, and use KL transform. The average
time of recognition was about 1.81 seconds, and the average
recognition rate became 89.25 percent. If the Gini value is set to 0.77 ~
0.81, the average recognition rate can be increased to 90.91%, the
average time was 2.06 seconds. Here we can find, the recognition rate
is not significantly improved, but the speed of recognition gets very
good effect.

並列關鍵字

face recognition ； K-L transform ； Gini ； Otsu

參考文獻

數位影像處理(Digital Image Processing, 2nd ed.）(繆紹綱譯)。(2003)。

recognition by independent component analysis. IEEE Transactions

on Neural Networks, 13(6), 1450-1464.

of an Object Under All Possible Lighting Conditions?. In Proc.

IEEE Conference on Computer Vision and Pattern Recognition.