Free Essay

Ir Face Recognition

In:

Submitted By jimhaung
Words 2058
Pages 9
專題製作成果報告

專題研究計畫名稱

近紅外線影像人臉辨識

組員:
姓名: 黃偉倫 學號: B9921243
姓名: 林稟軒 學號: B9921230

指導教師: 謝堯洋教授

中華民國102年6月28日

摘要 人臉辨識系統的應用非常的廣泛,人臉辨識的技術在這幾十年來也已經有不少的研究成果,因此有不少增加辨識率的方法被提出。但要如何處理當姿勢角度改變、表情改變、照片裡人臉的大小不一,甚至光所造成的亮度問題所產生的影像變化量,這是人臉辨識的一大難題。我們將針對目前提出的眾多增加辨識率之方法進行文獻探討,藉此分析比較各種人臉辨識方法之優缺點,以提供未來進行相關研究時的依據與參考,以有效增加人臉的辨識率。 而我們所做的的方面是在於紅外線的人臉辨識,而人臉辨識的步驟通常都是先把照片做處理,才會進行拿去進行辨識的工作。在紅外線人臉識別裡,有許許多多人做過了很多種方式去辨識,我們看了其中的6種辨識法來研讀,並且挑了其中的PCA、LDA、SVM來實際製作跟測試。 1. Principal Component Analysis(PCA) 2. Linear Discriminant Analysis(LDA) 3. Independent Component Analysis (ICA) 4. Support Vector Machines (SVM) 5. ARENA (基於Memory-based) 6. Unified Bayesian Framework

目錄 摘要 i 目錄 ii 第一章、緒論 1 1-1 研究動機 1 1-2 研究目的 2 第二章、理論基礎 3 2-1 研究方法 3 2-1-1 Principal Component Analysis (PCA) 3 2-1-2 Linear Discriminant Analysis (LDA) 3 2-1-3 Independent Component Analysis (ICA) 3 2-1-4 Support Vector Machines (SVM) 3 2-1-5 ARENA (基於 Memory based) 4 2-1-6 Unified Bayesian Framework 4 2-2 臉部辨識的基本過程 4 2-3 MATLAB介紹 7 第三章、實作或模擬過程 8 3-1 建立人臉資料庫以及其內部樣式 8 3-2 資料庫取樣方法 9 3-3 PCA辨識 9 3-3-1 SVD在PCA上的應用 9 3-3-2 PCA訓練方法 10 3-3-3 PCA測試方法 11 3-4 LDA辨識 11 3-5 SVM辨識 12 3-5-1 SVM訓練方法 12 3-5-2 SVM測試方法 12 3-5-3 multi SVM 12 四、測試或模擬分析結果與討論 13 4-1 測試結果 13 4-1-1 測試PCA、LDA及SVM_Linear第一階段 13 4-1-2 測試PCA、LDA及SVM_Linear第二階段 14 4-1-3 SVM影像降維度與未降維度 15 4-1-4 基於PCA第一階段之線性SVM與多項式SVM 16 4-1-5 測試結果比較圖 18 4-2 測試討論 19 五、結論與自評 20 5-1 專題結論 20 5-2 自我評定 20 六、參考資料 22 七、指導老師意見(簽名) 23

第一章、緒論
1-1 、研究動機 由於電腦科技的急速發展、網路的普及化,發達的資訊已經融入了我們現實的生活中,因此必須要更顧慮到資訊安全上的問題,目前有許多犯罪事件都是屬於資訊犯罪的狀況,例如金融帳號或個人電腦資訊被盜用、個人資料外流的種種問題,這意味著光靠一組密碼並無法保護個人安全。 早期都是靠卡片(身份證、學生證、駕照、職員證、金融卡等等)加上密碼,或是單純只以密碼或卡片來加以保護,只是一旦卡片不見了,或是密碼忘掉了,又或是被他人偽造、竊取,都將造成使用者極大的不便,甚至是巨大的損失。
現今由於科技的進步和電腦運算速度的增加,加上為了提供給使用者更大的便利和可靠性,所以漸漸開始使用以人類生理特徵如人臉、指紋、掌紋、血管、虹膜、語音等,或行為模式,如簽名、步伐、敲打鍵盤(keystrokes)為辨識依據的方法,統稱為生物認證系統(Biometrics),透過電腦或DSP等硬體運算處理設備計算,讓使用者不需要再去牢記一大堆煩人的密碼,即可透過最方便的方式去做身份的辨認。
其中又以人臉辨識最廣為被企業和教學單位所使用和研究,因為他不需要穿戴額外的裝置,更不需要和受測裝置有任何的接觸,只要透過簡單的攝影照相裝置,即可取得辨識所需的資料,因此可說是最方便的辨識方法之一。但是由於影像辨識易受外界環境所干擾,例如:光源、角度和遮罩物等,加上所需處理的資料量大,因此訓練和辨識需要的時間長,所以如何排除這些外在因素,又能以最短的時間達到辨識的效果,即是人臉辨識的一大挑戰。 人臉辨識的應用也可以融入在一般生活中,諸如門禁管制系統、考勤身份辨識系統、影像監視系統、提款卡身份辨識系統,乃至於前科犯資料庫辨識系統的建檔等,皆對人類長久以來的日常生活習慣產生革命性的影響。譬如利用人臉的特徵來做為門禁安全的控管,除了更有效外,可省去感應卡的管理建置費用,也能補足員工忘記帶卡的困擾,更甚者在雙手無暇刷卡或按密碼時,特別能感受到人臉辨識技術的好處。又例如提款機也可以加入人臉辨識的技術,結合晶片、晶片密碼、人臉密碼三種,能夠讓整體安全性更高,就算卡片遺失也不用太擔心會被盜領的情形。 人臉辨識的技術在這幾十年來已經有不少的研究成果,因此有不少增加辨識率的方法被提出。但事實上提升辨識率是一個很困難的問題,而如何能快速的辨識又能讓辨識率能保持在良好的辨視率上,這又是另一項課題了。
1-2、研究目的
人臉辨識系統的應用非常的廣泛,但要如何處理當姿勢角度改變、表情改變,甚至光所造成的亮度問題所產生的影像變化量,進而判斷影像中人臉的身分,這是人臉辨識的一大難題。雖然人臉辨識的發展不論在準確度或速度上,都有大幅的進步,相關的研究論文與應用產品的數量也不可勝數,但卻甚少論文能將這眾多的研究成果做個統整。 現今人臉辨識系統的辨識流程通常為,給一組人臉影像資料庫,訓練人臉辨識系統,使其能從一張新的人臉影像找出其在影像資料庫中符合的個人。而目前所見的人臉辨識系統大都是採用"他是誰?" 的方式,亦即是說,辨識系統會將輸入的影像與資料庫中的所有人臉進行比對,找出最為相似的人作為輸出的結果。且其辨識速度與資料庫人臉數成反比,隨著資料庫成員的增加,其辨識速度會跟著下降。 人臉辨識系統設計的目標在於使辨識性能盡可能與環境條件無關,使正確接受率和錯誤拒絕率同時達到最大化。有鑑於此,我們將針對目前所看的那六種增加辨識率之方法進行文獻探討,並進一步做歸納整理,且進行實際測試,藉此分析比較各種人臉辨識方法之優缺點,以提供未來進行人臉辨識相關研究時,在研究方法使用上的依據與參考。並將藉著統整分析的結果,找出人臉辨識的最佳解決方案,以期待能夠提升人臉辨識的效率與正確性。

第二章、理論基礎
2-1、研究方法
2-1-1、 Principal Component Analysis (PCA) PCA的意義是試圖用最少的變數來表示原始(最大)的影像,也就是說,假設有一張照片經過PCA轉換,取其少數明顯彼此不相關(uncorrelated)的特徵參數,便能以最少的變數來表示原本的照片。所以PCA的兩個優點,第一是降低資料維度,第二是讓資料變成互不相關的。 2-1-2、Linear Discriminant Analysis(LDA) 在PCA 的轉換過程中,其方法不具類別概念,在訓練的過程中,所有的人臉都一視同仁的被處理。因此為了改善此缺點而發展出來LDA的方法,在訓練的過程中,不同的人臉分成不同的類別來處理。LDA 轉換最主要是能將不同類別間的差異拉大,並將屬於同一類別內之影像分散程度縮小。因此解決了PCA轉換後之參數應用在分類與辨識時效果不佳的問題,所以LDA在其人臉分類及辨識上之應用非常廣泛。 2-1-3、Independent Component Analysis (ICA) 獨立成分分析是一種利用統計原理進行計算的方法,是一個線性轉換,這樣的轉換可將資料或訊號分離成獨立,是非高斯特性成分的線性組合。 在人臉辨識上面,Bartlett et al.(1998)首先提出利用ICA 做人臉辨識,將訓練影像以行向量排列,每個行向量代表不同的影像,利用ICA 估計出的獨立成分代表此特徵向量的結果,將辨識影像與訓練所得的解混合矩陣作運算,可得待測影像的結果,並計算待測影像的結果與訓練影像所得結果間的距離,最後利用此距離來辨識人臉。
ICA相比於PCA提供了更強大的數據表示法,它的目標是,提供一個獨立的,而不只是互不相關的圖像分解和表示。
2-1-4、Support Vector Machines (SVM) SVM的概念是在一個Rn的空間找出一個超平面(hyper-plane),使之將兩個不同的集合分開,由於實際資料可能是屬於高維度的資料,而超平面意指在高維中的平面。例如:有兩個集合A、B,集合A位於超平面的左側,則集合B位於另一側(右側)。而超平面的條件,並不只是介於兩集合中間而已,而是希望找到能將兩集合分的較開。 2-1-5、ARENA (基於Memory-based) 這種方法主要是將照片的像素降低,舉個例子由256*256降為16*16所以資料量會變得比較小,而且是相當容易的方式去做比較,沒甚麼特別複雜的,很適合拿來作為一般的用途 圖1. 降低像素之例子。 2-1-6、 Unified Bayesian Framework Bayesian framework使用bayes classifier雖然可以產生最佳化的結果,也就是錯誤率能降到最低,但條件是要先求出probability density functions(pdf),在低維度下可能很好計算,但是一般的圖像就算經過PCA降維度後,對於pdf來說還是算高維度的計算,所以要求出pdf是比較困難的。而Bayesian framework套用的是機率的概念,之前的方法都是將測試的圖像取最近似的距離(也就是最相似的圖像),而機率的概念就是將測試的圖像在已知資料庫中每一張圖像可能的機率表示出來(ex.為資料庫A的機率10%、B的機率15%…),取其機率最大的。
2-2、臉部辨識的基本過程
在這裡,我們主要是做了其中的3種辨識方法,有PCA、LDA以及SVM,而這3種的基本過程都大同小異,因為人臉照片辨識就是先建立一個人臉資料庫,之後就可以進行人臉資料的各個方法轉換,最後進行辨識的步驟。 以下就為3種方法的辨識流程圖。

圖2. PCA及LDA流程架構。

圖3. SVM流程架構。

2-3、MATLAB介紹 MATLAB是MATrix LABoratory的縮寫,是一款由美國The MathWorks公司出品的商業數學軟體。MATLAB是一種用於演算法開發、資料視覺化、資料分析以及數值計算的高階技術計算語言和互動式環境。除了 矩陣運算、繪製函式/資料影像等常用功能外,MATLAB還可以用來建立使用者介面及與呼叫其它語言(包括C,C++和FORTRAN)編寫的程式。 儘管MATLAB主要用於數值運算,但利用為數眾多的附加工具箱(Toolbox)它也適合不同領域的應用,例如控制系統設計與分析、影像處理、訊號處理 與通訊、金融建模和分析等。這些工具箱大多是用開放式的MATLAB語言寫成,使用者不但可以檢視原始碼,還可以可根據自己的需要進行修改以及建立自訂函式。 利用其強大的功能,我們實驗模擬主要以MATLAB來執行,我們也利用其提供的source code來執行我們所須的程式碼。使用的指令如下:
1. [ COEFF, SCORE, latent, tsquare ] = princomp ( X )
2. SVMStruct = svmtrain ( Training, Group )
3. Group = svmclassify ( SVMStruct, Sample )

第三章、實作或模擬過程
3-1、建立人臉資料庫以及其內部樣式
我們所建立的人臉辨識資料庫,總人數為22人,每一個人有11張照片共有242張照片,其中分為訓練(train)以及測試(test)兩部分,train裡面每個人有6張照片,test裡面每個人有5張照片。每個人的11張照片共分為±5、±4、±3、±2、±1以及0(0是正面其餘分別為±90度到±18度)。其中train裡面的每個人6張為±5、±3以及±1,test裡面為±4、±2跟0。

圖4. 第一個人的訓練(train)圖。(±5、±3、±1)

圖5. 第一個人的測試(test)圖。(±4、±2、0)

3-2、資料庫取樣方法 我們的訓練(train)方法是建立從2人到22人的資料庫,也就是我們本實驗提到的2 ~22 class,而每一class又有很多種可能;例如建立2人的資料庫,我們可能取第一人及第二人來建立一組,又另外取第一人及第三人建立另一組資料庫。其組合數量很大,所以我們以隨機取樣的方式,只取其中1000組來建立資料庫,未達1000組則以最大組合數目來建立,所以測試結果大約為1000*5組數據之平均。

圖6. 資料庫取樣方法。
3-3、PCA辨識
3-3-1、SVD在PCA上的應用 由於利用PCA運算方式所得出的特徵空間維度非常大,所以在建立資料庫時,需要非常大的記憶體空間,所以在硬體設備的考量下,我們選用另一種運算方式,奇異值分解(Singular value decomposition),也稱作SVD,來建立我們所需要的特徵空間。利用SVD所計算出來的特徵空間可能跟PCA原始運算方式有所差異,不過誤差很小,重要的是,在樣本數比影像解析度大小還小時,特徵空間維度減少非常多,在運算時間上也相對降低。在Matlab上使用的princomp指令,就是SVD的運算方式。PCA運算與SVD運算結果比較如下圖。

圖7. SVD運算與PCA運算比較。

3-3-2、PCA訓練方法 我們首先將影像以PCA的方法降維度,並取主成分的維度在超過80%停下,但是維度如果超過21,所以取21當作是我們所取的主成分維度。第二次則是將所有的主成分的維度都取在超過80%停下。兩次的所取的特徵數(維度)如下表。 所有的資料全部隨機取1000個組合來做實驗 不滿1000組的class就使用全部的組合。

表1. PCA降維超過21維度取為21。
(第一階段)
class | eigen | 2 | 7 | 3 | 10 | 4 | 13 | 5 | 15 | 6 | 18 | 7 | 21 | 8 | 21 | 9 | 21 | 10 | 21 | 11 | 21 | 12 | 21 | 13 | 21 | 14 | 21 | 15 | 21 | 16 | 21 | 17 | 21 | 18 | 21 | 19 | 21 | 20 | 21 | 21 | 21 | 22 | 21 |
表2. PCA降維取到80%主成分。
(第二階段)
class | eigen | 2 | 7 | 3 | 10 | 4 | 13 | 5 | 15 | 6 | 18 | 7 | 21 | 8 | 23 | 9 | 26 | 10 | 28 | 11 | 30 | 12 | 31 | 13 | 34 | 14 | 37 | 15 | 38 | 16 | 40 | 17 | 42 | 18 | 44 | 19 | 46 | 20 | 48 | 21 | 50 | 22 | 51 |

3-3-3、PCA測試方法 由PCA訓練方法所得出來的結果為一特徵空間及一組訓練影像權重,其中訓練影像為原始訓練影像乘上特徵空間所得。當一張新的影像輸入時,乘上特徵空間,也會得到一個測試影像權重,再與訓練影像權重比較,計算與每個訓練影像權重的歐氏距離,取其最接近的為辨識結果。
3-4、LDA辨識
在測試LDA中,我們參考[7]上的作法,首先用PCA降維度的方式來取得我們計算LDA所需要的參數,其中將維度方法也是使用PCA所做的兩種實驗測試(第一階段與第二階段)。接下來再用LDA的運算方式去建立資料庫,得出結果類似PCA的特徵空間及訓練影像權重。測試方法也如同PCA,比較歐氏距離,最接近的為辨識結果。
3-5、SVM辨識
3-5-1、SVM訓練方法 在測試SVM的方法中,使用了PCA降維度的影像(包含第一階段與第二階段)以及未降維度的影像去建立資料庫。Matlab提供了svmtrain指令,其中包含了線性、Gaussian Radial Basis Function(RBF) 、polynomial等等的方法,我們取了線性及polynomial來做測試。
3-5-2、SVM測試方法
利用Matlab提供的svmclassify指令,將SVM訓練所得出的資料庫以及測試影像輸入,便會得出辨識結果。
3-5-3、multi SVM 由於SVM是一個二元分類的方法,所以在多class需要用新的策略讓我們能夠實現multi SVM。而我們使用的策略如下圖,是先分第一類與其他類,若為第一類,則為辨識結果;若為其他類,則將第一類刪除,繼續下一個分法:第二類與其他類,以此類推。裡用此方法便可以實現multi class在SVM上應用。

圖8. multi SVM辨識方法。

第四章、測試或模擬分析結果與討論
4-1、測試結果
4-1-1、測試PCA、LDA及SVM_linear第一階段 第一階段測試基於PCA,影像降維度(特徵數)設定在21。

表3. 第一階段PCA、LDA及SVM_linear辨識率。 PCA | LDA | SVM | class | accuracy | class | accuracy | class | accuracy | 2 | 98.10 | 2 | 98.96 | 2 | 99.30 | 3 | 96.80 | 3 | 98.11 | 3 | 99.47 | 4 | 95.80 | 4 | 97.61 | 4 | 96.55 | 5 | 95.28 | 5 | 97.30 | 5 | 95.44 | 6 | 95.33 | 6 | 97.06 | 6 | 94.93 | 7 | 93.83 | 7 | 96.92 | 7 | 94.69 | 8 | 93.43 | 8 | 95.95 | 8 | 92.60 | 9 | 92.89 | 9 | 95.04 | 9 | 91.42 | 10 | 91.94 | 10 | 94.44 | 10 | 90.48 | 11 | 90.71 | 11 | 93.47 | 11 | 89.33 | 12 | 90.50 | 12 | 93.12 | 12 | 87.90 | 13 | 90.08 | 13 | 92.25 | 13 | 87.94 | 14 | 88.03 | 14 | 91.89 | 14 | 87.57 | 15 | 87.68 | 15 | 91.31 | 15 | 85.52 | 16 | 87.16 | 16 | 90.97 | 16 | 84.68 | 17 | 85.89 | 17 | 90.55 | 17 | 84.19 | 18 | 85.72 | 18 | 89.86 | 18 | 82.85 | 19 | 85.07 | 19 | 89.32 | 19 | 81.85 | 20 | 84.50 | 20 | 88.74 | 20 | 81.53 | 21 | 83.42 | 21 | 87.75 | 21 | 80.95 | 22 | 81.82 | 22 | 85.45 | 22 | 80.91 |

4-1-2、測試PCA、LDA及SVM_linear第二階段 第二階段測試基於PCA,影像降維度(特徵數)設定在80%主成分。

表4. 第二階段PCA、LDA及SVM_linear辨識率。 PCA | LDA | SVM | class | accuracy | class | accuracy | class | accuracy | 2 | 97.80 | 2 | 98.60 | 2 | 99.00 | 3 | 95.73 | 3 | 98.67 | 3 | 98.13 | 4 | 96.40 | 4 | 97.30 | 4 | 96.80 | 5 | 95.20 | 5 | 97.52 | 5 | 96.00 | 6 | 92.93 | 6 | 96.60 | 6 | 94.33 | 7 | 93.77 | 7 | 96.86 | 7 | 94.11 | 8 | 93.30 | 8 | 96.50 | 8 | 93.55 | 9 | 92.98 | 9 | 97.20 | 9 | 92.89 | 10 | 94.36 | 10 | 96.28 | 10 | 92.36 | 11 | 93.56 | 11 | 96.98 | 11 | 93.13 | 12 | 93.03 | 12 | 96.83 | 12 | 91.97 | 13 | 92.25 | 13 | 96.06 | 13 | 91.75 | 14 | 92.54 | 14 | 96.51 | 14 | 91.94 | 15 | 92.05 | 15 | 96.80 | 15 | 91.68 | 16 | 92.20 | 16 | 95.95 | 16 | 90.53 | 17 | 92.47 | 17 | 96.40 | 17 | 90.94 | 18 | 92.31 | 18 | 96.16 | 18 | 90.71 | 19 | 92.15 | 19 | 96.51 | 19 | 90.82 | 20 | 91.96 | 20 | 96.26 | 20 | 90.78 | 21 | 91.56 | 21 | 96.32 | 21 | 91.21 | 22 | 91.82 | 22 | 97.27 | 22 | 93.64 |

4-1-3、SVM影像降維度與未降維度

表5. SVM影像降維度與未降維度辨識率。 SVM基於PCA第一階段 | SVM基於PCA第二階段 | SVM影像未降維度 | class | accuracy | class | accuracy | class | accuracy | 2 | 99.30 | 2 | 99.00 | 2 | 99.60 | 3 | 99.47 | 3 | 98.13 | 3 | 98.67 | 4 | 96.55 | 4 | 96.80 | 4 | 99.00 | 5 | 95.44 | 5 | 96.00 | 5 | 95.20 | 6 | 94.93 | 6 | 94.33 | 6 | 97.33 | 7 | 94.69 | 7 | 94.11 | 7 | 95.71 | 8 | 92.60 | 8 | 93.55 | 8 | 93.25 | 9 | 91.42 | 9 | 92.89 | 9 | 91.78 | 10 | 90.48 | 10 | 92.36 | 10 | 91.80 | 11 | 89.33 | 11 | 93.13 | 11 | 91.27 | 12 | 87.90 | 12 | 91.97 | 12 | 90.17 | 13 | 87.94 | 13 | 91.75 | 13 | 91.85 | 14 | 87.57 | 14 | 91.94 | 14 | 89.86 | 15 | 85.52 | 15 | 91.68 | 15 | 90.00 | 16 | 84.68 | 16 | 90.53 | 16 | 90.25 | 17 | 84.19 | 17 | 90.94 | 17 | 89.29 | 18 | 82.85 | 18 | 90.71 | 18 | 88.89 | 19 | 81.85 | 19 | 90.82 | 19 | 89.47 | 20 | 81.53 | 20 | 90.78 | 20 | 88.00 | 21 | 80.95 | 21 | 91.21 | 21 | 88.10 | 22 | 80.91 | 22 | 93.64 | 22 | 88.18 |

4-1-4、基於PCA第一階段之線性SVM與多項式SVM 首先我們先嘗試做多項式SVM不同參數變之辨識率。

表6. 多項式SVM辨識率。 多項式SVM | 參數維度=1 | 參數維度=2 | 參數維度=3 | class | accuracy | class | accuracy | class | accuracy | 2 | 99.00 | 2 | 91.90 | 2 | 75.40 | 3 | 97.80 | 3 | 86.87 | 3 | 72.89 | 4 | 96.65 | 4 | 85.50 | 4 | 67.83 | 5 | 95.56 | 5 | 83.40 | 5 | 66.80 | 6 | 94.77 | 6 | 82.80 | 6 | 62.43 | 7 | 94.43 | 7 | 81.60 | 7 | 60.86 | 8 | 93.48 | 8 | 81.75 | 8 | 61.35 | 9 | 91.98 | 9 | 81.89 | 9 | 61.22 | 10 | 90.52 | 10 | 82.52 | 10 | 63.54 | 11 | 89.71 | 11 | 83.24 | 11 | 63.64 | 12 | 88.87 | 12 | 83.12 | 12 | 65.63 | 13 | 87.88 | 13 | 83.14 | 13 | 64.31 | 14 | 85.94 | 14 | 82.40 | 14 | 66.09 | 15 | 85.85 | 15 | 82.23 | 15 | 66.17 | 16 | 84.99 | 16 | 82.49 | 16 | 66.20 | 17 | 83.78 | 17 | 82.99 | 17 | 66.69 | 18 | 83.07 | 18 | 82.29 | 18 | 66.97 | 19 | 82.32 | 19 | 82.66 | 19 | 68.39 | 20 | 81.66 | 20 | 82.64 | 20 | 69.18 | 21 | 80.95 | 21 | 82.42 | 21 | 70.17 | 22 | 80.91 | 22 | 81.82 | 22 | 70.91 |

結果發現參數維度設維1時,辨識率最高,但是參數維度1時,其實就是線性SVM,顧我們取次高的參數維度=2之辨識率做為比較,辨識率如下表。

表7. 線性SVM與多項式SVM(參數=2)辨識率。 線性SVM | 多項式SVM(參數=2) | class | accuracy | class | accuracy | 2 | 99.30 | 2 | 91.90 | 3 | 99.47 | 3 | 86.87 | 4 | 96.55 | 4 | 85.50 | 5 | 95.44 | 5 | 83.40 | 6 | 94.93 | 6 | 82.80 | 7 | 94.69 | 7 | 81.60 | 8 | 92.60 | 8 | 81.75 | 9 | 91.42 | 9 | 81.89 | 10 | 90.48 | 10 | 82.52 | 11 | 89.33 | 11 | 83.24 | 12 | 87.90 | 12 | 83.12 | 13 | 87.94 | 13 | 83.14 | 14 | 87.57 | 14 | 82.40 | 15 | 85.52 | 15 | 82.23 | 16 | 84.68 | 16 | 82.49 | 17 | 84.19 | 17 | 82.99 | 18 | 82.85 | 18 | 82.29 | 19 | 81.85 | 19 | 82.66 | 20 | 81.53 | 20 | 82.64 | 21 | 80.95 | 21 | 82.42 | 22 | 80.91 | 22 | 81.82 |

4-1-5、測試結果比較圖

4-2、測試討論
1.利用PCA降維度,在第一階段測試中,維度設定在21(第七組)後,發現辨識率下降程度明顯提高,可見維度固定,在class逐漸增加的情況下,辨識率的下降程度會增加。
2. PCA降維度,在第二階段測試中,特徵數固定在80%的主成分,辨識率的下降程度就會較為穩定,不再像第一階段會有一個明顯的轉折點,在第7個class的地方辨識率開始下滑。PCA第一階段及第二階段,如下圖所示。

圖14. PCA第一階段與第二階段比較圖。
3.由於LDA是基於PCA的資料庫去做運算,更是加大了各個組別的差異性,所以在辨識方面混淆的機率較小更不容易誤判。由辨識結果可知LDA的表現在3種辨識方法裡是最突出的。
4.首先比較降維度影像與未降維度影像之SVM,雖然在降維度過後,會失去一部分原始影像的特徵,但是降維度固定在80%主成分其辨識率,卻高於未降維度的辨識率。由此可知適當的降維度,能夠去掉影像比較不重要的特徵(雜訊),對於SVM來說,也能夠減少過度擬合的發生。
5.多項式SVM隨著參數的增加,辨識率卻越來越低,不過在參數設定為3時,在越多class辨識率有逐漸上升的趨勢。

第五章、結論與自評
5-1、專題結論
本次的專題實驗,實現了我們當初所定下的初步目標,藉由各個文章所提及的方法,利用我們所使用的人臉辨識資料庫去實現並且驗證其真實性,才可以進一步的討論各方法的優劣。 使用PCA的方法,取主成分80%可以將一些雜訊給去除,其中對於PCA中的另一種形式SVD來說,其辨識率是控制在9成以上是相當正常的,因為我們的database基礎上而言只有臉的左右旋的角度有差異(也就是目前做了角度的差別),而使用了交叉的方式去做比對,只要不要是跟原本的臉部要太大的變動差異,就能保持在優異的辨識率。 對於LDA來說,先使用了PCA降維度之後,再以分類的形式,加大每一個類別的差異性,對於辨識方式是使用歐式距離來看,加大每一類的差異性就更不容易混淆,所以LDA的表現也一直是優於PCA的,也可以說在人臉辨識上,LDA是優化了PCA的缺點。 而SVM的辨識,我們是使用了Matlab所附帶的運算Code,在尋找超平面上,我們並沒有從指令去控制他的所畫的平面的內容,而是由程式碼預設的判斷。而在於結果的顯示上而言,SVM的成果的確是如同資料所顯示,class是少數的時候,他能夠有比較好的表現,但是當class越多的時候,反而連PCA也不如。
5-2、自我評定
B9921243 黃偉倫 在這一次的專題之中,學習如何去看一篇英文的論文,是一個很麻煩的困難,如何去使用查詢所需要的意思,在一開始看論文取得資訊的時候,有太多的事物完全沒辦法了解他其中的含意,直到開始了實作,藉由慢慢地摸索、討論,聽取老師的敘述,才能夠一點一滴地去瞭解到每個辨識方法的真義,或許在太多方面我能力還不夠,但是經過了這一年的專題磨練,我相信我各方面所需的知識有著絕大的進步。
B9921230 林稟軒 在剛開始接觸專題時,先是以閱讀論文的方式來學習,印象很深刻的是,很次看完一種方法,都覺得很抽象,甚至不太懂算式及原理。直到開始實作,實際接觸過發現原理其實不難,一旦了解方法,算式就慢慢的看懂了。而我們使用的辨識軟體是Matlab,在弄懂原理之後,就是要寫迴圈,不管是建立資料庫,讀取數據,測試影像...都需要用到,要如何寫的精簡,能夠花最短的時間完成一次測試,是我在跑過幾次測試後,最能體會到的,有時候會因為某一行指令放錯位置,導致辨識速率非常慢,時常的修正,才能節省運算時間。最後完成了3種方法,雖然只是依照論文上方法去實作並加以驗證,卻已學到不少。

第六章、參考資料
[1] Matthew A. Turk and Alex P. Pentland,”eigenfaces for recognition”,1991,P71-P86.
[2] Kamran Etemad and Rama Chellappa,”Discriminant analysis for recognition of human face images”,1997,P1724-P1733.
[3] Chengjun Liu and Harry Wechsler,”Comparative Assessment of Independent Component
Analysis (ICA) for Face Recognition”,1999,P1-P6.
[4] K. Jonsson, J. Matas, J. Kittlerand Y. P. Li ,”Learning Support Vectors for Face Verification and Recognition”.
[5] Terence Sim, Rahul Sukthankar, Matthew Mullin and ShumeetBaluja,”Memory-based Face Recognition for Visitor Identification”.
[6] Chengjun Liu and Harry Wechsler,”A Unified Bayesian Framework for FaceRecognition”,1998,P1-P5.
[7] Theodoridis, Sergios/ Pikrakis, Aggelos/ Koutroumbas, Konstantinos/ Cavouras, Dionisis, ”Introduction to Pattern Recognition: A Matlab Approach”.
[8] 鄭依貞,”人臉辨識方法之研究”,2008,P5-P7.

第七章、指導老師意見(簽名)

Similar Documents

Free Essay

Face Recognition

...Intel® Technology Journal | Volume 18, Issue 4, 2014 HETERogEnEoUs FAcE REcognITIon: An EmERgIng TopIc In BIomETRIcs Contributor Guodong Guo West Virginia University An emerging topic in biometrics is matching between heterogeneous image modalities, called heterogeneous face recognition (HFR). This emerging topic is motivated by the advances in sensor technology development that make it possible to acquire face images from diverse imaging sensors, such as the near infrared (NIR), thermal infrared (IR), and three-dimensional (3D) depth cameras. It is also motivated by the demand from real applications. For example, when a subject’s face can only be acquired at night, the NIR or IR imaging might be the only modality for acquiring a useful face image of the subject. Another example is that no imaging system was available to capture the face image of a suspect during a criminal act. In this case a forensic sketch, drawn by a police artist based on a verbal description provided by a witness or the victim, is likely to be the only available source of a face of the suspect. Using the sketch to search a large database of mug-shot face photos is also a heterogeneous face recognition problem. Thus it is interesting to study the HFR as a relatively new topic in biometrics. In this article, several specific HFR problems are presented, and various approaches are described to address the heterogeneous face matching problems. Some future research directions are discussed as well to advance...

Words: 7690 - Pages: 31

Free Essay

Avoiding Gain While Transferring Property and Liability to a Controlled Corporation

...liability attached to it, the liability assumed by the corporation is not a recognized gain to the transferor – it instead reduces the adjusted basis of the transferred property. If the liability assumed in the transfer exceeds the property adjusted basis – Section 357(c) provides that the excess liability is a recognized gain to the transferor. The prospect that business persons who incorporate their businesses, and in so doing might incur sizable income taxes due to transferred property liabilities exceeding adjusted basis, has no doubt led many to consider ways to increase the adjusted basis of the property transferred to avoid the tax. I suspect that research will uncover IRS Revenue Rulings, court cases, articles, etc., that will detail attempts by property transferors to avoid the gain recognition due because of Section 357(c) in Section 351 transactions. That will be the focus of this paper. Research Lipton and Bender3 describe...

Words: 1737 - Pages: 7

Premium Essay

Mnc and Bangladesh

...and trade union activities in multinational subsidiaries, the present study tried to investigate extent of the trade union activities and IR practices of MNCs in a developing country context, i.e. Bangladesh, as well as influence of the MNCs on trade unions and IR activities within the existing social and economic contexts. The study revealed that comparative power dominance of multinationals over the host country industrial relations actors emerge as the vital factor in employer-union bargaining process, as well as their capacity to manipulate the existing bargaining process also helped them to implement their desired practices in Bangladesh. However, multinationals were not in confronting mode at all the circumstances, but their potential lobbying power created difficulties for the employees and the trade unions to develop countervailing power in the collective bargaining process and IR activities. Keywords: Convergence-Divergence, Bangladesh, Multinationals, Trade Unions Introduction Empirical studies of international human resource management (IHRM) often reveal that industrial relations practices of multinational subsidiaries face a ‘challenge from established home country labour and IR institutions’ to comply with the local practices (Clark, and Almond, 2004; Monks, 1996; Muller-Camen, et al., 1998) and they need to adjust their IR practices according to the local circumstances. Sometimes, because of flexible and dynamic institutional arrangement, they also find less...

Words: 8880 - Pages: 36

Free Essay

Gaze Control

...psychological research as a tool for recording and studying human visual behavior. Real-time gaze-based text entry can also be a powerful means of communication and control for people with physical disabilities. Following recent technological advances and the advent of affordable eye trackers, there is a growing interest in pervasive attention-aware systems and interfaces that have the potential to revolutionize mainstream human technology interaction. This report provides an introduction to eye tracking technology and gaze estimation. There are a number of principles used in measuring eye movements. This paper focuses on two methods 1.tracking a number of visual features in the image of the eye 2. Measuring relative reflection of infra-red (IR) light. Finally, it includes a light on some of the applications of the eye tracking technique. Keywords—Cognitive processes, Point of Gaze (POG), Saccades I. Introduction The eye has a lot of communicative power. Eye contact and gaze direction are central and very important cues in human communication. The eye has also been said to be a mirror to the soul or window into the brain. Gaze behaviour reflects cognitive processes and can give hints of our thinking and intentions. We often...

Words: 2849 - Pages: 12

Premium Essay

Nt1320 Unit 2 Assignment

...service to CU students, and the day before the project proposal was due I was speaking with my boss about how everyone we work with (mostly the two of us) could improve their language in the office. From this conversation, I arrived at the idea of creating an interactive “swear jar” that would sense when someone swore and alert them to put money in the jar. VOICE RECOGNITION/CODE To start my project, I needed to figure out how to do voice recognition so I looked around online at many different pieces of voice recognition hardware, since this task is not currently possible purely through software. I decided that the VeeaR EasyVR Shield 3.0 would work well for my purposes and I ordered it through SparkFun. The shield arrived and I soldered all the necessary connections and looked at a demo arduino sketch to get started. Once I had the shield setup, I began to think about what words I would...

Words: 744 - Pages: 3

Premium Essay

Biometrics

...been conducted in the last decade in order to investigate how different physiological or behavioral human characteristics can be used as identity evidence to prove the individuality of each person. Some of these characteristics are: face, eyes, ears, teeth, fingers, hands, feet, veins, voice, signature, typing style and gait. Since the first biometric security systems appeared in the market, an increasing demand for novel techniques that will cover all different scenarios, has been observed. Every new method appears to outmatch some of its competitors but, at the same time, presents disadvantages compared to others. However, there is still no method that consists a single panacea to all different scenarios and demands for security. This is the reason for which researchers are on a continuous effort for more efficient and generic biometric modalities that can be used in various applications. In this chapter, emerging biometric modalities that appeared in the last years in order to improve the performance of biometric recognition systems, are presented. The presented methods are divided in two major categories, intrusive and non-intrusive ones, according to the level of user nuisance that each system sets off. 1 Introduction Biometric recognition is a well-known research area that aims to provide more efficient solutions to everyday growing human need for security. Biometrics refers to methods that can be used for uniquely recognizing humans based upon one or more intrinsic...

Words: 12845 - Pages: 52

Premium Essay

Critically Evaluate Best Fit and Best Practice of Human Resources Management

...Equality, Diversity and Inclusion: An International Journal Equality and diversity in employment relations: do we practise what we preach? Jane Holgate Sue Abbott Nicolina Kamenou Josie Kinge Jane Parker Susan Sayce Jacqueline Sinclair Laura Williams Article information: To cite this document: Jane Holgate Sue Abbott Nicolina Kamenou Josie Kinge Jane Parker Susan Sayce Jacqueline Sinclair Laura Williams, (2012),"Equality and diversity in employment relations: do we practise what we preach?", Equality, Diversity and Inclusion: An International Journal, Vol. 31 Iss 4 pp. 323 - 339 Permanent link to this document: http://dx.doi.org/10.1108/02610151211223021 Downloaded on: 01 December 2014, At: 02:26 (PT) References: this document contains references to 46 other documents. To copy this document: permissions@emeraldinsight.com The fulltext of this document has been downloaded 3118 times since 2012* Downloaded by UNIVERSITY OF LEICESTER At 02:26 01 December 2014 (PT) Users who downloaded this article also downloaded: C.W. Von Bergen, Barlow Soper, John A. Parnell, (2005),"Workforce diversity and organisational performance", Equal Opportunities International, Vol. 24 Iss 3/4 pp. 1-16 http:// dx.doi.org/10.1108/02610150510788033 Vidhi Agrawal, Vidhi Agrawal, (2012),"Managing the diversified team: challenges and strategies for improving performance", Team Performance Management: An International Journal, Vol. 18 Iss 7/8 pp. 384-400 http://dx.doi.org/10.1108/13527591211281129...

Words: 10252 - Pages: 42

Premium Essay

India

...BACKGROUND Capital: New Delhi Population: 1.3 milliard Currency: Indian Rupee GDP Value: 3.33% of the world economy Political system: Federal parliamentary democratic republic – President (head of state), prime minister (head of government) Became independent in 1947 * Major economic reforms in the 1990s paved the way for high economic growth in India and involved the considerable liberalisation of the expansive labour laws. * There is a large rural sector and a large ‘informal’ sector in which unions and collective bargaining are rare. Trade union membership overall is low but membership and collective bargaining coverage is higher in the public sector and large enterprises. There are weak laws regarding trade union recognition and representation and poor enforcement of labour laws and collective agreement provisions. Current concerns include lack of protection for workers who have been made worse-off by the market-based labour reforms, poor growth prospects in key parts of the labour market, high rates of contract and casual workers, increasing downsizing and a low-value added IT sector. The Indian industrial relations system is rooted in British common law * The unionisation rate is low at 5% of the total workforce due in part to large rural and informal sectors, which are not unionised. Most union are concentrated in large enterprises and government-related sectors. Introduction: employment relations in India—old narratives and new perspectives ...

Words: 2505 - Pages: 11

Premium Essay

How Technology Has Changed Tax Accounting

...How Technology Has Changed Tax Accounting How Technology Has Changed Tax Accounting The profession of tax accounting has been around for almost a century with the ratification of the 16th Amendment in 1913, which allowed the United States Federal Government to levy taxes on individuals’ income (Rettig, 2006). At that time, the tax laws were fairly simple and most people could complete their tax returns and the required calculations without the help of an accountant. Since then, the tax laws have increased so vastly in number and complexity that the average person cannot understand the laws enough to prepare their own return without assistance. This may come in the form of a tax accountant or software that walks the taxpayer through various rules that may apply to them. This paper focuses on how technology has changed the tax accounting profession throughout the years, highlighting the various innovations that have had the most significant impact. First, I will focus on how technology has changed the various aspects of tax accounting from recruiting clients and communicating with them, to filing completed tax returns and interacting with the Internal Revenue Service. Following, I will describe what tax accounting is likely to look like in the future due to some of the latest technological advances. Finally, I will summarize the major changes in the tax accounting profession over the years, and how tax accounting will continue to evolve moving forward. ...

Words: 3993 - Pages: 16

Free Essay

Tax Research Case

...TO: J Corporation FROM: Yizhen Gong RE: §357(c) gain IRS examination DATE: October 13, 2014 Facts Joe owns 100% of the stock of J Corporation. Joe plans on contributing a parcel of land to J co. having a fair market value of $1.5 million and a basis of $350,000. Further, the land is subject to a mortgage of $1.2 million. To avoid the gain that would result on this transaction, due to §357(c), Joe plans on contributing a promissory note to J Corporation of with a face value of $850,000. He claims that this note will have a basis equal to its face value thus eliminating the gain caused by §357(c). Issues Whether Joe’s transfer of a promissory note to its wholly owned corporation, in an amount equal to the excess of liabilities over the basis of assets contributed in a §351 transfer, avoids §357(c) gain recognition? Can he increase the basis of the assets by transferring his own promissory note to the corporation? Conclusion 1. Joe received stock valued at $300,000 and thus realized a gain of $1,150,000 ($300,000+$1,200,000-$350,000). §357(c) requires Joe recognize a gain of $850,000 ($1,200,000-$$350,000). 2. It depends on whether Joe’s note has a basis in Joe’s hands for purpose of section §357(c). Analysis 1. Under §357(c), no gain or loss is if property is transferred to a corporation solely in exchange for stock of that corporation if, immediately upon the transfer, the transferors are in control of the corporation. Joe transferred...

Words: 743 - Pages: 3

Premium Essay

Globalization in International Relation

...Summary 2. Introduction 3. Industrial Relations (IR) in the Global Context (a) IR - a definition (b) IR and HRM (c) Liberalization, globalization and IR (i) Overview (ii) The relevance of globalization to industrial relations - a summary (iii) The role of multinational corporations (MNC's) (iv) Information technology and IR (v) The impact of other trends (d) The changing nature of IR - a re-definition? 4. IR in Asia and the Pacific (a) The historical context (i) General (ii) Economic development and industrialization (iii) Industrialization and IR policies (iv) IR arrangements - regional divergence (v) Other considerations (b) The current environment (i) Overview 2 (ii) Current responses to liberalization and globalization (iii) Important regional trends 5. Responding to Current and Prospective IR challenges in Asia and the Pacific developing an agenda for employers and their organizations to facilitate transition to the Twenty-first century (a) Overview (b) General issues (i) Achieving competitiveness (ii) Effective IR is a key to improved enterprise performance (iii) The importance of diffusion (c) Specific issues (i) The employers' response (ii) The roles of employers' organizations (iii) The roles of government and trade unions 6. Conclusion References [Top] [Contents] [Previous] [Next] 1 Executive Summary To understand the range of industrial relations (IR) challenges that employers and their organizations are likely to face in Asia and the Pacific...

Words: 16447 - Pages: 66

Free Essay

Comparison of Various Bio-Metric Methods

...usually offered regarding commonly used biometrics, including, Face, Iris, Fingerprint, Finger Vein, Lips, Voice. The comparison criteria list introduced is restricted to accuracy, size of template, cost, security level, and long term stability. Keywords Biometric systems, technique, face, Iris, finger, lips, voice. 1. Introduction "Biometrics" implies "life measurement" however the term associated with the utilization of unique physiological characteristics to distinguish an individual. It’s a new way to verify authenticity. Biometrics utilizes biological characteristics or behavioral features to recognize an individual. In real a Biometrics system is a pattern identification system that uses various patterns such as iris patterns, retina design and biological characteristics like fingerprints, facial geometry, voice recognition and hand recognition and so forth. Biometric recognition system provides possibility to verify one’s identity simply by determining “who these people are” instead of “what these people possess or may be remembered” [4]. The very fact that makes it really interesting is that the various security codes like the security passwords and the PIN number could be interchanged among people but the physical traits cannot be. The principle use of Biometric security is to change the existing password 2. Biometric Techniques Jain et al. describe four operations stages of a uni-modal biometric recognition system [1]: ...

Words: 2857 - Pages: 12

Premium Essay

Amba 610 Individual Paper Assignment Essay

...the Marketability of an Organization? Abstract Every type of organization both small and large needs marketing to survive and continue in business. Even though marketing may not be the reason why an organization was established, it stands as the core of business behaviors. Any type of organization can become a not-for-profit organization, but why? Why compete with for-profit businesses or go through the hassle of becoming established with the IRS? Not-for-profit organizations continue to show growth in its market of the US economy, equaling or even rivaling with their for-profit counter parts. How Does Establishing Not-For-Profit (501c(3)) Status Benefit the Marketability of An Organization? A Review of the Literature Not-for-profit organizations (NPO’s) serve a very important role in society and thus, are often given a special status. The Internal Revenue Service (IRS) is very selective when granting Section 501 status. Any organization can establish itself as a NPO, as long as it meets the metrics laid out by the IRS. The Section 501 status grants tax-exempt status within 24 different categories of organizations. (Andreasen & Kotler, 2008)     1. Are Not-For-Profit Organization Business Savvy? 2. What are the perceptions of Not-For-Profit Organizations? 3. How do Not-For-Profit’s differ from For-Profit Organizations? 4. Why become a Not-For-Profit Organization? This review of the literature on Not-for-profit organizations marketability focuses...

Words: 1916 - Pages: 8

Free Essay

Challenges of an Australian Hr Manager May Face in London

...Central Queensland University Australia | Challenges of an Australian HR manager may face in London | | Subject: HRM in the Global Environment HRMT20022 Subject: HRM in the Global Environment HRMT20022 Over the last few decades the number of expatriates is increasing worldwide. These men and women represent a very important asset for organisations which are developing new business across the world. These expatriates are generally sending abroad to develop strategies or resolve problems in the host country. Consequently, the organisation has to be informed of any issue that expatriates might face and affect the work abroad in terms of international human resource management (IHRM) and industrial relations (IR) issues. Deb (2009) define IHRM as ‘a strategic process of managing a global and diverse pool of talented people in such a way as to achieve aims and objectives of the organisation both at parent and subsidiary level successfully, follow ethical practices on human and business issues and be adaptable to local cultural norms and other human needs and aspirations’. And Dowling, Festing and Engle (2013, p.242) define IR as ‘The board field of study that looks at wider issues of work and employment’. The purposes of this essay are to identify and discuss the IR and HRM issues an expatriate from Melbourne could face while living and working in London. In relation to Industrial relations issues, the role of unions, employment contracts, regulations, visas, taxation...

Words: 2911 - Pages: 12

Premium Essay

Tax Accounting

...discussion of the previous material as it applies to other entities and provides a discussion of accounting periods and accounting methods as they apply to all entities. Discussions of specifi c provisions as they apply to other entities (e.g., corporations, partnerships, etc.) are contained in subsequent chapters. The term “fi nancial accounting” refers to the reporting of the fi nancial data of an enterprise through fi nancial statements prepared in accordance with generally accepted accounting principles. Income tax accounting, hereafter referred to as “tax accounting,” is concerned with the reporting of fi nancial data to satisfy the requirements of the Internal Revenue Code, the Regulations which interpret the Code, rulings by the IRS which further interpret the Code and Regulations, and the decisions of the courts on litigated issues. Tax accounting is statutory. It is concerned with the determination of taxable income as the base to which tax rates are applied to establish the tax liability for a period, usually one year. Basically, taxable income or taxable loss is the net result of summarizing revenues, gains, expenses, and losses. This result is determined in considerable degree in accordance with...

Words: 30831 - Pages: 124