201023053 六、發明說明: 【發明所屬之技術領域】 本發明關於一種生物特徵驗證系統,特別是關於一種多生 物特徵驗證系統及其評分階段融合方法。 【先前技術】 近年來,生物特徵驗證技術已逐漸廣泛應用於各種需要辨 識個人身分的場合。有別於傳統資訊技術以個人持有晶片卡、 密碼或者錄匙的方式較易遭冒用或竊取,生物特徵(Biometrics) _ 驗證技術是以個人與生倶來、獨一無二的生理特徵或行為進行 φ 辨識,因此更具準確性。並且,許多不同的生物特徵皆可被用 以辨識個人身分,個人的生物特徵例如為:指紋特徵、臉部特 徵、手指靜脈特徵、前聲紋或簽名等。前述各種生物辨識之方 式,係以逐漸應用於私人物件持有者之身分辨識(如筆電)、銀 行保全、家庭保全、職場人事出席管理等不勝枚舉,足見生物 辨識技術具有廣大的市場需求,在科技一日千里的發展趨勢 下,生物特徵驗證系統已成為必然之趨勢。 然而現今所見多為單一生物特徵驗證系統之應用,而此類 系統有幾種實際應用上常見發生之嚴重問題,例如:擷取器雜 訊干擾產生不正常資訊、普遍性缺乏、辨識錯誤率過高、或者 由於其單一性導致易於被欺瞒。 因此,同時應用多種生物特徵之多生物徵驗證系統概念能 有效解決前述實際應用上的諸多問題。現今已有諸多相關研究 正蓬勃發展中,然而多生物特徵驗證系統中,需要一有效之評 分階段融合方法,以整合自多個單一生物特徵系統所得之資 訊。本發明之目的即在提出一於評分階段具高效能之融合方 法。 3 201023053 【發明内容】 • 本發明之目的在於提供一種多生物特徵驗證系統及其評 分階段融合方法,可整合自多個單一生物特徵系統所得之資 訊,於評分階段具高效能之融合方法,進行辨識時可獲得相當 高之準確性。 本發明之多生物特徵驗證系統包含一生物特徵擷取模 組、一生物特徵資料庫以及一判斷模組。生物特徵擷取模組, 用於擷取個體之生物特徵,形成對應之特徵資料。生物特徵資 ® 料庫,具有對應生物特徵各別之複數個特徵資料庫。判斷模組 根據生物特徵資料庫,對特徵資料進行比對,用以驗證個體之 生物特徵是否相符。 判斷模組係利用訓練資料組,產生正規化參數以及用於判 斷模組之核心參數。預先調整核心參數,以決定驗證個體真偽 之門檻值。接著,判斷模組係利用測試資料組及正規化模組, 調整正規化參數,以正規化模組對測試資料組所得之評分進行 正規化。分類單元根據正規化後之評分以及核心參數,得出錯 © 誤接受率(FAR)以及正確接受率(GAR),用以建立判斷模組。根 據錯誤接受率(FAR)以及正確接受率(GAR)結果可得出,以複數 個生物特徵驗證個體之正確性係高於以單一生物特徵驗證個 體之正確性。 本發明亦提供多生物特徵驗證系統之評分階段融合方 法,該方法包含下列步驟: 利用訓練資料組,產生正規化參數以及用於調整判斷模組 之核心參數, 利用測試資料組,以正規化模組對測試資料組所得之評分 進行正規化;以及 4 201023053 根據正規化後之評分以及核心參數’得出錯誤接受率(FAR) 以及正確接受率(GAR),調整正规化參數及核心參數,決定驗 證生物特徵是否相符之門檻值,用以建立判斷模組,以融合驗 證個體之生物特徵。 【實施方式】201023053 VI. Description of the Invention: [Technical Field] The present invention relates to a biometric verification system, and more particularly to a multi-biometric feature verification system and a scoring phase fusion method thereof. [Prior Art] In recent years, biometrics verification techniques have been widely used in various situations where it is necessary to recognize an individual's identity. Different from traditional information technology, it is easy to be fraudulently used or stolen by individuals holding wafer cards, passwords or keyloggers. Biometrics _ verification technology is based on individual and raw, unique physiological characteristics or behavior. φ identification, so more accurate. Moreover, many different biometric features can be used to identify an individual's identity, such as fingerprint characteristics, facial features, finger vein features, pre-soundprints, or signatures. The above-mentioned various methods of biometric identification are gradually applied to the identification of personal object holders (such as notebooks), bank preservation, family preservation, workplace personnel attendance management, etc., which shows that biometric technology has broad market demand. Under the development trend of science and technology, the biometric verification system has become an inevitable trend. However, most of the problems seen today are the application of a single biometric verification system. There are several serious problems that are common in practical applications. For example, the noise of the picker noise is abnormal, the lack of universality, and the recognition error rate. High, or because of its singularity, it is easy to be bullied. Therefore, the application of multiple biomarker verification system concepts with multiple biometrics can effectively solve many problems in the aforementioned practical applications. A number of related studies are currently flourishing. However, in multi-biometric verification systems, an effective scoring method is needed to integrate information from multiple single biometric systems. The object of the present invention is to propose a fusion method with high efficiency in the scoring phase. 3 201023053 [Description of the Invention] The object of the present invention is to provide a multi-biometric verification system and a scoring phase fusion method thereof, which can integrate information obtained from a plurality of single biometric systems, and have a high-performance fusion method in the scoring phase. A high degree of accuracy can be obtained when identifying. The multi-biometric verification system of the present invention comprises a biometric capture module, a biometric database, and a determination module. The biometric extraction module is configured to capture the biological characteristics of the individual and form corresponding feature data. Biometrics ® library with multiple feature databases for each biometric. The judging module compares the feature data according to the biometric database to verify whether the biometric characteristics of the individual match. The judging module uses the training data set to generate normalized parameters and core parameters for judging the module. The core parameters are pre-adjusted to determine the threshold for verifying the authenticity of the individual. Next, the judging module uses the test data set and the normalization module to adjust the normalization parameters, and normalizes the scores obtained by the normal data module to the test data set. The classification unit has an error based on the normalized score and the core parameters. The false acceptance rate (FAR) and the correct acceptance rate (GAR) are used to establish the judgment module. Based on the false acceptance rate (FAR) and the correct acceptance rate (GAR) results, it is concluded that the correctness of individuals with multiple biometrics is higher than the correctness of individuals with a single biometric. The invention also provides a scoring phase fusion method for a multi-biometric verification system, the method comprising the following steps: using a training data set, generating normalization parameters and adjusting core parameters of the judging module, using the test data set to normalize the model The group normalizes the scores obtained from the test data set; and 4 201023053 adjusts the normalized parameters and core parameters based on the normalized score and the core parameters 'failed acceptance rate (FAR) and correct acceptance rate (GAR). Verify the biometric characteristics of the threshold, to establish a judgment module to fuse the biometrics of the individual. [Embodiment]
參 請參考第1圖,係依據本發明多生物特徵驗證系統一實施 例之簡單示意圖。本發明多生物特徵驗證系統包含一生物特徵 擷取模組100、一判斷模組200以及一生物特徵資料庫3〇〇。 於此實施例中,生物特徵資料庫300可包含一指紋特徵資料庫 3〇2(尚可更進一步細分為左手指紋特徵資料庫及右手指紋特徵 資料庫)、一臉部特徵資料庫304以及一手指靜脈特徵資料庫 306。但本發明並非以此為限,亦可包含更多種生物特徵資料 庫,如虹膜特徵資料庫、聲紋資料庫或簽名資料庫。判斷模組 2〇〇更包含-分類單元’利用加法原理(sumrule)或支援向量機 (Support Vector Machine)作為分類機對前述生物特徵資料進行 分類。但本發《未以此為^為實現本發明,生物特徵搁取 模組10G當然包含操取多種生物特徵之元件,例如掃描指紋裝 置、攝影及臉部圖像分析裝置以及红外線攝影靜脈圖像分析裝 置等。 如第1圖中所示’生物特徵揭取模組⑽用於摘取個體之 各種生物特徵,形成對應之特徵資料後,傳送㈣斷模m 判斷模組即根據生物特徵資料庫_之指㈣徵資料庫 3〇2、臉部特徵資料庫304 β及手指靜脈特徵資料庫遍,對此 些特徵資料進行比對,用以驗證個體之生物特徵是否相符。根 據本發明之評分階段融合方法’透過以複數個生物特徵驗證個 體之正確性可高於以單-生物特徵驗證個體之正姐(於後詳 5 201023053 述)。 , 請參照第2圖,係本發明多生物特徵驗證系統評分階段具 高效能融合方法之示意圖。首先,利用訓練資料組190,產生 正規化參數應用於正規化模組194,產生核心參數,應用於調 整判斷模組200。接著,利用測試資料組192,以正規化模組 194對測試資料組所得之評分進行正規化。以及,將評分正規 化後送至判斷模組200。判斷模組200即可得出錯誤接受率 (FAR)以及正確接受率(GAR)。接著,調整正規化參數及核心參 ® 數,用以決定驗證生物特徵是否相符之門檻值。如此反覆測試 後即可建立一高效能、驗證正確性高之判斷模組200。 正規化模組能以最小-最大正規化方式(min-max normalization)對測試資料組192所得之評分進行正規化。最小 -最大正規化係以下列方程式對測試資料組所得之評分進行正 規化 X - min(X) X —- max(X)-min(X) (Eql) © 再者,亦能以修正過後的降低高評分效果正規化方式 (reduction of high-scores effect normalization),對測試資料組 192所得之評分進行正規化。前述降低高評分效果之正規化係 以下列方程式對測試資料組所得之評分進行正規化: ,一 X - min( X) X {mean (X*) + std(X*)} - min( X) (Eq2) 其中x為該測試資料組所得之評分,X為評分分布,X* 為驗證該個體為真之評分分布,X'則為正規化後之評分。利用 放程式中之{mean(X*) + std(X*)}項次,能使低真實評分(Low genuine scores)的真實性提高,以提高多種生物特徵驗證時,經 6 201023053 常出現的低真實評分,提高評分階段之效能及正確性。 以下以一簡單之評分及融合方法為例說明如何計算far 及GAR ’以建立本發明之判斷模組。本發明已比對了 5個人(a, B,C,D,E)的左手指紋特徵(每人均拍攝imagel以及image2 ) 及右手指紋特徵(每人亦均拍攝imagel以及image2 )並進行評 分,其結果如下: 表1 圓—------ ---紋 image 1 ------- A B c D E A 29 4 6 4 4 image 2 B 3 16 3 6 9 C 5 4 9 4 6 D 10 6 9 37 5 E 4 ---J -L_ 6 6 4 -〜紋 image i A B C D E A 84 5 5 4 8 image 2 B 4 —-------- -_ 5 3 7 4 C 15 8 10 7 5 D 3 4 4 4 10 E 6 9 7 3 4 7 201023053 理論上自我比對應為真實評分(genuine scores),因此,表 中粗體之為真實評分,其他的則均為虛假評分(impostor scores)。為簡單說明起見,假定前述為訓練資料組亦為測試資 料組,且使用最小-最大正規化方式,左手指紋特徵評分中最大 值為37,最小值為3 ;右手指紋特徵評分中最大值為84,最小 值為3。正規化後為: 表3 左手指紋 image 1 A B C D E A 0.764706 0.029412 0.088235 0.029412 0.029412 B 0 0.382353 0 0.088235 0.176471 image 2 C 0.058824 0.029412 0.176471 0.029412 0.088235 D 0.205882 0.088235 0.176471 1 0.058824 E 0.029412 0.147059 0.088235 0.088235 0.029412Reference is made to Figure 1 which is a simplified schematic diagram of an embodiment of a multi-biometric feature verification system in accordance with the present invention. The multi-biometric verification system of the present invention comprises a biometric capture module 100, a determination module 200 and a biometric database. In this embodiment, the biometric database 300 can include a fingerprint feature database 3〇2 (which can be further subdivided into a left-hand fingerprint feature database and a right-hand fingerprint feature database), a facial feature database 304, and a Finger vein feature database 306. However, the present invention is not limited thereto, and may include more kinds of biometric databases, such as an iris feature database, a voiceprint database, or a signature database. The judging module 2 further includes a sorting unit ‘using the sumrule or the support vector machine as a classifier to classify the biometric data. However, in the present invention, the biometrics surviving module 10G of course includes components for taking various biological features, such as a scanning fingerprint device, a photographic and facial image analyzing device, and an infrared photographic vein image. Analysis device, etc. As shown in Fig. 1, the biometric extraction module (10) is used to extract various biological features of an individual, and after forming the corresponding feature data, the transmission (4) template is determined according to the biometric database _ (4) The database 3〇2, the facial feature database 304β and the finger vein characteristic database are used to compare the characteristic data to verify whether the biological characteristics of the individual match. The scoring phase fusion method according to the present invention can be used to verify the correctness of an individual by a plurality of biometrics than to verify the individual's elder sister with a single-biological feature (described later in detail in 201023053). Please refer to FIG. 2, which is a schematic diagram of a high-efficiency fusion method in the scoring phase of the multi-biometric verification system of the present invention. First, the training data set 190 is used to generate normalization parameters for the normalization module 194 to generate core parameters for use in the adjustment determination module 200. Next, the test data set 192 is used to normalize the scores obtained by the test data set by the normalization module 194. And, the score is normalized and sent to the judgment module 200. The judgment module 200 can obtain a false acceptance rate (FAR) and a correct acceptance rate (GAR). Next, adjust the normalization parameters and the core parameters to determine the threshold for verifying that the biometrics match. After the test is repeated, a high-performance, high-verification judgment module 200 can be established. The normalization module can normalize the scores obtained by the test data set 192 in a minimum-max normalization manner. The minimum-maximum normalization normalizes the scores obtained from the test data set by the following equation X - min(X) X --- max(X)-min(X) (Eql) © Furthermore, it can be corrected The reduction of high-scores effect normalization was reduced and the scores obtained from the test data set 192 were normalized. The aforementioned normalization to reduce the effect of high scoring is to normalize the scores of the test data set by the following equation: , X - min( X ) X {mean (X*) + std(X*)} - min( X) (Eq2) where x is the score obtained by the test data set, X is the score distribution, X* is the score distribution that verifies that the individual is true, and X' is the score after normalization. By using the {mean(X*) + std(X*)} term in the program, the authenticity of Low Genuine scores can be improved to improve the verification of multiple biometrics, which often occurs in 6 201023053. Low real scores improve the effectiveness and correctness of the scoring phase. The following is an example of how to calculate far and GAR' by a simple scoring and fusion method to establish the judgment module of the present invention. The present invention has compared and scored the left hand fingerprint features of five people (a, B, C, D, E) (imagel and image2 per person) and the right hand fingerprint feature (imagel and image2 are also taken per person). The results are as follows: Table 1 Circle ------- --- Grain image 1 ------- AB c DEA 29 4 6 4 4 image 2 B 3 16 3 6 9 C 5 4 9 4 6 D 10 6 9 37 5 E 4 ---J -L_ 6 6 4 -~ pattern image i ABCDEA 84 5 5 4 8 image 2 B 4 —-------- -_ 5 3 7 4 C 15 8 10 7 5 D 3 4 4 4 10 E 6 9 7 3 4 7 201023053 Theoretically, the self-correspondence is the true scores, so the bold in the table is the true score, and the others are the impostor scores. . For the sake of simplicity, it is assumed that the aforementioned training data set is also a test data set, and the minimum-maximum normalization method is used. The maximum value of the left-hand fingerprint feature score is 37, and the minimum value is 3; the maximum value of the right-hand fingerprint feature score is 84, the minimum value is 3. After normalization: Table 3 Left hand fingerprint image 1 A B C D E A 0.764706 0.029412 0.088235 0.029412 0.029412 B 0 0.382353 0 0.088235 0.176471 image 2 C 0.058824 0.029412 0.176471 0.029412 0.088235 D 0.205882 0.088235 0.176471 1 0.058824 E 0.029412 0.147059 0.088235 0.088235 0.029412
表4 右手指紋 image 1 A B C D E A 1 0.024691 0.024691 0.012346 0.061728 B 0.012346 0.024691 0 0.049383 0.012346 image 2 C 0.148148 0.061728 0.08642 0.049383 0.024691 D 0 0.012346 0.012346 0.012346 0.08642 E 0.037037 0.074074 0.049383 0 0.012346 201023053 接著,為簡單說明起見’以加法原理(sum rule)將前述兩生 物特徵之評分進行融合得出結果為: 表5 A B C D E A 1.764706 0.054103 0.112927 0.041757 0.09114 B 0.012346 0.407044 0 0.137618 0.188816 C 0.206972 0.09114 0.26289 0.078794 0.112927 D 0.205882 0.100581 0.188816 1.012346 0.145243 E 0.066449 0.221133 0.137618 0.088235 0.041757 為計算錯誤接受率(FAR)以及正確接受率(GAR),設定門 φ 檻值為 0.26(表中之 C on C),則 FAR=0/20=0 ; GAR=4/5=0.8。 如希望GAR為1(表示100%成功辨識率),則可設定門檻 值為0.041(表中之E on E),但將會有18個虛假評分(impostor scores)超過 0.041,則 FAR=18/20=0.9。亦即 FAR 將高達 90%。 一般而言,門檻值之設定係將FAR值壓至約低越好。 請參照第3圖至第6圖,分別係依據本發明以加法原理為 基礎,與各種單一生物特徵效能之比較。係以降低高評分效果 正規化方式(reduction of high-scores effect normalization)對評 φ 分進行正規化處理,並以加法原理為基礎進行融合。FAR與 GAR之關係曲線可稱之為接收器運作指標曲線(ROC curve ; Receiver Operating Characteristic Curve)。可表示本發明多生物 特徵驗證系統相較於單一生物特徵驗證之優越效能(正確性)。 值得一提的是,若仔細地選取核心參數以及正規化參數,SVM 融合方法較sum rule融合方法可以獲得更好的效果。 請參照第3圖,單一臉部特徵1之R〇C curve 301以及單 一臉部特徵2之ROC curve 302,均低於融合處理兩者後之ROC curve 303 〇 請參照第4圖,左手指紋特徵之R〇c curve 401以及手指 201023053 紋特徵之ROC curve 402,亦均低於融合處理左右手指紋特徵 兩者後之ROC curve 403。請參照第5囷,單一臉部特徵1之 ROC curve 501低於單一臉部特徵2之ROC curve 502 ;單一臉 部特徵2之ROC curve 502低於左手指紋特徵之ROC curve 503;左手指紋特徵之ROC curve 503低於右手指紋特徵之R〇c curve 504,而前述四種單一生物特徵之ROC curve均低於融合 處理四者後之ROC curve 505。 請參照第6圖,單一臉部特徵1之ROC curve 601低於右 ® 手指紋特徵之ROC curve 602 ;右手指紋特徵之ROC curve 602 低於手指靜脈特徵之ROC curve 603,而前述三種單一生物特 徵之ROC curve均低於融合處理三者後之ROC curve 604。 雖然本發明已就較佳實施例揭露如上,然其並非用以限定 本發明。本發明所屬技術領域中具有通常知識者,在不脫離本 發明之精神和範圍内,當可作各種之變更和潤飾。因此,本發 明之保護範圍當視後附之申請專利範圍所界定者為準。 【圖式簡單說明】 © 第1圖係依據本發明多生物特徵驗證系統一實施例之簡單 示意圖。 第2圖係本發明多生物特徵驗證系統評分階段具高效能融 合方法之示意圖。 第3圖係依據本發明以加法原理為基礎的臉部特徵評分階 段之效能說明圖。 第4圖係依據本發明以加法原理為基礎的指紋特徵評分階 段之效能說明圖。 第5圖係依據本發明以加法原理為基礎的一多生物特徵評 分階段實施例之效能說明圖。 201023053 第6圖係依據本發明以加法原理為基礎的另一多生物特徵 評分階段實施例之效能說明圖。 【主要元件符號說明】 100 生物特徵擷取模組 190 訓練資料組 192 測試資料組 194 正規化模組 196 加法原理/支援向量機 198 得出 FAR/GAR 200 判斷模組 300 生物特徵資料庫 302 指紋特徵資料庫 304 臉部特徵資料庫 306 手指靜脈特徵資料庫 ❹ 11Table 4 Right hand fingerprint image 1 ABCDEA 1 0.024691 0.024691 0.012346 0.061728 B 0.012346 0.024691 0 0.049383 0.012346 image 2 C 0.148148 0.061728 0.08642 0.049383 0.024691 D 0 0.012346 0.012346 0.012346 0.08642 E 0.037037 0.074074 0.049383 0 0.012346 201023053 Next, for the sake of simplicity, the principle of addition (sum rule) The results of the above two biometric scores were combined to obtain the following results: Table 5 ABCDEA 1.764706 0.054103 0.112927 0.041757 0.09114 B 0.012346 0.407044 0 0.137618 0.188816 C 0.206972 0.09114 0.26289 0.078794 0.112927 D 0.205882 0.100581 0.188816 1.012346 0.145243 E 0.066449 0.221133 0.137618 0.088235 0.041757 To calculate the error acceptance rate (FAR) and the correct acceptance rate (GAR), set the gate φ 槛 value to 0.26 (C on C in the table), then FAR=0/20=0; GAR=4/5=0.8. If you want GAR to be 1 (representing 100% successful recognition rate), you can set the threshold to 0.041 (E on E in the table), but there will be 18 false scores (inpostor scores) over 0.041, then FAR=18/ 20=0.9. That is, the FAR will be as high as 90%. In general, the threshold value is set to press the FAR value to about as low as possible. Please refer to Figures 3 to 6 for comparison with various single biometric efficiencies based on the principle of addition in accordance with the present invention. The evaluation φ points are normalized by the reduction of high-scores effect normalization, and the fusion is based on the addition principle. The relationship between FAR and GAR can be called the receiver operating characteristic curve (ROC curve; Receiver Operating Characteristic Curve). It can represent the superior performance (correctness) of the multi-biometric verification system of the present invention over a single biometric verification. It is worth mentioning that if the core parameters and the normalization parameters are carefully selected, the SVM fusion method can obtain better results than the sum rule fusion method. Referring to FIG. 3, the R〇C curve 301 of the single facial feature 1 and the ROC curve 302 of the single facial feature 2 are both lower than the ROC curve 303 after the fusion processing. Please refer to FIG. 4, the left hand fingerprint feature. The ROC curve 402 of the R〇c curve 401 and the finger 201023053 pattern feature are also lower than the ROC curve 403 after the fusion processing of the left and right hand fingerprint features. Referring to FIG. 5, the ROC curve 501 of the single facial feature 1 is lower than the ROC curve 502 of the single facial feature 2; the ROC curve 502 of the single facial feature 2 is lower than the ROC curve 503 of the left-hand fingerprint feature; The ROC curve 503 is lower than the R〇c curve 504 of the right hand fingerprint feature, and the ROC curve of the above four single biometrics is lower than the ROC curve 505 after the fusion process. Referring to Figure 6, the ROC curve 601 of the single facial feature 1 is lower than the ROC curve 602 of the right hand fingerprint feature; the ROC curve 602 of the right hand fingerprint feature is lower than the ROC curve 603 of the finger vein feature, and the aforementioned three single biometrics The ROC curve is lower than the ROC curve 604 after the fusion process. While the invention has been described above in terms of preferred embodiments, it is not intended to limit the invention. Various changes and modifications can be made without departing from the spirit and scope of the invention. Therefore, the scope of protection of the present invention is defined by the scope of the appended claims. BRIEF DESCRIPTION OF THE DRAWINGS Fig. 1 is a simplified schematic view of an embodiment of a multi-biometric authentication system in accordance with the present invention. Fig. 2 is a schematic diagram showing a high-performance fusion method in the scoring stage of the multi-biometric verification system of the present invention. Fig. 3 is a diagram showing the performance of the facial feature scoring stage based on the principle of addition according to the present invention. Fig. 4 is a diagram showing the performance of the fingerprint feature scoring stage based on the principle of addition according to the present invention. Fig. 5 is a diagram showing the performance of a multi-biometric stage evaluation stage based on the principle of addition according to the present invention. 201023053 Figure 6 is a diagram illustrating the performance of another multi-biometric scoring phase embodiment based on the principle of addition in accordance with the present invention. [Main component symbol description] 100 Biometric capture module 190 Training data set 192 Test data set 194 Normalization module 196 Addition principle / Support vector machine 198 Get FAR/GAR 200 judgment module 300 Biometric database 302 Fingerprint Feature Database 304 Face Feature Database 306 Finger Vein Feature Database ❹ 11