201120765 六、發明說明： 【發明所屬之技術領域】 本發明係關於-種兩階段人臉辨識系統與方法，特別為一 種利用分類與驗證兩階段的方法，透過運算複雜度較低的支 持向量機（SVM)分類方法以加速辨識時間，再利用較精確的特 徵比對方法做驗證，以提升人臉辨識準確性。 【先前技術】 人臉辨識是近幾年來新興起的一項研究技術，其非接觸性 及高方便性的優勢，廣泛的受到研究學者與產業界的高度重 • 視’期盼能於治安或門禁管理上有優異的表現；一個成功的 人臉辨識线可以應用於大樓出人口處身份識別、犯罪嫌疑 人身份識別，甚至可以應用在維護航空安全的出入境身份識 別。當前的人臉辨識系統中，辨識程序為人員站在定位準備 辨識，由影像擷取模組同步的擷取臉部正面或側面影像，經 由特徵抽取後與人員資料庫的全部臉部特徵資料進行比對， 進而找出差異度最低的人員資料庫成員，最後藉由一差昱度 門檻值來判定待辨識影像是已註冊/未註冊人員，這種做法$ 揭露於專利「應用於設定環境下人臉偵測及辨識之方法與裝 φ 置J (專利公開號：200709075)中。 、 另外，如論文[1]中，將臉部區域分割成若干個小區域， 以LBP(L〇Cal Binary Pattern)方法對臉部區域的每個像素進行 影像編碼，計算其t心像素與八鄰域像素值的相關性，藉以 產生一二位元代碼，經由各小區域二位元代碼之hist〇^a叫直 方圖）計數得到臉部特徵，使用Chi_square distance[2]計算測 試影像與訓練資料臉部特徵差異度作為辨識分類器而產出 辨識結果；而論文[3]中，使用MBLBP(Multi_scaie B1〇ck [沉以 Binary Patterns)方法作臉部特徵計算，再以AdaB〇〇st模組⑷ 作為辨識分類器；以上這些辨識系統與方法都有一個共通 201120765 田測°式〜像輸入欲辨識時，都必須和人員資料庫（訓練 資料）的全部人員之臉部特徵進行―對—比對，進而得到差里 度，低的辨識人W果，但隨著f料庫的減，將造成 f异：雜度與貧料比對時間的增加，進而降低了人臉辨識商 品之實用性。 ▲由此可見’上述習用方式仍有諸多不足，實非一良善之設 計’而亟待加以改良。 本案i月人l於上述習用方式所衍生的各項缺點，乃亟思 力二改良創新’並經多年苦心孤錯潛心研究後，終於成功研 發元成本件一種兩階段人臉辨識系統與方法。 【發明内容】 本七月之目的即在於提供一種兩階段人臉辨識系統盥方 法，主要係利用分類與驗證兩階段的方法整合力: 辨識時間，以及提升人臉辨識準確性之目的。 】力速 I達成上述發明目的之兩階段人臉韻㈣與方法，係改 當㈣人臉辨識方法需花費魔大的運算複雜度和資料比對 =間:亚加人驗證的程序，用以提升人臉辨識之準確性。每 ^貝在辨識前需進行註冊，拍攝—連串多臉部角度之人臉 衫像，經程式軟體篩選角度適合的影像後，抽取人臉 將其儲存於人員f料庫内，透過支持向量 :法[5]對人員資料庫中的人員臉部 類別㈣產生支持向量機(SVM)分類模組，每一類別由= 二：夕：部角度特徵資料所組成；當一待辨識影像輸入欲辨 識時’會計算其臉料徵㈣，此臉 ==分類模組的分類，可得到—最相似於已註= 貝之分類結果；再以特徵比對方法做人臉驗證，將 5 。卩特徵貧料與該已註冊人員之臉部特徵資料進行一 201120765 證比對，驗證比對方式係使用⑶― 计鼻兩者臉部特徵之差異度，若 右比對兩者臉部特徵差異度低 於-門檻值時，即認為待辨識影像與該已註冊人員為同二 人，輸出該人員姓名；若比對兩者臉部特徵差里产言於 播值時，即認為待辨識影像與該已註冊人貝不為^人 出非註冊人員之辨識結果。 m 【實施方式】 架構7::，，係本發明兩階段人臉辨識系統與方法之 徵抽2臉擷取模、组1G’其係用以擷取人臉之影像，再由特 徵抽取杈組11計算其臉部特徵； 一特徵抽取模組u，其係用特 徵’獲得臉部特徵㈣12; 丨抽取方Μ异出臉部特 其^ ’該㈣抽取模組u的㈣抽取方法以特徵 -r (Local Binary Patterns . LBP) ^ ^ ^ , ° 區塊局部二元化計算(Mu==k=析度 細咖，MBLBP)影像編碼方法，再加上A u咖3^ (AdaB〇ostlearning)方法； 再加上 AdaB〇ost 學習 —支持向量機（SVM)分類模組13 料u計算並得到一最相似之已註冊n"寺徵貧 訓練所得到之人員1類=— 料庫14做多類別的 度特徵資料所組成；、果，.且母類別由一人員之多臉部角 其中，：貝一貝人=14 ’其係儲存全部人員之臉部特徵資料； 的特徵資料二資料包含多張與多臉部角度影像 了，·二耘式軟體或人工篩選適合的影像； —資料庫抽取模組15，其係透 ， 抽取模組11 人臉#料組1〇與特徵 的臉部特徵資料12’配合支持向量機™)分類 201120765 模組13的分類結果，從已註冊人員資料庫14中選取出特定 成員臉部特徵資料16，即是選取出分類結果之成員臉部 資料，以做為驗證比對時使用； 4 一臉部特徵驗證模組17,其係用以對臉部特徵資料12與 特定成員臉部特徵資料16做一對多的驗證比對；驗證比對^ 式係使用卡方距離（Chi-square distance)計算兩者臉部特徵差 異度，右差異度低過一門檻值時，即認為待辨識影像與該成 員為同人，並輸出s玄成員姓名；若比對後兩者臉部特徵差 異度咼於一門檻值時，即認為待辨識影像與該成員不為同一 人’並輸出非註冊人員； 其中，門檻值係經由計算人員資料庫14中人員臉部特徵 間之Chi-Square distance，包含計算本人和本人臉部特徵i201120765 VI. Description of the Invention: [Technical Field] The present invention relates to a two-stage face recognition system and method, and in particular to a method for utilizing two stages of classification and verification, through a support vector machine with low computational complexity (SVM) classification method to accelerate the identification time, and then use the more accurate feature comparison method to verify, in order to improve the face recognition accuracy. [Prior Art] Face recognition is a research technology emerging in recent years. Its non-contact and high-convenience advantages are widely recognized by researchers and industry. Excellent access control management; a successful face recognition line can be applied to the identity of the building's population, the identification of criminal suspects, and even the identification of entry and exit identification for maintaining aviation safety. In the current face recognition system, the identification program is for the person standing in the position to prepare for identification, and the face or side image of the face is synchronized by the image capturing module, and is extracted through the feature and all facial features of the personnel database. Compare, and then find the member database with the lowest degree of difference, and finally determine the image to be recognized as a registered/unregistered person by a threshold value. This method is disclosed in the patent application. The method of face detection and recognition and the installation of φ set J (Patent Publication No.: 200709075). In addition, as in the paper [1], the face area is divided into several small areas, LBP (L〇Cal Binary The Pattern method encodes each pixel of the face region, calculates the correlation between the t-pixel and the eight-neighbor pixel value, thereby generating a two-digit code, via the small-bit two-bit code of hist〇^ a is called histogram) to get the facial features, using Chi_square distance[2] to calculate the difference between the test image and the training data facial feature as the identification classifier to produce the identification result; and in the paper [3], Use MBLBP (Multi_scaie B1〇ck [Binary Patterns] method for facial feature calculation, then use AdaB〇〇st module (4) as identification classifier; above these identification systems and methods have a common 201120765 field measurement ° type ~ When input is to be recognized, it is necessary to perform a "pair-alignment" with the facial features of all personnel in the personnel database (training data), thereby obtaining a poor degree, a low recognition person, but with the f-repository If it is reduced, it will cause f-differentiation: the increase of the time between the miscellaneous and the poor material will reduce the practicability of the face recognition product. ▲It can be seen that 'the above-mentioned methods of use still have many shortcomings, which is not a good design'. Urgently needed to be improved. In this case, the shortcomings derived from the above-mentioned conventional methods are the two innovations of innovation and innovation. After years of painstaking research, they finally succeeded in developing a two-stage face recognition. System and method. [Summary] The purpose of this July is to provide a two-stage face recognition system method, which mainly uses the two-stage method of classification and verification to integrate forces: Time, and the purpose of improving the accuracy of face recognition. 】 Force speed I achieves the two-stage facial rhyme (four) and method of the above-mentioned invention purpose, is to change (4) face recognition method requires expensive computing complexity and data comparison =Between: The Accredited program is used to improve the accuracy of face recognition. Each cup needs to be registered before being identified, and a series of face images with multiple face angles can be used. After the image is taken, the face is extracted and stored in the staff f-base, and the support vector machine (SVM) classification module is generated by the support vector: method [5] for the person face category (4) in the personnel database, each category It consists of = 2: eve: part angle feature data; when an image input to be recognized is to be recognized, 'the face sign will be calculated (4), the face == classification module classification, can be obtained - most similar to the note = The classification result of Bay; and then the face comparison method by feature comparison method, will be 5 .卩Characteristics and poor materials are compared with the facial features of the registered person for a 201120765 test. The verification comparison method uses (3) ― the difference between the facial features of the nose, if the right contrasts the difference between the two facial features When the degree is lower than the - threshold value, it is considered that the image to be recognized is the same as the registered person, and the name of the person is output; if the difference between the facial features is compared to the value of the face, the image to be recognized is considered to be recognized. The identification result of the non-registered person who is not registered with the registered person. [Embodiment] Architecture 7::, is the two-stage face recognition system and method of the present invention. The 2 face extraction mode and the group 1G' are used to capture the image of the face, and then extract the feature. Group 11 calculates its facial features; a feature extraction module u, which uses the feature 'to obtain facial features (4) 12; 丨 丨 Μ Μ 脸 特 特 ' 该 该 该 该 该 该 该 该 该 该 该 该 该-r (Local Binary Patterns . LBP) ^ ^ ^ , ° Block localization calculation (Mu==k=resolution fine coffee, MBLBP) image coding method, plus A uB〇ostlearning ) method; plus AdaB〇ost learning - support vector machine (SVM) classification module 13 material u calculation and get a most similar registered n" temple levy training personnel obtained class 1 = - material library 14 The multi-category degree characteristic data consists of; the fruit, and the parent category consists of a multi-faceted corner of a person: Beiyibei = 14 'The department stores the facial features of all personnel; Contains multiple and multi-face angle images, two-dimensional software or manual screening suitable images; The extraction module 15 is permeable, and the extraction module 11 face #料组1〇 and the feature facial feature data 12' cooperate with the support vector machine TM) classification 201120765 module 13 classification result from the registered personnel database In the 14th, the specific member facial feature data 16 is selected, that is, the member facial data of the classification result is selected for use as the verification comparison; 4 a facial feature verification module 17 is used for the facial The feature data 12 is compared with the specific member facial feature data 16 for one-to-many verification; the verification comparison method uses the Chi-square distance to calculate the difference of the facial features, and the right difference is lower. When a threshold is deemed, the image to be recognized is considered to be the same person as the member, and the name of the member of the sinus member is output; if the difference between the facial features of the latter two is compared to a threshold value, the image to be recognized is not considered to be the member. For the same person' and output non-registered personnel; wherein, the threshold value is calculated by calculating the Chi-Square distance between the facial features of the personnel in the personnel database 14, including calculating the facial features of the person and the person i
Chi square distance與本人和非本人臉部特徵之chi_squai>e tanCe再根據此兩類Chi-square distance取平均值所決定 其中，可利用一種内儲程式之電腦可讀取媒體，當電腦載 、私式並執行後可完成本案所述兩階段人臉辨識系統與方 一另外，為了測試本案兩階段人臉辨識系統與方法，係以本 實，室所拍攝的人臉資料庫進行實驗，人臉資料庫$ ^位 人：所組成’每-人員連續拍攝了 60張包含多臉部角度之人 :影像’其影像解析度為32〇χ24〇，.經程式軟體篩選適合的 張影像將其分成兩個部分’ 15張為訓練 識之測試影像； 張為待辨 、—此實驗之實施方式為亂數從100位人員中選取80位人員 主冊’即是將此8G位人員訓練資料之臉部特徵儲存於乂 8V類=中，且利用支持向量機(SVM)演算法對^員資料庫你 0類別的训練產生人員分類模組，每— 部角度特徵資料所組成； 人貝之多廢 201120765 透過兩階段人臉辨識方法之步驟，將1〇〇位人員待 測試影像輸人，若-人員在15_試影像中超過5張被^ 某一註冊人貝，則輸出該註冊人員姓名；反之若一二 15張測試影像中不超過5張被認為 貝 註冊人員之結果； ^ A冊人貝，則輸出非 利用分類與驗證兩階段的程序可大幅降低辨識時 於和人員資料庫的全部人員之臉部特徵進行—對 [3]’辨識時間由每—人員識秒降低至每一人員〇5卿. 力旦料人㈣^準雜，與錢^持向 里辨减方法比較，人員辨識率均為98.75% ’但人 員拒絕率從8 8 %提升至9 8 %。 步驟==示，為本發明兩階段人臉辨識系統與方法之 步驟一：人臉擷取模組擷取待辨識人臉影像2〇1 ; 7 矛丨用知'徵抽取模組抽取出臉部特徵資料202 · 二最T:特徵資料透過支持向量機(s vm)分類模組計 ^付到一,最相似之已註冊人員資料庫某成員203; 部特S :，貧Λ庫抽取模組於人員資料庫選取特定成員之臉 :驟五·偽I疋選取分類結果之成員臉部特徵資料2〇4; ’广.臉部特徵驗證模組比對人員資料庫之臉邛特:¾及 特定^臉部特徵是否符合2〇5; 之“特徵及 若严合，則驗證成功2。6,並輸出成員姓名斯； 之訊息209:右不符合，則驗證失敗28°’並輸出非註冊人員 本發明所提供之兩階段人臉辨識系 技術相互比較時，更具備下列優點： 肖其他^ 大幅降低運算複雜度與加快辨識時間，使其可 〜用於較大型人員資料庫之人臉辨識系統。 201120765 2.本發明可提升人臉辨識準確性，使其應用於人臉辨識系 統更可確保居家與門禁管理安全。 上列詳細說明係針對本發明之一可行實施例之具體說 明，惟該實施例並非用以限制本發明之專利範圍，凡未脱離 本發明技藝精神所為之等效實施或變更，均應包含於本案之 專利範圍中。 ^ 物 綜上所述，本案不但在技術思想上確屬創新，並能較習用 增進上述多項功效，應、已充分符合新穎性及進步性之法 ^明^利要in法提出中請，懇請t局核准本件發 月專利申請案，以勵發明’至感德便。 【圖式簡單說明】 統與方法之架構圖； 統與方法之步驟流程圖 圖一為本發叼兩階段人臉辨識系 圖二為本發明兩階段人臉辨識系 【主要元件符號說明】 10 人臉擷取模組 11 特徵抽取模組 12 臉部特徵資料 13 支持向量機（SVM)分類 14 人員資料庫 15 資料庫抽取模組 16 特定成員臉部特徵資料 17 臉部特徵驗證模組 201120765 參考文獻 1. T. Ahonen, A. Hadid, and M.Pietikainen, Face Recognition with Local Binary Patterns, Proc. Eighth European Conf. Computer Vision, pp. 469-481, 2004. 2. Rodriquez C , Chi-Square Test for Independence , last accessed 21 February 2006. 3. S. Liao, X. Zhu, Z. Lei, L. Zhang, and S. Z. Li., Learning Multi-scale Block Local Binary Patterns for Face Recognition, In International Conference on Biometrics (ICB), pages 828-837, 2007. 4. Friedman, J., Hastie, T., Tibshirani, R., Additive Logistic Regression: A Statistical View of Boosting, Technical report, Department of Statistics, Sequoia Hall, Stanford University (July 1998).Chi square distance and my own and non-self facial features of chi_squai>e tanCe are determined by the average of the two types of Chi-square distance, which can be read by a computer with a built-in program. After the implementation and execution, the two-stage face recognition system and the method described in the present case can be completed. In order to test the two-stage face recognition system and method of the present case, the face database of the real and the room is used for experiment, face Database $^People: The group consisting of 'every-persons continuously shot 60 people with multiple face angles: the image' has an image resolution of 32〇χ24〇. The program image is divided into suitable images to divide it into Two parts '15 sheets are test images for training; Zhang is to be identified, and the implementation method of this experiment is to select 80 people's main books from 100 people in random numbers', that is, the face of this 8G personnel training materials The features are stored in the 乂8V class=, and the support vector machine (SVM) algorithm is used to generate the personnel classification module for each of the 0 categories of the training database, and each of the angle characteristics data is composed; Waste 201 120765 Through the steps of the two-stage face recognition method, one person to be tested is input to the test image, and if the person is more than 5 in the 15_ test image, the registered person name is output; On the other hand, if no more than 5 of the 15 test images are considered to be the result of the registered personnel; ^ A book of people, the output of the non-utilization classification and verification two-stage procedure can greatly reduce the identification and the personnel database. The facial features of the personnel are carried out - the recognition time for [3]' is reduced from each person to the second person. The force is calculated by the person who is in the middle of the period. The recognition rate is 98.75% 'but the personnel rejection rate has increased from 88% to 98%. Step == shows, step one of the two-stage face recognition system and method of the present invention: the face capture module captures the face image to be recognized 2〇1; 7 the spear uses the knowledge extraction module to extract the face Departmental characteristics data 202 · Two most T: Feature data is paid to the supporter vector machine (s vm) classification module by one, the most similar member of the registered personnel database 203; Department special S:, poor library extraction mode The group selects the face of a specific member in the personnel database: Step 5: Pseudo I疋 selects the member's facial feature data of the classification result 2〇4; 'Guang. Face feature verification module compares the face of the personnel database: 3⁄4 And whether the specific ^face feature meets 2〇5; "the feature and if it is strict, the verification succeeds 2.6, and the member name is output; the message 209: the right does not match, the verification fails 28°' and outputs non- Registered personnel The two-stage face recognition system technology provided by the present invention has the following advantages when compared with each other: Xiao other ^ greatly reduces the computational complexity and speeds up the identification time, so that it can be used for the face of a larger personnel database. Identification system. 201120765 2. The invention can improve the accuracy of face recognition It is applied to the face recognition system to ensure the security of home and access control management. The above detailed description is specific to a possible embodiment of the present invention, but the embodiment is not intended to limit the scope of the patent of the present invention. Equivalent implementations or changes from the spirit of the invention should be included in the scope of the patent of the present invention. ^ In summary, the case is not only innovative in terms of technical thinking, but also can enhance the above-mentioned functions more effectively. Should be, has fully complied with the novelty and progressiveness of the law ^ Ming ^ Lie in the law, please ask the bureau to approve the patent application for this month, in order to invent the 'to the sense of virtue. The structure of the method and the method of the method flow chart Figure 1 is the two-stage face recognition system of the present invention. Figure 2 is the two-stage face recognition system of the present invention. [Main component symbol description] 10 Face capture module 11 Feature extraction module 12 Face feature data 13 Support vector machine (SVM) classification 14 Personnel database 15 Database extraction module 16 Specific member facial feature data 17 Face feature verification Group 201120765 References 1. T. Ahonen, A. Hadid, and M. Pietikainen, Face Recognition with Local Binary Patterns, Proc. Eighth European Conf. Computer Vision, pp. 469-481, 2004. 2. Rodriquez C , Chi- Square Test for Independence , last accessed 21 February 2006. 3. S. Liao, X. Zhu, Z. Lei, L. Zhang, and SZ Li., Learning Multi-scale Block Local Binary Patterns for Face Recognition, In International Conference on Biometrics (ICB), pages 828-837, 2007. 4. Friedman, J., Hastie, T., Tibshirani, R., Additive Logistic Regression: A Statistical View of Boosting, Technical report, Department of Statistics, Sequoia Hall, Stanford University (July 1998).
5. J.C. Burges, A Tutorial on Support Vector Machines for Pattern Recognition, Data Mining and Knowledge Discovery, vol. 2(2), 1998, pp 121-167.5. J.C. Burges, A Tutorial on Support Vector Machines for Pattern Recognition, Data Mining and Knowledge Discovery, vol. 2(2), 1998, pp 121-167.