1325568 九、發明說明 【發明所屬之技術領域】 -綠t =疋有關於"'種人臉驗證之方法，且特別是有關於 夕數決規則與動態臨界值來辨識人臉之人臉驗證方 【先前技術】1325568 IX. Description of the invention [Technical field to which the invention belongs] - Green t = 疋 There is a method for "face verification", and in particular, there are face verification rules and dynamic thresholds for face recognition of faces. [Prior Art]
由於經濟的蓬勃發展及犯罪率的提升，使得保全系統越 來越又重視。過去常用的保全系統係採用如密碼、磁卡、作 業系統的權限密碼來作為身份之確認，然而，這些方法皆有 其缺點’例如密碼被破解、卡片遺失盜用或遺忘密碼等風險， 因此k可靠的方式係使用生理特徵來作為保全機制如人 臉、指紋、虹膜、語音等。 將生理特徵作為保全系統的認證機制已逐漸成為市場急 迫的需求，而人臉驗證係為生物驗證技術當中最埶 二 題’：此有許多學者在此領域提出相關之演算法，例如：支 # ^ t ^ (Support Vector Machine； SVM)^ 0 ^ ^ (GeneticDue to the booming economy and the increase in crime rates, the security system has become more and more important. In the past, the security system used in the past used passwords, magnetic cards, and operating system passwords as identifications. However, these methods have their shortcomings, such as passwords being cracked, cards being stolen or forgotten, etc., so k is reliable. The method uses physiological features as a security mechanism such as face, fingerprint, iris, voice, and the like. The use of physiological characteristics as the authentication mechanism of the security system has gradually become an urgent demand in the market, and the face verification system is the second most problem in the biometric verification technology': many scholars have proposed relevant algorithms in this field, for example: ^ t ^ (Support Vector Machine; SVM)^ 0 ^ ^ (Genetic
Algorithm ; GA)' 類神經網路（Ne〇c〇gnitr〇n Neurai; nn)以及 隱藏式馬可夫模型（Hidden Markov Model ; HMM)。支持向量 機方法係把人臉的特徵當作高維度的向量，進而尋找—個超 平面（Hyper Plane)，並針對不同類的人臉來進行分類而得到 不同的分類器。使用這些分類器，可以決定哪些人臉係屬於 哪個分類群組，同一個分類群組的人臉可以視為相似的人 臉。但是，此方法若需辨識N張人臉，需要N-1個分類器才 丄⑵:)⑽ 能完成人臉的辨識，所 糞法則#因兔户 所以此方法的運算複雜度太高。基因演 具則係因為在不同解 ia Λ, ^ , 丁号衣最佳解，因此其處理時 =相：：較長。類神經網路係—種利用許多輸入與輸出的資 &成模擬數學模型的訓練與學習，並經過不斷地 與學習資料：來推估、預測與辨識測試資料，但是類神經網 路之訓練方法較複雜。隱藏式馬可夫模型係—種使用少量不 同角度的人臉影像，來辨識大量且多種角度的人臉影像，但 疋除非人臉的影像完全在同一個角度，要不然則需更多角 度的人臉影像，才能適應各種角度的人臉辨識，並且具有下 列缺點.使用者必需刻意的面對攝影機來操取人臉影像、驗 也之方法/又有考慮到即時（〇n Une)處理與後置⑴汀Lhe)處理 同時運用之；^法創丨，練的複雜度亦較高。所謂㉛時處理係在 待測者輸入後’馬上進行辨識並知道結果，而後置處理則係 一疋數量之待測者輸入累積到一定數量時，亦或過一段時間 後，才將此批次做一辨識處理。於人臉驗證時亦有採用主成 份分析（Principal Component Analysis ; pCA)之統計方法來計 算人臉的特徵值，傳統的主成份分析因為其計算方式係求特 徵值（Eigenvalue)、特徵向量（Eigenvect〇r)以及共變異數 (Covariance)，因此需要大量的記憶空間來計算。 又，於人臉驗證系統中，通常採用一種固定之臨界值來 驗證人臉，此臨界值係資料庫中同一人之數個人臉影像所計 鼻出之特徵值’當待測人臉影像之特徵值大於此臨界值時， 即代表不相似’則允許其進入。然而，即使為同一張人臉， 每天都會有稍許不同’例如可能帶有不同表情的臉孔、臉上 1325568 多了些傷跡或長鬍子，因茈 〜’右僅依賴固定臨界值來驗證人 臉，誤判的機率會相當高。 因此’有必要提出一種 馆此提升人臉驗證準確性、降低運 鼻複雜度、減少樣本測試時P弓 吁間、降低光線對人臉之影響、減 少因人臉表情之稍微差異而 士 向誤判之機率，且能同時運用於即 時處理與後置處理情況下之^ υ人臉驗證系統與方法。 【發明内容】Algorithm; GA) 'Nerve network (Ne〇c〇gnitr〇n Neurai; nn) and Hidden Markov Model (HMM). The support vector machine method treats the features of the face as high-dimensional vectors, and then looks for a Hyper Plane, and classifies different types of faces to obtain different classifiers. Using these classifiers, you can decide which face belongs to which category group, and faces of the same category group can be treated as similar faces. However, if this method needs to identify N faces, it needs N-1 classifiers. (2) :) (10) It can complete the face recognition, and the manure method is because of the rabbit family. Therefore, the computational complexity of this method is too high. The gene model is based on the different solutions of ia Λ, ^, and the best solution, so the processing time = phase:: longer. Neural network-like system—using many input and output resources to develop and train mathematical models, and to continuously evaluate and predict and identify test data, but to train neural networks. The method is more complicated. The hidden Markov model is a face image that uses a small number of different angles to identify a large number of angles of face images, but unless the face image is at the same angle, otherwise a more angled face is required. The image can be adapted to face recognition from various angles, and has the following shortcomings: the user must deliberately face the camera to fuck the face image, the method of the test / also consider the instant (〇n Une) processing and the post (1) Ting Lhe) treatment at the same time; ^ method of creation, the complexity of training is also higher. The so-called 31-hour processing is to identify and know the result immediately after the input of the test subject, and the post-processing is performed when a certain number of test subject inputs accumulate to a certain amount, or after a period of time, the batch is made. An identification process. In the face verification, the statistical method of Principal Component Analysis (pCA) is also used to calculate the eigenvalue of the face. The traditional principal component analysis is based on the calculation method to obtain the eigenvalue (Eigenvalue) and the feature vector (Eigenvect). 〇r) and Covariance, so a large amount of memory space is needed to calculate. Moreover, in the face verification system, a fixed threshold value is usually used to verify the face, and the critical value is the characteristic value of the same person's personal face image in the database, when the image of the face to be tested is When the eigenvalue is greater than this threshold, it means that it is not similar, then it is allowed to enter. However, even for the same face, there will be a slight difference every day. For example, a face with a different expression may have more faces or 1325568 more scars or long beards, because the right only relies on a fixed threshold to verify the face. The probability of misjudgment will be quite high. Therefore, it is necessary to propose a kind of pavilion to improve face verification accuracy, reduce nose complexity, reduce P-bends during sample testing, reduce the effect of light on the face, and reduce the slight difference in facial expression. The probability and ability can be applied to both face processing systems and methods in the case of instant processing and post processing. [Summary of the Invention]
一種人臉驗證方法，能同 因此’本發明的目的係在提供 時運用於即時處理與後置處理情況下，並採用動態臨界值， 可以時常更新影像資料庫，以避免因時間的改變所造成人臉 特徵的變化，而造成誤判的情形，進而提高辨識度。 本發明的另一目的係在提供一種人臉驗證方法，利用複 合主成份分析（Composite Principal c〇mp〇nent Analysis; CPCA)來減少樣本測試時間、降低光線對人臉之影響以及降低 運算複雜度。 ^ 本發明的又一目的係在提供一種人臉驗證方法，藉由利 用灸數決規則（Majority Rule)之多數表決之方法，來共同辨識 人臉影像，進而可以防止人臉驗證之誤判。 根據本發明之上述目的，提出一種人臉驗證之方法，至 少包含下列步驟。首先，輸入驗證資料。接著，判斷驗證資 料是否正確，並產生第一結果；若第一結果為否，則拒絕驗 證資料之使用者進入；若第一結果為是，則將驗證資料健存 至資料處理與儲存系統，其中驗證資料係對應資料處理與儲 7 1325568 存系統中之數個人臉樣本影像。接著，啟&影像擷取器，來 擷取人臉待測影像，並儲存人臉待測影像至資料處理與儲存 系統。接下來，對人臉待測影像與這些人臉樣本影像進行人 辨識之步驟，以计鼻出人臉特—徵相包。隨後，進行動雜 臨界值比對’判斷人臉特徵相似值是否大玲臨界值，並產生 第一結果’其中臨界值係先藉由計算資料處理與儲存系統中 之°卩伤人臉樣本影像間之數個第一特徵距離，再根據這此第 φ 特徵距離所決定。若第二結果為否，則允許驗證資料之使 用者進入；若第二結果為是’則進行多數決規則，並產生允 許票數和拒絕票數，並判斷允許票數是否大於拒絕票數而 產生第四結果。若第四結果為否，則拒絕驗證資料之使用者 ‘進入；若第四結果為是’則允許驗證資料之使用者進入，並 .進行更新人臉樣本影像步驟。 根據本發明之另一目的，提出一種人臉驗證之方法，至 少包含下列步驟。首先，輸入數個驗證資料，其中每一驗證 _ 資料對應一使用者》接著，判斷每一驗證資料是否正確，並 拒絕驗證資料不正確之使用者進入。接下來，儲存驗證資料 至-貝料處理與儲存系統，其中每一驗證資料係對應資料處理 與儲存系統内之數個人臉樣本影像。隨後，啟動影像擷取器， 對每一驗證資料之使用者擷取人臉待測影像，並儲存人臉待 測影像至資料處理與儲存系統。接著，對人臉待測影像與人 臉樣本影像進行人臉辨識之步驟，以計算出人臉特徵相似 值°然後，進行動態臨界值比對，判斷人臉特徵相似值是否 大於臨界值’並產生第二結果，其中臨界值係先藉由計算資 8 1325568 =理與儲存系統中之部份人臉樣本影像間之數個第一特徵 二:再根據這些第一特徵距離所決定。若第二結果為否， :後置處理；若第二結果為是，則進行多數決規則，並 允許票數和拒絕票數’並判斷允許票數是否大於拒絕票 1而產生第四結果。若第四結果為否’則進行後置處理； :::結果為是’則進行後置處理’並進行更新人臉樣本影 像步驟》A method for face verification, which can be used for the purpose of real-time processing and post-processing, and using dynamic thresholds, can constantly update the image database to avoid time changes. The change of the face features causes a misjudgment, thereby improving the recognition. Another object of the present invention is to provide a face verification method, which uses Composite Principal c〇mp〇nent Analysis (CPA) to reduce sample test time, reduce the influence of light on a human face, and reduce computational complexity. . Another object of the present invention is to provide a face verification method for jointly recognizing a face image by using a majority voting method of a Majority Rule, thereby preventing false positives of face verification. According to the above object of the present invention, a method of face verification is proposed, which comprises at least the following steps. First, enter the verification data. Then, determining whether the verification data is correct, and generating a first result; if the first result is no, the user who refuses to verify the data enters; if the first result is yes, the verification data is saved to the data processing and storage system, The verification data is corresponding to the data processing and storage of the personal face sample images in the 7 1325568 storage system. Then, the image capture device is used to capture the image to be tested and to store the image to be measured to the data processing and storage system. Next, the step of recognizing the face image to be tested and the face sample images is performed to count the face-specific signs. Subsequently, the comparison of the motion and the criticality value is performed to determine whether the similarity value of the facial feature is a threshold value and a first result is generated. The critical value is obtained by calculating the image of the facial sample in the data processing and storage system. The first characteristic distance between the two is determined according to the φ characteristic distance. If the second result is no, the user of the verification data is allowed to enter; if the second result is yes, the majority rule is executed, and the number of allowed votes and the number of rejected votes are generated, and it is judged whether the number of permitted votes is greater than the number of rejected votes. Produce a fourth result. If the fourth result is no, the user who refuses to verify the data is 'entered; if the fourth result is yes', the user of the verification data is allowed to enter, and the step of updating the face sample image is performed. According to another object of the present invention, a method of face verification is provided, comprising at least the following steps. First, a plurality of verification data are input, wherein each verification _ data corresponds to a user. Then, it is judged whether each verification data is correct, and the user who refuses to verify the data is refused to enter. Next, the verification data is stored to the bedding processing and storage system, wherein each verification data corresponds to a plurality of personal face sample images in the data processing and storage system. Then, the image capture device is activated to capture the image of the face to be tested for each user of the verification data, and store the image of the face to be processed to the data processing and storage system. Then, the steps of face recognition on the face image to be tested and the face sample image are performed to calculate the similarity value of the face feature, and then the dynamic critical value comparison is performed to determine whether the similarity value of the face feature is greater than the critical value' A second result is generated, wherein the threshold value is first determined by calculating a number of first feature two between the image of the face sample in the storage system: and then determining according to the first feature distance. If the second result is no, the post processing is performed; if the second result is YES, the majority decision rule is executed, and the number of votes and the number of rejected votes are allowed, and it is judged whether the number of permitted tickets is greater than the rejected ticket 1 to produce a fourth result. If the fourth result is no, then post processing is performed; ::: the result is yes, then post processing is performed and the face sample image step is updated.
【實施方式】 ’可參照下列描述 為了使本發明之敘述更加詳盡盘穿 並配合第！圖之圖示。 圖 睛參照第 -"仪丨土員犯例您人臉 ^方法之_流程圖H進好驟細，制者輸入驗證 貝枓（例如：員工卡號、ID編號）至輸入裝置中，其中，輸入 裝置係選自於由感應式讀卡機、1(:卡讀卡機、磁卡 式刷卡機、條碼刷卡機、條碼掃描器以及個人密 輸入器所組成之一族群。 接著，進行步驟202,判斷驗證資料是否正確，並產生— 第、。果。右第一結果為否，代表此驗證資料非儲存至資料 處理與儲存系統中的設定者，則拒絕此驗證資料之使用者進 入此門禁系統(步驟224);反之若第一結果為是，則輸入裝 置將此驗證資料儲存至資料處理與儲存系統（步驟204)，且此 驗證資料對應到此資料處理與儲存系統之欄位，此欄位 數個人臉樣本影像。接著，進行步驟施，藉由資料處理與健 9 1325568 存系統來啟動影像擷取器’此影像擷取器將對使用者之人臉 擷取一人臉待測影像，並將此人臉待測影像儲存至資料處理 。與儲存系統（㈣2G8)，此步驟係將人臉待測影像儲存至驗證 '貧料所對應之攔位中，其令影像擷取器可為攝影機、照相 機或其它具有影像擷取功能之裝置。 接著，進行人臉辨識之步驟，此步驟係分為二部分，一 為人臉偵測方法（步驟210)與人臉比對方法（步驟212)。人臉 籲❹J方法（步驟210)係從人臉待測影像中取出人臉區域影像， 此人臉偵測方法係藉由將人臉待測影像與人臉樣本影像進行 人臉膚色區域比對，並按照頭形的比例，取出此人臉區域影 像。此人臉債測已為此技街領㉟中具有通常知識者所= - 知，故不另贅述。 …、 - 接著，將人臉區域影像與一部分人臉樣本影像透過人臉 比對方法（步驟212)來取得人臉特徵，並計算出人臉待測影像 之第二特徵值與-部分人臉樣本影像之第一特徵值。對第一 籲特徵值與數個第二特徵值進行特徵值距離運算，以求出數個 特徵距離，再自這些特徵距離中取最大值或者取其平均值， 來獲得一人臉特徵相似值。此人臉特徵相似值即人臉待測影 像與每個人臉樣本影像之相似程度，人臉特徵相似值愈小愈 相似#中，人臉比對方法可利用複合主成份分析演算法， 」而矛J用複合主成份分析演算法當成一種人臉比對方法， 僅用以舉例說明，本發明並不在此限。以下進—步說明本發 月較佳實施例所使用之複合主成份分析演算法。 假叹有一組資料包含了許多交互相關（Cr〇ss c〇rreUti〇n) ^25568[Embodiment] The following description can be referred to in order to make the description of the present invention more detailed and cooperate with the first! Diagram of the figure. The eye of the figure refers to the first-" 丨 丨 犯 犯 犯 您 ^ ^ ^ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ The input device is selected from the group consisting of an inductive card reader, a card reader, a magnetic card reader, a barcode card reader, a barcode scanner, and a personal secret input device. Next, proceed to step 202. Determining whether the verification data is correct, and generating - the first, the result. If the first first result is no, the user who refuses to verify the verification data is not stored in the data processing and storage system, and the user who rejects the verification data enters the access control system. (Step 224); if the first result is YES, the input device stores the verification data to the data processing and storage system (step 204), and the verification data corresponds to the field of the data processing and storage system, this column The number of digits of the personal face sample image. Then, the steps are performed, and the image capture device is activated by the data processing and the health 9 1325568 storage system. The image capture device will capture a face to be measured on the user's face. Image, and store the face image to be processed to the data processing. And the storage system ((4) 2G8), this step is to store the face image to be tested to verify the trap of the poor material, which makes the image picker It can be a camera, a camera or other device with image capturing function. Next, a step of face recognition is performed, which is divided into two parts, one is a face detection method (step 210) and a face matching method ( Step 212) The face-recalling method (step 210) extracts a face area image from the image to be tested of the face, and the face detection method performs the face by using the face-to-measure image and the face sample image. The skin color area is compared, and the face area image is taken out according to the ratio of the head shape. This face debt test has been known to those who have the usual knowledge in the skill street 35. Therefore, no further details are given. ..., - Then And the face region image and a part of the face sample image are obtained by the face matching method (step 212) to obtain the face feature, and the second feature value of the face image to be tested and the part of the face sample image are calculated. a feature value The eigenvalue distance operation is performed with a plurality of second eigenvalues to obtain a plurality of feature distances, and then a maximum value or an average value is taken from the feature distances to obtain a facial feature similarity value. That is, the similarity between the image to be tested and the image of each face sample, the smaller the similarity value of the face feature is, the more similar the face feature is, the face matching method can use the composite principal component analysis algorithm, and the spear J uses the composite principal component. The analysis algorithm is regarded as a face matching method, which is only used for exemplification, and the present invention is not limited thereto. The following is a step-by-step description of the composite principal component analysis algorithm used in the preferred embodiment of the present month. The data contains many interactions (Cr〇ss c〇rreUti〇n) ^25568
的變數，主成份分析演算法的中心觀念係在儘量保持這組資 料中之變異量（Variance)情況下，來縮減這組資料的維度。而 複合主成份分析演算法係將原始L*L矩陣的影像，分割成N 個較小的（L/N^(L/N)的區塊影像，利用更小的影像區塊，重 新組合成一個|χΛ^的新矩陣。如此一來，可以比習知的主成 刀析（PCΑ)運算時的计鼻量還少。雖然複合主成份分析做樣 =訓練的時間比主成份分析還慢，但是因為複合主成份分析 分割成較小的區塊影像，因此測試資料的時間比主成份分析 還决此外，有別於習知❸主成份分析，複合主成份分析在 處理影像區塊時，會取得影/像區塊之平均值，因&，可以降 低光線對人臉的影響，且效能比主成份分析更佳。 本發明能同時整合即時處理作業與後置處理作業，因此 可依照使用者需求，選擇即時處理作業或後置處理作業。若 選擇即時處理作業，則進行之步驟為步驟22〇與步驟224，若 選擇後置處理作業，則進行之步驟為步驟222、步驟226以及 步驟228。 接著’再進行動態臨界值比對 力mm對之步驟214,將人臉特徵相 似值與一臨界值做·一比較，类生丨齡· I 1 <1+ ψ 來判斷人臉特徵相似值是否大於 臨界值’並產生—第-{士 J® . 、'Ό果。此fe界值係由資料處理與儲存 系統中取一部分人臉樣本影傻， 像再计异彼此之特徵距離，並 由這些特徵距離中取其最夬估+、T ^ , 值或平均值而得之。若第二結果 為否’代表此待測人臉影傻盘A gx 冢/、人臉樣本影像相似，則進行步 驟220或步驟222。若第二妹果 、果為疋’則代表目前輸入驗證資 料之使用者有可能不是錄存於咨制_占 疋诚许於貧枓處理與儲存系統内之設定 11 1325568 / 者但也有可能為誤判》雖然動態臨界值可降低人臉辨識之 誤差’惟有些誤差可能將錯誤^人臉誤判成正確的人臉，因 此’本發明更採用多數決規則216來進行人臉驗證，即利用 y數服從多數之精髓來減少人臉辨識之誤差，於下文將詳細 描述多數決規則。 在步驟216中，進行多數決規則，將人臉待測影像與一 部分人臉樣本影像進行比較’而獲得數個多數決相似值。此 步驟更=斷這些多數決相似值是否大於上述之臨界值，並產 生第一結果。若第三結果為是，則將拒絕票數加一；若第三 結f為否，則將允許票數加一。接著，步驟218判斷允許票 數疋否大於拒絕票數，並產生第四結果。若第四結果為是， 則進行步驟220或步驟222;若第四結果為否，則進行步驟 224或步肆226。’亦即’若第四結果為是，且選擇即時處理作 業時’則允許使用者進人（步驟22G)，並更新人臉樣本影像（步 驟234)。右第四結果為是，但選擇後置處理作業時，則進行 步驟222之後置處理，並更新人臉樣本影像（步驟234)。 步驟234係將人臉待測影像取代資料處理 驗證這些驗證資料之人臉待測影像:後 本影像之其中一者……法可為先進先二: ut ’ FIFO)法，亦即將此人臉待測影像取 ”臉樣本影像，或為其他的取代方法，例如可將此= 測#像取代資料處理與儲存㈣中，人臉樣本影像與此人臉 4測影像差異最大的那—個人臉樣本影像。後置處理（步轉 22^批次處理’冑置處^之方式係使驗證資料之數量累積至 12 1325568 置處理方式.亦可裳3主 i H . 專待一特疋時間後，才驗證這些驗證資 人臉待測影像。 貝科之 右第四結果為^，且選擇料處理㈣時，則進行 224,立即拒絕驗證資料之使用者進入。 右第四結果為否，但選擇後置處理作業時，則進行步226 之再確$步驟°此㈣係由仲裁者判斷人臉待測影像是否盘 人臉樣本影像相同，祐吝笛 並產生第五結果，其中仲裁者可為監 此人臉驗證系統之一使用去，蔬士仙吞土主養 便用者藉由仲裁者手動判斷人臉待測 影像是否正確4第五結果為否，則於資料處理與儲存系統 中標。己使用者（步驟228)之驗證資料，並進行步驟出之後置 處理；若第五結果為是’則進行步驟㈣更新人臉樣本影像。 於步驟234更新人臉樣本影像之後，接著進行步驟236 :::練樣本’此步驟係重新計算資料處理與儲存系統中之 邛刀或全部之人臉樣本影像彼此間之特徵值距離，並從這 些特徵值距離中拙甚 士 ^古+ i f 匕選最大值或者平均值當成一新臨界值，並 於下次當此驗證資料輸入時’將此新臨界值取代步冑214中 之臨界值，此即為動態臨界值之方法。 由上述本發明較佳實施例可知，應用本發明具有下列優 點:首先’本發明可同時運用於即時處理與後置處理情況下。 ，、人本發明可利用動態臨界值比對，並可時常更新影像資 料庫，避免因時間的改變所造成人臉特徵的變化，== 判的隋形。再者，本發明利用複合主成份分析來減少樣本測 試時間、降低光線對人臉之料，進而降料算複雜度。此 外，本發明亦採用動態臨界值之方法，藉由時常更新的臨界 13 1325568 值’來驗證人臉，可v 乂減、因人臉之稍微變化而造成之誤判， 進而可以提高辨識度。本發明更利用多數決規則來辨識人臉 影像，使人臉驗證能進一步地防止被誤判。 臉 —’、、：發月已以—較佳實施例揭露如上，然其並非用以 限疋本發明’任何熱習此技藝者，在不脫離本發明之精神和 :圍内’當可作各種之更動與满飾’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】 ▲為讓本發明之上述和其他目的、特徵、和優點能更明顯 易懂’下文特舉-較佳實施例’並配合所附圖式，作詳細說 明如下： 、° 第1圖係繪示依照本發明之較佳實施例之人臉驗證方法 的一種流程圖。 【主要元件符號說明】 200 :輸入驗證資料 202:判斷驗證資料是否正確 2〇4 :儲存驗證資料 206 :啟動影像擷取器 208 :儲存人臉待測資料 210 :人臉偵測 2 12 ·•人臉比對 2 14 :動態臨界值比對 14 1325568 2 1 6 :多數決規則 218:允許票數是否大於拒絕票數 220 :允許使用者進入 222 :後置處理 224 :拒絕使用者進入 226 :再確認是否為正確使用者 228 :標記使用者 234 :更新人臉影像 236 :重新訓練樣本The central idea of the principal component analysis algorithm is to reduce the dimensions of this set of data while maintaining the Variance in this set of data. The composite principal component analysis algorithm divides the image of the original L*L matrix into N smaller (L/N^(L/N) block images, and recombines them into smaller image blocks. A new matrix of |χΛ^. In this way, the amount of the nose can be less than that of the conventional master analysis (PCΑ). Although the composite principal component analysis is done = the training time is slower than the principal component analysis. However, because the composite principal component analysis is divided into smaller block images, the time of testing the data is more than the principal component analysis. In addition, unlike the conventional principal component analysis, the composite principal component analysis is used to process the image block. The average value of the image/image block will be obtained, because &, the effect of light on the face can be reduced, and the performance is better than the principal component analysis. The invention can integrate both the instant processing operation and the post processing operation, so The user needs to select an instant processing job or a post processing job. If the instant processing job is selected, the steps are step 22 and step 224, and if the post processing job is selected, the steps are step 222, step 226, and Step 228. Then, 'the dynamic threshold value comparison force mm is paired with step 214, and the facial feature similarity value is compared with a critical value, and the class age is I 1 < 1+ ψ to judge the face Whether the feature similarity value is greater than the critical value' and produces - the first -{士J®., 'Ό果. This fe boundary value is taken from the data processing and storage system to take a part of the face sample silly, like re-counting each other's characteristics The distance is obtained by taking the most estimated value of +, T ^ , or the average of these feature distances. If the second result is no, it represents the face of the person to be tested, A gx 冢 /, face sample If the image is similar, proceed to step 220 or step 222. If the second girl's fruit is 疋', it means that the user who currently inputs the verification data may not be recorded in the consultation system. _ 疋 疋 许 枓 枓 枓 枓 枓 枓 枓 枓 枓The setting inside 11 1325568 / but it may also be a false positive" although the dynamic threshold can reduce the error of face recognition 'only some errors may mistake the face ^ wrong face into the correct face, so the invention uses the majority rule 216 to perform face verification, that is, use y number to obey the majority The essence is to reduce the error of face recognition, and the majority rule will be described in detail below. In step 216, the majority rule is used to compare the face image to be tested with a part of the face sample image to obtain a plurality of majority decisions. Similar value. This step is to determine whether the majority of the similarity values are greater than the above threshold value and produce the first result. If the third result is yes, the number of rejected votes is increased by one; if the third node f is no, then The number of allowed votes is increased by one. Then, step 218 determines whether the number of allowed tickets is greater than the number of rejected votes, and generates a fourth result. If the fourth result is yes, proceed to step 220 or step 222; if the fourth result is no, Then, step 224 or step 226 is performed. That is, if the fourth result is YES, and the immediate processing job is selected, the user is allowed to enter (step 22G), and the face sample image is updated (step 234). The fourth result in the right is YES, but when the post-processing job is selected, the post-processing is performed in step 222, and the face sample image is updated (step 234). Step 234 is to replace the face test image with the data processing to verify the face image of the verification data: one of the latter images... the method can be advanced second: ut 'FIFO method, and the face is also about to be The image to be tested is taken as a "face sample image, or other substitution method. For example, this = measurement # image replaces the data processing and storage (4), the face sample image and the face 4 image difference the most - personal face Sample image. Post-processing (step-to-step 22^ batch processing) is based on the method of accumulating the amount of verification data to 12 1325568. It can also be 3 main i H. After a special time The verification of the face of the verification person is to be tested. The fourth result of Beko's right is ^, and when the material is processed (4), 224 is performed, and the user who immediately rejects the verification data enters. The fourth result is no, but When the post-processing job is selected, the re-determination step of step 226 is performed. (4) The arbitrator judges whether the image to be tested on the face is the same as the face sample image, and the whistle flute produces a fifth result, wherein the arbitrator can To monitor one of the face verification systems If the user uses the arbitrator to manually determine whether the image to be tested is correct or not, the fifth result is no, then the data processing and storage system is marked. The user (step 228) Verify the data and perform the step-out processing; if the fifth result is yes, perform step (4) to update the face sample image. After updating the face sample image in step 234, proceed to step 236::: training sample 'this step Recalculate the characteristic value distance between the file of the file processing and storage system or all the face sample images, and from these eigenvalue distances, 拙士士古+ if 匕Select the maximum or average value as a new The critical value, and the next time this verification data is input, 'this new threshold value is substituted for the critical value in step 214, which is the method of dynamic critical value. From the above preferred embodiment of the present invention, the present invention is applied. The utility model has the following advantages: Firstly, the invention can be applied to both the instant processing and the post processing. The human invention can utilize the dynamic critical value comparison and can update the image database from time to time. Avoid the change of facial features caused by the change of time, == the shape of the judgment. Moreover, the invention uses the composite principal component analysis to reduce the sample test time, reduce the light to the face material, and then reduce the complexity of the calculation. In addition, the present invention also adopts a dynamic threshold method, which verifies the face by the frequently updated critical value of 13 1325568', which can reduce the degree of misjudgment caused by a slight change of the face, thereby improving the recognition degree. The invention further utilizes the majority rule to identify the face image, so that the face verification can further prevent the misjudgment. The face-',,: the month has been disclosed in the preferred embodiment, but it is not limited to The invention is not limited to the spirit of the invention and may be used as a variety of modifications and accessories. Therefore, the scope of the invention is defined by the scope of the appended claims. . BRIEF DESCRIPTION OF THE DRAWINGS The above and other objects, features, and advantages of the present invention will become more apparent <RTIgt; </ RTI> <RTIgt; 1 is a flow chart showing a face verification method in accordance with a preferred embodiment of the present invention. [Main component symbol description] 200: Input verification data 202: Determine whether the verification data is correct 2〇4: Store verification data 206: Start image capture device 208: Store face data to be tested 210: Face detection 2 12 ·• Face alignment 2 14 : Dynamic threshold comparison 14 1325568 2 1 6 : Majority rule 218: Whether the number of allowed votes is greater than the number of rejected votes 220: Allow the user to enter 222: Post processing 224: Reject user access 226: Reconfirm whether it is the correct user 228: Mark user 234: Update face image 236: Retrain sample
1515