TWI807851B - A feature disentanglement system, method and computer-readable medium thereof for domain generalized face anti-spoofing - Google Patents
A feature disentanglement system, method and computer-readable medium thereof for domain generalized face anti-spoofing Download PDFInfo
- Publication number
- TWI807851B TWI807851B TW111121276A TW111121276A TWI807851B TW I807851 B TWI807851 B TW I807851B TW 111121276 A TW111121276 A TW 111121276A TW 111121276 A TW111121276 A TW 111121276A TW I807851 B TWI807851 B TW I807851B
- Authority
- TW
- Taiwan
- Prior art keywords
- domain
- living body
- loss
- contour
- image data
- Prior art date
Links
Images
Abstract
Description
本發明關於一種人臉防偽的特徵解析技術,尤其是指一種領域泛化之人臉防偽的特徵解析系統、方法及電腦可讀媒介。 The present invention relates to a face anti-counterfeiting feature analysis technology, in particular to a field generalized face anti-counterfeiting feature analysis system, method and computer readable medium.
隨人臉辨識系統逐漸成熟,越來越多系統或裝置(如手機及大樓門禁)等皆以人臉辨識作為基礎進行驗證,但這也面臨新的安全性問題,例如:惡意攻擊者試圖透過印刷出之人臉或螢幕顯示出人臉來欺騙人臉辨識模型,故人臉防偽(face anti-spoofing)成為提升資訊安全的一項重要技術。 With the gradual maturity of the face recognition system, more and more systems or devices (such as mobile phones and building access control) are based on face recognition for verification, but this also faces new security issues. For example, malicious attackers try to deceive the face recognition model through printed faces or faces displayed on the screen. Therefore, face anti-spoofing has become an important technology to improve information security.
然而,既有的人臉防偽模型,常會因為領域的差異影響到辨識的效果,例如:某人臉防偽模型使用相機拍攝的人臉資料進行訓練,但在將模型應用於監視器影像的人臉辨識時,會因為資料的光線、拍攝角度、 設備的焦距或解析度等領域的差異,而影響到辨識的效果,此問題被稱為領域差異性。 However, existing face anti-counterfeiting models often affect the recognition effect due to differences in fields. For example, a face anti-counterfeiting model is trained using face data captured by a camera, but when the model is applied to face recognition on monitor images, it may be affected by the light of the data, shooting angle, Differences in fields such as the focal length or resolution of devices affect the recognition effect. This problem is called field differences.
因此,如何避免領域差異性之問題,並實現領域泛化之人臉防偽,以精準地達到領域泛化之人臉防偽的效果,遂成為業界亟待解決的課題。 Therefore, how to avoid the problem of domain differences, and realize domain generalized face anti-counterfeiting, so as to accurately achieve the effect of domain generalized face anti-counterfeiting, has become an urgent problem to be solved in the industry.
為解決前述習知的技術問題或提供相關之功效,本發明提供一種領域泛化之人臉防偽的特徵解析系統,係包括:一活體特徵抽取模組,係接收一欲判斷是否具有活體之人臉影像之影像資料,以從該影像資料中抽取出活體特徵;一活體分類器模組,係通訊連接該活體特徵抽取模組,以接收該影像資料之活體特徵,再由該活體分類器模組依據該影像資料之活體特徵計算出該影像資料之活體分類結果;以及一處理模組,係通訊連接該活體分類器模組,以接收該影像資料之活體分類結果,再由該處理模組依據該影像資料之活體分類結果判斷該影像資料中之人臉影像是否為活體。 In order to solve the aforementioned conventional technical problems or provide related effects, the present invention provides a feature analysis system for face anti-counterfeiting generalized in the field, which includes: a living body feature extraction module, which receives image data of a human face image to judge whether there is a living body, so as to extract the living body features from the image data; The living body classification result of the image data; and a processing module, which is communicatively connected to the living body classifier module to receive the living body classification result of the image data, and then the processing module judges whether the face image in the image data is a living body according to the living body classification result of the image data.
本發明復提供一種領域泛化之人臉防偽的特徵解析方法,係包括:由活體特徵抽取模組接收一欲判斷是否具有活體之人臉影像之影像資料,以從該影像資料中抽取出活體特徵;由活體分類器模組接收該影像資料之活體特徵後,依據該影像資料之活體特徵計算出該影像資料之活體分類結果;以及由處理模組接收該影像資料之活體分類結果後,依據該影像資料之活體分類結果判斷該影像資料中之人臉影像是否為活體。 The present invention further provides a feature analysis method for anti-counterfeiting of a generalized human face, which includes: a living body feature extraction module receives an image data of a human face image for judging whether there is a living body, so as to extract living body features from the image data; after receiving the living body features of the image data, the living body classifier module calculates the living body classification result of the image data according to the living body characteristics of the image data; Whether the face image in the image is live.
於一實施例中,該領域泛化之人臉防偽的特徵解析系統及方法更包括:一通訊連接該活體分類器模組之輪廓特徵抽取模組、一通訊連接該輪廓特徵抽取模組之臉部輪廓重建模組、一通訊連接該活體分類器模組之領域特徵抽取模組,以及一通訊連接該活體特徵抽取模組、該輪廓特徵抽取模組及該領域特徵抽取模組之領域分類器模組。 In one embodiment, the face anti-counterfeiting feature analysis system and method of generalization in the field further includes: a contour feature extraction module communicating with the living body classifier module, a face contour reconstruction module communicating with the contour feature extraction module, a domain feature extraction module communicating with the living body classifier module, and a domain classifier module communicating with the living body feature extraction module, the contour feature extraction module and the domain feature extraction module.
於一實施例中,該處理模組更計算出一活體分類損失、一活體混淆損失、一輪廓重建損失、一輪廓混淆損失、一領域分類損失及一領域混淆損失,以由該處理模組依據該活體分類損失、該活體混淆損失、該輪廓重建損失、該輪廓混淆損失、該領域分類損失及該領域混淆損失之其中至少一者更新該活體特徵抽取模組、該輪廓特徵抽取模組、該領域特徵抽取模組、該活體分類器模組、該臉部輪廓重建模組及該領域分類器模組。 In one embodiment, the processing module further calculates a living body classification loss, a living body confusion loss, a contour reconstruction loss, a contour confusion loss, a domain classification loss, and a domain confusion loss, so that the processing module can update the living body feature extraction module, the contour feature extraction module, the domain feature extraction module, and the living body classifier module according to at least one of the living body classification loss, the living body confusion loss, the contour reconstruction loss, the contour confusion loss, the domain classification loss, and the domain confusion loss , the face contour reconstruction group and the domain classifier module.
於一實施例中,由該活體特徵抽取模組接收至少一訓練影像資料,以從該訓練影像資料中抽取出該訓練影像資料之活體特徵,再由該活體分類器模組及該領域分類器模組依據該訓練影像資料之活體特徵分別計算出一第一活體分類結果及一第一領域分類結果,以令該處理模組依據該第一活體分類結果及該第一領域分類結果分別計算出該活體分類損失及該活體混淆損失。 In one embodiment, the living body feature extraction module receives at least one training image data to extract the living body features of the training image data from the training image data, and then the living body classifier module and the domain classifier module calculate a first living body classification result and a first domain classification result according to the living body features of the training image data, so that the processing module calculates the living body classification loss and the living body confusion loss according to the first living body classification result and the first domain classification result.
於一實施例中,由該輪廓特徵抽取模組接收至少一訓練影像資料,以從該訓練影像資料中抽取出人臉影像之輪廓特徵,再由該臉部輪廓重建模組依據該輪廓特徵重建一臉部輪廓,以令該處理模組依據經重建之該臉部輪廓以計算出該輪廓重建損失。 In one embodiment, the contour feature extraction module receives at least one training image data to extract the contour features of the face image from the training image data, and then the facial contour reconstruction group reconstructs a facial contour according to the contour features, so that the processing module calculates the contour reconstruction loss based on the reconstructed facial contour.
於一實施例中,由該活體分類器模組及該領域分類器模組依據該輪廓特徵分別計算出一第二活體分類結果及一第二領域分類結果,以令該處理模組依據該第二活體分類結果及該第二領域分類結果計算出該輪廓混淆損失。 In one embodiment, the living body classifier module and the domain classifier module respectively calculate a second living body classification result and a second domain classification result according to the contour feature, so that the processing module calculates the contour confusion loss according to the second living body classification result and the second domain classification result.
於一實施例中,由該領域特徵抽取模組接收至少一訓練影像資料,以從該訓練影像資料中抽取出領域特徵,再由該活體分類器模組及該領域分類器模組依據該領域特徵分別計算出一第三活體分類結果及一第三領域分類結果,以令該處理模組依據該第三領域分類結果及該第三活體分類結果分別計算出該領域分類損失及該領域混淆損失。 In one embodiment, the domain feature extraction module receives at least one training image data to extract domain features from the training image data, and then the living body classifier module and the domain classifier module respectively calculate a third living body classification result and a third domain classification result according to the domain features, so that the processing module calculates the domain classification loss and the domain confusion loss according to the third domain classification result and the third living body classification result respectively.
本發明又提供一種電腦可讀媒介,應用於具有處理器及/或記憶體的電腦或計算裝置中,該電腦或該計算裝置透過處理器及/或記憶體執行一目標程式及電腦可讀媒介,並用於執行電腦可讀媒介時執行如上所述之領域泛化之人臉防偽的特徵解析方法。 The present invention further provides a computer-readable medium, which is applied to a computer or computing device with a processor and/or memory, and the computer or the computing device executes a target program and the computer-readable medium through the processor and/or memory, and is used to perform the feature analysis method for face anti-counterfeiting as described above when the computer-readable medium is executed.
由上述可知,本發明之領域泛化之人臉防偽的特徵解析系統、方法及其電腦可讀媒介,藉由計算出的損失函數(即活體分類損失、活體混淆損失、輪廓重建損失、輪廓混淆損失、領域分類損失及領域混淆損失)對上述該些模組進行更新,並直至該些模組達到收斂,進而完成其神經網路之訓練。是以,本發明之領域泛化之人臉防偽的特徵解析系統經訓練後能精準地從一影像中抽取出領域泛化之人臉之活體特徵,以於對影像中的人臉進行活體判斷時,不會受到領域或環境因素的干擾,故本發明可應用於任意的領域中,以達到領域泛化(Domain Generalization)之效果。 As can be seen from the above, the domain generalized face anti-counterfeiting feature analysis system, method and computer-readable medium of the present invention update the above-mentioned modules through the calculated loss functions (i.e., living body classification loss, living body confusion loss, contour reconstruction loss, contour confusion loss, domain classification loss, and domain confusion loss), and complete the training of the neural network until the modules reach convergence. Therefore, the feature analysis system of the domain generalized face anti-counterfeiting of the present invention can accurately extract the living features of the domain generalized human face from an image after training, so that the liveness judgment of the human face in the image will not be disturbed by the domain or environmental factors, so the present invention can be applied to any field to achieve the effect of domain generalization.
1:領域泛化之人臉防偽的特徵解析系統 1: Feature analysis system for face anti-counterfeiting in domain generalization
11:活體特徵抽取模組 11: Living body feature extraction module
12:輪廓特徵抽取模組 12: Contour feature extraction module
13:領域特徵抽取模組 13: Domain feature extraction module
14:活體分類器模組 14: Living classifier module
15:臉部輪廓重建模組 15: Facial contour reconstruction group
16:領域分類器模組 16: Domain classifier module
17:處理模組 17: Processing module
S21A~S23A及S21B~S23B:步驟 S21A~S23A and S21B~S23B: steps
S31A~S33A及S31B~S33B:步驟 S31A~S33A and S31B~S33B: steps
S41A~S43A及S41B~S43B:步驟 S41A~S43A and S41B~S43B: steps
S51~S511:步驟 S51~S511: steps
圖1及圖1-1係為本發明之領域泛化之人臉防偽的特徵解析系統。 Fig. 1 and Fig. 1-1 are the feature analysis system of the generalized face anti-counterfeiting in the field of the present invention.
圖2A係為本發明之活體分類損失之計算流程示意圖。 FIG. 2A is a schematic diagram of the calculation flow of the living body classification loss in the present invention.
圖2B係為本發明之活體混淆損失之計算流程示意圖。 FIG. 2B is a schematic diagram of the calculation flow of the living body confusion loss of the present invention.
圖3A係為本發明之輪廓重建損失之計算流程示意圖。 FIG. 3A is a schematic diagram of the calculation flow of the contour reconstruction loss in the present invention.
圖3B係為本發明之輪廓混淆損失之計算流程示意圖。 FIG. 3B is a schematic diagram of the calculation flow of the contour confusion loss of the present invention.
圖4A係為本發明之領域分類損失之計算流程示意圖。 FIG. 4A is a schematic diagram of the calculation flow of the domain classification loss of the present invention.
圖4B係為本發明之領域混淆損失之計算流程示意圖。 FIG. 4B is a schematic diagram of the calculation flow of the domain confusion loss of the present invention.
圖5係為本發明之領域泛化之人臉防偽的特徵解析系統中之模組的神經網路之訓練方法流程示意圖。 Fig. 5 is a schematic flow chart of the training method of the neural network of the module in the feature analysis system of face anti-counterfeiting generalized in the field of the present invention.
以下藉由特定的具體實施例說明本發明之實施方式,熟悉此技藝之人士可由本說明書所揭示之內容輕易地瞭解本發明之其他優點及功效。 The implementation of the present invention is described below through specific specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification.
須知,本說明書所附圖式所繪示之結構、比例、大小等,均僅用以配合說明書所揭示之內容,以供熟悉此技藝之人士之瞭解與閱讀,並非用以限定本發明可實施之限定條件,故不具技術上之實質意義,任何結構之修飾、比例關係之改變或大小之調整,在不影響本發明所能產生之功效及所能達成之目的下,均應仍落在本發明所揭示之技術內容得能涵蓋之範圍內。同時,本說明書中所引用之如「一」、「第一」、「第二」、 「上」及「下」等之用語,亦僅為便於敘述之明瞭,而非用以限定本發明可實施之範圍,其相對關係之改變或調整,在無實質變更技術內容下,當視為本發明可實施之範疇。 It should be noted that the structures, proportions, and sizes shown in the drawings attached to this specification are only used to match the content disclosed in the specification for the understanding and reading of those who are familiar with this technology, and are not used to limit the conditions for the implementation of the present invention, so they have no technical significance. At the same time, references in this specification such as "a", "first", "second", Terms such as "above" and "below" are only used for clarity of description, and are not used to limit the applicable scope of the present invention. Changes or adjustments in their relative relationships are deemed to be within the applicable scope of the present invention if there is no substantial change in the technical content.
圖1係為本發明之領域泛化之人臉防偽的特徵解析系統1,包括:一活體特徵抽取模組11、一輪廓特徵抽取模組12、一領域特徵抽取模組13、一活體分類器模組14、一臉部輪廓重建模組15及一領域分類器模組16,其中,活體特徵抽取模組11、輪廓特徵抽取模組12、領域特徵抽取模組13、活體分類器模組14、臉部輪廓重建模組15及領域分類器模組16皆由人工神經網路(Artificial Neural Network,簡稱:ANN)(或稱:神經網路、類神經網路(Neural Network,簡稱:NN))所構建而成。在一實施例中,領域泛化之人臉防偽的特徵解析系統1更包括一通訊連接上述該些模組之處理模組17(如圖1-1所示),係用於計算上述該些模組之損失函數(Loss Function)。
Fig. 1 is the
在一實施例中,領域泛化之人臉防偽的特徵解析系統1係可建立於相同(或不同)伺服器(如通用型伺服器、檔案型伺服器、儲存單元型伺服器等)及電腦等具有適當演算機制之電子設備中,其中,領域泛化之人臉防偽的特徵解析系統1中之各個模組均可為軟體、硬體或韌體;若為硬體,則可為具有資料處理與運算能力之處理單元、處理器、電腦或伺服器;若為軟體或韌體,則可包括處理單元、處理器、電腦或伺服器可執行之指令,且可安裝於同一硬體裝置或分布於不同的複數硬體裝置。
In one embodiment, the face anti-counterfeiting
所述之活體特徵抽取模組11係為一個輸入維度為H×W×3,而輸出維度為D的神經網路,其中,H及W分別為領域泛化之人臉防偽的特徵解析系統1所接收之影像資料的高度及寬度,3代表彩色影像的三個顏色
通道,以及D為活體特徵的向量長度(或稱特徵長度)。於此實施例中,活體特徵抽取模組11所使用之神經網路係為扣除最後一層的ResNet18,且特徵長度D為512。具言之,領域泛化之人臉防偽的特徵解析系統1接收所輸入之一張影像資料(包含人臉影像),由活體特徵抽取模組11從複雜的影像資料中抽取出長度為D的活體特徵,以作為用來判斷該影像資料中之人臉是否為活體的資訊。
The living body
於本實施例中,為了使從該影像資料中抽取出的活體特徵不包含無關的領域資訊,而透過處理模組17計算出一活體分類損失及一活體混淆損失,以依據活體分類損失及活體混淆損失訓練活體特徵抽取模組11,其中,活體分類損失及活體混淆損失之計算方式如下所述:
In this embodiment, in order to make the living body features extracted from the image data not contain irrelevant domain information, a living body classification loss and a living body confusion loss are calculated through the
1.活體分類損失: 1. Live classification loss:
如圖2A所示,活體分類損失之計算流程包含: As shown in Figure 2A, the calculation process of the living body classification loss includes:
於步驟S21A中,輸入一訓練影像資料,且訓練影像資料包含人臉影像。 In step S21A, a training image data is input, and the training image data includes a face image.
於步驟S22A中,由活體特徵抽取模組11從訓練影像資料中抽取出活體特徵後,再由活體分類器模組14依據訓練影像資料之活體特徵計算出一第一活體分類結果。詳言之,活體分類器模組14為一種類神經網路,且輸入維度為D之活體特徵,以及輸出與活體標籤y i 相同維度之預測值。以活體分類器模組14為一層全連接層為例,活體分類器模組14所預測出之第一活體分類結果為θL×E L (x i,j )+bL,其中,θL及bL為活體分類器模組14之參數。
In step S22A, after the living body features are extracted from the training image data by the living body
於步驟S23A中,由處理模組17利用活體分類損失之損失函數(1)計算出第一活體分類結果與訓練影像資料之活體標籤之間的差距,以得到活體分類損失。詳言之,活體分類損失係為確保活體特徵中包含輸入影像中可供活體判斷的重要資訊,其中,活體分類損失之損失函數(1)之方程式例如下所示:
In step S23A, the
其中,x i,j 為輸入的訓練影像資料;E L 為活體特徵抽取模組11;C L 為活體分類器模組14;y i 為訓練影像資料之活體標籤;S為訓練時使用到的領域數量,在此實施例中為訓練影像資料集的數量;N為神經網路平行運算時,一次輸入的訓練影像資料數量(即batch size);以及L live 為活體分類損失,即活體特徵經由活體分類器模組14計算後所得到之第一活體分類結果與活體標籤之間的差距。
Among them, x i, j are the input training image data; E L is the living body
2.活體混淆損失: 2. Live body confusion loss:
如圖2B所示,活體混淆損失之計算流程包含: As shown in Figure 2B, the calculation process of the living body confusion loss includes:
於步驟S21B中,輸入一訓練影像資料。 In step S21B, a training image data is input.
於步驟S22B中,由活體特徵抽取模組11從訓練影像資料中抽取出活體特徵,再由領域分類器模組16依據訓練影像資料之活體特徵計算出一第一領域分類結果。詳言之,領域分類器模組16為一種類神經網路,且輸入維度為D之活體特徵,以及輸出為與領域標籤m i 相同維度之預測值。以領域分類器模組16為一層全連接層為例,領域分類器模組16所預測出的
第一領域分類結果為θD×E L (x i,j )+bD,其中,θD及bD為領域分類器模組16之參數。
In step S22B, the living body
於步驟S23B中,由處理模組17利用活體混淆損失之損失函數(2)計算出第一領域分類結果與均勻分佈之間的差距,以得到活體混淆損失。詳言之,活體混淆損失係為確保訓練影像資料之活體特徵中不包含訓練影像資料之領域資訊(如拍攝時的光線、使用設備的焦距與解析度等),故領域分類器模組16辨識的第一領域分類結果應接近均勻分佈,其中,活體混淆損失之損失函數(2)之方程式例如下所示:
In step S23B, the
其中,x i,j 為輸入的訓練影像資料;E L 為活體特徵抽取模組11;C D 為領域分類器模組16;S為訓練影像資料集的數量;N為一次輸入的訓練影像資料數量;以及為活體混淆損失,即活體特徵經由領域分類器模組16計算後所得到之第一領域分類結果與均勻分佈之間的差距。
Among them, x i, j is the input training image data; E L is the living body
在一實施例中,第一活體分類結果及第一領域分類結果可為0~1之間的機率值,或是以百分比呈現之數值,且於此不限活體分類結果之數值呈現方式。 In one embodiment, the first living body classification result and the first field classification result may be a probability value between 0 and 1, or a numerical value expressed as a percentage, and the numerical presentation of the living body classification result is not limited here.
所述之輪廓特徵抽取模組12亦為一個輸入維度為H×W×3,而輸出維度為D的神經網路。於此實施例中,活體特徵抽取模組11所使用之神經網路係為扣除最後一層的ResNet18,且特徵長度D為512。具言之,領域泛化之人臉防偽的特徵解析系統1接收所輸入之一張影像資料,由輪廓特徵
抽取模組12從複雜的影像資料中抽取出長度為D的輪廓特徵,以作為用來重建該影像資料中之臉部輪廓的資訊。
The contour
於本實施例中,為了使從該影像資料中抽取出的輪廓特徵不包含無關的活體資訊及領域資訊,而透過處理模組17計算出一輪廓重建損失及一輪廓混淆損失,其中,輪廓重建損失及輪廓混淆損失之計算方式如下所述:
In this embodiment, in order to prevent the contour features extracted from the image data from including irrelevant living body information and domain information, a contour reconstruction loss and a contour confusion loss are calculated through the
1.輪廓重建損失: 1. Contour reconstruction loss:
如圖3A所示,輪廓重建損失之計算流程包含: As shown in Figure 3A, the calculation process of contour reconstruction loss includes:
於步驟S31A中,輸入一訓練影像資料。 In step S31A, a training image data is input.
於步驟S32A中,由輪廓特徵抽取模組12從訓練影像資料中抽取出輪廓特徵,再由臉部輪廓重建模組15依據訓練影像資料之輪廓特徵重建出一臉部輪廓。
In step S32A, the contour
詳言之,臉部輪廓重建模組15為一種類神經網路,且輸入維度為D之輪廓特徵,而輸出的預測值之維度與預先訓練之臉部輪廓生成模組Φ所生成的輪廓Φ(x i,j )之輪廓維度相同。以臉部輪廓重建模組15為一層全連接層且輪廓維度為H×W為例,臉部輪廓重建模組15所重建出之臉部輪廓為θC×E c (x i,j )+bC,其中,θC及bC為領域分類器模組16之參數,θC之維度為H×W×D,bC之維度為H×W,故重建出之輪廓為大小H×W的預測值。
In detail, the facial
於步驟S33A中,由處理模組17利用輪廓重建損失之損失函數(3)計算出經重建之臉部輪廓與訓練影像資料之臉部輪廓之間的差距,以得到輪廓重建損失。詳言之,輪廓重建損失係為確保輪廓特徵中包含輸入
的訓練影像資料之中可供重建出臉部輪廓的重要資訊,其中,輪廓重建損失之損失函數(3)之方程式例如下所示:
In step S33A, the
其中,x i,j 為輸入的訓練影像資料;E C 為輪廓特徵抽取模組12;D C 為臉部輪廓重建模組15;Φ為預先訓練之臉部輪廓生成模組(圖中未示),以從輸入的訓練影像資料抽取出臉部輪廓,俾供與經重建之臉部輪廓計算出輪廓重建損失;S為訓練影像資料集的數量;N為一次輸入的訓練影像資料數量;以及L cont 為輪廓重建損失,即輪廓特徵經由臉部輪廓重建模組15重建後所得到之臉部輪廓與預先訓練之臉部輪廓生成模組所抽取之臉部輪廓之間的差距。
Among them, x i, j are the input training image data; E C is the contour
在一實施例中,預先訓練之臉部輪廓生成模組係為利用PRNet所建立的輸入維度為D,而輸出維度為H×W×3的神經網路。 In one embodiment, the pre-trained facial contour generation module is a neural network with an input dimension of D and an output dimension of H×W×3 established by PRNet.
2.輪廓混淆損失: 2. Contour confusion loss:
如圖3B所示,輪廓混淆損失之計算流程包含: As shown in Figure 3B, the calculation process of the contour confusion loss includes:
於步驟S31B中,輸入一訓練影像資料。 In step S31B, a training image data is input.
於步驟S32B中,由輪廓特徵抽取模組12從訓練影像資料中抽取出輪廓特徵後,再由活體分類器模組14及領域分類器模組16依據訓練影像資料之輪廓特徵分別計算出一第二活體分類結果及一第二領域分類結果。具言之,以活體分類器模組14及領域分類器模組16皆為一層全連接層為例,活體分類器模組14所預測出之第二活體分類結果為θL×E C (x i,j )+bL,而領域分類器模組16所預測出的第二領域分類結果為θD×E C (x i,j )+
bD,其中,θL及bL為活體分類器模組14之參數,而θD及bD為領域分類器模組16之參數。
In step S32B, after the contour
於步驟S33B中,由處理模組17利用輪廓混淆損失之損失函數(4)計算出第二活體分類結果、第二領域分類結果與均勻分佈之間的差距,以得到輪廓混淆損失。詳言之,輪廓混淆損失係為確保輪廓特徵中不包含輸入影像中的活體資訊或領域資訊,故第二活體分類結果及第二領域分類結果皆應接近均勻分佈,其中,輪廓混淆損失之損失函數(4)之方程式例如下所示:
In step S33B, the
其中,x i,j 為輸入的訓練影像資料;E C 為輪廓特徵抽取模組12;C L 為活體分類器模組14;C D 為領域分類器模組16;S為訓練影像資料集的數量;N為一次輸入的訓練影像資料數量;以及為輪廓混淆損失,即輪廓特徵經由活體分類器模組14及領域分類器模組16計算後所得到之第二活體分類結果及第二領域分類結果相較於均勻分佈之間的差距之總和。
Wherein, x i, j is the input training image data; E C is the contour
在一實施例中,第二活體分類結果及第二領域分類結果可為0~1之間的機率值,或是以百分比呈現之數值,且於此不限活體分類結果之數值呈現方式。 In one embodiment, the second living body classification result and the second field classification result may be a probability value between 0 and 1, or a numerical value expressed as a percentage, and the numerical presentation of the living body classification result is not limited here.
所述之領域特徵抽取模組13亦為一個輸入維度為H×W×3,而輸出維度為D的神經網路。於此實施例中,領域特徵抽取模組13所使用之神經網路係為扣除最後一層的ResNet18,且特徵長度D為512。具言之,領域泛化之人臉防偽的特徵解析系統1接收所輸入之一張影像資料,由領域特徵抽
取模組13從複雜的影像資料中抽取出長度為D的領域特徵,以作為用來分辨該影像資料中之領域資訊(如拍攝時的光線、使用設備的焦距與解析度等)。
The domain
於本實施例中,為了使從該影像資料中抽取出的領域特徵不包含無關的活體資訊,而透過處理模組17計算出一領域分類損失及一領域混淆損失,其中,領域分類損失及領域混淆損失之計算方式如下所述:
In this embodiment, in order to prevent the domain features extracted from the image data from including irrelevant living body information, a domain classification loss and a domain confusion loss are calculated through the
1.領域分類損失: 1. Domain classification loss:
如圖4A所示,領域分類損失之計算流程包含: As shown in Figure 4A, the calculation process of domain classification loss includes:
於步驟S41A中,輸入一訓練影像資料。 In step S41A, a training image data is input.
於步驟S42A中,由領域特徵抽取模組13從訓練影像資料中抽取出領域特徵後,再由領域分類器模組16依據訓練影像資料之領域特徵計算出一第三領域分類結果。具言之,以領域分類器模組16為一層全連接層為例,領域分類器模組16所預測出的第三領域分類結果為θD×E D (x i,j )+bD,其中,θD及bD為領域分類器模組16之參數。
In step S42A, after domain features are extracted from the training image data by the domain
於步驟S43A中,由處理模組17利用領域分類損失之損失函數(5)計算出第三領域分類結果與訓練影像資料之領域標籤之間的差距,以得到領域分類損失。詳言之,領域分類損失係為確保領特徵中包含輸入影像中可供領域判斷的重要資訊,其中,領域分類損失之損失函數(5)之方程式例如下所示:
In step S43A, the
其中,x i,j 為輸入的訓練影像資料;E D 為領域特徵抽取模組13;C D 為領域分類器模組16;m i 為訓練影像資料之領域標籤;S為訓練影像資料集的數
量;N為一次輸入的訓練影像資料數量;以及L dom 為領域分類損失,即領域特徵經由領域分類器模組16計算後所得到之第三領域分類結果與領域標籤之間的差距。
Among them , x i,j is the input training image data; E D is the domain
2.領域混淆損失: 2. Domain confusion loss:
如圖4B所示,領域混淆損失之計算流程包含: As shown in Figure 4B, the calculation process of domain confusion loss includes:
於步驟S41B中,輸入一訓練影像資料。 In step S41B, a training image data is input.
於步驟S42B中,由領域特徵抽取模組13從訓練影像資料中抽取出領域特徵,再由活體分類器模組14依據訓練影像資料之領域特徵計算出一第三活體分類結果。具言之,以活體分類器模組14為一層全連接層為例,活體分類器模組14所預測出的第三活體分類結果為θL×E D (x i,j )+bL,其中,θL及bL為活體分類器模組14之參數。
In step S42B, the domain
於步驟S43B中,由處理模組17利用領域混淆損失之損失函數(6)計算出第三活體分類結果與均勻分佈之間的差距,以得到領域混淆損失。詳言之,領域混淆損失係為確保訓練影像資料之領域特徵中不包含訓練影像資料之活體資訊,故活體分類器模組14辨識的第三活體分類結果應接近均勻分佈,其中,領域混淆損失之損失函數(6)之方程式例如下所示:
In step S43B, the
其中,x i,j 為輸入的訓練影像資料;E D 為領域特徵抽取模組13;C L 為活體分類器模組14;S為訓練影像資料集的數量;N為一次輸入的訓練影像資料數量;以及為領域混淆損失,即領域特徵經由活體分類器模組14計算後所得到之第三活體分類結果與均勻分佈之間的差距。
Among them , x i, j is the input training image data; E D is the domain
在一實施例中,第三領域分類結果及第三活體分類結果可為0~1之間的機率值,或是以百分比呈現之數值,且於此不限活體分類結果之數值呈現方式。 In one embodiment, the third domain classification result and the third living body classification result may be a probability value between 0 and 1, or a numerical value expressed as a percentage, and the numerical presentation method of the living body classification result is not limited here.
於另一實施例中,以於藉由輸入的訓練影像資料而得到各項損失函數(1)~(6)之損失(即活體分類損失、活體混淆損失、輪廓重建損失、輪廓混淆損失、領域分類損失及領域混淆損失)後,由處理模組17利用類神經網路的反向傳播(Backpropagation,BP),並結合梯度下降法(Gradient descent)計算出各項損失函數(1)~(6)之梯度,以藉由各個損失函數(1)~(6)之梯度回饋最佳化方法,俾對活體特徵抽取模組11、輪廓特徵抽取模組12、領域特徵抽取模組13、活體分類器模組14、臉部輪廓重建模組15及領域分類器模組16中之權重進行更新,進而優化各項損失函數(1)~(6),亦即得到更小的損失函數。再者,各項損失函數(1)~(6)的權重可根據訓練影像資料及神經網路之設計進行調整。
In another embodiment, after obtaining the losses of various loss functions (1)-(6) (i.e., living body classification loss, living body confusion loss, contour reconstruction loss, contour confusion loss, domain classification loss, and domain confusion loss) through the input training image data, the
詳言之,於活體特徵抽取模組11、輪廓特徵抽取模組12、領域特徵抽取模組13、活體分類器模組14、臉部輪廓重建模組15及領域分類器模組16進行更新時,所分別使用到的損失函數如下所述:
In detail, when the living body
1.活體特徵抽取模組11:活體分類損失及活體混淆損失 1. Living body feature extraction module 11: living body classification loss and living body confusion loss
2.活體分類器模組14:活體分類損失 2. Live Classifier Module 14: Live Classification Loss
3.輪廓特徵抽取模組12:輪廓重建損失及輪廓混淆損失 3. Contour Feature Extraction Module 12: Contour Reconstruction Loss and Contour Confusion Loss
4.臉部輪廓重建模組15:輪廓重建損失 4. Facial Contour Remodeling Group 15: Contour Reconstruction Loss
5.領域特徵抽取模組13:領域分類損失及領域混淆損失 5. Domain Feature Extraction Module 13: Domain Classification Loss and Domain Confusion Loss
6.領域分類器模組16:領域分類損失 6. Domain Classifier Module 16: Domain Classification Loss
接著,持續輸入複數訓練影像資料至活體特徵抽取模組11、輪廓特徵抽取模組12、領域特徵抽取模組13、活體分類器模組14、臉部輪廓重建模組15及領域分類器模組16中,以對該些模組進行如上述實施例所述之訓練過程,直至各項損失函數(1)~(6)達到收斂,藉此完成活體特徵抽取模組11、輪廓特徵抽取模組12、領域特徵抽取模組13、活體分類器模組14、臉部輪廓重建模組15及領域分類器模組16之神經網路的訓練。另一方面,通常以觀察損失函數下降的趨勢做為訓練時模型收斂與否的參考。
Next, continuously input multiple training image data into the living body
在一實施例中,判斷活體特徵抽取模組11、輪廓特徵抽取模組12、領域特徵抽取模組13、活體分類器模組14、臉部輪廓重建模組15及領域分類器模組16之神經網路的訓練是否達到收斂,以由各項損失函數(1)~(6)的總和大小來進行判斷。具言之,由於實作時各項損失函數(1)~(6)之大小會來回震盪,故需要將各項損失函數(1)~(6)隨更新次數的變化以折線圖表示,並將折線做平滑化,由處理模組17計算各項損失函數(1)~(6)在一定的更新次數內(如:500次)的下降數值是否小於一門檻值,以於下降數值小於一門檻值時判斷該些模組之神經網路的訓練達到收斂,而此門檻值可依據訓練準確度的需求設定之,故於此不限門檻值。
In one embodiment, it is judged whether the neural network training of the living body
在一實施例中,亦可由處理模組17從臉部輪廓重建模組15所重建之臉部輪廓的品質、活體分類器模組14或/及領域分類器模組16所得到之分類結果的正確率,來判斷活體特徵抽取模組11、輪廓特徵抽取模組12、領域特徵抽取模組13、活體分類器模組14、臉部輪廓重建模組15及領域分類器模組16之神經網路的訓練是否完成。
In one embodiment, the quality of the facial contour reconstructed by the
圖5係為本發明之領域泛化之人臉防偽的特徵解析系統中之模組的神經網路之訓練方法流程示意圖,且一併參閱圖1~圖4-2說明之。此外,本實施例與上述實施例相同處不再贅述,且此訓練方法流程包含以下步驟S51至步驟S511: FIG. 5 is a schematic flow chart of the neural network training method of the modules in the face anti-counterfeiting feature analysis system generalized in the field of the present invention, and it is explained with reference to FIGS. 1 to 4-2. In addition, the similarities between this embodiment and the above embodiment will not be repeated, and the training method flow includes the following steps S51 to S511:
於步驟S51中,輸入一訓練影像資料,其包含人臉影像,以及人臉影像以外的領域資訊(如拍攝時的光線、使用設備的焦距與解析度等)。 In step S51, a training image data is input, which includes a face image and domain information other than the face image (such as the light when shooting, the focal length and resolution of the equipment used, etc.).
於步驟S52中,由活體特徵抽取模組11從訓練影像資料中抽取出有關於人臉影像之活體特徵。
In step S52, the living body
於步驟S53中,由活體分類器模組14依據活體特徵計算出一第一活體分類結果,進而計算出第一活體分類結果與訓練影像資料之活體標籤之間的差距,以得到活體分類損失。
In step S53, the living
於步驟S54中,由領域分類器模組16依據訓練影像資料之活體特徵計算出一第一領域分類結果,進而計算出第一領域分類結果與均勻分佈之間的差距,以得到活體混淆損失。
In step S54, the
於步驟S55中,由輪廓特徵抽取模組12從訓練影像資料中抽取出有關於人臉影像之輪廓特徵。
In step S55, the contour
於步驟S56中,由臉部輪廓重建模組15依據訓練影像資料之輪廓特徵重建一臉部輪廓,進而計算出經重建之臉部輪廓與訓練影像資料之臉部輪廓之間的差距,以得到輪廓重建損失。
In step S56, the facial
於步驟S57中,由活體分類器模組14及領域分類器模組16依據訓練影像資料之輪廓特徵分別計算出一第二活體分類結果及一第二領域
分類結果,進而計算出第二活體分類結果、第二領域分類結果與均勻分佈之間的差距,以得到輪廓混淆損失。
In step S57, the living
於步驟S58中,由領域特徵抽取模組13從複雜的影像資料中抽取出領域特徵。
In step S58, the domain
於步驟S59中,由領域分類器模組16依據訓練影像資料之領域特徵計算出一第三領域分類結果,進而計算出第三領域分類結果與訓練影像資料之領域標籤之間的差距,以得到領域分類損失。
In step S59, the
於步驟S510中,由活體分類器模組14依據訓練影像資料之領域特徵計算出一第三活體分類結果,進而計算出第三活體分類結果與均勻分佈之間的差距,以得到領域混淆損失。
In step S510 , the living
於步驟S511中,以依據活體分類損失、活體混淆損失、輪廓重建損失、輪廓混淆損失、領域分類損失及領域混淆損失其中至少一者對活體特徵抽取模組11、輪廓特徵抽取模組12、領域特徵抽取模組13、活體分類器模組14、臉部輪廓重建模組15及領域分類器模組16中之權重進行更新,直至上述該些模組達到收斂。
In step S511, the weights in the living body
此外,本發明還揭示一種電腦可讀媒介,係應用於具有處理器(例如,CPU、GPU等)及/或記憶體的計算裝置或電腦中,且儲存有指令,並可利用此計算裝置或電腦透過處理器及/或記憶體執行此電腦可讀媒介,以於執行此電腦可讀媒介時執行上述之方法及各步驟。 In addition, the present invention also discloses a computer-readable medium, which is applied to a computing device or computer having a processor (for example, CPU, GPU, etc.) and/or memory, and stores instructions, and the computing device or computer can be used to execute the computer-readable medium through the processor and/or memory, so as to execute the above-mentioned method and steps when executing the computer-readable medium.
下列係為本發明之本發明之領域泛化之人臉防偽的特徵解析系統1應用實施例,且一併參閱圖1說明之。
The following is an application embodiment of the face anti-counterfeiting
於本實施例中,本發明之領域泛化之人臉防偽的特徵解析系統1中之模組已經上述實施例之訓練完成,以於接收到一欲判斷是否為活體的影像資料後,由活體特徵抽取模組11從影像資料抽取出活體特徵,且因經訓練後領域泛化之人臉防偽的特徵解析系統1能將與活體無關的輪廓特徵與領域特徵抽離,故活體特徵不會受到人臉輪廓及領域差異的影響。
In this embodiment, the modules in the domain generalized face anti-counterfeiting
再者,活體分類器模組14接收由活體特徵抽取模組11所抽取之活體特徵,以計算出一活體分類結果,即可藉由活體分類結果判斷該影像是否為活體。具言之,可以取一閾值,若活體分類結果高於該閾值,則處理模組17判定影像資料為異常,即影像資料中不具備活體;若低於該閾值,則理模組判定影像資料為正常,即影像資料中具備活體。
Furthermore, the living
在一實施例中,活體分類結果及閾值可為0~1之間的機率值,或是以百分比呈現之數值,且閾值可依據需求設定之,故於此不限活體分類結果及閾值之數值呈現方式,以及閾值之大小。 In one embodiment, the living body classification result and the threshold value can be a probability value between 0 and 1, or a value expressed as a percentage, and the threshold value can be set according to requirements, so there is no limitation on the numerical presentation method of the living body classification result and the threshold value, and the size of the threshold value.
綜上所述,本發明之領域泛化之人臉防偽的特徵解析系統、方法及其電腦可讀媒介,藉由計算出的損失函數(即活體分類損失、活體混淆損失、輪廓重建損失、輪廓混淆損失、領域分類損失及領域混淆損失)對活體特徵抽取模組、輪廓特徵抽取模組、領域特徵抽取模組、活體分類器模組、臉部輪廓重建模組及領域分類器模組等模組進行更新,並直至前述該些模組達到收斂,進而完成該些模組之神經網路之訓練。因此,本發明之領域泛化之人臉防偽的特徵解析系統經訓練後能精準地從一影像中抽取出領域泛化之人臉之活體特徵,以於對影像中的人臉進行活體判斷時,不 會受到領域或環境因素的干擾,故本發明可應用於任意的領域中,以達到領域泛化之人臉防偽之效果。 To sum up, the feature analysis system, method and computer-readable medium of the domain-generalized face anti-counterfeiting of the present invention update modules such as the living body feature extraction module, the contour feature extraction module, the domain feature extraction module, the living body classifier module, the facial contour reconstruction group, and the domain classifier module through the calculated loss functions (i.e., the living body classification loss, the living body confusion loss, the contour reconstruction loss, the contour confusion loss, the domain classification loss, and the domain confusion loss), until the aforementioned modules reach convergence, and then Complete the training of the neural network of these modules. Therefore, the generalized human face anti-counterfeiting feature analysis system of the present invention can accurately extract the living body features of the human face with generalized domain from an image after training, so that when the human face in the image is judged live, it will not It will be interfered by field or environmental factors, so the present invention can be applied in any field to achieve the effect of field generalized face anti-counterfeiting.
再者,相較於先前技術利用單一特徵抽取器,以抽取並判斷人臉的活體特徵,導致其應用受限於訓練時所使用的領域,以及單一特徵抽取器所抽取之活體特徵更包含無關的人臉輪廓特徵及領域特徵(如拍攝時的光線、使用設備的焦距與解析度等),進而影響到人臉防偽的判斷效果。 然而,本發明分別使用輪廓特徵抽取模組及領域特徵抽取模組,能將與活體特徵無關的人臉輪廓資訊及領域特徵抽離,藉此讓活體特徵不受人臉輪廓差異與領域差異的干擾,能夠專注於人臉的活體特徵判讀。 Furthermore, compared to the prior art using a single feature extractor to extract and judge live features of the face, its application is limited to the field used in training, and the live features extracted by the single feature extractor include irrelevant face contour features and domain features (such as the light when shooting, the focal length and resolution of the equipment used, etc.), which in turn affects the judging effect of face anti-counterfeiting. However, the present invention uses the contour feature extraction module and the domain feature extraction module respectively, which can separate the facial contour information and domain features that are not related to the living body features, so that the living body features will not be interfered by the differences in the face contour and domain differences, and can focus on the interpretation of the living body features of the human face.
是以,本發明能應用於不同的領域中,以進行人臉的活體特徵判斷,藉此避免有心人士利用非活體之人臉照片或影片進行驗證,進而達到領域泛化之人臉防偽之目的。 Therefore, the present invention can be applied in different fields to judge the living body features of human faces, so as to prevent interested people from using non-living human face photos or videos for verification, and then achieve the purpose of face anti-counterfeiting in generalized fields.
上述實施形態僅例示性說明本發明之原理及其功效,而非用於限制本發明。任何熟習此項技藝之人士均可在不違背本發明之精神及範疇下,對上述實施形態進行修飾與改變。因此,本發明之權利保護範圍應如申請專利範圍所列。 The above-mentioned embodiments are only illustrative to illustrate the principles and effects of the present invention, and are not intended to limit the present invention. Anyone skilled in the art can modify and change the above-mentioned embodiments without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the rights of the present invention should be listed in the scope of the patent application.
1:領域泛化之人臉防偽的特徵解析系統 1: Feature analysis system for face anti-counterfeiting in domain generalization
11:活體特徵抽取模組 11: Living body feature extraction module
12:輪廓特徵抽取模組 12: Contour feature extraction module
13:領域特徵抽取模組 13: Domain feature extraction module
14:活體分類器模組 14: Living classifier module
15:臉部輪廓重建模組 15: Facial contour reconstruction group
16:領域分類器模組 16: Domain classifier module
Claims (13)
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
TW111121276A TWI807851B (en) | 2022-06-08 | 2022-06-08 | A feature disentanglement system, method and computer-readable medium thereof for domain generalized face anti-spoofing |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
TW111121276A TWI807851B (en) | 2022-06-08 | 2022-06-08 | A feature disentanglement system, method and computer-readable medium thereof for domain generalized face anti-spoofing |
Publications (2)
Publication Number | Publication Date |
---|---|
TWI807851B true TWI807851B (en) | 2023-07-01 |
TW202349261A TW202349261A (en) | 2023-12-16 |
Family
ID=88149292
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
TW111121276A TWI807851B (en) | 2022-06-08 | 2022-06-08 | A feature disentanglement system, method and computer-readable medium thereof for domain generalized face anti-spoofing |
Country Status (1)
Country | Link |
---|---|
TW (1) | TWI807851B (en) |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
TW202026948A (en) * | 2018-12-29 | 2020-07-16 | 大陸商北京市商湯科技開發有限公司 | Methods and devices for biological testing and storage medium thereof |
CN112287765A (en) * | 2020-09-30 | 2021-01-29 | 新大陆数字技术股份有限公司 | Face living body detection method, device and equipment and readable storage medium |
US20210200995A1 (en) * | 2017-03-16 | 2021-07-01 | Beijing Sensetime Technology Development Co., Ltd | Face anti-counterfeiting detection methods and systems, electronic devices, programs and media |
CN113255511A (en) * | 2021-05-21 | 2021-08-13 | 北京百度网讯科技有限公司 | Method, apparatus, device and storage medium for living body identification |
-
2022
- 2022-06-08 TW TW111121276A patent/TWI807851B/en active
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20210200995A1 (en) * | 2017-03-16 | 2021-07-01 | Beijing Sensetime Technology Development Co., Ltd | Face anti-counterfeiting detection methods and systems, electronic devices, programs and media |
TW202026948A (en) * | 2018-12-29 | 2020-07-16 | 大陸商北京市商湯科技開發有限公司 | Methods and devices for biological testing and storage medium thereof |
CN112287765A (en) * | 2020-09-30 | 2021-01-29 | 新大陆数字技术股份有限公司 | Face living body detection method, device and equipment and readable storage medium |
CN113255511A (en) * | 2021-05-21 | 2021-08-13 | 北京百度网讯科技有限公司 | Method, apparatus, device and storage medium for living body identification |
Also Published As
Publication number | Publication date |
---|---|
TW202349261A (en) | 2023-12-16 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
TWI686774B (en) | Human face live detection method and device | |
US11106920B2 (en) | People flow estimation device, display control device, people flow estimation method, and recording medium | |
WO2019109743A1 (en) | Url attack detection method and apparatus, and electronic device | |
WO2021232985A1 (en) | Facial recognition method and apparatus, computer device, and storage medium | |
US9633044B2 (en) | Apparatus and method for recognizing image, and method for generating morphable face images from original image | |
JP5899472B2 (en) | Person attribute estimation system and learning data generation apparatus | |
JP6678246B2 (en) | Semantic segmentation based on global optimization | |
CN105225222B (en) | Automatic assessment of perceptual visual quality of different image sets | |
Li et al. | Face spoofing detection with image quality regression | |
WO2019056503A1 (en) | Store monitoring evaluation method, device and storage medium | |
WO2019200702A1 (en) | Descreening system training method and apparatus, descreening method and apparatus, device, and medium | |
WO2016172923A1 (en) | Video detection method, video detection system, and computer program product | |
WO2018078857A1 (en) | Line-of-sight estimation device, line-of-sight estimation method, and program recording medium | |
WO2022078168A1 (en) | Identity verification method and apparatus based on artificial intelligence, and computer device and storage medium | |
CN107316029A (en) | A kind of live body verification method and equipment | |
Parde et al. | Face and image representation in deep CNN features | |
CN114925748A (en) | Model training and modal information prediction method, related device, equipment and medium | |
CN111680544B (en) | Face recognition method, device, system, equipment and medium | |
CN113139462A (en) | Unsupervised face image quality evaluation method, electronic device and storage medium | |
US20210150238A1 (en) | Methods and systems for evaluatng a face recognition system using a face mountable device | |
WO2021042544A1 (en) | Facial verification method and apparatus based on mesh removal model, and computer device and storage medium | |
Parde et al. | Deep convolutional neural network features and the original image | |
CN111382791A (en) | Deep learning task processing method, image recognition task processing method and device | |
TWI807851B (en) | A feature disentanglement system, method and computer-readable medium thereof for domain generalized face anti-spoofing | |
CN110147740B (en) | Face recognition method, device, equipment and storage medium |