KR20130035849A - Single image-based fake face detection - Google Patents

Abstract

PURPOSE: A forgery face detection method using a single image is provided to be effectively applied to a face recognition system and an ATM(Automated Teller Machine) system through only addition of software, thereby providing a high ripple effect in comparison with methods of needing additional hardware. CONSTITUTION: A color or black-and-white user face image is obtained. Each pixel of the face image is converted into a frequency domain through Fourier transform. A location of a low frequency component is converted using a result value with Fourier transform in order to be in the center. One dimensional feature vector is generated by calculating an average energy value of a concentric circle. Forgery of the face image is determined by comparing one dimensional feature vector with a reference forgery face feature vector based on a support vector machine.

Description

Fake face detection method using single image {Single image-based fake face detection}

The present invention relates to a technology for determining the authenticity of a face image obtained from a camera attached to a face recognizer and an ATM. More specifically, frequency analysis, local binary pattern, and a combination of these methods are applied to face images acquired from commonly used color or black and white cameras without any special sensors or lighting. The present invention relates to a method for determining whether or not obtained from a fake face.

Due to the recent demand for safety, efforts are being actively made to increase the security level by using technologies such as face recognition and face masking in access control systems and cash dispenser systems. Accordingly, in order to hide their identity, criminals attempt to attack the security systems as described above by presenting various kinds of fake faces as shown in FIG.

Researches for detecting such attacks using fake faces can be divided into coercive and noncoercive methods. The coercive method requires a user's cooperation, such as movement of the lips or head rotation, and can cope with an attack using a two-dimensional mask, a video, and a three-dimensional mask. This method has the disadvantage that the system can be highly reluctant because it requires the cooperation of the user, but the advantage of the system is high accuracy. On the contrary, the non-coercive method does not require the cooperation of the user, so the user's rejection is less, but the accuracy of the system may not be higher than that of the coercive method.

Non-coercive methods detect masks by detecting three-dimensional information or eye blinks of faces based on multiple images, and detect masks by decomposing images into components or analyzing them in the frequency domain based on a single image. Can be divided. The single image-based approach is very useful because it can be applied to both a single image and a video without requiring a separate sensor and lighting.

The occlusion determination system used in the existing face recognition system or cash dispenser checks whether the face image acquired through the image acquisition device is the same as the face of the user who is stored or whether there is severe occlusion on the face. However, these systems have a disadvantage in that even when a fake face is input, if the input fake face image is the same as the stored user or if there is no severe obstruction on the face, the system recognizes the face as a normal user and allows financial transactions. Therefore, it is necessary to study a fake face detection method that can protect against attacks using such fake faces.

In addition, in order to easily apply a fake face detection method to an existing face recognition system or cash machine operation situation, a separate sensor such as a thermal infrared camera other than a general color or black and white camera or an infrared light, etc. should not be required. It should be able to operate even when the video is not secured.

The present invention proposes a method of determining whether or not a fake face using one or more face images obtained from a general color camera or a black and white camera so as to be immediately applicable to an existing face recognition system and an ATM machine. The proposal is composed of three main methods.

First, a fake face is a method of detecting a fake face by the difference of the entire frequency band components of the acquired face image by using the characteristic that the detail and the three-dimensional feeling are reduced compared to the real face.

Second, since the fake face has different characteristics in terms of the material and the actual face, the difference in the local binary pattern, which is easy for analyzing the texture of the image, is used by using the characteristic that the texture is different in the texture. It is a method to detect fake face using.

Third, a fake face is detected by combining the aforementioned feature extraction method using the frequency component and the local binary pattern in a feature level, score level, or decision level method.

The above-described method can be implemented as follows.

First, a fake face detection method using a single image according to an embodiment includes: obtaining a user face image in color or black and white; Fourier transforming each pixel of the face image into a frequency domain; Using the Fourier transformed result to position transform the low frequency component to be centered; Partitioning the position-converted result into a plurality of concentric circles and calculating an average energy value of each concentric circle to generate a one-dimensional feature vector; And comparing the calculated one-dimensional feature vector with a reference fake face feature vector using a support vector machine to determine whether the user face image is a fake face.

The reference fake face feature vector may further include preparing one or more fake face images; Generating each one-dimensional feature vector from the fake face images; The generated one-dimensional feature vectors are generated through a step of learning as a reference forged face feature vector in the support vector machine.

Further, the step of Fourier transforming each pixel of the face image into a frequency domain may include: dividing the face image into a plurality of pieces; Fourier transforming each pixel for each of the pieces.

In addition, a fake face detection method using a single image according to another embodiment of the present invention, the method comprising the steps of obtaining a user face image of color or black and white; Calculating a local binary pattern code from the face image; Generating a one-dimensional feature vector by histogramming the calculated local binary pattern code; And comparing the calculated one-dimensional feature vector with a reference fake face feature vector using a support vector machine to determine whether the user face image is a fake face.

The reference fake face feature vector may further include preparing one or more fake face images; Generating each one-dimensional feature vector from the fake face images; The generated one-dimensional feature vectors are generated through a step of learning as a reference forged face feature vector in the support vector machine.

The calculating of the local binary pattern code from the face image may include: dividing the face image into a plurality of pieces; Calculating a local binary pattern code for each of the pieces.

The method proposed in the present invention extracts various differences in the image acquired from the real face and the image acquired from the fake face in a single image, and determines whether the face image is forged through them. There is no need for separate sensors such as infrared cameras and separate lights such as infrared light.

Therefore, it can be effectively applied only through the addition of software without adding additional devices to the existing face recognition system and the ATM machine, and thus have a high ripple effect compared to methods requiring additional hardware. .

1 shows an example of a real face and various kinds of fake faces. (a) real face, (b) photo fake face, (c) print fake face, (d) magazine fake face, (e) caricature fake face, and (f) three-dimensional fake face.
2 shows a feature extraction process for the fake face detection method using the full frequency band. (a) An example of a face image, (b) A face image is changed into a frequency domain and separated along the concentric circles.
3 shows a feature extraction process for a fake face detection method using a local binary pattern. (a) An example of a face image, (b) A drawing of the local binary pattern code extracted from the face image to image it, and (c) A histogram of the local binary pattern code expressed as a one-dimensional feature vector.
4 illustrates an example of an arbitrarily divided face image in order to apply the frequency-based feature extraction method and the local binary pattern-based feature extraction method to pieces of the face image.

The present invention relates to a method for detecting an attack using a fake face (two-dimensional or three-dimensional mask) that may occur in a face recognizer and an ATM operation situation. The present invention proposes a method for detecting a fake face using a single image acquired from a general color or monochrome camera without using any additional sensor or lighting at all. The proposal is composed of three main methods.

First, we propose a method of detecting a fake face from the entire frequency band components of a face image obtained by using the property that the detail and the three-dimensional feeling are reduced compared to the real face.

Second, we propose a method for detecting forged faces using local binary patterns, which are advantageous for image texture analysis, using the feature that fake faces have different characteristics in terms of material and material.

Finally, a method of detecting a fake face by combining the aforementioned feature extraction method using the frequency component and the local binary pattern will be described.

The method proposed in the present invention is easy to be applied to a face recognizer or cash dispenser that is currently constructed because it shows an improved fake face detection performance and does not require a separate sensor and lighting other than a general camera.

The method for practicing the present invention is described in more detail.

The first method is to use the power spectrum of the entire frequency band. To this end, the input face image is first transformed into a frequency domain by Fourier transforming.

As an example of the Fourier transform, the following equation can be used.

Here, f (x, y) is a brightness value at (x, y) coordinates of an image having a size of H × W, and u, v are frequency variables.

F is the result of the Fourier transform, where the frequency zero component is shifted to the center.

2 (b) shows the absolute values of the Fourier transformed results in log-scale.

Thereafter, the low frequency value is converted to be located at the center of the Fourier transformed result, and then, as shown in FIG. 2C, a predetermined number of concentric circles are drawn based on the center of the Fourier transformed result, and the average between the concentric circles Calculate the energy value. Finally, the calculated average energy value is expressed as a one-dimensional vector, which undergoes a min-max normalization process.

Through such processes, the feature vector as shown in FIG. 2 (d) is extracted.

This power spectrum-based feature vector may be extracted from the entire face image as shown in FIG. 2 (a), but randomly splits the face image as shown in FIG. 4, and extracts the feature vector from each of the divided face image pieces and combines them into one. It may be. By extracting these feature vectors from real face images and fake face images and learning with a classifier such as a support vector machine, a fake face detector can be generated.

The second method is to use a local binary pattern. To this end, first, after calculating the code of the local binary pattern from the input face image as shown in FIG. 3 (b), the codes are histogramized and normalized so that the sum of the histograms is 1, FIG. We can get feature vectors like). The local binary pattern-based feature vector may be extracted from the entire face image as shown in FIG. 3A, but may be extracted from each of the randomly divided face image fragments as shown in FIG. 4 and merged into one. By extracting these feature vectors from real face images and fake face images and learning with a classifier such as a support vector machine, a fake face detector can be generated.

The third method is a combination of the power spectrum-based method and the local binary pattern-based method described above. The two methods above can be combined in three approaches: First, a classifier such as a support vector machine is used by combining a feature vector extracted based on a power spectrum as shown in FIG. 2 (d) and a feature vector extracted based on a local binary pattern as shown in FIG. 3 (c) at a feature vector level. Can be used for counterfeit face detection. Second, scores obtained by passing feature vectors extracted on the basis of power spectrum to classifiers such as support vector machines and scores obtained by passing feature vectors extracted on the basis of local binary patterns through classifiers such as support vector machines are score levels. Can be used for counterfeit face detection by combining with and learning through classifiers such as support vector machines. Third, the decision result obtained by passing the feature vector extracted based on the power spectrum through a classifier such as a support vector machine and the decision result obtained by passing the feature vector extracted based on a local binary pattern through a classifier such as a support vector machine. It can be used to detect fake faces in combination at the judgment result level.

In summary, the fake face detection method using a single image according to the present invention converts a face image photographed by a color or monochrome camera into a frequency domain, converts a position of the low frequency component to a center, and then converts the image into a plurality of concentric circles. By dividing, one-dimensional feature vectors are generated with average energy values between concentric circles, and a fake face detector is generated by learning them using a support vector machine. The power spectrum-based feature vectors may be extracted from the entire face image, or may be extracted from the fragments of the arbitrarily divided face images and merged into one.

In addition, after generating a local binary pattern code from a face image photographed by a color or monochrome camera, histogram is generated to generate a one-dimensional feature vector, and a fake face detector is generated by learning using a support vector machine. The local binary pattern-based feature vectors may be extracted from the entire face image, or may be extracted from the fragments of the arbitrarily divided face images and merged into one.

Lastly, the power spectrum based fake mask detection method and the local binary pattern based fake mask detection method can be used to detect the fake face by combining at three levels as follows. First, forgery face detectors can be made by combining feature vectors extracted on the basis of power spectrum and feature vectors extracted on the basis of local binary pattern at a feature vector level and learning using a classifier such as a support vector machine. Second, the score obtained by passing the feature vector extracted on the basis of the power spectrum through a classifier such as a support vector machine and the score obtained by passing the feature vector extracted on the basis of a local binary pattern through a classifier such as a support vector machine are combined at a score level. Can be used. Third, the judgment result obtained by passing the feature vector extracted on the basis of the power spectrum through a classifier such as a support vector machine and the result of passing the feature vector extracted on the basis of a local binary pattern through a classifier such as a support vector machine. It can be used in combination with to determine if a fake face is present.

Claims (6)

Obtaining a user face image in color or black and white;
Fourier transforming each pixel of the face image into a frequency domain;
Using the Fourier transformed result to position transform the low frequency component to be centered;
Partitioning the position-converted result into a plurality of concentric circles and calculating an average energy value of each concentric circle to generate a one-dimensional feature vector;
And comparing the calculated one-dimensional feature vector with a reference fake face feature vector using a support vector machine to determine whether the user's face image is a fake face. The method of claim 1,
The reference forged face feature vector,
Preparing at least one fake face image;
Generating each one-dimensional feature vector from the fake face images;
And a step of learning the generated one-dimensional feature vectors as a reference forged face feature vector in the support vector machine. The method of claim 1,
The step of Fourier transforming each pixel of the face image to a frequency domain may include:
Partitioning the face image into a plurality of pieces;
And performing a Fourier transform on each pixel of each of the pieces. Obtaining a user face image in color or black and white;
Calculating a local binary pattern code from the face image;
Generating a one-dimensional feature vector by histogramming the calculated local binary pattern code;
And comparing the calculated one-dimensional feature vector with a reference fake face feature vector using a support vector machine to determine whether the user's face image is a fake face. 5. The method of claim 4,
The reference forged face feature vector,
Preparing at least one fake face image;
Generating each one-dimensional feature vector from the fake face images;
And a step of learning the generated one-dimensional feature vectors as a reference forged face feature vector in the support vector machine. 5. The method of claim 4,
Computing a local binary pattern code from the face image:
Partitioning the face image into a plurality of pieces;
And calculating a local binary pattern code for each of the pieces.

Images