Living body face detection system and method based on hyperspectrum

Abstract

The invention provides a hyperspectrum-based living body face detection system and a hyperspectrum-based living body face detection method. The detection system comprises an objective lens, a diffraction grating, a focusing lens, a mechanical slit, a sensor and a data processing device, the objective lens focuses light reflected by a human face on the surface of the diffraction grating, the diffraction grating disperses the light into different wave bands, the focusing lens converges the dispersed light onto the sensor, and the mechanical slit is formed in the sensor. The mechanical slit is positioned between the focusing lens and the sensor and is used for screening wavebands; the sensor transmits the received information to the data processing device, and the data processing device carries out feature discrimination on the input information to obtain a detection result. The living body face detection system is low in cost, simple in light path and high in practicability.

Description

Technical Field

[0001] The present invention relates to the technical field of computer vision, and in particular to a living human face detection system and method based on hyperspectral. Background Art

[0002] In today's era of big data, users often find it difficult to tolerate the various and redundant forms of identity identification. Face recognition technology uses the face, a biological feature that is not easily forgotten or changed, as its object, and has the advantages of simple operation, intuitive results, non-invasiveness, and non-contact. At present, in the field of live face detection, spectral domain imaging authentication technology has made some progress.

[0003] With the continuous development of physiology and optics, researchers have found that hemoglobin in the dermis of human skin has the ability to absorb incident light, and as the wavelength changes, its reflection characteristics show unique characteristics that distinguish other objects, and stably show a "W"-shaped trend at the 550nm band, which provides a theoretical basis for spectral domain imaging liveness detection. On this basis, Kim et al. proposed to use the spectral data at 850nm and 685nm as the characteristic bands to distinguish true and false faces and different races by analyzing a large amount of facial spectral information. First, the image information of the forehead area of ​​the person is obtained under the selected band, and then the area of ​​fixed size is selected and its grayscale mean is calculated. The calculated results are used to form a two-dimensional feature vector for liveness detection. This method sacrifices the flexibility of the system in order to obtain high-quality facial skin information. Although good experimental results have been achieved, the mandatory requirement to expose the forehead area and the strictly fixed measurement distance have led to a very unfriendly user experience. As an improvement, Zhang Zhiwei et al. (Zhang, Zhiwei, et al. "Face liveness detection by learning multispectral reflectance distributions." 2011IEEE International Conference on Automatic Face & Gesture Recognition (FG). IEEE, 2011.) set the active light source for illuminating the face to 850nm and 1450nm LED lights, and used a photosensitive device to obtain the light reflected by the image. Finally, the value obtained by the photoelectric sensor was used as a feature vector to use the LDA method to identify the face. Although this solution has improved human-computer interaction, the close acquisition and magnification distance of 20-34cm inevitably brings about the user's rebellious psychology, reducing the naturalness and non-intrusiveness of the acquisition.

[0004] In order to realize live face detection using spectral information, it is necessary to obtain hyperspectral information. Among the existing solutions, the more mature ones are SD-CASSI (Single Disperser-Coded Aperture Spectral Imaging) and DD-CASSI (Double Disperser-Coded Aperture Spectral Imaging). However, these systems have large amounts of spectral data and high computational complexity, long calculation time, and lack of real-time performance, and cannot be implemented in live face detection applications. Summary of the invention

[0005] In view of various deficiencies existing in the current practical applications of live face detection, the present invention provides a live face detection system and method based on hyperspectral spectrum to further improve the simplicity of spectral detection devices and the accuracy of inspection.

[0006] The technical solution adopted by the present invention is as follows:

[0007] A live face detection system based on hyperspectral includes an objective lens, a diffraction grating, a focusing lens, a mechanical slit, a sensor and a data processing device. The objective lens focuses the light reflected by the face on the surface of the diffraction grating, the diffraction grating disperses the light into different bands, the focusing lens converges the dispersed light onto the sensor, the mechanical slit is located between the focusing lens and the sensor and is used to screen the bands; the sensor transmits the received information to the data processing device, and the data processing device performs feature discrimination on the input information to obtain a detection result.

[0008] Furthermore, the mechanical slit is located on the R plane, i.e., the rainbow plane. The scattered light with a specific wavelength emitted from all points in the scene on the rainbow plane will reach a specific area on the R plane, thereby converting the wavelength into a spatial position interval, and then the mechanical slit performs spectral screening and modulation.

[0009] Furthermore, the data processing device performs feature discrimination on the input information by specifically using a single-class support vector machine. The single-class support vector machine maps high-density areas in sample data of a single category to a high-dimensional space using specific geometric shapes, and obtains the optimal support domain by finding a geometric shape that can surround most of the samples and has the smallest volume.

[0010] The present invention also provides a focusing method for a living face detection system based on a hyperspectral spectrum, the method comprising the following steps:

[0011] S1, first adjust the focus so that the objective lens is focused on the sensor to determine the imaging distance;

[0012] S2, remove the sensor and place the diffraction grating at the sensor focus position determined in step S1, then adjust the sensor angle and the distance from the diffraction grating to determine the dispersion direction of the first-order dispersion surface of the scene;

[0013] S3, according to the dispersion direction determined in step S2, adjusting the mechanical slit gating position to select the band, placing a focusing lens behind the slit and adjusting the position of the sensor;

[0014] S4, the sensor receives the imaging information and transmits it to the data processing device, and the data processing device determines the focusing result through a feature recognition algorithm.

[0015] The present invention also provides a detection method for a live face detection system based on a hyperspectral spectrum, which specifically includes: firstly, the detection system performs band screening on the scene, the sensor extracts feature points from the data after collecting information, selects the face area as the feature extraction area, and then calculates the grayscale difference-ratio feature, and performs classification by a single-class support vector machine, and finally outputs the result.

[0016] The above technical solution of the present invention has the following advantages compared with the prior art:

[0017] (1) The CASSI system is simplified according to the actual application scenario, and an efficient spectral image acquisition optical path is designed along the design idea of ​​agile spectral imaging. The living face detection system of the present invention has low cost, simple optical path and strong practicality.

[0018] (2) The spectral image with a wavelength of 500nm±10nm was selected for detection, and its performance in identifying real and fake faces is superior to that of the bands near 550nm and 600nm.

[0019] (3) The present invention selects the cheek position which is less covered in the face identification situation. Compared with the method of forcibly detecting in the forehead area in the prior art, removing the cover of the cheek position is more convenient and easier, and less invasive. In addition, the skin on the cheek is thicker, which can more obviously reflect the spectral reflectance characteristics of the skin.

[0020] (4) The face recognition detection measurement distance in the prior art is strictly fixed and the collection distance is only 20-34cm. The method of the present invention proposes the grayscale difference-ratio of a specific area as a face feature for classifier training in view of the changes in distance and light in real scenes, and uses the obtained classifier to classify true and false faces. The improved scheme can not only reduce the pollution of ambient light to the CCD, but also ensure that the distance between the subject and the system is not excessively restricted, thereby reducing the user's deliberate interaction and improving the comfort of use. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1Schematic diagram of the optical path principle of the system of the present invention;

[0022] Figure 2 It is a diagram of the actual optical path of the system of the present invention;

[0023] Figure 3 Schematic diagram of the images at each optical path component (to demonstrate the optical path spectrum selection capability in a more specific and obvious way, the images were collected by a PointGrey Grasshopper3 GS3-U3-51S5C-C industrial color camera equipped with a Sony IMX250 sensor);

[0024] Figure 4 It is a schematic diagram of the detection structure of the system of the present invention;

[0025] Figure 5 The reflectance spectra of real human faces, human models and masks;

[0026] Figure 6 Schematic diagram of feature locations for faces. DETAILED DESCRIPTION

[0027] The present invention is further described below in conjunction with the accompanying drawings.

[0028] The technical concept of the hyperspectral-based live face detection system of the present invention is as follows: by introducing a diffraction grating and a series of lenses in the optical path, a plane is generated in which the light rays of all points in the scene of a specific wavelength intersect at a certain point. According to the one-to-one correspondence between the wavelength and the spatial position of the light rays on the plane, a mechanical slit is introduced on the plane to selectively block different wavelengths. Then, the light rays are refocused on the sensor of the camera through the lens, so that the spectrum of all points in the image is modulated by the slit to generate the final image.

[0029] The design principle of the system of the present invention is as follows Figure 1 As shown in Figure 1, lens L1 focuses scene point X precisely on diffraction grating plane P. Each image point X on grating plane P P , each corresponds to a cone of incident light with a cone angle of θ. The diffraction grating disperses each light ray into its component wavelengths. For each ray in the cone of incident light from a scene point, the grating creates a cone of outgoing light rays, each with a different wavelength and a dispersion angle of α (determined by the model of the diffraction grating). For ease of understanding, Figure 1 θ is magnified in the figure. In the real system, θ<<α. Under the focusing effect of lens L2, the sensor plane S (Sensor) and the diffraction grating plane P (Diffraction Grating) are conjugate, so the scene point is imaged on the sensor plane X S Place.

[0030] In this imaging system, Figure 1 The R plane (Rainbow plane) is a rainbow plane. The scattered light with a specific wavelength emitted by all points in the scene will reach a specific area on the R plane, thereby converting the wavelength into a spatial position interval. By placing slits in the area corresponding to the specific wavelength on this R plane, it is possible to filter the band from the spectrum of the entire image to achieve high spectral modulation, that is, to achieve high spatial resolution spectral imaging by applying rainbow plane coding.

[0031] pass Figure 1 The analysis of the finite aperture case yields:

[0032]

[0033] Where R θ is the width of the same wavelength in the rainbow plane on the R plane, a1 is the aperture diameter of lens L1, f2 is the focal length of L2, d and p are the distance from L1 to the P plane and the distance from the P plane to L2 respectively. Obviously, the smaller R is, the smaller the overlap of the regions of different bands in the R plane is, the more accurate the band screening is, and the more accurate the spectral modulation can be. Furthermore, considering that the model of the diffraction grating P fixes the grating constant α, the system design requires θ1<<d. A small aperture objective lens L1 can achieve this requirement.

[0034] The optical path structure of the face detection system built based on the above finite aperture situation is as follows Figure 2 As shown, in actual application, the system uses an industrial black and white camera CCD equipped with a sensor. In order to expand the field of view, the light of the scene is focused on the grating surface by the industrial lens L1. L1 in this embodiment is a 25mm industrial high-definition fixed-focus lens. The transmission diffraction grating model is THORLABSGT50-03, 300 grooves / mm, groove angle 17.5°, and size is 50mm x 50mm. The industrial lens L2 is a megapixel fixed-focus lens. In this embodiment, considering the easy adjustability and manufacturing difficulty in the actual operation process, the THORLABS VA100C / M30 mechanical slit is selected to screen spectra of different wavelengths. The GCM-TD50MX gear rack translation stage is used to move the slit, and the size of the slit is adjusted to achieve spectrum selection.

[0035] When the above optical path is constructed in this embodiment, the specific process of focusing is as follows:

[0036] S1, first adjust the focus so that the lens L1 is focused on the CCD of the industrial camera and determine the distance of the image formed by the objective lens.

[0037] S2, remove the CCD and place the diffraction grating on the CCD focus position determined in step S1. At this time, the light path is deflected due to the dispersion effect of the grating, so the subsequent light path needs to be deflected by a certain angle accordingly. The dispersion direction of the first-order dispersion surface of the scene is determined by adjusting the angle and the distance from the grating using an industrial camera equipped with a lens, such as Figure 3 As shown, this process is to make the dispersion bands as clear as possible and without aliasing.

[0038] S3, according to the dispersion direction determined in step S2, the slit is fixed to select the band in combination with formulas 1 and 2. Lens L2 is placed behind the slit and the CCD position is adjusted. The dispersed image is focused by the eyepiece and formed on the sensor target surface of the grayscale camera. The image is received by the SpinView professional software, and the focus result is determined by the feature recognition algorithm.

[0039] The structure of the system using this embodiment during detection is as follows: Figure 4 As shown, a face 101 to be detected is placed in front of the detection system, a fixed-focus lens 102 is used to focus the light reflected by the face on the surface of a grating 103, and the grating 103 disperses the collected light into different bands. A pixel fixed-focus lens 104 collects the dispersed image onto the sensor target surface of a CCD camera 106. A mechanical slit 105 is placed between the pixel fixed-focus lens 104 and the CCD camera 106 to filter the bands. The CCD camera 106 receives light path information as a sensor and transmits the information to a computer 107. The computer 107 performs data analysis on the collected spectral information and finally draws an identification conclusion. An optical platform 108 is used to fix the above optical devices.

[0040] The overall system live face detection process is as follows: first, the hardware optical path directly performs band screening on the scene. After CCD collects information, it uses the open source face_recogition library to extract feature points, selects the face area as the feature extraction area, and then calculates the grayscale difference-ratio feature, and uses ONE-CLASS-SVM (single-class support vector machine) for classification, and finally outputs the result.

[0041] Figure 5The reflectance spectra of the real face, PVC mannequin and two silicone masks in the 380nm to 1100nm band obtained by the Avantes AvaSpec-128 ultrafast fiber optic spectrometer. It can be more clearly observed that in the 470nm to 550nm band and the 700nm to 900nm band, the reflectivity of the silicone mask and the PVC mannequin is different from that of human skin. In addition, although the skin reflectivity in the 470nm to 550nm band is not high, the difference between each counterfeit material and human skin is more significant, so this embodiment selects the 470nm to 550nm band for live face detection. However, in the actual operation process, it is limited by the accuracy of optical components, and the selection of the spectrum cannot achieve the ideal situation. Because the band near 500nm is selected in the 470nm to 550nm band.

[0042] Considering that the spectral reflectance curves of various face materials in the selected band of 490nm-550nm do not intersect, that is to say, given the same light source, the concrete expression of the integral of the spectral dimension of the light reflected by faces of different materials - grayscale, will have a unique and definite size relationship. Therefore, taking the grayscale average of the cheek area in the 490nm-510nm band of the face image as the original feature not only simplifies the experimental operation, reduces restrictions and requirements, but also has good practicality.

[0043] In actual applications, the detection system of this embodiment will be contaminated by the ambient light spectrum. On the other hand, different acquisition distances and exposure times will also cause different responses of the CCD to the same reflectivity. Assuming that the ambient light is a globally consistent light source during the system acquisition process, that is, it does not change with the spatial position, the CCD response at the face point a is I a (λ i ), then the formula can be written as follows:

[0044]

[0045] where λ i Indicates the minimum wavelength λ imin To the maximum wavelength λ imax is the wavelength of band i, H() is a step function, c(λ), r(λ) and p(λ) are the response of the image sensor, the reflectivity of the object surface and the global light source intensity at wavelength λ, respectively. min and λ max is the minimum and maximum value of the wavelength λ range controlled by the mechanical slit. Considering that the reflectivity of the face varies at different locations due to the difference in the proportion of skin and subcutaneous tissue, we further select points b and c so that the reflectivity of points a, b, and c are different, and perform the following calculation to obtain the grayscale difference-ratio feature:

[0046]

[0047] Obviously, in the grayscale difference ratio formula, by comparing I(λ i ) is divided after subtraction to solve the impact of CCD response differences and ambient light source differences on grayscale characteristics, ensuring that the environment of the subject and rigid requirements such as collection distance and time are not overly restricted, which can improve the comfort of use.

[0048] For the classifier, this embodiment chooses to use a single-class support vector machine. This is because in the live face detection task, since any non-human face can become a negative sample, the number of positive and negative samples is seriously unequal. At this time, the classification hyperplane of the SVM will deviate, greatly reducing the classification accuracy. Therefore, a single-class support vector machine is chosen here. It can map the high-density area in the sample data of a single category to a high-dimensional space using a specific geometric shape, and obtain the optimal support domain by finding a geometric shape that can surround most of the samples and has the smallest volume. For a given test sample, by calculating the value of the optimal classification function, it can be determined whether it belongs to this category. This processing method can have good classification accuracy in the case of limited samples and noise interference, and is suitable for the task drive of human live face detection.

[0049] In specific testing, such as Figure 6 As shown, it is necessary to reasonably select different groups of regions with different spectral reflectance characteristics of different objects. This embodiment takes into account that the thickness of the skin on the cheeks and the bridge of the nose is inconsistent, the reflectivity is different, and the skin thickness and reflectivity of the area near the tip of the nose and far from the tip of the nose are also different. At the same time, the face photo and the silicone face will not cause a large change in reflectivity due to the uniform material. Therefore, four areas on the left and right cheeks and two areas from the center of the eyebrows to the tip of the nose are selected, a total of 10 areas to form the feature extraction area. The average grayscale values ​​of the feature extraction area of ​​the left cheek are I1, I2, I3, and I4, respectively. Similarly, the average grayscale values ​​of the feature extraction area of ​​the right cheek are I5, I6, I7, and I8, respectively. The average grayscale values ​​of the features from the center of the eyebrows to the tip of the nose, that is, the average grayscale values ​​of the extraction area of ​​the bridge of the nose far from the tip of the nose and near the tip of the nose are I9, I10, I110, I120, I130, I140, I150, I160, I170, I180, I190, I20 ... 10 Further analysis shows that I1 to I8 all belong to the cheek area, and their reflectance can be considered consistent. 10 There are two areas with inconsistent reflectivity, and the reflectivity of the two areas is also inconsistent. 10 Combine to generate features:

[0050]

[0051] Finally, the input classifier is normalized.

[0052] In the parameter setting of the classifier, the popular REF kernel function, Linear kernel function and Poly kernel parameters are analyzed and selected through experimental comparison. The result comparison shows that the Poly kernel function has better performance, its real face recognition rate decays slowly and the fake face recognition rate of the two materials reaches more than 0.65 when the training error is 0.2, and the recognition rate is evenly distributed, so the Poly kernel function is selected for parameter setting here.

[0053] In this embodiment, when the Poly kernel function is selected and the training error is set to 0.27, the system can recognize real faces, face photos, and silicone masks with an accuracy of 0.75, 0.71, and 0.80, respectively. In addition, through testing, this system can achieve 18 frames of images per second under the condition of using an industrial black and white camera, and achieve high-definition sampling of scenes with an image resolution of 2016×2016, so this system meets the requirements of live face detection tasks. In summary, the system of the present invention has strong practicality in live face detection and has unique advantages compared to other current live face detection systems.

[0054] The present invention is based on the inconvenience of the existing spectral imaging system. Considering that face liveness detection only needs to check the characteristic values ​​of a few bands and does not need to reconstruct all hyperspectral spectra, a hyperspectral face detection system based on rainbow face coding and support vector machine is designed and built. The system skips the spectral reconstruction link and directly uses optical equipment to screen the scene bands. The grayscale images sampled by the two-dimensional image sensor directly extract features to judge true and false faces, so that what you see is what you get.

Claims

1. A live face detection system based on hyperspectral, comprising an objective lens, a diffraction grating, a focusing lens, a mechanical slit, a sensor and a data processing device, characterized in that: The objective lens focuses the light reflected from the face onto the surface of the diffraction grating, and the diffraction grating disperses the light into different bands. The focusing lens converges the dispersed light onto the sensor. The mechanical slit is located between the focusing lens and the sensor and is used to screen the bands. The sensor transmits the received information to the data processing device, and the data processing device performs feature discrimination on the input information to obtain a detection result.

2. The hyperspectral-based live face detection system according to claim 1, characterized in that: The mechanical slit is located on the R plane, i.e., the rainbow plane. The scattered light with a specific wavelength emitted from all points in the scene on the rainbow plane will reach a specific area on the R plane, thereby converting the wavelength into a spatial position interval, and then the mechanical slit performs spectral screening and modulation.

3. The hyperspectral-based live face detection system according to claim 1, characterized in that: The data processing device specifically uses a single-class support vector machine to perform feature discrimination on the input information. The single-class support vector machine maps the high-density area in the single-category sample data to the high-dimensional space using a specific geometric shape, and obtains the optimal support domain by finding the geometric shape that can surround most of the samples and has the smallest volume.

4. The focusing method of a live face detection system based on a hyperspectral spectrum as claimed in claim 1, characterized in that: The method comprises the following steps: S1, first adjust the focus so that the objective lens is focused on the sensor to determine the imaging distance; S2, remove the sensor and place the diffraction grating at the sensor focus position determined in step S1, then adjust the sensor angle and the distance from the diffraction grating to determine the dispersion direction of the first-order dispersion surface of the scene; S3, according to the dispersion direction determined in step S2, adjusting the mechanical slit gating position to select the band, placing a focusing lens behind the slit and adjusting the position of the sensor; S4, the sensor receives the imaging information and transmits it to the data processing device, and the data processing device determines the focusing result through a feature recognition algorithm.

5. The detection method using the hyperspectral-based live face detection system as claimed in claim 1, characterized in that: The method specifically includes: firstly, the detection system performs band screening on the scene, the sensor extracts feature points from the data after collecting information, selects the face area as the feature extraction area, then calculates the grayscale difference-ratio feature, and classifies it by a single-class support vector machine, and finally outputs the result.

6. The detection method according to claim 5, characterized in that: The selected wavelength band is 490-510nm.

7. The detection method according to claim 5, characterized in that: The feature extraction area specifically selects the cheek position of the face.

8. The detection method according to claim 7, characterized in that: The feature extraction area specifically selects four areas on the left and right cheeks of the face and two areas from the center of the eyebrows to the tip of the nose.

9. The detection method according to claim 5, characterized in that: The calculation formula of grayscale difference-ratio feature is: Among them, λ i Indicates the minimum wavelength λ imin To the maximum wavelength λ imax The wavelength of band i; λ min and λ max The minimum and maximum values ​​of the wavelength λ range controlled by the mechanical slit; I a (λ i ), I b (λ i ) and I c (λ i ) is the sensor response at different face feature extraction areas a, b, c; H() is the step function; r a (λ), r b (λ) and r c (λ) is the surface reflectivity at different feature extraction areas a, b, and c of the face, and these three reflectivities are different from each other.

10. The detection method according to claim 9, characterized in that: where c(λ) and p(λ) are the sensor response and global light source intensity at wavelength λ, respectively.

Images