WO2022041627A1 - 一种活体人脸检测方法及系统 - Google Patents

一种活体人脸检测方法及系统 Download PDF

Info

Publication number
WO2022041627A1
WO2022041627A1 PCT/CN2020/141998 CN2020141998W WO2022041627A1 WO 2022041627 A1 WO2022041627 A1 WO 2022041627A1 CN 2020141998 W CN2020141998 W CN 2020141998W WO 2022041627 A1 WO2022041627 A1 WO 2022041627A1
Authority
WO
WIPO (PCT)
Prior art keywords
face
hyperspectral
image
depth
depth image
Prior art date
Application number
PCT/CN2020/141998
Other languages
English (en)
French (fr)
Inventor
王兆民
黄源浩
肖振中
Original Assignee
奥比中光科技集团股份有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 奥比中光科技集团股份有限公司 filed Critical 奥比中光科技集团股份有限公司
Publication of WO2022041627A1 publication Critical patent/WO2022041627A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
    • G06V40/16—Human faces, e.g. facial parts, sketches or expressions
    • G06V40/161—Detection; Localisation; Normalisation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
    • G06F17/10—Complex mathematical operations
    • G06F17/16—Matrix or vector computation, e.g. matrix-matrix or matrix-vector multiplication, matrix factorization
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/30—Determination of transform parameters for the alignment of images, i.e. image registration
    • G06T7/33—Determination of transform parameters for the alignment of images, i.e. image registration using feature-based methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/50—Depth or shape recovery
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/80—Analysis of captured images to determine intrinsic or extrinsic camera parameters, i.e. camera calibration
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
    • G06V40/16—Human faces, e.g. facial parts, sketches or expressions
    • G06V40/168—Feature extraction; Face representation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
    • G06V40/16—Human faces, e.g. facial parts, sketches or expressions
    • G06V40/172—Classification, e.g. identification
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/40—Spoof detection, e.g. liveness detection
    • G06V40/45—Detection of the body part being alive
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10032—Satellite or aerial image; Remote sensing
    • G06T2207/10036—Multispectral image; Hyperspectral image

Definitions

  • the present application relates to the technical field of image processing, and in particular, to a method and system for detecting a living face.
  • face-based authentication has been widely used, such as face-swiping payment, face-recognition unlocking and other functions have been widely used in people's daily life, greatly improving the convenience of people's lives sex.
  • face recognition technology has greatly improved the convenience of people's lives, its security problems have gradually been exposed, especially with the emergence of high-fidelity masks, many criminals have caused visual deception through realistic camouflage to carry out A series of criminal acts have also attacked ordinary face recognition systems. Therefore, face living anti-counterfeiting technology has attracted widespread attention.
  • hyperspectral imaging technology is the current research hotspot.
  • Hyperspectral imaging technology is based on a very wide range of narrow-band image data technology. It combines imaging technology and spectral technology to detect a two-dimensional combination of targets. Spatial and one-dimensional spectral information can be obtained to obtain continuous and narrow-band image data with high spectral resolution, thereby greatly improving the accuracy of target detection.
  • due to the large number of bands in this technology and the inevitable correlation between adjacent bands there is a certain degree of redundancy in hyperspectral image data, and the amount of data is large, which increases the pressure on post-processing of images.
  • the purpose of the present application is to provide a method and system for detecting a live human face, so as to solve at least one of the above-mentioned background technical problems.
  • An embodiment of the present application provides a method for detecting a living face, including the following steps:
  • S2 analyze the depth image, and detect the face region in the depth image; if a human face is detected, perform positioning of key feature points of the face on the face region in the depth image to obtain The location information of the key feature points of the face;
  • step S1 the depth image is acquired by a depth camera, and the hyperspectral image is acquired by a hyperspectral camera; wherein the hyperspectral image includes a plurality of hyperspectral images of different wavelength bands.
  • step S1 the depth camera and the hyperspectral camera acquire the depth image and the hyperspectral image synchronously, respectively; or, the depth camera and the hyperspectral camera follow a certain time sequence
  • the depth image and the hyperspectral image are acquired at intervals, respectively.
  • step S2 includes the following steps:
  • S20 Calculate the curvature value and orientation of the curvature of each vertex in the depth image
  • step S21 Based on the curvature value and the orientation obtained in step S20, the depth image is divided, the human body and the background are distinguished, and the face region is detected;
  • step S22 According to the face region detected in step S21, locate the key feature points of the face by using the curvature value obtained in step S20.
  • the curvature includes at least one of a principal curvature, a Gaussian curvature, and an average curvature.
  • the Gaussian curvature and the average curvature are employed to determine the local face shape.
  • step S3 the depth image and the hyperspectral image are matched to extract spectral information of the key feature points of the face from the depth image.
  • matching the depth image to the hyperspectral image includes the steps of:
  • step S30 According to the depth image of the face area detected in step S2, obtain a point cloud data set of the depth image of the area, and project all points on a spatial plane to confirm the second depth image of the face area.
  • step S31 Based on the part of the face region detected in step S2, intercept the hyperspectral images of different wavebands as registration data, extract grayscale images with different wavebands but the same size, and obtain the pixel range of the face region;
  • step S32 According to the pixel range of the hyperspectral image corresponding to the face area, grid the point cloud data obtained in step S30, The information value is matched with the cloud data of the corresponding points of the face region after gridding, and the spectral information of each point of the two-dimensional point cloud data is obtained;
  • step S33 Based on the two-dimensional point cloud data corresponding to the hyperspectral image obtained in step S32, according to the label of each point in the two-dimensional point cloud, match the spectral information to the corresponding original three-dimensional face area to obtain the spectral information of each point in the point cloud data of the original face area;
  • Embodiments of the present application further provide a living face detection system, including a depth camera, a hyperspectral camera, and a processor; wherein the depth camera is used to acquire a depth image of a target area; the hyperspectral camera is used to acquire all hyperspectral images of different wavebands of the target area; the processor includes a depth processor and a hyperspectral processor, the depth processor is used to analyze the acquired depth image and detect the human face in the depth image area, and locate the key feature points of the face in the face area to obtain the position information of the key feature points of the face; the hyperspectral processor extracts the position information of the key feature points of the face according to the position information The spectral information of the key feature points of the face corresponding to the hyperspectral image is used for living body detection to determine whether it is a living body.
  • the depth camera is used to acquire a depth image of a target area
  • the hyperspectral camera is used to acquire all hyperspectral images of different wavebands of the target area
  • the processor includes a depth processor
  • the depth camera and the hyperspectral camera are configured on the same image sensor array to acquire depth images and hyperspectral images, respectively; alternatively, the depth camera and the hyperspectral camera are provided separately.
  • An embodiment of the present application provides a method for detecting a living face, including the steps of: S1: acquiring a depth image and a hyperspectral image of a target area; S2: analyzing the depth image, and detecting a face area in the depth image; If a human face is detected, position the key feature points of the face in the face region in the depth image to obtain the position information of the key feature points of the face; S3: According to all the key features in the depth image The location information of the key feature points of the face is extracted, the spectral information corresponding to the key feature points of the face in the hyperspectral image is extracted, and living body detection is performed to determine whether it is a living body.
  • the face region is identified in the depth image and the key feature points of the face are located.
  • the hyperspectral image only the spectral information of the key feature points of the face is extracted for live detection to complete the face recognition.
  • the live body detection can be completed without the user's cooperation to perform corresponding actions. The experience is excellent, and the detection time is saved, the image processing calculation amount is reduced, and the efficiency and accuracy of live face detection are improved. .
  • FIG. 1 is a schematic flowchart of a method for detecting a living face according to an embodiment of the present application.
  • FIG. 2 is a flow chart of detecting a face region based on curvature in a method for detecting a living face according to an embodiment of the present application.
  • FIG. 3 is a flowchart of matching a depth image and a hyperspectral image in a method for detecting a living face according to an embodiment of the present application.
  • FIG. 4 is a schematic plane projection diagram of three-dimensional point cloud data in a method for detecting a living face according to an embodiment of the present application.
  • FIG. 5 is a schematic diagram of a live face detection system according to another embodiment of the present application.
  • connection can be used for either a fixing function or a circuit connecting function.
  • first and second are only used for descriptive purposes, and should not be construed as indicating or implying relative importance or implying the number of indicated technical features. Thus, a feature defined as “first”, “second” may expressly or implicitly include one or more of that feature. In the description of the embodiments of the present application, “plurality” means two or more, unless otherwise expressly and specifically defined.
  • FIG. 1 is a schematic flowchart of a method for detecting a living face according to an embodiment of the present application. The method includes the following steps:
  • the method for acquiring a depth image includes acquiring by using a structured light depth camera or acquiring by using a time-of-flight depth camera, or acquiring by using a binocular depth camera, which is not particularly limited in the embodiments of the present application.
  • a hyperspectral image of the target area is acquired using a hyperspectral camera, and the hyperspectral image includes a plurality of hyperspectral images of different wavelength bands.
  • the spectral images of different wavelength bands include 550nm band hyperspectral images, 685nm band hyperspectral images and 850nm band hyperspectral images; due to the influence of special substances such as skin melanin, the skin reflection curve has a "W" feature in the 550nm band, that is, in the 550nm band The skin reflection curve of real skin near this band forms a "W" shape, which promotes skin recognition and can distinguish materials that mimic the diffraction of human skin, which is helpful for more realistic modeling and rendering of human skin; for the 850nm band, it is suitable for In vivo detection; for the 685nm band, different races can be distinguished.
  • the above wavebands are only exemplary descriptions, and the embodiments of the present application are not limited to the above wavebands.
  • the processor in a method for acquiring a depth image and a hyperspectral image, provides a synchronous trigger signal to the depth camera and the hyperspectral camera, so that when the depth camera acquires the depth image, the hyperspectral camera also simultaneously acquires the hyperspectral image images for real-time computing.
  • the acquisition of the depth image and the hyperspectral image is asynchronous, and the depth image and the hyperspectral image are acquired separately in a certain time series (ie, at a certain time interval), which can reduce the need for the processor storage and computing power requirements.
  • S2 Analyze the depth image to detect the face area in the depth image; if a face is detected, locate the key feature points of the face in the face area in the depth image, and obtain the position information of the key feature points of the face .
  • step S2 includes the following steps:
  • S20 Calculate the curvature value and orientation of each vertex in the depth image; preferably, the curvature includes at least one of principal curvature, Gaussian curvature and average curvature.
  • the face surface in order to analyze the curvature of a three-dimensional human face, is defined as S, and S is defined by a quadratic differentiable function f, let f: U ⁇ R, Then the surface S can be expressed as:
  • f xx can obtain the contour of the face in the X-axis direction
  • f yy can obtain the contour of the face in the Y-axis direction
  • f xy can obtain the contour of the face in the XY space
  • the Gaussian curvature and the average curvature are used to determine the partial shape of the human face, and the use of the Gaussian curvature and the mean curvature to determine the partial shape of the human face will be described in detail later.
  • the average curvature (H) and Gaussian curvature (K) at each point (x, y, f(x, y)) on the three-dimensional surface can be calculated by the following formulas:
  • f x , f y , f xx , f yy , f xy are the first and second differentials of f at (x,y). Since the second derivative is very sensitive to noise, the surface is smoothed.
  • the depth data is processed with a Gaussian filter before calculating the curvature.
  • a point P on a curved surface S there are an infinite number of planes passing through the normal vector of the point P.
  • the intersection of these planes and surface S is a series of curves.
  • the curvature of these curves is called the normal curvature of point P on the surface S.
  • the normal vector of the surface determines the sign of the normal curvature.
  • the maximum and minimum values of the normal curvatures are called principal curvatures, and the directions corresponding to the two principal curvatures are called the principal directions of point P on the surface S.
  • the degree of curvature of the surface S can be measured by the shape operator L.
  • the shape operator describes the tangent variation of each point on the surface, and its eigenvalues and eigenvectors can represent the principal curvature and principal direction, namely:
  • L is the matrix expression of the shape operator of the surface S
  • D represents the second form of the surface
  • G represents the first form of the surface
  • x, y represent partial derivatives to variables x, y.
  • step S21 Divide the depth image based on the curvature value and orientation in the curvature obtained in step S20, distinguish the human body from the background, and detect the face region.
  • the Gaussian curvature uniquely determines the shape of the convex surface, while the average curvature, under certain auxiliary conditions, uniquely determines the shape of the graphic surface.
  • geometric features in the neighborhood of a point on the surface can be derived to detect faces.
  • the Gaussian curvature is positive, zero, or negative
  • the local curved surfaces correspond to ellipsoids, parabolas, and hyperboloids, respectively.
  • the sign of the average curvature indicates the concave and convex characteristics of the neighborhood surface.
  • the average curvature is non-negative, the neighborhood surface is convex; otherwise, when the average curvature is negative, the neighborhood surface is concave.
  • the face and background are divided by computing the eigenvalues (curvature values of the principal curvatures) and eigenvectors (the principal curvature orientations) of the shape operator.
  • the curvature map contains the principal curvature of each pixel, that is, the eigenvalues with larger absolute values and the corresponding curvature orientations.
  • the original curvature may be positive or negative, where positive curvature corresponds to a convex surface pattern, negative curvature corresponds to a concave surface pattern, body parts such as the head are convex in nature and have a strong positive curvature.
  • pixels with positive curvature values have light gray shades in the curvature map, while those with negative curvature values are dark gray, and the transition from positive to negative curvature is a good indication of the edge of the face part, so that it can be clearly divided Human face and background.
  • positive and negative curvature here is arbitrary, and it can also be defined as a convex surface figure with negative curvature, which is not limited here.
  • the curved shape of the face area is relatively complex due to the eyes, nose, mouth and eye sockets, the common feature of these three directions is the curved surface from the forehead to the top of the head, from the cheeks to the sides, and the chin to the neck. There is a change from flatter to more curved. Therefore, this change can be measured by Gaussian curvature, mean curvature and shape operators.
  • step S22 According to the face region detected in step S21, the key feature points of the face are located by the curvature value obtained in step S20.
  • the face key feature points include nose, eyes and mouth.
  • the regions with higher principal curvature values are the inner corner regions of the nose and eyes. Eyes and noses can be quickly detected by setting the curvature threshold.
  • the curvature value is positive, the part is convex, and the area is identified as the nose; in the HK classification table with a threshold, the part that is elliptical is identified as the area. for eyes and mouth. Based on the area of the key feature points of the face obtained by the above method, the position information of the area in the depth image is obtained.
  • S3 According to the position information of the key feature points of the face in the depth image, extract the spectral information corresponding to the key feature points of the face in the hyperspectral image, and perform living body detection to determine whether it is a living body. If it is determined to be a living body, the face depth information is matched with the face information in the preset database to obtain a face recognition result.
  • depth image data is matched with hyperspectral image data to extract spectral information of key feature points of the face from the depth image.
  • the depth image of the face is a kind of irregular spatial three-dimensional data
  • the hyperspectral image is a two-dimensional image data.
  • the two kinds of data need to be transformed in the spatial dimension.
  • a point cloud data set of a depth image is obtained, and then the point cloud data is imaged in two dimensions, so that the difficulty of data processing and the amount of calculation can be reduced.
  • matching the depth image with the hyperspectral image specifically includes the following steps:
  • step S30 According to the depth image of the face area detected in step S2, obtain the point cloud data set of the depth image of this area, and project all points P(x, y, z) on a space plane W to confirm the face The coordinate range of the point cloud after the 2D projection of the regional depth image.
  • the point cloud data is projected onto the spatial plane W through the matrix T, where:
  • C [x c , y c , z c ] is the camera center, I is the identity matrix, and R is the rotation matrix:
  • ⁇ , ⁇ , ⁇ are Euler triangles: Yaw, Pitch, Roll.
  • the camera calibration matrix is:
  • f x and f y represent the focal lengths in the X, Y axis directions.
  • step S31 Based on the part of the face region detected in step S2, intercept hyperspectral images of different bands as registration data, extract grayscale images of different bands but the same size, and obtain the pixel range of the face region, such as the The pixel range is: X: 356-843, Y: 50-962. After cropping, the data range of the required face area can be obtained, that is, 488 ⁇ 913.
  • step S33 Based on the two-dimensional point cloud data corresponding to the hyperspectral image obtained in step S32, according to the label of each point in the two-dimensional point cloud, the spectral information is matched to the point cloud of the corresponding original three-dimensional face area, To obtain the spectral information of each point of the point cloud data of the original face area.
  • step S2 based on the position information of the key feature points of the face of the depth image obtained in step S2, extract the spectral information of the corresponding key feature points of the face in the hyperspectral image, and calculate the area of the key feature points of the face based on the spectral information.
  • the light reflection information and the gray value of the key feature point area of the face are obtained, and then it is judged whether the detected face is a living body. If it is a living body, the face recognition result is obtained by matching the face depth image obtained in step S1 with the face image in the preset database. If the matching is successful, the target object is a matching person, otherwise it is a suspicious person.
  • the images used for face recognition are not limited to depth images, but can also be color images or infrared images, or a combination of both, as long as they match the face images saved in the preset database. There is no particular limitation in the embodiments of the present application.
  • FIG. 5 is a schematic diagram of a live face detection system according to another embodiment of the present application.
  • the system 500 includes a depth camera 501, a hyperspectral camera 502, a processor 503, and a face detector 504; wherein, the depth camera 501 is used to obtain a depth image of the target area; Spectral image; the processor 503 includes a depth processor 5030 and a hyperspectral processor 5031, the depth processor 5030 is used to analyze the acquired depth image, detect the face area in the depth image, and analyze the face in the face area.
  • the key feature points are located, and the position information of the key feature points of the face is obtained and then sent to the hyperspectral processor 5031, and the hyperspectral processor 5031 receives the position information of the key feature points of the face, and extracts the hyperspectral image corresponding to the person according to the information.
  • the spectral information of the key feature points of the face is used for living body detection to determine whether it is a living body; if it is a living body, the depth image obtained by the depth camera 501 is transmitted to the face detector 504, and the face detector 504 compares the depth image with the preset database. The face image in the match is matched, and the face recognition result is output.
  • the depth camera 501 may be a depth camera based on technical solutions such as structured light, binocular, TOF (time-of-flight algorithm).
  • the depth camera is taken as an example of a structured light depth camera for description.
  • a structured light depth camera includes a transmitting module and a receiving module.
  • the structured light pattern emitted by the transmitting module is an infrared speckle image
  • the receiving module is an infrared camera
  • the structured light pattern is collected by the infrared camera and then output to the processor 503, and the processor 503 passes the structured light The pattern is calculated to obtain the depth image of the target person.
  • depth camera 501 and hyperspectral camera 502 may be configured to acquire depth images and hyperspectral images on the same image sensor array, the image sensor array including filters.
  • the filter is arranged over the image sensor array so as to optically cover the image sensor array. The filter selectively transmits light in a specific wavelength band and blocks light outside the specific wavelength band from reaching the image sensor array. The blocked light can be absorbed, reflected and/or scattered, depending on the implementation of the filter, which is not considered here. limit.
  • the filter In one filtering state, if the depth image is generated based on infrared light, the filter can project infrared light and block light outside the infrared wavelength band; in another filtering state, the hyperspectral image is based on In contrast to the generation of depth images using different wavelength bands, the filter can transmit beams in the wavelength bands contained in the hyperspectral image and block beams outside the wavelength bands. It should be understood that a hyperspectral image contains multiple wavelength bands of light beams, and the filter can be configured to switch between multiple filter states for different wavelength bands of light. It will be appreciated that the filter can be switched between any suitable number of different filtering states to transmit any suitable wavelength band of light while blocking light outside that wavelength band.
  • the depth camera and the hyperspectral camera are set separately. Before acquiring the depth image and the hyperspectral image, the depth camera and the hyperspectral camera need to be calibrated to obtain the relative positional relationship between the two cameras (R ,T), R is the rotation matrix, and T is the translation matrix.
  • Q G is the spatial coordinate of the same point Q on the reference object in the color camera coordinate system
  • q G is the projected coordinate of this point on the color image
  • the relationship between Q G and Q D is determined by the depth camera and the hyperspectral camera.
  • the external parameter matrix between Q G and Q D is represented, and the external parameter matrix includes two parts: the rotation matrix R and the translation matrix T; the conversion relationship between Q G and Q D is as follows:
  • R D and T D are the rotation matrix and translation matrix of the internal parameter matrix of the depth camera respectively
  • R G and T G are the rotation matrix and translation matrix of the internal parameter matrix of the hyperspectral camera respectively
  • the extrinsic parameter matrix of the depth camera and the hyperspectral camera that is, the expressions of the rotation matrix R and the translation matrix T in the extrinsic parameter matrix are:
  • the external parameter matrices of the depth camera and the hyperspectral camera can be obtained, and the depth image and the hyperspectral image can be aligned one by one. .
  • the present application also proposes a computer-readable storage medium, where the computer-scaled storage medium stores a computer program, and when the computer program is executed by a processor, implements the living face detection method according to the solution of the above embodiment.
  • the storage medium may be implemented by any type of volatile or non-volatile storage device, or a combination thereof.
  • Embodiments of the present application may include or utilize a special purpose or general purpose computer including computer hardware, as discussed in more detail below. Embodiments within the scope of the present application also include physical and other computer-readable media for carrying or storing computer-executable instructions and/or data structures. Such computer-readable media can be any available media that can be accessed by a general purpose or special purpose computer system. Computer-readable media that store computer-executable instructions are physical storage media. Computer-readable media carrying computer-executable instructions are transmission media. Thus, by way of example and not limitation, embodiments of the present application may include at least two distinct computer-readable media: physical computer-readable storage media and transmission computer-readable media.
  • An embodiment of the present application further provides a computer device, the computer device includes a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor executes the computer During the program, at least the living face detection method described in the foregoing embodiments is implemented.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Oral & Maxillofacial Surgery (AREA)
  • Human Computer Interaction (AREA)
  • Multimedia (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Health & Medical Sciences (AREA)
  • Mathematical Physics (AREA)
  • General Health & Medical Sciences (AREA)
  • Computational Mathematics (AREA)
  • Mathematical Analysis (AREA)
  • Mathematical Optimization (AREA)
  • Pure & Applied Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Computing Systems (AREA)
  • Algebra (AREA)
  • Databases & Information Systems (AREA)
  • Software Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Length Measuring Devices By Optical Means (AREA)
  • Image Analysis (AREA)

Abstract

本申请公开了一种活体人脸检测方法及系统，包括：S1：获取目标区域的深度图像和高光谱图像；S2：对深度图像进行分析，检测深度图像中的人脸区域；若检测到人脸，则对深度图像中的人脸区域进行人脸关键特征点的定位，得到人脸关键特征点的位置信息；S3：根据深度图像中人脸关键特征点的位置信息，提取所述高光谱图像中对应人脸关键特征点的光谱信息，并进行活体检测，判断是否为活体。本申请结合了深度图像和高光谱图像，不需要用户配合执行相应的动作就能够完成活体检测，体验效果佳，且节省检测时间，减少图像处理计算量，提高了活体人脸检测的效率及准确率。

Description

一种活体人脸检测方法及系统
本申请要求于2020年8月31日提交中国专利局，申请号为202010899445.3，发明名称为“一种活体人脸检测方法及系统”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本申请涉及图像处理技术领域，尤其涉及一种活体人脸检测方法及系统。
背景技术
随着电子商务等技术的发展，基于人脸的身份验证已经得到广泛的应用，如刷脸支付、人脸识别解锁等功能已广泛应用于人们的日常生活中，极大提高了人们生活的便利性。然而，在人脸识别技术极大提高了人们生活便利性的同时，其安全性问题也逐渐暴露出来，尤其是随着高仿真面具的出现，使得许多不法分子通过逼真的伪装引起视觉欺骗以进行一系列犯罪行为，也对普通的人脸识别系统造成了攻击。因此，人脸活体防伪技术引发了广泛的关注。
在相关活体人脸检测技术中，高光谱成像技术是当前的研究热点，高光谱成像技术是基于非常多窄波段的影像数据技术，它将成像技术和光谱技术相结合，探测目标的二维结合空间及一维光谱信息，以获取高光谱分辨率的连续、窄波段的图像数据，进而极大提高目标探测的准确性。但由于该技术的波段较多，且相邻波段间必然具有相关性，使得高光谱图像数据存在一定程度的冗余现象，并且数据量大，给图像的后期处理增加压力。
发明内容
本申请的目的在于提供一种活体人脸检测方法及系统，以解决上述背景技 术问题中的至少一种问题。
本申请实施例提供一种活体人脸检测方法，包括如下步骤：
S1：获取目标区域的深度图像和高光谱图像；
S2：对所述深度图像进行分析，检测所述深度图像中的人脸区域；若检测到人脸，则对所述深度图像中的所述人脸区域进行人脸关键特征点的定位，得到所述人脸关键特征点的位置信息；
S3：根据所述深度图像中的所述人脸关键特征点的所述位置信息，提取所述高光谱图像中对应所述人脸关键特征点的光谱信息，并进行活体检测，判断是否为活体。
在一些实施例中，步骤S1中，通过深度相机获取所述深度图像，通过高光谱相机获取所述高光谱图像；其中，所述高光谱图像包括多个不同波段的高光谱图像。
在一些实施例中，步骤S1中，所述深度相机和所述高光谱相机同步分别获取所述深度图像和所述高光谱图像；或者，所述深度相机和所述高光谱相机以一定的时序间隔分别采集所述深度图像和所述高光谱图像。
在一些实施例中，步骤S2包括如下步骤：
S20：计算所述深度图像中每一个顶点的曲率的曲率值和取向；
S21：基于步骤S20得到所述曲率值和所述取向划分所述深度图像，区分人体和背景，检测出人脸区域；
S22：根据步骤S21检测的所述人脸区域，通过基于步骤S20得到的所述曲率值对所述人脸关键特征点进行定位。
在一些实施例中，所述曲率包括主曲率、高斯曲率和平均曲率中的至少一种。
在一些实施例中，采用所述高斯曲率和所述平均曲率确定人脸局部形状。
在一些实施例中，步骤S3中，将所述深度图像与所述高光谱图像进行匹配，以从所述深度图像中提取所述人脸关键特征点的光谱信息。
在一些实施例中，将所述深度图像与所述高光谱图像进行相匹配包括以下步骤：
S30：根据步骤S2检测所得的所述人脸区域的所述深度图像，获取该区域深度图像的点云数据集，将所有点投影到一个空间平面上，以确认所述人脸区域深度图像二维投影后的点云坐标范围；
S31：基于步骤S2检测所得的所述人脸区域部分，截取不同波段的所述高光谱图像作为配准数据，提取不同波段但尺寸相同的灰度图，得到人脸区域的像素范围；
S32：根据所述人脸区域对应的高光谱图像像素范围，将步骤S30得到的所述点云数据网格化，将所述人脸区域对应的所述高光谱图像中的每个像元的信息值与网格化后的相对应的所述人脸区域的点所述云数据进行匹配，得到二维点云数据每一点的光谱信息；
S33：基于步骤S32得到的与所述高光谱图像对应的所述二维点云数据，根据其二维点云中每个点的标号，将所述光谱信息匹配至对应的原始三维人脸区域的点云中，以获得原始人脸区域的点云数据的每一点的光谱信息；
S34：根据所述深度图像与所述原始三维人脸区域点云数据的坐标映射关系，确定所述深度图像的每一点像素在所述高光谱图像中匹配到相应的光谱信息。
本申请实施例还提供一种活体人脸检测系统，包括深度相机、高光谱相机、以及处理器；其中，所述深度相机用于获取目标区域的深度图像；所述高光谱相机用于获取所述目标区域的不同波段的高光谱图像；所述处理器包括深度处理器和高光谱处理器，所述深度处理器用于对获取的所述深度图像进行分析，检测所述深度图像中的人脸区域，并对所述人脸区域中的人脸关键特征点进行定位，得到所述人脸关键特征点的位置信息；所述高光谱处理器根据所述人脸关键特征点的位置信息，提取所述高光谱图像对应的所述人脸关键特征点的光谱信息以进行活体检测，判断是否为活体。
在一些实施例中，所述深度相机和所述高光谱相机被配置于同一图像传感器阵列上以分别获取深度图像和高光谱图像；或者，所述深度相机和所述高光谱相机为分开设置。
本申请实施例提供一种活体人脸检测方法，包括步骤：S1：获取目标区域的深度图像和高光谱图像；S2：对所述深度图像进行分析，检测所述深度图像中的人脸区域；若检测到人脸，则对所述深度图像中的所述人脸区域进行人脸关键特征点的定位，得到所述人脸关键特征点的位置信息；S3：根据所述深度图像中的所述人脸关键特征点的所述位置信息，提取所述高光谱图像中对应所述人脸关键特征点的光谱信息，并进行活体检测，判断是否为活体。通过获取深度图像和高光谱图像，在深度图像中识别人脸区域并进行人脸关键特征点定位，在高光谱图像中只提取人脸关键特征点的光谱信息进行活体检测以完成人脸识别，结合了深度图像和高光谱图像，不需要用户配合执行相应的动作就能够完成活体检测，体验效果极佳，且节省检测时间，减少图像处理计算量，提高了活体人脸检测的效率及准确率。
附图说明
为了更清楚地说明本申请实施例或现有技术中的技术方案，下面将对实施例或现有技术描述中所需要使用的附图作简单地介绍，显而易见地，下面描述中的附图仅仅是本申请的一些实施例，对于本领域普通技术人员来讲，在不付出创造性劳动性的前提下，还可以根据这些附图获得其他的附图。
图1是根据本申请一个实施例活体人脸检测方法的流程示意图。
图2是根据本申请一个实施例活体人脸检测方法中基于曲率检测人脸区域的流程图示。
图3是根据本申请一个实施例活体人脸检测方法中匹配深度图像和高光谱图像的流程图。
图4是根据本申请一个实施例活体人脸检测方法中三维点云数据的平面投影示意图。
图5是根据本申请另一个实施例活体人脸检测系统示意图。
具体实施方式
为了使本申请实施例所要解决的技术问题、技术方案及有益效果更加清楚明白，以下结合附图及实施例，对本申请进行进一步详细说明。应当理解，此处所描述的具体实施例仅仅用以解释本申请，并不用于限定本申请。
需要说明的是，当元件被称为“固定于”或“设置于”另一个元件，它可以直接在另一个元件上或者间接在该另一个元件上。当一个元件被称为是“连接于”另一个元件，它可以是直接连接到另一个元件或间接连接至该另一个元件上。另外，连接即可以是用于固定作用也可以是用于电路连通作用。
需要理解的是，术语“长度”、“宽度”、“上”、“下”、“前”、“后”、“左”、“右”、“竖直”、“水平”、“顶”、“底”“内”、“外”等指示的方位或位置关系为基于附图所示的方位或位置关系，仅是为了便于描述本申请实施例和简化描述，而不是指示或暗示所指的装置或元件必须具有特定的方位、以特定的方位构造和操作，因此不能理解为对本申请的限制。
此外，术语“第一”、“第二”仅用于描述目的，而不能理解为指示或暗示相对重要性或者隐含指明所指示的技术特征的数量。由此，限定有“第一”、“第二”的特征可以明示或者隐含地包括一个或者更多该特征。在本申请实施例的描述中，“多个”的含义是两个或两个以上，除非另有明确具体的限定。
参照图1所示，图1为根据本申请实施例提供的一种活体人脸检测方法流程示意图，方法包括以下步骤：
S1：获取目标区域的深度图像和高光谱图像；
在一些实施例中，深度图像获取的方法包括采用结构光深度相机获取或采用飞行时间深度相机获取，或者采用双目深度相机获取，在本申请实施例中不 作特别限制。
在一些实施例中，利用高光谱相机获取目标区域的高光谱图像，高光谱图像包括多个不同波段的高光谱图像。优选地，不同波段的光谱图像包括550nm波段高光谱图像、685nm波段高光谱图像及850nm波段高光谱图像；由于皮肤黑色素等特殊物质的影响，皮肤反射曲线在550nm波段具有“W”特征，即在该波段附近真实皮肤的皮肤反射曲线形成“W”形状，对皮肤识别起促进作用，能够区分模仿人体皮肤衍射的材料，有助于更真实的建模和渲染人体皮肤；对于850nm波段，适合进行活体检测；对于685nm波段，可区分不同的人种。可以理解的是，以上波段仅是示例性说明，本申请实施例中并不仅限于以上波段。
在一些实施例中，在获取深度图像和高光谱图像获取方法中，通过处理器向深度相机和高光谱相机提供同步触发信号，使得深度相机获取深度图像的同时，高光谱相机也同时获取高光谱图像，以实现实时运算。当然，对于同步获取深度图像和高光谱图像的情形，其对整体系统性能的要求高。在另一个实施例中，深度图像和高光谱图像的获取是不同步的，采取以一定的时序的方式(即以一定的时序间隔)分开采集深度图像和高光谱图像，如此可降低对处理器的存储及运算能力的要求。
S2：对深度图像进行分析，检测深度图像中的人脸区域；若检测到人脸，则对深度图像中的人脸区域进行人脸关键特征点的定位，得到人脸关键特征点的位置信息。
具体的，参照图2所示，在一些实施例中，步骤S2包括如下步骤：
S20：计算深度图像中每一个顶点的曲率值和取向；优选地，曲率包括主曲率、高斯曲率和平均曲率中的至少一种。
在一些实施例中，为了分析三维人脸的曲率，定义人脸曲面为S，通过二次可微的函数f定义S，令f：U→R，
Figure PCTCN2020141998-appb-000001
则曲面S可以表示为：
S＝{(x,y,z)|(x,y)∈U，z∈R，f(x，y)＝z}
对于曲面S上的每一点(x，y，f(x，y))，其沿x，y两个坐标轴方向的一 阶导数为f x和f y，因为三维人脸的深度数据是离散的，所以在计算曲面的一阶导数时用差分来代替：
f x(x,y)＝f(x+1,y)-f(x,y)
f y(x,y)＝f(x,y+1)-f(x,y)
同样，离散曲面上的二阶导数也用差分表示：
f xx(x,y)＝f x(x+1,y)-f x(x,y)
f yy(x,y)＝f y(x,y+1)-f y(x,y)
f xy(x,y)＝f x(x,y+1)-f x(x,y)
根据上述离散曲面的二阶导数可知，f xx可以获取人脸在X轴方向上的轮廓，f yy可以获取人脸在Y轴方向上的轮廓，f xy可以获取人脸在X-Y空间上的轮廓，但是二阶偏微分对噪声十分敏感，容易受到各种各样的噪声干扰，其结果不稳定。因此，在一些实施例中，采用高斯曲率和平均曲率确定人脸局部形状，后面对采用高斯曲率和平均曲率确定人脸局部形状进行详细说明。
由微分几何的定义可得，三维曲面上各点(x，y，f(x，y))处的平均曲率(H)和高斯曲率(K)可用如下公式计算：
高斯曲率：
Figure PCTCN2020141998-appb-000002
平均曲率：
Figure PCTCN2020141998-appb-000003
其中，f x，f y，f xx，f yy，f xy是f在(x,y)处的一阶和二阶微分。由于二阶微分对于噪声十分敏感，所以要对曲面进行平滑。在计算曲率之前，用高斯滤波器对深度数据进行处理。
在一实施例中，对于一个曲面S上的一点P，有无数个经过P点法向量的平面。这些平面和曲面S的交线是一系列曲线，这些曲线的曲率称之为P点在曲面S的法曲率，曲面的法向矢量决定了法曲率的符号。法曲率中的最大值和 最小值称之为主曲率，两个主曲率对应的方向称为曲面S上P点的主方向。当两个主曲率不相等时，相应的两个主方向完全确定，且相互正交；而当两个主曲率相等时，主方向不能唯一确定，此时曲面S在该点的任一切向都是主曲率的方向。
曲面S的弯曲程度可以通过形状算子L进行衡量，形状算子描述了曲面上各个点的切线变化量，其特征值和特征向量可以表示主曲率和主方向，即：
L＝DG -1
其中，L为曲面S的形状算子的矩阵表达式，D表示曲面的第二形式，G表示曲面的第一形式，则：
Figure PCTCN2020141998-appb-000004
Figure PCTCN2020141998-appb-000005
其中，x，y表示对变量x，y的偏导。
S21：基于步骤S20得到的曲率中的曲率值和取向划分深度图像，区分人体和背景，检测出人脸区域。
在一些实施例中，高斯曲率唯一地确定凸表面的形状，而平均曲率在一定的辅佐条件下，可唯一地确定图形表面的形状。通过分析高斯曲率和平均曲率的符号，可得出曲面上的某一点的邻域内的几何特征，以检测出人脸。具体的，当高斯曲率为正数、零、负数时，局部曲面分别对应为椭球面、抛物线面以及双曲面。平均曲率的符号指示了邻域曲面的凹凸特性，当平均曲率为非负数时，邻域曲面为凸面；否则平均曲率为负数时，邻域曲面为凹面。应当理解的是，根据高斯曲率的符号和平均曲率符号，可以有很多种不同的组合，分别对应于不同特性的曲面，能够对三维表面的顶点进行初步归类，从而划分人脸与背景。
在一些实施例中，通过计算形状算子的本征值(主曲率的曲率值)和本征向量(主曲率取向)划分人脸和背景。曲率图包含每个像素的主曲率，即具有 较大绝对值的本征值和对应的曲率取向。原始曲率可以是正的或者是负的，其中，正曲率对应凸形表面图形，负曲率对应凹形表面图形，人体部分诸如头部本质上是凸形的，且具有强正曲率。此外，具有正曲率值的像素在曲率图中具有浅的灰影，而负曲率值的像素是暗灰的，从正曲率向负曲率的转变为人脸部分的边缘的良好指示，从而可以清楚划分人脸和背景。应当理解的是，此处的正负曲率的定义是任意的，也可以定义成负曲率为凸形表面图形，此处不做限制。
尽管人脸区域因眼、鼻、嘴以及眼眶造成了曲面形状比较复杂，但从额头向头顶的走向、从脸颊向两侧走向和下巴向颈部走向，这三个方向的共同特征点是曲面有一个从较平坦向较弯曲变化的过程。因此，通过高斯曲率、平均曲率和形状算子可以衡量出这种变化。
S22：根据步骤S21检测的人脸区域，通过基于步骤S20得到的曲率值对人脸的关键特征点进行定位。
在一些实施例中，人脸关键特征点包括鼻子、眼睛及嘴巴。假设三维人脸曲面是光滑的曲面，则具有较高主曲率值的区域为鼻子和眼睛的内角区域。通过设置曲率阈值，可快速检测出眼睛和鼻子。优选地，设置阈值T h＝0.04，T k＝0.0005，将具有较低曲率值的点集剔除，即K≥T h，G≥T k。在具有阈值的平均曲率图中，曲率值是正值的，该部分表现为凸起，则认定该区域为鼻子；在具有阈值的HK分类表中，表现为椭圆形的部分，则认定该区域为眼睛及嘴巴。基于上述方法得到的人脸关键特征点的区域，获取该区域在深度图像中的位置信息。
S3：根据深度图像中人脸关键特征点的位置信息，提取高光谱图像中对应人脸关键特征点的光谱信息，并进行活体检测，判断是否为活体。若判断为活体，则将人脸深度信息与预设数据库中的人脸信息进行匹配，获取人脸识别结果。
在一些实施例中，将深度图像数据与高光谱图像数据进行匹配，以从深度 图像中提取人脸关键特征点的光谱信息。人脸的深度图像是一种无规则的空间三维数据，而高光谱图像是一种二维图像数据，在数据配准之前，需要将两种数据在空间维度上进行转换。优选地，在本申请实施例中，获取深度图像的点云数据集，再将点云数据二维图像化，从而可以减少数据处理难度和计算量。
如图3所示，将深度图像与高光谱图像进行相匹配具体包括以下步骤：
S30：根据步骤S2检测所得的人脸区域的深度图像，获取该区域深度图像的点云数据集，将所有点P(x,y,z)投影到一个空间平面W上，以确认该人脸区域深度图像二维投影后的点云坐标范围。
如图4所示，在一些实施例中，利用相机模型原理，通过矩阵T将点云数据投影到空间平面W上，其中：
T＝KR[I|-C]
C＝[x c,y c,z c]为相机中心，I为单位矩阵，R为旋转矩阵：
Figure PCTCN2020141998-appb-000006
α、β、γ分别为欧拉三角：Yaw,Pitch,Roll。
相机定标矩阵为：
Figure PCTCN2020141998-appb-000007
f x和f y表示在X，Y轴方向上的焦距长度。
在一些实施例中，获取点云数据时，相机被配置为以Y轴正向指向被测目标物体。假设空间投影平面与三维坐标系Y轴垂直，设相机坐标系O为圆点，三维点云数据中的任意点P＝(x，y，z)，对应投影平面点为P'＝(x′，y′，z′)。从而：
y′＝y c
Figure PCTCN2020141998-appb-000008
Figure PCTCN2020141998-appb-000009
待完全计算完毕后，确定二维投影后人脸区域点云数据坐标范围，如X：-0.05786m～0.07055m，Y：-0.05844m～0.14559m；X的长度Lx＝0.12841m，Y的长度为：Ly＝0.20403m。
S31：基于步骤S2检测所得的人脸区域部分，截取不同波段的高光谱图像作为配准数据，提取不同波段但尺寸相同的灰度图，得到人脸区域的像素范围，如该人脸区域的像素范围为：X：356～843，Y：50～962，裁剪后可得到所需人脸区域部分的数据范围，即488×913。
S32：根据人脸区域对应的高光谱图像像素范围，将步骤S30得到的点云数据网格化，如：选取步长为：m＝Lx/488＝0.00026，n＝Ly/913＝0.00022，将人脸区域对应的高光谱图像中的每个像元的信息值与网格化后的相对应的人脸区域的点云数据进行匹配，得到二维点云数据每一点的光谱信息。
S33：基于步骤S32得到的与高光谱图像对应的的二维点云数据，根据其二维点云中每个点的标号，将光谱信息匹配至对应的原始三维人脸区域的点云中，以获得原始人脸区域的点云数据的每一点的光谱信息。
S34：根据深度图像与原始三维人脸区域点云数据的坐标映射关系，确定深度图像的每一点像素在高光谱图像中匹配到相应的光谱信息。
更具体地，对于不同的材质或结构，光反射信息存在差异，在图像获取的过程中，不同的特征区的灰度值也存在差异，通过上述的差异，可以判断目标对象是否为活体。
在一些实施例中，基于步骤S2获得的深度图像的人脸关键特征点的位置信息，提取高光谱图像中对应的人脸关键特征点的光谱信息，基于该光谱信息计算人脸关键特征点区域的光反射信息及获取人脸关键特征点区域的灰度值，进而判断检测到的人脸是否为活体。若为活体，则根据步骤S1中获取的人脸深度图像与预设数据库中的人脸图像进行匹配，获取人脸识别结果，若匹配成功，则该目标对象为匹配人员，否则为可疑人员。应当理解的是，人脸识别使用的 图像不限于深度图像，还可以是彩色图像或红外图像，或者是两两的组合，只需与预设数据库所保存的人脸图像相匹配即可，在本申请实施例中不作特别限制。
图5为根据本申请另一实施例提供的一种活体人脸检测系统的示意图。系统500包括深度相机501、高光谱相机502、处理器503以及人脸检测器504；其中，深度相机501用于获取目标区域的深度图像；高光谱相机502用于获取目标区域的不同波段的高光谱图像；处理器503包括深度处理器5030和高光谱处理器5031，深度处理器5030用于对获取的深度图像进行分析，检测深度图像中的人脸区域，并对人脸区域中的人脸关键特征点进行定位，得到人脸关键特征点的位置信息后将其发送至高光谱处理器5031，高光谱处理器5031接收人脸关键特征点的位置信息，并根据该信息提取高光谱图像对应人脸关键特征点的光谱信息以进行活体检测，判断是否为活体；若为活体，则将深度相机501获取的深度图像传输至人脸检测器504，人脸检测器504将深度图像与预设数据库中的人脸图像进行匹配，输出人脸识别结果。
在一些实施例中，深度相机501可以基于结构光、双目、TOF(时间飞行算法)等技术方案的深度相机。优选地，以深度相机为结构光深度相机为例进行说明，一般地，结构光深度相机包括发射模组和接收模组。在一些实施例中，发射模组发射的结构光图案为红外散斑图像，接收模组为红外相机，通过红外相机采集到所述结构光图案后输出至处理器503，处理器503通过结构光图案计算从而得到目标人物的深度图像。
在一些实施例中，深度相机501和高光谱相机502可以被配置为同一图像传感器阵列上获取深度图像和高光谱图像，图像传感器阵列包括滤光器。可选地，滤光器布置在图像传感器阵列上方，以便光学覆盖图像传感器阵列。滤光器可选择地透射特定波段的光束，并阻止特定波段外的光到达图像传感器阵列，被阻挡的光可以被吸收、反射和/或散射，取决于滤光器的实现方式，此处不作限制。
在一种滤光状态下，若深度图像是基于红外光生成，在则滤光器可以投射红外光并阻挡红外光波段之外的光；在另一种滤光状态下，高光谱图像是基于与深度图像使用不同波段的生成，则滤光器可以透射高光谱图像包含的波段的光束并阻挡波段之外的光束。应当理解的是，高光谱图像含有多个波段的光束，滤光器可以被配置为在不同波段光的多个滤光状态之间切换。应当理解的是，滤光器可以在任何合适数量的不同滤光状态之间切换，以透射任何合适的波段的光束，同时阻挡该波段之外的光束。
在一些实施例中，深度相机和高光谱相机为分开设置的，在获取深度图像与高光谱图像前，需要将深度相机和高光谱相机进行标定，获取两个相机之间的相对位置关系(R,T)，R为旋转矩阵，T为平移矩阵。
更具体地，假设选取参考物上的一点Q在深度图像上的投影坐标为q D，在深度相机坐标系下的空间坐标为Q D，深度相机的内参数矩阵为H G，q D与Q D之间的转换关系如下所示：
q D＝H D·Q D
同理，假设Q G为参考物上同一点Q在彩色相机坐标系的空间坐标，q G为该点在彩色图像上的投影坐标，Q G与Q D之间关系通过深度相机与高光谱相机之间的外参数矩阵表示，外参数矩阵包括旋转矩阵R和平移矩阵T两部分；Q G与Q D之间的转换关系如下：
Q G＝RQ D+T
假设参考物同一点Q在彩色相机坐标系下的空间坐标为Q G，在彩色图像上的投影坐标为q G，Q G与q G之间的变换关系由彩色相机的内参数矩阵H G表示，即
q G＝H G·Q G
则点Q在深度相机坐标系和彩色相机坐标系下的空间坐标Q D、Q G可下式求得：
Figure PCTCN2020141998-appb-000010
其中，R D、T D分别为深度相机内参数矩阵的旋转矩阵和平移矩阵，R G、T G分 别为高光谱相机内参数矩阵中的旋转矩阵和平移矩阵；在式子Q D＝R DQ+T D中，点Q将由R D、Q D、T D表示，带入式子Q G＝RQ D+T，得：
Figure PCTCN2020141998-appb-000011
整理上式，得深度相机和高光谱相机的外参数矩阵，即外参数矩阵中的旋转矩阵R和平移矩阵T的表达式为：
Figure PCTCN2020141998-appb-000012
因此，在同一场景下，通过相机标定获取深度相机和高光谱相机的内参数矩阵后，由此求得深度相机和高光谱相机的外参数矩阵，即可将深度图像和高光谱图像一一对齐。
本申请还提出了一种计算机可读存储介质，计算机刻度存储介质存储有计算机程序，该计算机程序被处理器执行时实现上述实施例方案的活体人脸检测方法。所述存储介质可以由任何类型的易失性或非易失性存储设备、或者它们的组合来实现。
本申请的实施例可以包括或利用包括计算机硬件的专用或通用计算机，如下面更详细讨论的。在本申请的范围内的实施例还包括用于携带或存储计算机可执行指令和/或数据结构的物理和其他计算机可读介质。这样的计算机可读介质可以是可以被通用或专用计算机系统访问的任何可用介质。存储计算机可执行指令的计算机可读介质是物理存储介质。携带计算机可执行指令的计算机可读介质是传输介质。因此，作为示例而非限制，本申请的实施例可以包括至少两种截然不同的计算机可读介质：物理计算机可读存储介质和传输计算机可读介质。
本申请实施例还提供一种计算机设备，所述计算机设备包括存储器、处理器以及存储在所述存储器上并可在所述处理器上运行的计算机程序，其中，所述处理器执行所述计算机程序时至少实现前述实施例方案中所述的活体人脸检测方法。
可以理解的是，以上内容是结合具体/优选的实施方式对本申请所作的进一 步详细说明，不能认定本申请的具体实施只局限于这些说明。对于本申请所属技术领域的普通技术人员来说，在不脱离本申请构思的前提下，其还可以对这些已描述的实施方式做出若干替代或变型，而这些替代或变型方式都应当视为属于本申请的保护范围。在本说明书的描述中，参考术语“一种实施例”、“一些实施例”、“优选实施例”、“示例”、“具体示例”、或“一些示例”等的描述意指结合该实施例或示例描述的具体特征、结构、材料或者特点包含于本申请的至少一个实施例或示例中。
在本说明书中，对上述术语的示意性表述不必须针对的是相同的实施例或示例。而且，描述的具体特征、结构、材料或者特点可以在任一个或多个实施例或示例中以合适的方式结合。此外，在不相互矛盾的情况下，本领域的技术人员可以将本说明书中描述的不同实施例或示例以及不同实施例或示例的特征进行结合和组合。尽管已经详细描述了本申请的实施例及其优点，但应当理解，在不脱离由所附权利要求限定的范围的情况下，可以在本文中进行各种改变、替换和变更。
此外，本申请的范围不旨在限于说明书中所述的过程、机器、制造、物质组成、手段、方法和步骤的特定实施例。本领域普通技术人员将容易理解，可以利用执行与本文所述相应实施例基本相同功能或获得与本文所述实施例基本相同结果的目前存在的或稍后要开发的上述披露、过程、机器、制造、物质组成、手段、方法或步骤。因此，所附权利要求旨在将这些过程、机器、制造、物质组成、手段、方法或步骤包含在其范围内。

Claims (10)

  1. 一种活体人脸检测方法，其特征在于，包括如下步骤：，
    S1：获取目标区域的深度图像和高光谱图像；
    S2：对所述深度图像进行分析，检测所述深度图像中的人脸区域；若检测到人脸，则对所述深度图像中的所述人脸区域进行人脸关键特征点的定位，得到所述人脸关键特征点的位置信息；
    S3：根据所述深度图像中的所述人脸关键特征点的所述位置信息，提取所述高光谱图像中对应所述人脸关键特征点的光谱信息，并进行活体检测，判断是否为活体。
  2. 如权利要求1所述的活体人脸检测方法，其特征在于：在步骤S1中，通过深度相机获取所述深度图像，通过高光谱相机获取所述高光谱图像；其中，所述高光谱图像包括多个不同波段的高光谱图像。
  3. 如权利要求2所述的活体人脸检测方法，其特征在于：在步骤S1中，所述深度相机和所述高光谱相机同步分别获取所述深度图像和所述高光谱图像；或者，所述深度相机和所述高光谱相机以一定的时序间隔分别采集所述深度图像和所述高光谱图像。
  4. 如权利要求1所述的活体人脸检测方法，其特征在于：步骤S2包括如下步骤：
    S20：计算所述深度图像中每一个顶点的曲率的曲率值和取向；
    S21：基于步骤S20得到所述曲率值和所述取向划分所述深度图像，区分人体和背景，检测出人脸区域；
    S22：根据步骤S21检测的所述人脸区域，通过基于步骤S20得到的所述曲率值对所述人脸关键特征点进行定位。
  5. 如权利要求4所述的活体人脸检测方法，其特征在于：所述曲率包括主曲率、高斯曲率和平均曲率中的至少一种。
  6. 如权利要求4所述的活体人脸检测方法，其特征在于：采用所述高斯曲 率和所述平均曲率确定人脸局部形状。
  7. 如权利要求1所述的活体人脸检测方法，其特征在于：步骤S3中，将所述深度图像与所述高光谱图像进行匹配，以从所述深度图像中提取所述人脸关键特征点的光谱信息。
  8. 如权利要求7所述的活体人脸检测方法，其特征在于：将所述深度图像与所述高光谱图像进行相匹配包括以下步骤：
    S30：根据步骤S2检测所得的所述人脸区域的所述深度图像，获取该区域深度图像的点云数据集，将所有点投影到一个空间平面上，以确认所述人脸区域深度图像二维投影后的点云坐标范围；
    S31：基于步骤S2检测所得的所述人脸区域部分，截取不同波段的所述高光谱图像作为配准数据，提取不同波段但尺寸相同的灰度图，得到人脸区域的像素范围；
    S32：根据所述人脸区域对应的高光谱图像像素范围，将步骤S30得到的所述点云数据网格化，将所述人脸区域对应的所述高光谱图像中的每个像元的信息值与网格化后的相对应的所述人脸区域的所述点云数据进行匹配，得到二维点云数据每一点的光谱信息；
    S33：基于步骤S32得到的与所述高光谱图像对应的所述二维点云数据，根据其二维点云中每个点的标号，将所述光谱信息匹配至对应的原始三维人脸区域的点云中，以获得原始人脸区域的点云数据的每一点的光谱信息；
    S34：根据所述深度图像与所述原始三维人脸区域点云数据的坐标映射关系，确定所述深度图像的每一点像素在所述高光谱图像中匹配到相应的光谱信息。
  9. 一种活体人脸检测系统，其特征在于：包括深度相机、高光谱相机、以及处理器；其中，
    所述深度相机用于获取目标区域的深度图像；
    所述高光谱相机用于获取所述目标区域的不同波段的高光谱图像；
    所述处理器包括深度处理器和高光谱处理器，所述深度处理器用于对获取 的所述深度图像进行分析，检测所述深度图像中的人脸区域，并对所述人脸区域中的人脸关键特征点进行定位，得到所述人脸关键特征点的位置信息；所述高光谱处理器根据所述人脸关键特征点的位置信息，提取所述高光谱图像对应的所述人脸关键特征点的光谱信息以进行活体检测，判断是否为活体。
  10. 如权利要求9所述的活体人脸检测系统，其特征在于：所述深度相机和所述高光谱相机被配置于同一图像传感器阵列上以分别获取深度图像和高光谱图像；或者，所述深度相机和所述高光谱相机为分开设置。
PCT/CN2020/141998 2020-08-31 2020-12-31 一种活体人脸检测方法及系统 WO2022041627A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202010899445.3A CN112232109B (zh) 2020-08-31 2020-08-31 一种活体人脸检测方法及系统
CN202010899445.3 2020-08-31

Publications (1)

Publication Number Publication Date
WO2022041627A1 true WO2022041627A1 (zh) 2022-03-03

Family

ID=74115792

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/141998 WO2022041627A1 (zh) 2020-08-31 2020-12-31 一种活体人脸检测方法及系统

Country Status (2)

Country Link
CN (1) CN112232109B (zh)
WO (1) WO2022041627A1 (zh)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115278031A (zh) * 2022-07-29 2022-11-01 盛泰光电科技股份有限公司 基于高光谱的摄像头模组及其图像处理方法和应用
CN115797995A (zh) * 2022-11-18 2023-03-14 北京的卢铭视科技有限公司 人脸活体检测方法、电子设备及存储介质

Families Citing this family (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112926498B (zh) * 2021-03-20 2024-05-24 杭州知存智能科技有限公司 基于多通道融合和深度信息局部动态生成的活体检测方法及装置
CN112926497B (zh) * 2021-03-20 2024-07-05 杭州知存智能科技有限公司 基于多通道数据特征融合的人脸识别活体检测方法和装置
CN113139465A (zh) * 2021-04-23 2021-07-20 北京华捷艾米科技有限公司 人脸识别方法及装置
CN113408414A (zh) * 2021-06-18 2021-09-17 北京理工大学 基于高光谱生物特征的人脸防伪识别方法和装置
CN113658155A (zh) * 2021-08-24 2021-11-16 凌云光技术股份有限公司 一种基于光度立体的物体表面瑕疵检测分析方法及装置
CN113963425B (zh) * 2021-12-22 2022-03-25 北京的卢深视科技有限公司 人脸活体检测系统的测试方法、装置及存储介质
CN114827561B (zh) * 2022-03-07 2023-03-28 成都极米科技股份有限公司 投影控制方法、装置、计算机设备和计算机可读存储介质
CN115273249A (zh) * 2022-08-01 2022-11-01 思博特集成科技(珠海横琴)有限公司 光谱三维识别方法、识别方法装置及可读存储介质
CN116718271B (zh) * 2023-08-07 2023-11-14 北京理工大学 高速高光谱成像系统和方法

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104881632A (zh) * 2015-04-28 2015-09-02 南京邮电大学 高光谱人脸识别方法
CN106204601A (zh) * 2016-07-15 2016-12-07 华东师范大学 一种基于波段扫描形式的活体高光谱序列图像并行配准方法
US20190138807A1 (en) * 2015-11-02 2019-05-09 Fotonation Limited Iris liveness detection for mobile devices
CN111046703A (zh) * 2018-10-12 2020-04-21 杭州海康威视数字技术股份有限公司 人脸防伪检测方法、装置及多目相机

Family Cites Families (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102622588B (zh) * 2012-03-08 2013-10-09 无锡中科奥森科技有限公司 双验证人脸防伪方法及装置
CN105023010B (zh) * 2015-08-17 2018-11-06 中国科学院半导体研究所 一种人脸活体检测方法及系统
CN105335722B (zh) * 2015-10-30 2021-02-02 商汤集团有限公司 一种基于深度图像信息的检测系统及方法
CN105975926B (zh) * 2016-04-29 2019-06-21 中山大学 基于光场相机的人脸活体检测方法
CN107451510B (zh) * 2016-05-30 2023-07-21 北京旷视科技有限公司 活体检测方法和活体检测系统
KR102317180B1 (ko) * 2017-06-08 2021-10-27 (주)이더블유비엠 3차원 깊이정보 및 적외선정보에 기반하여 생체여부의 확인을 행하는 얼굴인식 장치 및 방법
WO2019056310A1 (en) * 2017-09-22 2019-03-28 Qualcomm Incorporated SYSTEMS AND METHODS FOR DETECTING FACIAL ACTIVITY
FR3089036B1 (fr) * 2018-11-28 2020-10-30 In Idt dispositif et procédé d’authentification d’un individu

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104881632A (zh) * 2015-04-28 2015-09-02 南京邮电大学 高光谱人脸识别方法
US20190138807A1 (en) * 2015-11-02 2019-05-09 Fotonation Limited Iris liveness detection for mobile devices
CN106204601A (zh) * 2016-07-15 2016-12-07 华东师范大学 一种基于波段扫描形式的活体高光谱序列图像并行配准方法
CN111046703A (zh) * 2018-10-12 2020-04-21 杭州海康威视数字技术股份有限公司 人脸防伪检测方法、装置及多目相机

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
"Master Thesis", 1 March 2017, BEIJING JIAOTONG UNIVERSITY, CN, article LIU YIFEI: "Face Liveness Detection Based on Spectrum Analysis and Depth Information", pages: 1 - 68, XP055903281 *
FENG JIAO; FANG XIAOJING; LI SHOUFENG; WANG YONGJIN: "High-emulation mask recognition with high-resolution hyperspectral video capture system", PROCEEDINGS OF SPIE, IEEE, US, vol. 9273, 4 November 2014 (2014-11-04), US , pages 92733H - 92733H-7, XP060043417, ISBN: 978-1-62841-730-2, DOI: 10.1117/12.2071819 *

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115278031A (zh) * 2022-07-29 2022-11-01 盛泰光电科技股份有限公司 基于高光谱的摄像头模组及其图像处理方法和应用
CN115797995A (zh) * 2022-11-18 2023-03-14 北京的卢铭视科技有限公司 人脸活体检测方法、电子设备及存储介质
CN115797995B (zh) * 2022-11-18 2023-09-01 北京的卢铭视科技有限公司 人脸活体检测方法、电子设备及存储介质

Also Published As

Publication number Publication date
CN112232109B (zh) 2024-06-04
CN112232109A (zh) 2021-01-15

Similar Documents

Publication Publication Date Title
WO2022041627A1 (zh) 一种活体人脸检测方法及系统
EP3866052B1 (en) Face anti-spoof detection method, device and multi-view camera
US11017205B2 (en) Three-dimensional finger vein recognition method and system
Kang et al. Study of a full-view 3D finger vein verification technique
US7876931B2 (en) Face recognition system and method
CN108549886A (zh) 一种人脸活体检测方法及装置
CN110309782A (zh) 一种基于红外与可见光双目系统的活体人脸检测方法
US20140043329A1 (en) Method of augmented makeover with 3d face modeling and landmark alignment
CN105956515A (zh) 基于极成像的立体-高光谱人脸识别方法
CN104036546A (zh) 一种基于自适应形变模型的任意视角人脸三维重构方法
CN109670390A (zh) 活体面部识别方法与系统
CN112052831A (zh) 人脸检测的方法、装置和计算机存储介质
CN106155299B (zh) 一种对智能设备进行手势控制的方法及装置
CN107491744A (zh) 人体身份识别方法、装置、移动终端和存储介质
CN108596193A (zh) 一种搭建针对人耳识别的深度学习网络结构的方法及系统
Smith et al. Facial shape-from-shading and recognition using principal geodesic analysis and robust statistics
CN112257641A (zh) 一种人脸识别活体检测方法
CN111291730A (zh) 人脸防伪检测方法、服务器和存储介质
Parziale et al. Advanced technologies for touchless fingerprint recognition
Bastias et al. A method for 3D iris reconstruction from multiple 2D near-infrared images
Chen et al. 3d face mask anti-spoofing via deep fusion of dynamic texture and shape clues
CN106778491A (zh) 人脸3d特征信息的获取方法及设备
Bogo et al. Automated detection of new or evolving melanocytic lesions using a 3D body model
Ambika et al. Periocular authentication based on FEM using Laplace–Beltrami eigenvalues
US20180047206A1 (en) Virtual mapping of fingerprints from 3d to 2d

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20951293

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20951293

Country of ref document: EP

Kind code of ref document: A1