WO2015090126A1 - 人脸特征的提取、认证方法及装置 - Google Patents

人脸特征的提取、认证方法及装置 Download PDF

Info

Publication number
WO2015090126A1
WO2015090126A1 PCT/CN2014/091046 CN2014091046W WO2015090126A1 WO 2015090126 A1 WO2015090126 A1 WO 2015090126A1 CN 2014091046 W CN2014091046 W CN 2014091046W WO 2015090126 A1 WO2015090126 A1 WO 2015090126A1
Authority
WO
WIPO (PCT)
Prior art keywords
face
dimensional
face image
target
image
Prior art date
Application number
PCT/CN2014/091046
Other languages
English (en)
French (fr)
Inventor
江武明
张祥德
王宁
郑金增
李倩颖
张芹芹
Original Assignee
北京天诚盛业科技有限公司
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 北京天诚盛业科技有限公司 filed Critical 北京天诚盛业科技有限公司
Publication of WO2015090126A1 publication Critical patent/WO2015090126A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F21/00—Security arrangements for protecting computers, components thereof, programs or data against unauthorised activity
    • G06F21/30—Authentication, i.e. establishing the identity or authorisation of security principals
    • G06F21/31—User authentication
    • G06F21/32—User authentication using biometric data, e.g. fingerprints, iris scans or voiceprints
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/40—Extraction of image or video features
    • G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
    • G06V10/443—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components by matching or filtering
    • G06V10/449—Biologically inspired filters, e.g. difference of Gaussians [DoG] or Gabor filters
    • G06V10/451—Biologically inspired filters, e.g. difference of Gaussians [DoG] or Gabor filters with interaction between the filter responses, e.g. cortical complex cells
    • 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

Definitions

  • the present invention relates to the field of image processing and pattern recognition, and in particular to a method and apparatus for extracting and authenticating facial features.
  • Face authentication is a form of biometric recognition. By effectively characterizing a face and obtaining the characteristics of two face photos, it is determined whether the two photos are the same person. Compared with other biometric authentication technologies, it is friendly, convenient and non-invasive. Therefore, in recent years, face authentication technology has become a research hotspot of many scientific research and commercial institutions.
  • the face authentication process is mainly divided into three parts: face detection, eye positioning and normalization, feature extraction and authentication. Since the face is a three-dimensional deformation model, and the face authentication is based on the photo taken by the camera imaging model, the result of the authentication is easily affected by external factors such as illumination, posture, expression, and occlusion.
  • face recognition technology involves many interdisciplinary subjects such as pattern recognition, statistical learning, machine vision, applied mathematics and information science, and its wide application prospects have received more and more attention.
  • the original face recognition algorithm is based on the difference of pixels between face feature points. This method has very poor effect on background illumination. In recent years, the research on face recognition has made rapid progress, and there are a large number of new ones. Face recognition algorithm is proposed. Different people have different classification methods. According to the representation method of face in recognition process, face recognition methods can be roughly divided into: geometric feature based method, global feature based method, local based The method of texture feature.
  • Face representation based on geometric features is mainly to extract key points of the face, such as eyes, nose, mouth, chin, eyebrows and so on. Then use the relative positions of these key points to calculate the distance, the ratio of the area, etc., and then use these ratios as a representation of the face.
  • This method is robust to illumination changes on the premise of accurate positioning. Sex, and the feature quantity is relatively small. However, it has a natural disadvantage, that is, it depends on the positioning of key points, and the relative position of key points is easily affected by expressions and gestures.
  • a major representation of face representation based on global features is a method based on sparse representation.
  • the main idea is to directly use a large database as a set of faces of the human face, and then perform a sparse projection of the faces to be compared to the set of bases, and obtain the projection coefficients of each face under the set of bases, and then use This set of projection coefficients is used to represent the face.
  • a sparse linear combination of the set of bases is used to characterize a face to be recognized.
  • This method can achieve quite good recognition when the database used for the base is very large and closed-loop test.
  • the training library is relatively small, or when the open-collection test is performed, the projection effect on the outsider is It is not very good. That is to say, the generalization of the algorithm is not strong.
  • a typical example based on local texture features is a face representation based on Gabor features.
  • the Gabor kernels of different scales and directions are used to filter on the image, and then the filtered images are compared.
  • the Gabor feature effectively balances the representation of the time and frequency domains of the signal. It is one of the most popular feature representations at present. However, the biggest problem with this method is that the amount of data is very large.
  • a Gabor core with 8 scales and 5 directions will change a photo into 40 features, which increases the complexity of storage and calculation.
  • the main object of the present invention is to provide a face feature extraction and authentication scheme to solve at least the above problems.
  • a method for extracting a face feature comprising: performing a two-dimensional Hilbert transform on the acquired face image; and a face that has undergone the above two-dimensional Hilbert transform
  • the image is represented by a two-dimensional analytical signal.
  • the formula for performing a two-dimensional Hilbert transform on the acquired face image is:
  • the time domain formula of the two-dimensional Hilbert transform is:
  • the two-dimensional analytical signal representation of the face image after the two-dimensional Hilbert transform includes: expressing the two-dimensional analysis signal of the face image by the three components of the local amplitude A, the local phase ⁇ and the local direction ⁇ , among them,
  • sign( ⁇ ) is a symbolic function
  • ⁇ arctan(f y (x, y)/f x (x, y)) for representing geometric information of the two-dimensional analytical signal
  • f(x, y) is the real part of the two-dimensional analytical signal
  • f x (x, y) and f y (x, y) are the two imaginary parts of the two-dimensional analytical signal, respectively.
  • the method before performing the two-dimensional Hilbert transform on the acquired face image, the method further comprises: respectively filtering the acquired face image by a band pass filter, wherein the band pass filter comprises log-Gabor filtering Device.
  • the band pass filter comprises log-Gabor filtering Device.
  • the method further comprises: passing the two-dimensional analysis signal of the face image after the two-dimensional Hilbert transform
  • the preset sparse self-encoding neural network obtains the corresponding optimal value, wherein the weight matrix and the offset matrix in the preset sparse self-coding neural network are obtained by the face training sample.
  • a method for authenticating facial features using the above-described method for extracting facial features comprising: respectively representing a first face image and a second face image using a face feature extraction method; Performing a similarity calculation on the first face image and the second face image after the extraction method of the face feature; and completing the face authentication process according to the result of the similarity calculation described above.
  • performing similarity calculation on the first facial image and the second facial image after the facial feature extraction method is expressed includes: a first facial image and a second representation after the facial feature extraction method is expressed Each component of the face image is subjected to similarity calculation; the weighted average of each component similarity is obtained as the total similarity between the first face image and the second face image represented by the face feature extraction method.
  • a device for extracting facial features comprising: a two-dimensional Hilbert transform module, performing a two-dimensional Hilbert transform on the acquired face image; A face image representation module for representing a two-dimensional Hilbert transform face image with a two-dimensional analysis signal.
  • an apparatus for authenticating a face feature using the above-described facial feature extraction device comprising: a feature extraction module, configured to respectively represent the first using the face feature extraction device a face image and a second face image; a similarity calculation module, performing similarity calculation on the first face image and the second face image represented by the face feature extraction device; and an authentication module for using the similarity degree The result of the calculation completes the face authentication process.
  • the method of "two-dimensional Hilbert transform on the acquired face image and the two-dimensional analysis signal represented by the two-dimensional Hilbert transform" is used to solve the related art.
  • the problem of poor performance and high complexity of the face recognition method simplifies the complexity of the system implementation and improves the accuracy and accuracy of the system.
  • FIG. 1 is a flowchart of a method for extracting facial features according to an embodiment of the present invention
  • FIG. 2 is a flowchart of a method for authenticating a face feature according to an embodiment of the present invention
  • FIG. 3 is a structural block diagram of an apparatus for extracting facial features according to an embodiment of the present invention.
  • FIG. 4 is a structural block diagram of an apparatus for extracting facial features in accordance with a preferred embodiment of the present invention.
  • FIG. 5 is a structural block diagram of an authentication apparatus for a face feature according to an embodiment of the present invention.
  • FIG. 6 is a schematic flow chart of a face authentication method based on a two-dimensional Hilbert transform according to a preferred embodiment of the present invention
  • FIG. 7 is a schematic diagram of a convolution template coordinate system based on a two-dimensional Hilbert transform, in accordance with a preferred embodiment of the present invention.
  • FIG. 8 is a three-dimensional view of a frequency domain representation of a three-scale log-Gabor filter in accordance with a preferred embodiment of the present invention.
  • FIG. 9 is a schematic diagram showing a two-dimensional analytical representation of a face photo at a scale in accordance with a preferred embodiment of the present invention.
  • FIG. 10 is a schematic diagram of a self-encoding neural network in accordance with another preferred embodiment of the present invention.
  • FIG. 11 is a schematic diagram of an ROC curve on a DupI database in accordance with a preferred embodiment of the present invention.
  • the embodiment of the invention provides a face authentication method, and the face authentication method mainly includes the following steps A to D:
  • Step A performing a two-dimensional Hilbert transform on the first face image and the second face image to extract the first face feature of the first face image and the second face feature of the second face image
  • the face feature extraction method may be performed on the first face image and the second face image by using any one of the face feature extraction methods provided by the foregoing content in the embodiment of the present invention.
  • Step B Calculate the similarity between the first facial feature and the second facial feature.
  • Step C Determine whether the similarity reaches a preset threshold, wherein the preset threshold may be actually set according to the authentication accuracy requirement.
  • Step D In the case that it is determined that the similarity reaches the preset threshold, it is determined that the face represented by the first face image and the face represented by the second face image are the same face.
  • the face authentication method provided by the embodiment of the present invention extracts a face feature by performing a two-dimensional Hilbert transform on the face image, and then performs face authentication based on the extracted similarity of the face feature.
  • the authentication is based on the local statistical features of the face. Because local features have good robustness to illumination, pose, expression, etc., it can improve the performance and accuracy of face authentication, and authenticate based on local features.
  • the process and the complexity of the time and space of the algorithm are also relatively low. Therefore, the face authentication method provided by the embodiment of the present invention solves the problem of poor performance and high complexity of the face recognition method in the prior art, and simplifies the system implementation. The complexity increases the accuracy and accuracy of the system.
  • the first face image and the second face image are extracted in the same manner. If any of the face images in the first face image and the second face image is the target face image,
  • the specific method of extracting the target face features of the target face image is as follows:
  • the target image matrix f(x, y) is subjected to two-dimensional Hilbert transform to obtain a target two-dimensional analysis signal, and the obtained target two-dimensional analysis signal is the target facial feature of the target human face image.
  • the specific transformation manner may adopt the following steps SE1 to SE3:
  • SE1 acquiring a convolution template for performing a two-dimensional Hilbert transform on the target image matrix f(x, y), which can be obtained by transforming the frequency domain representation of the two-dimensional Hilbert transform.
  • the frequency ⁇ can be expressed as ( ⁇ x , ⁇ y ), where ⁇ x and ⁇ y are respectively two components of ⁇ (ie, abscissa and ordinate), therefore, two-dimensional
  • , H y -j ⁇ y /
  • the convolution template is a time domain
  • Convolution operation is performed on the target image matrix f(x, y) and the convolution template (h x , h y ) to obtain a convolution result.
  • the target image matrix f(x, y) is separately convoluted.
  • the template (h x , h y ) h x and h y are respectively convoluted, and the result of the convolution operation is used as the two imaginary parts of the target two-dimensional analysis signal f x (x, y) and f y (x, y ),which is:
  • i, j are integers.
  • u and v represent the spatial position of (H x , H y ) in the Cartesian coordinate system with (x, y) as the origin, where u and v are in the range of w is a positive integer, which is the size of the convolution window.
  • the size of the convolution template is w ⁇ w.
  • the face authentication method provided by the embodiment of the present invention may also determine the target two.
  • the energy information of the dimensionally analyzed signal is the local amplitude A
  • the structural information of the target two-dimensional analytical signal is determined to be the local phase ⁇
  • the geometric information of the target two-dimensional analytical signal is determined to be the local direction ⁇
  • sign( ⁇ ) is a symbolic function; where f(x, y) is the real part of the target two-dimensional analytical signal, and f x (x, y) and f y (x, y) are the target two-dimensional respectively. Parse the two imaginary parts of the signal.
  • the two-dimensional analytical signal adds geometric information indicating the local main direction, and different components describe a signal from different angles, so that these components can be used to represent a signal, and a person will be realized.
  • the face image signal can be represented by a local amplitude A, a local phase ⁇ , and a local direction ⁇ .
  • the embodiment of the present invention further provides a specific manner for calculating the similarity between the first facial feature and the second facial feature, which mainly includes the following steps SF1 to SF3:
  • the image matrix of the face image is subjected to two-dimensional analysis signals obtained by two-dimensional Hilbert transform, and the second two-dimensional analysis signal is two-dimensional analysis obtained by performing two-dimensional Hilbert transform on the image matrix of the second face image.
  • the signal, the component of the two-dimensional analytical signal mainly includes the local amplitude A, the local phase ⁇ and the local direction ⁇ .
  • the weight values of the local amplitude A, the local phase ⁇ , and the local direction ⁇ may be sequentially set to 0.25, 0.45, and 0.3, respectively.
  • SF3 Determine the weighted average as the similarity between the first facial feature and the second facial feature.
  • the face authentication method provided by the embodiment of the present invention further includes: filtering the first face image and the second face image by using a band pass filter, and then, filtering the first face image and the second person
  • the face image is subjected to a two-dimensional Hilbert transform to extract a first face feature of the first face image and a second face feature of the second face image, wherein the band pass filter comprises a log-Gabor filter.
  • the frequency domain response of the log-Gabor filter can be expressed as:
  • ⁇ 0 is the center frequency and ⁇ is the scale factor of the bandwidth.
  • ⁇ / ⁇ 0 is set to be a constant.
  • log-Gabor is a band-pass filter, in order to more fully describe a signal, it is necessary to extract different frequency components, thereby requiring log-Gabor filters of different scales.
  • the parameters ⁇ and ⁇ 0 can be rewritten as:
  • ⁇ min is the shortest wavelength
  • [mu] is a multiplier on the wavelength
  • s is an index scale
  • ⁇ ratio is the ratio ⁇ 0 and [sigma] of ⁇ / ⁇ 0.
  • F( ⁇ ) and F -1 ( ⁇ ) represent Fourier and inverse Fourier transform, respectively.
  • the embodiment of the present invention further provides a face authentication device, which can be used to perform the face authentication method provided by the foregoing content of the embodiment of the present invention, and the face authentication device provided by the embodiment of the present invention
  • the method mainly includes an extracting unit, a calculating unit, a judging unit and a first determining unit, wherein:
  • the extracting unit is configured to perform a two-dimensional Hilbert transform on the first face image and the second face image to extract the first face feature of the first face image and the second face feature of the second face image
  • the extracting unit may perform face feature extraction on the first face image and the second face image by using any one of the face feature extraction methods provided by the foregoing content of the embodiment of the present invention.
  • the calculation unit is configured to calculate the similarity between the first facial feature and the second facial feature.
  • the determining unit is configured to determine whether the similarity reaches a preset threshold, wherein the preset threshold may be actually set according to the authentication accuracy requirement.
  • the first determining unit is configured to determine, in the case that the determining unit determines that the similarity reaches the preset threshold, determining that the face represented by the first face image and the face represented by the second face image are the same face.
  • the face authentication device extracts a face feature by performing a two-dimensional Hilbert transform on the face image, and then performs face authentication based on the extracted similarity of the face feature.
  • the authentication is based on the local statistical features of the face. Because local features have good robustness to illumination, pose, expression, etc., it can improve the performance and accuracy of face authentication, and authenticate based on local features.
  • the process and the complexity of the time and space of the algorithm are also relatively low. Therefore, the face authentication method provided by the embodiment of the present invention solves the problem of poor performance and high complexity of the face recognition method in the prior art, and simplifies the system implementation. The complexity increases the accuracy and accuracy of the system.
  • the extracting unit mainly includes an acquiring sub-unit and a transform sub-unit, wherein the extracting unit performs the same feature extraction on the first facial image and the second facial image, and the first facial image and the first Any face image in the two-face image is the target face image, and the structure and function of the acquisition sub-unit and the transformation sub-unit are as follows:
  • the transform subunit is used to perform a two-dimensional Hilbert transform on the target image matrix f(x, y) to obtain a target two-dimensional analytical signal, and the obtained target two-dimensional analytical signal is the target facial feature of the target facial image. .
  • the transformation subunit mainly includes an acquisition module, an operation module, and a determination module, wherein:
  • the obtaining module is configured to obtain a convolution template for performing a two-dimensional Hilbert transform on the target image matrix f(x, y), and the convolution module can transform the frequency domain representation of the two-dimensional Hilbert transform.
  • , H y - J ⁇ y /
  • the operation module is configured to perform a convolution operation on the target image matrix f(x, y) and the convolution template (h x , h y ) to obtain a convolution result, specifically, respectively, the target image matrix f(x, y)
  • h x and h y are respectively convoluted, and the result of the convolution operation is used as the two imaginary parts f x (x, y) and f y (x) of the target two-dimensional analysis signal.
  • y ie:
  • i, j are integers.
  • u and v represent the spatial position of (H x , H y ) in the Cartesian coordinate system with (x, y) as the origin, where u and v are in the range of w is a positive integer, which is the size of the convolution window.
  • the size of the convolution template is w ⁇ w.
  • the face authentication apparatus may further include a second determining unit, a third determining unit, and a fourth determining unit, where the first determining unit determines that the target image matrix is the real part of the target two-dimensional analysis signal. And determining that the convolution result is the imaginary part of the target two-dimensional analysis signal, the second determining unit is configured to determine that the energy information of the target two-dimensional analytical signal is the local amplitude A, and the third determining unit is configured to determine the target two-dimensional analytical signal.
  • the structural information is a local phase ⁇
  • the fourth determining unit is configured to determine that the geometric information of the target two-dimensional analytical signal is a local direction ⁇ , wherein
  • sign( ⁇ ) is a symbolic function; where f(x, y) is the real part of the target two-dimensional analytical signal, and f x (x, y) and f y (x, y) are the target two-dimensional respectively. Parse the two imaginary parts of the signal.
  • the two-dimensional analytical signal adds geometric information indicating the local main direction, and different components describe a signal from different angles, so that these components can be used to represent a signal, and a person will be realized.
  • the face image signal can be represented by a local amplitude A, a local phase ⁇ , and a local direction ⁇ .
  • the calculating unit in the face authentication device mainly includes a first calculating subunit, a second calculating subunit, and a determining subunit, wherein:
  • the first calculation subunit is configured to calculate a component similarity of each component of the first two-dimensional analysis signal and each corresponding component of the second two-dimensional analysis signal to obtain a plurality of component similarities, wherein the first two-dimensional analysis
  • the signal is a two-dimensional analytical signal obtained by performing a two-dimensional Hilbert transform on the image matrix of the first facial image
  • the second two-dimensional analytical signal is a two-dimensional Hilbert transform on the image matrix of the second facial image.
  • the obtained two-dimensional analytical signal, the components of the two-dimensional analytical signal mainly include a local amplitude A, a local phase ⁇ , and a local direction ⁇ .
  • the second calculating subunit is configured to calculate a weighted average value of the plurality of component similarities according to the preset weight value.
  • the weight values of the local amplitude A, the local phase ⁇ , and the local direction ⁇ may be sequentially set to 0.25, 0.45 and 0.3.
  • the determining subunit is configured to determine the weighted average as the similarity of the first facial feature and the second facial feature.
  • the face authentication device further includes a filtering unit: performing a two-dimensional Hilbert transform on the first face image and the second face image in the extracting unit to extract the first face image Before the first face feature and the second face feature of the second face image, the filtering unit filters the first face image and the second face image by using a band pass filter, and correspondingly, the extracted unit pairs the filtered The first face image and the second face image are subjected to two-dimensional Hilbert transform, and the first face feature of the first face image and the second face feature of the second face image are extracted, wherein the band pass The filter includes a log-Gabor filter.
  • the frequency domain response of the log-Gabor filter can be expressed as:
  • ⁇ 0 is the center frequency and ⁇ is the scale factor of the bandwidth.
  • ⁇ / ⁇ 0 is set to be a constant.
  • log-Gabor is a band-pass filter, in order to more fully describe a signal, it is necessary to extract different frequency components, thereby requiring log-Gabor filters of different scales.
  • the parameters ⁇ and ⁇ 0 can be rewritten as:
  • ⁇ min is the shortest wavelength
  • [mu] is a multiplier on the wavelength
  • s is an index scale
  • ⁇ ratio is the ratio ⁇ 0 and [sigma] of ⁇ / ⁇ 0.
  • F( ⁇ ) and F -1 ( ⁇ ) represent Fourier and inverse Fourier transform, respectively.
  • FIG. 1 is a flowchart of a method for extracting facial features according to an embodiment of the present invention. As shown in FIG. 1, the method includes:
  • Step S102 performing a two-dimensional Hilbert transform on the acquired face image
  • step S104 the face image subjected to the two-dimensional Hilbert transform is represented by a two-dimensional analysis signal.
  • the method of "two-dimensional Hilbert transform of the acquired face image and the two-dimensional analysis signal represented by the two-dimensional Hilbert transform" is used to solve the related art.
  • the face recognition method has poor performance and high complexity, which simplifies the complexity of the system implementation and improves the accuracy and accuracy of the system.
  • the one-dimensional Hilbert transform (ie 1D Hilbert) can be expressed as:
  • ⁇ represents the frequency and sign( ⁇ ) is the sign function, ie the sign (positive or negative) of the variable ⁇ .
  • sign( ⁇ ) is equal to 1
  • sign( ⁇ ) is equal to -1.
  • Its function is to reduce the phase of the positive frequency component of the signal by a quarter cycle while increasing the phase of the negative frequency component by a quarter cycle.
  • the 1D Hilbert transform is extended to two-dimensional, and applied to the field of image processing and pattern recognition, that is, a two-dimensional Hilbert transform based on signals to represent a human face.
  • a sparse self-encoding algorithm can also be utilized.
  • the transformed face is encoded to authenticate the face as the final feature.
  • the time domain representation of the two-dimensional Hilbert transform used in step S102 can be:
  • h x, h y are H x, H y the result of the inverse Fourier transform
  • H x, H y are the two-dimensional frequency domain formula Hilbert transform -j ⁇ /
  • the abscissa component and the ordinate component, u and v both take real numbers, and (u, v) represents the spatial position within the Cartesian coordinate system.
  • u is the abscissa of the spatial point (u, v) in the Cartesian coordinate system
  • v is the ordinate of the spatial point (u, v) in the Cartesian coordinate system.
  • the two-dimensional analysis signal of the face image after the two-dimensional Hilbert transform is:
  • f(x, y) is the real part of the two-dimensional analysis signal
  • f x (x, y) and f y (x, y) are respectively two imaginary parts of the two-dimensional analysis signal
  • the convolution template size is w ⁇ w, where w is a positive integer and is the size of the convolution window.
  • the two-dimensional analysis signal of the face image may be represented by three components of the local amplitude A, the local phase ⁇ and the local direction ⁇ , wherein
  • sign( ⁇ ) is a symbol function
  • f(x, y) represents the element a xy in the face image matrix of the xth row and the yth column.
  • the pixel of the xth row and the yth column in the image matrix of the face image can be represented. grayscale value. It can be seen that the two-dimensional analytical signal of the face image after two-dimensional Hilbert transform can be finally characterized by the extracted A, ⁇ , and ⁇ .
  • the acquired face image may be separately filtered by a band pass filter, wherein the band pass filter comprises a log-Gabor filter.
  • the two-dimensional analysis signal of the face image after the two-dimensional Hilbert transform can also be passed through a preset sparse self-coding neural network to obtain a corresponding optimal value, wherein the preset The weight matrix and bias matrix in the sparse self-coding neural network are obtained from face training samples.
  • the input of the preset sparse self-encoding neural network may be a direct two-dimensional analytical signal, or may be three components characterized by the real and imaginary parts of the two-dimensional analytical signal: A , ⁇ , ⁇ . That is, the corresponding output or encoding is obtained through a preset sparse self-encoding neural network.
  • FIG. 2 is a flowchart of a method for authenticating a face feature according to an embodiment of the present invention. As shown in FIG. 2, the method includes:
  • Step S202 performing a two-dimensional Hilbert transform on the acquired first face image and the second face image, respectively, and using the first face image and the second face image after the two-dimensional Hilbert transform Two-dimensional analytical signal representation;
  • Step S204 performing similarity calculation on the two-dimensional analysis signals of the first human face image and the second human face image after the two-dimensional Hilbert transform;
  • Step S206 the face authentication process is completed according to the result of the similarity calculation.
  • the first face image and the second face after the two-dimensional Hilbert transform are respectively performed by performing two-dimensional Hilbert transform on the acquired first face image and the second face image, respectively.
  • the image is represented by a two-dimensional analytical signal, and the similarity calculation is performed on the two-dimensional analysis signal of the first human face image and the second human face image after the two-dimensional Hilbert transform, and then the result is calculated according to the similarity calculation result.
  • the face authentication process solves the problem of poor performance and high complexity of the face recognition method in the related art, simplifies the complexity of the system implementation, and improves the accuracy and accuracy of the system.
  • each component of the two-dimensional analysis signal of the first face image and the second face image after the two-dimensional Hilbert transform may be firstly performed (ie, local amplitude A, local phase ⁇ And the three components of the local direction ⁇ are used for similarity calculation, and then the weighted average of each component similarity is obtained by two-dimensional analysis of the first face image and the second face image after two-dimensional Hilbert transform. The total similarity of the signals.
  • step S206 the result of the similarity calculation may be compared with a preset threshold, and then it is determined whether the first face image and the second face image are the same person according to the comparison result.
  • the acquired first face image and the second face image are respectively filtered by a band pass filter, wherein the band pass filter may be a log-Gabor filter.
  • the two-dimensional analysis signals of the first human face image and the second human face image after the two-dimensional Hilbert transform can be respectively passed through a preset sparse self-coding neural network (ie, after sparseness)
  • the self-encoding algorithm encodes the Hilbert transformed face, and obtains the corresponding optimal values.
  • the weight matrix and the offset matrix in the preset sparse self-coding neural network are obtained by the face training sample and passed.
  • the similarity between the optimal value corresponding to the face image after the preset sparse self-encoding neural network and the optimal value corresponding to the second face image is calculated.
  • FIG. 3 is a structural block diagram of a face feature extraction apparatus according to an embodiment of the present invention.
  • the extraction apparatus includes: a two-dimensional Hilbert transform module 32, which performs two-dimensional imaging on the acquired face image. a Berbert transform; and a face image representation module 34,
  • the two-dimensional Hilbert transform module 32 is coupled to the above-described face image after the two-dimensional Hilbert transform is represented by a two-dimensional analysis signal.
  • the two-dimensional Hilbert transform module 32 performs a two-dimensional Hilbert transform on the acquired face image
  • the face image representation module 34 uses the two-dimensional Hilbert transform face image.
  • the two-dimensional analytical signal representation solves the problem of poor performance and high complexity of the face recognition method in the related art, simplifies the complexity of the system implementation, and improves the accuracy and accuracy of the system.
  • the authentication device further includes an encoding module 42 coupled to the face image representation module 34 for The two-dimensional analytical signal of the face image after the Weihilbert transform is obtained by the preset sparse self-coding neural network, and the corresponding optimal value is obtained.
  • the weight matrix and offset in the preset sparse self-coding neural network are obtained.
  • the matrix is obtained from a face training sample.
  • the identification device further comprises: a filtering module 44 coupled to the two-dimensional Hilbert transform module 32 for respectively filtering the acquired face image by a band pass filter, wherein the band pass filter comprises a log -Gabor filter.
  • a filtering module 44 coupled to the two-dimensional Hilbert transform module 32 for respectively filtering the acquired face image by a band pass filter, wherein the band pass filter comprises a log -Gabor filter.
  • FIG. 5 is a structural block diagram of an authentication apparatus for a face feature according to an embodiment of the present invention.
  • the authentication apparatus includes: a feature extraction module 52, configured to respectively represent the first person by using the above-mentioned facial feature extraction device a face image and a second face image; the similarity calculation module 54 is coupled to the feature extraction module 52, and performs similarity calculation on the first face image and the second face image represented by the extraction device of the face feature;
  • the authentication module 56 is coupled to the similarity calculation module 54 for completing the face authentication process according to the result of the similarity calculation.
  • the feature extraction module 52 respectively displays the first face image and the second face image using the face feature extraction device, and the similarity calculation module 54 indicates the first person who has passed the face feature extraction device.
  • the face image and the second face image are similarly calculated, and the authentication module 56 completes the face authentication process according to the result of the similarity calculation described above, and solves the problem that the face recognition method has poor performance and high complexity in the related art, and simplifies the problem.
  • the complexity of the system implementation increases the accuracy and accuracy of the system.
  • FIG. 6 is a schematic flow chart of a face authentication method based on a two-dimensional Hilbert transform according to a preferred embodiment of the present invention.
  • a vertical dotted line divides the flow into two processes, and the left flow is called sparse self-encoding.
  • the training process of the neural network is trained to obtain the weight matrix W and the bias matrix b. This training process is independent. Once the training process is over, after getting W and b, you can only use W and b when doing authentication.
  • the above method of face authentication can be divided into two parts: feature extraction and feature comparison.
  • any analytical signal z(t) can be expressed by the following equation in the continuous time domain:
  • Z( ⁇ ) is a complex coefficient that controls the amplitude and phase of the sinusoidal e j ⁇ t of the positive frequency complex at frequency ⁇ .
  • a real-valued sinusoid for example, the expression A cos( ⁇ t+ ⁇ )
  • a cos( ⁇ t+ ⁇ ) can be converted to a positive-frequency complex simply by adding a phase-consistent component A sin( ⁇ t+ ⁇ ) as a complex part.
  • a real-valued signal can be transformed into an analytical signal by adding a phase-complex complex component.
  • the main function of the Hilbert transform is to reduce the phase of each positive frequency by a quarter cycle, while increasing the phase of each negative frequency by a quarter cycle.
  • ⁇ t ⁇ x ⁇ denote the output at time t after the Hilbert transform of the signal x
  • the signal z(t) becomes a complex analytical signal corresponding to the signal x(t). That is to say through the following formula:
  • the negative frequency of the represented signal z(t) is zero, in other words, by the equation (4), the negative frequency portion of x(t) is filtered out.
  • f(x, y) is used to represent the input image matrix of a face image of size M ⁇ N:
  • a xy is an element in the image matrix, x is a positive integer not greater than M, y is a positive integer not greater than N, M represents the number of rows of the image matrix, and N represents the number of columns of the image matrix, all being positive integers.
  • the frequency ⁇ can be expressed as ( ⁇ x , ⁇ y ), where ⁇ x and ⁇ y are the two components of ⁇ (ie, the abscissa and the ordinate), respectively, and the two-dimensional Hilbert transform can be performed.
  • decomposed into two components, H x -j ⁇ x /
  • and H y -j ⁇ y /
  • the face image f(x, y) is convoluted with h x and h y respectively to obtain two imaginary parts f x (x, y) and f y (x, y) of the two-dimensional analytical representation of the image. ),which is:
  • u and v represent the spatial position in the Cartesian coordinate system with (x, y) as the origin, where u and v have a range of values w is a positive integer, which is the size of the convolution window.
  • the size of the convolution template is w ⁇ w.
  • 7 is a schematic diagram of a convolution template coordinate system based on a two-dimensional Hilbert transform according to a preferred embodiment of the present invention. As shown in FIG. 7, taking w as an example, the figure is the above Cartesian coordinate system (planar right angle) Coordinate system), where the current point (x, y) is taken as the origin of the coordinate system.
  • f(x, y) is the real part of the two-dimensional analytical representation of the image
  • f x (x, y) and f y (x, y) are the two imaginary parts. Based on these real and imaginary parts, the original two-dimensional signal f(x, y) can be decomposed into the following three components: local amplitude A, local phase ⁇ and local direction ⁇ :
  • ⁇ describes the energy information of the signal
  • ⁇ describes the structural information of the signal
  • ⁇ describes the geometric information of the signal.
  • the two-dimensional bits add geometric information indicating the local main direction, and different components describe a signal from different angles, so that these components can be used to represent a signal, that is, a human face.
  • the image signal can be represented by local amplitude A, local phase ⁇ and local direction ⁇ , while local amplitude A, local phase ⁇ and local direction ⁇ are again f(x, y), f in f A (x, y)
  • the three components x (x, y) and f y (x, y) are calculated.
  • this embodiment extends the one-dimensional Hilbert transform to two-dimensional and applies it to the feature extraction of face recognition.
  • the 2D Hilbert transform the three components A, ⁇ and ⁇ of the face image can be extracted.
  • band-pass filtering helps maintain the "invariant-covariant" nature of signal decomposition, which represents energy (local amplitude) and structure (local phase and direction). Is independent information.
  • Gabor filters are a traditional choice for obtaining local frequency domain information, and they provide a better fusion of spatial domain location and frequency domain information.
  • they have two main limitations: on the one hand, its maximum bandwidth is limited to a range of approximately one frequency, on the other hand, if one wants to obtain the largest feature localization in a wide spectral range, Gabor is not the most. Excellent, and because the mean value of Gabor is not zero, it is susceptible to DC components.
  • Field proposed the log-Gabor feature.
  • the log-Gabor filter suppresses the DC component and can solve the bandwidth limitation of the traditional Gabor filter, while its response on the logarithmic frequency scale is still Gaussian, which will allow more high frequency information to be preserved. .
  • the frequency domain response of the log-Gabor filter can be expressed as:
  • ⁇ 0 is the center frequency and ⁇ is the scale factor of the bandwidth.
  • log-Gabor is a band-pass filter, in order to more fully describe a signal, it is necessary to extract different frequency components, thereby requiring log-Gabor filters of different scales.
  • the parameters ⁇ and ⁇ 0 can be rewritten as:
  • ⁇ min is the shortest wavelength
  • ⁇ is a multiplier on the wavelength
  • s is an index scale
  • ⁇ ratio is the ratio ⁇ 0 and [sigma] of ⁇ / ⁇ 0.
  • F ( ⁇ ) and F -1 ( ⁇ ) denote the Fourier and inverse Fourier transform.
  • FIG. 8 is a three-dimensional view of a frequency domain representation of a three-scale log-Gabor filter, as shown in FIG. 8, divided into upper and lower rows, the first row representing different scales from left to right, in accordance with a preferred embodiment of the present invention.
  • the top view of the log_Gabor function in the frequency domain can clearly see that after filtering by different scales of log_Gabor function, the image retains information of different frequency segments; the second row sequentially represents the three-dimensional representation of the image of the first line, which can be clearly Characterizes the bandpass characteristics of the log_Gabor function.
  • FIG. 9 is a schematic diagram showing a two-dimensional analytical representation of a face photo at a scale according to a preferred embodiment of the present invention. As shown in FIG. 9, through the above two steps, a two-dimensional analytical representation of a face can be obtained.
  • Sparse self-encoding learning can find the intrinsic essential features of these samples from a large number of unlabeled samples through learning algorithms, thus alleviating the small sample problem in statistical learning.
  • sparse self-encoding learning can automatically find existence.
  • the intrinsic feature in the data is essentially a process of feature learning.
  • a self-encoding neural network is an unsupervised learning process that utilizes a back propagation algorithm and makes the learning goal equal to the input value.
  • FIG. 10 is a schematic diagram of a self-encoding neural network, as shown in FIG. 10, which is an example of a self-encoding neural network, in accordance with another preferred embodiment of the present invention:
  • the goal of the self-encoding neural network is to obtain a hypothesis h w,b (x) ⁇ x from the input layer to the output layer by the stochastic gradient descent method, ie it attempts to approximate an identity function so that the output of the network Close to input x, where W is the weight matrix and b is the bias matrix.
  • the stochastic gradient descent method is an optimization problem for the required solution, randomly moving along the direction of the gradient reduction, and finally reaching the final local or global optimal solution of the problem to be optimized.
  • the number m of neurons of the hidden layer L2 is smaller than the dimension n of the training sample, it is still possible to recover the n-dimensional sample well using data smaller than n. That is to say, the m essential features of the n-dimensional samples can be obtained by learning, so that the m features can be utilized to characterize the n-dimensional training data.
  • the self-encoding neural network can well characterize the essential characteristics of the sample.
  • This constraint is the activation degree of the hidden layer neurons, which can add a sparsity limit to this activation, that is, to ensure that the hidden layer neurons are sparsely activated in one propagation, so that the sparse self-coding network can be used to obtain the training data. Structural characteristics.
  • Figure 10 is a typical neural network model showing a mapping from the input layer (Layer) L1 through the hidden layer (Layer) L2 to the output layer (Layer) L3.
  • the relationship from the input layer L1 to the hidden layer L2 can be passed. Equation (12) is obtained.
  • the hidden layer L2 is also the input layer of the output layer L3. Therefore, the relationship from the hidden layer L2 to the output layer L3 is also given by the formula (12).
  • the content of the output layer needs to be manually calibrated during the general neural network training process, but the sparse self-encoding neural network used in this embodiment does not need to be manually calibrated.
  • the layers are equal to the input layer, ie, the network output is used to simulate the essential characteristics of the input.
  • the weight matrix W and the offset b obtained by the sparse self-encoding learning algorithm that is, the minimum value of the defined loss function are solved by the stochastic gradient descent method
  • the three components in the formula (8) are respectively used as the input of the network, thereby Will get their respective output, which is the encoding of the two-dimensional face analysis.
  • x, y is the vector representation of the encoded feature
  • ⁇ x, y> represents the inner product of the vector x, y. That is, the cosine of its angle is used to measure the similarity degree of the two vectors. Then, the total similarity is obtained by the weighted average of each component similarity, wherein the weights are: amplitude 0.25, phase 0.45, direction 0.3.
  • the total similarity is compared with a predetermined threshold to give a determination as to whether it is the same person.
  • the two-dimensional Hilbert transform is used as the carrier, and the analytical representation of the face image is obtained, and the obtained analytical representation is encoded by the sparse self-encoding learning algorithm. It not only utilizes the local texture information to be robust to illumination, pose, expression, etc., but also has low time and space complexity.
  • the four sub-libraries Fb, Fc, DupI, and DupII achieved the preferred recognition rates of 99.7%, 99.5%, 93.6%, and 91.5%, respectively.
  • FIG. 11 is a schematic diagram of an ROC curve on a DupI database according to a preferred embodiment of the present invention.
  • the misrecognition rate is one thousand on a sub-library DupI with a large change in shooting time, illumination, and expression. In one of the cases, the recognition rate reached 93.21%.
  • the face can be first filtered by using a bandpass filter log-Gabor of different scales, and the filtered image is obtained by using a two-dimensional Hilbert transform to obtain a two-dimensional analytical representation at different resolutions;
  • the sparse self-learning coding is performed on the analytical representation of the face, and the essential features of the data are automatically searched by the learning method, thereby obtaining a more accurate representation of the face.
  • the two-dimensional analytical signal obtained by the two-dimensional Hilbert transform of the signal is used to represent the face, and the transformed face can be encoded by the sparse self-encoding algorithm as the final feature.
  • the sparse self-encoding algorithm to authenticate the face, that is, to use the local statistical features of the face, not only can solve the influence of the illumination attitude on the result to a certain extent, but the time and space complexity of the algorithm are lower than the Gabor feature, but the generalization And the recognition result is better than the Gabor feature.
  • modules or steps of the present invention described above can be implemented by a general-purpose computing device that can be centralized on a single computing device or distributed across a network of multiple computing devices. Alternatively, they may be implemented by program code executable by the computing device so that they may be stored in the storage device by the computing device, or they may be fabricated into individual integrated circuit modules, or Multiple modules or steps are made into a single integrated circuit module. Thus, the invention is not limited to any specific combination of hardware and software.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Health & Medical Sciences (AREA)
  • General Physics & Mathematics (AREA)
  • Multimedia (AREA)
  • Oral & Maxillofacial Surgery (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • General Health & Medical Sciences (AREA)
  • Computer Security & Cryptography (AREA)
  • Biomedical Technology (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Biodiversity & Conservation Biology (AREA)
  • Human Computer Interaction (AREA)
  • Molecular Biology (AREA)
  • Computer Hardware Design (AREA)
  • Software Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Image Analysis (AREA)
  • Collating Specific Patterns (AREA)

Abstract

本发明公开了一种人脸特征的提取、认证方法及装置，其中，人脸特征的提取方法，包括：对获取的人脸图像进行二维希尔伯特变换；以及将经过上述二维希尔伯特变换后的人脸图像用二维解析信号表示。通过本发明，解决了相关技术中人脸识别方法性能差、复杂度高的问题，简化了系统实现的复杂度，提高了系统的精确度和准确性。

Description

人脸特征的提取、认证方法及装置 技术领域
本发明涉及图像处理与模式识别领域，具体而言，涉及一种人脸特征的提取、认证方法及装置。
背景技术
随着网络的普及以及信息技术的飞速发展，信息安全问题越来越引起人们的重视，已经成为技术发展必须要解决的关键问题。其中，如何准确的认证一个人的身份信息是信息安全领域的重要组成部分。
人脸认证是生物特征识别的一种形式，通过有效地表征人脸，得到两幅人脸照片的特征，来判定这两张照片是否是同一个人。相比于其他的生物特征认证技术，具有友好、方便、非侵入性等特点。因此，近年来人脸认证技术成为众多科研和商业机构的研究热点。
通常情况下，人脸认证的流程主要分为三部分：人脸检测，眼睛定位与归一化，特征提取与认证。由于人脸是一个三维形变模型，而且人脸认证是以摄像机成像模型所成的照片为介质的，所以认证的结果容易受到光照、姿态、表情和遮挡等外界因素的影响。同时由于人脸认证技术涉及到了模式识别，统计学习，机器视觉，应用数学与信息科学等众多交叉学科，再加上其广泛的应用前景，受到了越来越多的关注。
最初的人脸识别算法是利用人脸特征点之间像素的差别来做的，这种方法对背景光照等效果非常差，近年来，人脸识别的研究取得了飞速的进展，有一大批新的人脸识别算法被提出，不同的人有不同的分类方法，按照识别过程中人脸的表征方法，可以把人脸识别方法大体分为：基于几何特征的方法、基于整体特征的方法、基于局部纹理特征的方法。
(1)基于几何特征的方法
基于几何特征的人脸表示主要是提取人脸的关键点，比如眼睛、鼻子、嘴巴、下巴、眉毛等。然后利用这些关键点的相对位置来计算距离、面积的比率等，然后利用这些比率作为人脸的表征。该方法在定位准确的前提下，对光照变化有很强的鲁棒 性，而且特征量比较小。但是，它有一个天生的缺点，那就是特别依赖于关键点的定位，而且关键点的相对位置容易受到表情以及姿态的影响。
(2)基于整体特征的方法
基于整体特征的人脸表示的一个主要的代表就是基于稀疏表示的方法。主要思想是直接利用一个很大的数据库作为人脸的一组基，然后将要进行比对的人脸向这组基进行稀疏投影，得到每一个人脸在这组基下的投影系数，然后利用这组投影系数来表征人脸。实质上就是利用这组基的稀疏线性组合来表征一个待识别的人脸。该方法当用来做基的数据库非常大且是闭集测试的时候，能取得相当不错的识别效果，但是，当训练库比较小的时候，或者进行开集测试时，对库外人的投影效果就不是很好。也就是说算法的泛化性不强。
(3)基于局部纹理特征的方法
基于局部纹理特征的一个典型的例子是基于Gabor特征的人脸表示。利用不同尺度和方向的Gabor核在图像上进行滤波，然后针对滤波后的图像做比对。Gabor特征有效的兼顾了信号的时域和频域的表示。是目前最为流行的特征表示之一。然而，该方法最大的一个问题就是数据量非常大，一个利用5个尺度8个方向的Gabor核就会把一幅照片变为40幅特征，提高了存储和计算的复杂度。
针对相关技术中人脸识别方法性能差、复杂度高的问题，目前尚未提出有效的解决方案。
发明内容
本发明的主要目的在于提供一种人脸特征的提取、认证方案，以至少解决上述问题。
根据本发明的一个方面，提供了一种人脸特征的提取方法，包括：对获取的人脸图像进行二维希尔伯特变换；以及将经过上述二维希尔伯特变换后的人脸图像用二维解析信号表示。
优选地，对获取的人脸图像进行二维希尔伯特变换的公式为：
Figure PCTCN2014091046-appb-000001
其中，f(x,y)＝axy，用于表示M行N列的人脸图像的图像矩阵，axy为图像矩阵中的元素，x为不大于M的正整数，y为不大于N的正整数，fx(x,y)和fy(x,y)为二维解析信号的两个虚部；hx、hy分别为Hx、Hy的逆傅立叶变换的结果，而Hx、Hy分别为二维希尔伯特变换的频域公式-jω/||ω||分解出的横坐标分量和纵坐标分量。
优选地，二维希尔伯特变换的时域公式为：
Figure PCTCN2014091046-appb-000002
其中，在hx、hy分别与f(x,y)卷积时，(u,v)用于表示以(x,y)为原点的笛卡尔坐标系内的空间位置，u和v的取值范围均为
Figure PCTCN2014091046-appb-000003
w为正整数，为卷积窗口的大小。
优选地，将经过二维希尔伯特变换后的人脸图像用二维解析信号表示包括：通过局部振幅A，局部相位φ和局部方向θ三个分量表示人脸图像的二维解析信号，其中，
Figure PCTCN2014091046-appb-000004
用于表示二维解析信号的能量信息；
Figure PCTCN2014091046-appb-000005
用于表示二维解析信号的结构信息，sign(·)为符号函数；
θ＝arctan(fy(x,y)/fx(x,y))，用于表示二维解析信号的几何信息；
且经过二维希尔伯特变换后的人脸图像的二维解析信号为：
fA(x,y)＝(f(x,y),fx(x,y),fy(x,y))，
其中，f(x,y)为二维解析信号的实部，fx(x,y)和fy(x,y)分别为二维解析信号的两个虚部。
优选地，对获取的人脸图像进行二维希尔伯特变换之前，该方法还包括：分别将获取的人脸图像经过带通滤波器进行滤波，其中，带通滤波器包括log-Gabor滤波器。
优选地，将经过二维希尔伯特变换后的人脸图像用二维解析信号表示之后，该方法还包括：将经过二维希尔伯特变换后的人脸图像的二维解析信号通过预设的稀疏自编码神经网络，得到相应的最优值，其中，预设的稀疏自编码神经网络中的权重矩阵和偏置矩阵由人脸训练样本获得。
根据本发明的另一方面，提供了一种采用上述人脸特征的提取方法进行人脸特征的认证方法，包括：分别使用人脸特征的提取方法表示第一人脸图像和第二人脸图像；对经过人脸特征的提取方法表示后的第一人脸图像和第二人脸图像进行相似度计算；以及根据上述相似度计算的结果完成人脸认证过程。
优选地，对经过人脸特征的提取方法表示后的第一人脸图像和第二人脸图像进行相似度计算包括：对经过人脸特征的提取方法表示后的第一人脸图像和第二人脸图像的每个分量进行相似度计算；将每个分量相似度的加权平均得到人脸特征的提取方法表示后的第一人脸图像和第二人脸图像的总的相似度。
根据本发明的又一方面，还提供了一种人脸特征的提取装置，该提取装置包括：二维希尔伯特变换模块，对获取的人脸图像进行二维希尔伯特变换；以及人脸图像表示模块，用于将经过二维希尔伯特变换后的人脸图像用二维解析信号表示。
根据本发明的再一方面，提供了一种采用上述人脸特征的提取装置进行人脸特征的认证装置，该认证装置包括：特征提取模块，用于分别使用人脸特征的提取装置表示第一人脸图像和第二人脸图像；相似度计算模块，对经过人脸特征的提取装置表示的第一人脸图像和第二人脸图像进行相似度计算；以及认证模块，用于根据相似度计算的结果完成人脸认证过程。
通过本发明，采用“对获取的人脸图像进行二维希尔伯特变换，将经过二维希尔伯特变换后的人脸图像用二维解析信号表示”的方式，解决了相关技术中人脸识别方法性能差、复杂度高的问题，简化了系统实现的复杂度，提高了系统的精确度和准确性。
附图说明
此处所说明的附图用来提供对本发明的进一步理解，构成本申请的一部分，本发明的示意性实施例及其说明用于解释本发明，并不构成对本发明的不当限定。在附图中：
图1是根据本发明实施例的人脸特征的提取方法的流程图；
图2是根据本发明实施例的人脸特征的认证方法的流程图；
图3是根据本发明实施例的人脸特征的提取装置的结构框图；
图4是根据本发明优选实施例的人脸特征的提取装置的结构框图；
图5是根据本发明实施例的人脸特征的认证装置的结构框图；
图6是根据本发明一个优选实施例的基于二维Hilbert变换的人脸认证方法的流程示意图；
图7是根据本发明一个优选实施例的基于二维Hilbert变换的卷积模板坐标系的示意图；
图8是根据本发明一个优选实施例的三个尺度的log-Gabor滤波器频域表示的三维视图；
图9是根据本发明一个优选实施例的人脸照片在一个尺度下的二维解析表示的示意图；
图10是根据本发明另一优选实施例的一个自编码神经网络的示意图；
图11是根据本发明一个优选实施例的在DupI数据库上的ROC曲线的示意图。
具体实施方式
下文中将参考附图并结合实施例来详细说明本发明。需要说明的是，在不冲突的情况下，本申请中的实施例及实施例中的特征可以相互组合。
本发明实施例提供了一种人脸认证方法，该人脸认证方法主要包括如下步骤A至步骤D：
步骤A：对第一人脸图像和第二人脸图像进行二维希尔伯特变换，以提取第一人脸图像的第一人脸特征和第二人脸图像的第二人脸特征，其中，可以采用本发明实施例上述内容所提供的任一种人脸特征的提取方法对第一人脸图像和第二人脸图像进行人脸特征提取。
步骤B：计算第一人脸特征和第二人脸特征的相似度。
步骤C：判断相似度是否达到预设阈值，其中，预设阈值可以根据认证精度要求进行实际设定。
步骤D：在判断出相似度达到预设阈值的情况下，确定第一人脸图像所表示的人脸和第二人脸图像所表示的人脸为同一人脸。
本发明实施例所提供的人脸认证方法，通过对人脸图像进行二维希尔伯特变换，来提取出人脸特征，进而基于提取出的人脸特征的相似度进行人脸认证，实现了以人脸的局部统计特征为基础进行认证，由于局部特征对光照、姿态、表情等具有良好的鲁棒性特点，因此能够提高人脸认证的性能和准确度，并且基于局部特征进行认证的过程，算法的时间和空间的复杂度也比较低，所以，本发明实施例所提供的人脸认证方法解决了现有技术中人脸识别方法性能差、复杂度高的问题，简化了系统实现的复杂度，提高了系统的精确度和准确性。
在本发明实施例中，对第一人脸图像和第二人脸图像进行特征提取的方式相同，假设第一人脸图像和第二人脸图像中任一人脸图像为目标人脸图像，则对目标人脸图像的目标人脸特征提取的具体方式如下：
首先，获取表示目标人脸图像的目标图像矩阵，其中，可以以f(x,y)表示一幅大小为M×N的目标人脸图像的目标图像矩阵，f(x,y)＝axy，axy为目标图像矩阵中的元素，x为不大于M的正整数，y为不大于N的正整数，M表示图像矩阵的行数，N表示图像矩阵的列数，M和N均为正整数。
然后，对目标图像矩阵f(x,y)进行二维希尔伯特变换，得到目标二维解析信号，所得到的目标二维解析信号即是目标人脸图像的目标人脸特征。具体变换方式可以采用如下步骤SE1至步骤SE3：
SE1：获取对目标图像矩阵f(x,y)进行二维希尔伯特变换的卷积模板，该卷积模块可以通过对二维希尔伯特变换的频域表示式进行变换得到，在本发明实施例中，由于频率ω可以表示为(ωx,ωy)，其中，ωx和ωy分别为ω的两个分量(即横坐标和纵坐标)，因此，可以将二维希尔伯特变换的频域表示式分解为第一分量和第二分量，其中，Hx＝-jωx/||ω||，Hy＝-jωy/||ω||，-jω/||ω||为频域表示式，Hx为第一分量，Hy为第二分量，再对第一分量和第二分量做逆傅立叶变换，得到二维希尔伯特变换的时域表示式，其中，卷积模板为时域表示式。第一分量和第二分量的逆傅立叶变换如下：
Figure PCTCN2014091046-appb-000006
其中，u和v均取实数，(u,v)用于表示(Hx,Hy)在笛卡尔坐标系内的空间位置。
SE2：将目标图像矩阵f(x,y)和卷积模板(hx,hy)进行卷积运算，得到卷积结果，具体地，将目标图像矩阵f(x,y)分别和卷积模板(hx,hy)中hx和hy分别进行卷积运算，卷积运算的结果作为目标二维解析信号的两个虚部fx(x,y)和fy(x,y)，即：
Figure PCTCN2014091046-appb-000007
其中，卷积运算的过程为：
Figure PCTCN2014091046-appb-000008
Figure PCTCN2014091046-appb-000009
i，j均为整数。
在卷积过程中，u和v表示(Hx,Hy)在以(x,y)为原点的笛卡尔坐标系内的空间位置，这里u和v的取值范围均为
Figure PCTCN2014091046-appb-000010
w为正整数，为卷积窗口的大小，即做卷积时，卷积模板的大小为w×w。
SE3：确定目标图像矩阵为目标二维解析信号的实部，并确定卷积结果为目标二维解析信号的虚部，即，确定目标二维解析信号fA(x,y)＝(f(x,y),fx(x,y),fy(x,y))，其中，f(x,y)是目标二维解析信号fA(x,y)的实部，fx(x,y)和fy(x,y)是目标二维解析信号fA(x,y)的两个虚部。
进一步地，在确定目标图像矩阵为目标二维解析信号的实部，并确定卷积结果为目标二维解析信号的虚部之后，本发明实施例所提供的人脸认证方法还可以确定目标二维解析信号的能量信息为局部振幅A，确定目标二维解析信号的结构信息为局部相位φ，确定目标二维解析信号的几何信息为局部方向θ，其中，
Figure PCTCN2014091046-appb-000011
上式中，sign(·)为符号函数；其中，f(x,y)为目标二维解析信号的实部，fx(x,y)和fy(x,y)分别为目标二维解析信号的两个虚部。
可见，相比于一维的情况，二维解析信号增加了表明局部主方向的几何信息，不同的分量从不同的角度来描述一个信号，从而可以利用这些分量来表征一个信号，实现将一个人脸图像信号可以通过局部振幅A，局部相位φ和局部方向θ来表示。
基于上述进行人脸特征提取的方式，本发明实施例还提供了一种计算第一人脸特征和第二人脸特征的相似度的具体方式，主要包括如下步骤SF1至步骤SF3：
SF1：计算第一二维解析信号的每个分量和第二二维解析信号的对应的每个分量的分量相似度，得到多个分量相似度，其中，第一二维解析信号为对第一人脸图像的图像矩阵进行二维希尔伯特变换得到的二维解析信号，第二二维解析信号为对第二人脸图像的图像矩阵进行二维希尔伯特变换得到的二维解析信号，二维解析信号的分量主要包括局部振幅A，局部相位φ和局部方向θ。
SF2：按照预设权重值计算多个分量相似度的加权平均值，在本发明实施例中，可以将局部振幅A，局部相位φ和局部方向θ的权重值依次分别设置为0.25、0.45和0.3.
SF3：确定加权平均值为第一人脸特征和第二人脸特征的相似度。
进一步地，在对第一人脸图像和第二人脸图像进行二维希尔伯特变换，以提取第一人脸图像的第一人脸特征和第二人脸图像的第二人脸特征之前，本发明实施例所提供的人脸认证方法还包括：利用带通滤波器对第一人脸图像和第二人脸图像滤波，然后，对滤波后的第一人脸图像和第二人脸图像进行二维希尔伯特变换，提取出第一人脸图像的第一人脸特征和第二人脸图像的第二人脸特征，其中，带通滤波器包括log-Gabor滤波器。
在本发明实施例中，log-Gabor滤波器的频域响应可以表示为：
Figure PCTCN2014091046-appb-000012
其中，ω0是中心频率，σ是带宽的尺度因子，为了保证滤波器组有一个固定的形状，在本发明实施例中，设置σ/ω0为一个常数。
由于log-Gabor是带通滤波器，为了更充分地描述一个信号，则需要提取不同的频率分量，从而需要不同尺度的log-Gabor滤波器。在多尺度的log-Gabor滤波器中，参数σ和ω0可以被重写为：
σ＝σratioω0,ω0＝(λminμs-1)-1，
上式中，λmin的物理意义是最短的波长，μ是关于波长的一个乘子，s是尺度的索引，σratio是σ和ω0的比率σ/ω0。
针对一幅人脸图片f(x,y)，具体的滤波过程可以通过下式表示：
ffiltered＝F-1(F(f(x,y))*G(ω))，
其中，F(·)和F-1(·)分别表示傅立叶和逆傅立叶变换。
此外，本发明实施例还提供了一种人脸认证装置，该人脸认证装置可以用于执行本发明实施例上述内容所提供的人脸认证方法，本发明实施例所提供的人脸认证装置主要包括提取单元、计算单元、判断单元和第一确定单元，其中：
提取单元用于对第一人脸图像和第二人脸图像进行二维希尔伯特变换，以提取第一人脸图像的第一人脸特征和第二人脸图像的第二人脸特征，其中，提取单元可以采用本发明实施例上述内容所提供的任一种人脸特征的提取方法对第一人脸图像和第二人脸图像进行人脸特征提取。
计算单元用于计算第一人脸特征和第二人脸特征的相似度。
判断单元用于判断相似度是否达到预设阈值，其中，预设阈值可以根据认证精度要求进行实际设定。
第一确定单元用于在判断单元判断出相似度达到预设阈值的情况下，确定第一人脸图像所表示的人脸和第二人脸图像所表示的人脸为同一人脸。
本发明实施例所提供的人脸认证装置，通过对人脸图像进行二维希尔伯特变换，来提取出人脸特征，进而基于提取出的人脸特征的相似度进行人脸认证，实现了以人脸的局部统计特征为基础进行认证，由于局部特征对光照、姿态、表情等具有良好的鲁棒性特点，因此能够提高人脸认证的性能和准确度，并且基于局部特征进行认证的过程，算法的时间和空间的复杂度也比较低，所以，本发明实施例所提供的人脸认证方法解决了现有技术中人脸识别方法性能差、复杂度高的问题，简化了系统实现的复杂度，提高了系统的精确度和准确性。
在本发明实施例中，提取单元主要包括获取子单元和变换子单元，其中，提取单元对第一人脸图像和第二人脸图像进行特征提取的方式相同，以第一人脸图像和第二人脸图像中任一人脸图像为目标人脸图像，来说明获取子单元和变换子单元的结构和功能如下：
获取子单元用于获取表示目标人脸图像的目标图像矩阵，其中，可以以f(x,y)表示一幅大小为M×N的目标人脸图像的目标图像矩阵，f(x,y)＝axy，axy为目标图像矩阵中的元素，x为不大于M的正整数，y为不大于N的正整数，M表示图像矩阵的行数，N表示图像矩阵的列数，M和N均为正整数。
变换子单元用于对目标图像矩阵f(x,y)进行二维希尔伯特变换，得到目标二维解析信号，所得到的目标二维解析信号即是目标人脸图像的目标人脸特征。
具体地，变换子单元主要包括获取模块、运算模块和确定模块，其中：
获取模块用于获取对目标图像矩阵f(x,y)进行二维希尔伯特变换的卷积模板，该卷积模块可以通过对二维希尔伯特变换的频域表示式进行变换得到，在本发明实施例中，由于频率ω可以表示为(ωx,ωy)，其中，ωx和ωy分别为ω的两个分量(即横坐标和纵坐标)，因此，获取模块可以通过获取模块中的分解子模块将二维希尔伯特变换的频域表示式分解为第一分量和第二分量，其中，Hx＝-jωx/||ω||，Hy＝-jωy/||ω||，-jω/||ω||为频域表示式，Hx为第一分量，Hy为第二分量，再通过获取模块中的变化子模块对第一分量和第二分量做逆傅立叶变换，得到二维希尔伯特变换的时域表示式，其中，卷积模板为时域表示式。第一分量和第二分量的逆傅立叶变换如下：
Figure PCTCN2014091046-appb-000013
其中，u和v均取实数，(u,v)用于表示(Hx,Hy)在笛卡尔坐标系内的空间位置。
运算模块用于将目标图像矩阵f(x,y)和卷积模板(hx,hy)进行卷积运算，得到卷积结果，具体地，将目标图像矩阵f(x,y)分别和卷积模板(hx,hy)中hx和hy分别进行卷积运算，卷积运算的结果作为目标二维解析信号的两个虚部fx(x,y)和fy(x,y)，即：
Figure PCTCN2014091046-appb-000014
其中，卷积运算的过程为：
Figure PCTCN2014091046-appb-000015
Figure PCTCN2014091046-appb-000016
i，j均为整数。
在卷积过程中，u和v表示(Hx,Hy)在以(x,y)为原点的笛卡尔坐标系内的空间位置，这里u和v的取值范围均为
Figure PCTCN2014091046-appb-000017
w为正整数，为卷积窗口的大小，即做卷积时，卷积模板的大小为w×w。
确定模块用于确定目标图像矩阵为目标二维解析信号的实部，并确定卷积结果为目标二维解析信号的虚部，即，确定目标二维解析信号fA(x,y)＝(f(x,y),fx(x,y),fy(x,y))，其中，f(x,y)是目标二维解析信号fA(x,y)的实部，fx(x,y)和fy(x,y)是目标二维解析信号fA(x,y)的两个虚部。
进一步地，本发明实施例所提供的人脸认证装置还可以包括第二确定单元、第三确定单元和第四确定单元，在第一确定单元确定目标图像矩阵为目标二维解析信号的实部，并确定卷积结果为目标二维解析信号的虚部之后，第二确定单元用于确定目标二维解析信号的能量信息为局部振幅A，第三确定单元用于确定目标二维解析信号的结构信息为局部相位φ，第四确定单元用于确定目标二维解析信号的几何信息为局部方向θ，其中，
Figure PCTCN2014091046-appb-000018
上式中，sign(·)为符号函数；其中，f(x,y)为目标二维解析信号的实部，fx(x,y)和fy(x,y)分别为目标二维解析信号的两个虚部。
可见，相比于一维的情况，二维解析信号增加了表明局部主方向的几何信息，不同的分量从不同的角度来描述一个信号，从而可以利用这些分量来表征一个信号，实现将一个人脸图像信号可以通过局部振幅A，局部相位φ和局部方向θ来表示。
基于提取单元的上述结构，本发明实施例所提供的人脸认证装置中的计算单元主要包括第一计算子单元、第二计算子单元和确定子单元，其中：
第一计算子单元用于计算第一二维解析信号的每个分量和第二二维解析信号的对应的每个分量的分量相似度，得到多个分量相似度，其中，第一二维解析信号为对第一人脸图像的图像矩阵进行二维希尔伯特变换得到的二维解析信号，第二二维解析信号为对第二人脸图像的图像矩阵进行二维希尔伯特变换得到的二维解析信号，二维解析信号的分量主要包括局部振幅A，局部相位φ和局部方向θ。
第二计算子单元用于按照预设权重值计算多个分量相似度的加权平均值，在本发明实施例中，可以将局部振幅A，局部相位φ和局部方向θ的权重值依次分别设置为0.25、0.45和0.3.
确定子单元用于确定加权平均值为第一人脸特征和第二人脸特征的相似度。
进一步地，本发明实施例所提供的人脸认证装置还包括滤波单元：在提取单元对第一人脸图像和第二人脸图像进行二维希尔伯特变换，以提取第一人脸图像的第一人脸特征和第二人脸图像的第二人脸特征之前，滤波单元利用带通滤波器对第一人脸图像和第二人脸图像滤波，相应地，提取单元对滤波后的第一人脸图像和第二人脸图像进行二维希尔伯特变换，提取出第一人脸图像的第一人脸特征和第二人脸图像的第二人脸特征，其中，带通滤波器包括log-Gabor滤波器。
在本发明实施例中，log-Gabor滤波器的频域响应可以表示为：
Figure PCTCN2014091046-appb-000019
其中，ω0是中心频率，σ是带宽的尺度因子，为了保证滤波器组有一个固定的形状，在本发明实施例中，设置σ/ω0为一个常数。
由于log-Gabor是带通滤波器，为了更充分地描述一个信号，则需要提取不同的频率分量，从而需要不同尺度的log-Gabor滤波器。在多尺度的log-Gabor滤波器中，参数σ和ω0可以被重写为：
σ＝σratioω0,ω0＝(λminμs-1)-1，
上式中，λmin的物理意义是最短的波长，μ是关于波长的一个乘子，s是尺度的索引，σratio是σ和ω0的比率σ/ω0。
针对一幅人脸图片f(x,y)，具体的滤波过程可以通过下式表示：
ffiltered＝F-1(F(f(x,y))*G(ω))，
其中，F(·)和F-1(·)分别表示傅立叶和逆傅立叶变换。
根据本发明实施例，提供了一种人脸特征的提取方法。图1是根据本发明实施例的人脸特征的提取方法的流程图，如图1所示，该方法包括：
步骤S102，对获取的人脸图像进行二维希尔伯特(即2D Hilbert)变换；以及
步骤S104，将经过二维希尔伯特变换后的人脸图像用二维解析信号表示。
通过上述步骤，采用对“获取的人脸图像进行二维希尔伯特变换，将经过二维希尔伯特变换后的人脸图像用二维解析信号表示”的方式，解决了相关技术中人脸识别方法性能差、复杂度高的问题，化简了系统实现的复杂度，提高了系统的精确度和准确性。
一维希尔伯特(即1D Hilbert)变换可以被表示为：
H(ω)＝-jsign(ω)，
其中，ω代表的是频率，sign(ω)是符号函数，即取变量ω的符号(正或负)。当ω大于等于0时，sign(ω)等于1，反之，当ω小于0时，sign(ω)等于-1。
其作用是对信号的正频率分量相位减少四分之一周期，同时使负频率分量的相位增加四分之一周期。
本实施例是将1D Hilbert变换推广到二维，应用到图像处理与模式识别领域，即基于信号的二维Hilbert变换来表征人脸，之后在一个优选实施例中，还可以利用稀疏自编码算法对变换后的人脸进行编码，作为最终的特征来认证人脸。
优选地，在实施过程中，步骤S102中所采用的二维希尔伯特变换的时域表示可以为：
Figure PCTCN2014091046-appb-000020
其中，hx、hy分别为Hx、Hy的逆傅立叶变换的结果，Hx、Hy分别为二维希尔伯特变换的频域公式-jω/||ω||分解出的横坐标分量和纵坐标分量，u和v均取实数，(u,v)表示笛卡尔坐标系内的空间位置。其中，u为笛卡尔坐标系内空间点(u,v)的横坐标，v为笛卡尔坐标系内空间点(u,v)的纵坐标。
优选地，若经过二维希尔伯特变换后的人脸图像的二维解析信号为：
fA(x,y)＝(f(x,y),fx(x,y),fy(x,y))，
其中，f(x,y)为上述二维解析信号的实部，fx(x,y)和fy(x,y)分别为上述二维解析信号的两个虚部；
则对获取的人脸图像进行二维希尔伯特变换可以表示为：
Figure PCTCN2014091046-appb-000021
其中，f(x,y)＝axy，用于表示M行N列的人脸图像的图像矩阵，axy为该图像矩阵中的元素，x为不大于M的正整数，y为不大于N的正整数，fx(x,y)和fy(x,y)为 上述二维解析信号的两个虚部。优选地，若使用如上二维希尔伯特变换的时域表示式进行hx或hy与f(x,y)卷积，则(u,v)用于表示以(x,y)为原点的笛卡尔坐标系内的空间位置，u和v的取值范围均为
Figure PCTCN2014091046-appb-000022
即卷积模板大小为w×w，其中，w为正整数，为卷积窗口的大小。
优选地，在步骤S104中，可以通过局部振幅A，局部相位φ和局部方向θ三个分量来表示人脸图像的二维解析信号，其中，
Figure PCTCN2014091046-appb-000023
用于表示上述二维解析信号的能量信息；
Figure PCTCN2014091046-appb-000024
用于表示上述二维解析信号的结构信息，sign(·)为符号函数；
θ＝arctan(fy(x,y)/fx(x,y))，用于表示上述二维解析信号的几何信息；
这里的f(x,y)代表的是第x行第y列的人脸图像矩阵中的元素axy，在实际应用中，可以表示人脸图像的图像矩阵中第x行第y列的像素灰度值。可见，人脸图像经过二维希尔伯特变换后的二维解析信号最终可以通过提取出来的A、φ、θ来表征。
优选地，在步骤S102之前，可以分别将获取的人脸图像经过带通滤波器进行滤波，其中，带通滤波器包括log-Gabor滤波器。
优选地，在步骤S104之后，还可以将经过二维希尔伯特变换后的人脸图像的二维解析信号通过预设的稀疏自编码神经网络，得到相应的最优值，其中，预设的稀疏自编码神经网络中的权重矩阵和偏置矩阵由人脸训练样本获得。
需要说明的是，在实施过程中，预设稀疏自编码神经网络的输入可以是直接的二维解析信号，也可以是通过二维解析信号的实部和虚部表征出来的三个分量：A、φ、θ。即通过预设的稀疏自编码神经网络，得到相应的输出或者编码。
根据本发明实施例，还提供了一种人脸特征的认证方法。图2是根据本发明实施例的人脸特征的认证方法的流程图，如图2所示，该方法包括：
步骤S202，分别对获取的第一人脸图像和第二人脸图像进行二维希尔伯特变换，将经过二维希尔伯特变换后的第一人脸图像和第二人脸图像用二维解析信号表示；
步骤S204，对经过二维希尔伯特变换后的第一人脸图像和第二人脸图像的二维解析信号进行相似度计算；以及
步骤S206，根据相似度计算的结果完成人脸认证过程。
通过上述步骤，采用分别对获取的第一人脸图像和第二人脸图像进行二维希尔伯特变换，将经过二维希尔伯特变换后的第一人脸图像和第二人脸图像用二维解析信号表示，并对经过二维希尔伯特变换后的第一人脸图像和第二人脸图像的二维解析信号进行相似度计算，然后根据相似度计算的结果完成人脸认证过程的方式，解决了相关技术中人脸识别方法性能差、复杂度高的问题，简化了系统实现的复杂度，提高了系统的精确度和准确性。
优选地，在步骤S204中，可以首先对经过二维希尔伯特变换后的第一人脸图像和第二人脸图像的二维解析信号的每个分量(即局部振幅A，局部相位φ和局部方向θ这三个分量)进行相似度计算，然后将每个分量相似度的加权平均得到经过二维希尔伯特变换后的第一人脸图像和第二人脸图像的二维解析信号的总的相似度。
在实施时，在步骤S206中，可以将相似度计算的结果与预设阈值比较，然后根据比较结果判定第一人脸图像与第二人脸图像是否为同一个人。
优选地，在步骤S202之前，还可以分别将获取的第一人脸图像和第二人脸图像经过带通滤波器进行滤波，其中，带通滤波器可以为log-Gabor滤波器。
优选地，在步骤S204中，可以将经过二维希尔伯特变换后的第一人脸图像和第二人脸图像的二维解析信号分别通过预设的稀疏自编码神经网络(即经过稀疏自编码算法对Hilbert变换后的人脸进行编码)，得到各自对应的最优值，其中，预设的稀疏自编码神经网络中的权重矩阵和偏置矩阵由人脸训练样本获得，并对通过预设的稀疏自编码神经网络后的人脸图像对应的最优值与第二人脸图像对应的最优值进行相似度计算。
对应于上述识别方法，提供了一种人脸特征的提取装置。图3是根据本发明实施例的人脸特征的提取装置的结构框图，如图3所示，该提取装置包括：二维希尔伯特变换模块32，对获取的人脸图像进行二维希尔伯特变换；以及人脸图像表示模块34， 耦合至二维希尔伯特变换模块32，用于将经过上述二维希尔伯特变换后的上述人脸图像用二维解析信号表示。
通过上述提取装置，二维希尔伯特变换模块32对获取的人脸图像进行二维希尔伯特变换，人脸图像表示模块34将经过二维希尔伯特变换后的人脸图像用二维解析信号表示，解决了相关技术中人脸识别方法性能差、复杂度高的问题，简化了系统实现的复杂度，提高了系统的精确度和准确性。
图4是根据本发明优选实施例的人脸特征的提取装置的结构框图，如图4所示，该认证装置还包括：编码模块42，耦合至人脸图像表示模块34，用于将经过二维希尔伯特变换后的人脸图像的二维解析信号通过预设的稀疏自编码神经网络，得到相应的最优值，其中，预设的稀疏自编码神经网络中的权重矩阵和偏置矩阵由人脸训练样本获得。
优选地，该识别装置还包括：滤波模块44，耦合至二维希尔伯特变换模块32，用于分别将获取的人脸图像经过带通滤波器进行滤波，其中，带通滤波器包括log-Gabor滤波器。
对应于上述认证方法，还提供了一种人脸特征的认证装置。图5是根据本发明实施例的人脸特征的认证装置的结构框图，如图5所示，该认证装置包括：特征提取模块52，用于分别使用上述人脸特征的提取装置表示第一人脸图像和第二人脸图像；相似度计算模块54，耦合至特征提取模块52，对经过上述人脸特征的提取装置表示的第一人脸图像和第二人脸图像进行相似度计算；以及认证模块56，耦合至相似度计算模块54，用于根据相似度计算的结果完成人脸认证过程。
通过上述认证装置，特征提取模块52分别使用上述人脸特征的提取装置表示第一人脸图像和第二人脸图像，相似度计算模块54对经过上述人脸特征的提取装置表示的第一人脸图像和第二人脸图像进行相似度计算，以及认证模块56根据上述相似度计算的结果完成人脸认证过程，解决了相关技术中人脸识别方法性能差、复杂度高的问题，简化了系统实现的复杂度，提高了系统的精确度和准确性。
下面结合多个优选实施例和附图对上述实施例的实现过程进行详细说明。
图6是根据本发明一个优选实施例的基于二维Hilbert变换的人脸认证方法的流程示意图，如图6所示，竖虚线将其分为左右两个流程，左侧流程称为稀疏自编码神经网络的训练过程，经过训练得到权重矩阵W和偏置矩阵b。这个训练过程是独立的， 一旦训练过程结束，得到W和b之后，在做认证时，只用借助于W和b就可以。在实施过程中，上述人脸认证的方法可以分为特征提取与特征比对两部分。
(一)特征提取过程：
(a)Hilbert变换及其二维推广
由于不含负频率分量的信号被称为解析信号(Analytic Signal)，所以，在连续的时间域里面，任何一个解析信号z(t)都可以通过下式来表示：
Figure PCTCN2014091046-appb-000025
其中，Z(ω)是复系数，控制着正频率复形的正弦曲线ejωt在频率ω处的振幅和相位。
一个实值的正弦曲线，例如，表达式为A cos(ωt+φ)，可以简单地通过增加一个相正交的分量A sin(ωt+φ)作为复数部分，而转化为正频率复形的正弦曲线A exp[j(ωt+φ)]：
A exp[j(ωt+φ)]＝A cos(ωt+φ)+jA sin(ωt+φ)    (2)
也就是说一个实值的信号可以通过增加一个相正交的复数分量变换为一个解析信号。
对于更为复杂的信号来说，它们都可以表示为一些正弦曲线的和的形式。即可以建立一个这样的滤波器，它能够旋转每一个频率的正弦曲线四分之一周期，从而把任何一个信号转化为一个解析信号的形式。其中，这个变换就是Hilbert变换，其形式为：
H(ω)＝-jsign(ω)   (3)
其中，sign(ω)＝ω/||ω||，表示符号函数。
可见，Hilbert变换的主要作用就是为每一个正频率的相位减少四分之一周期，同时为每一个负频率的相位增加四分之一周期。
下面用Ηt{x}表示对信号x做Hilbert变换后在时间t的输出，则y(t)＝Ηt{x}表示对信号x做Hilbert变换之后的信号，即利用x(t)和y(t)就可以形成一个新的复信号z(t)＝x(t)+jy(t)。信号z(t)就成为对应于信号x(t)的复解析信号。也就是说通过下式：
z(t)＝x(t)+jΗt{x}   (4)
表示的信号z(t)的负频率为零，换句话说就是通过(4)式，x(t)的负频率部分被滤掉了。
为了利用解析信号的优良性质，同时结合人脸图像是一个二维信号的事实，下面对Hilbert变换向二维情况进行一个很自然的推广。
例如，首先，用f(x,y)表示输入的一幅大小为M×N的人脸图像的图像矩阵：
f(x,y)＝axy，
axy为图像矩阵中的元素，x为不大于M的正整数，y为不大于N的正整数，M表示图像矩阵的行数，N表示图像矩阵的列数，均为正整数。
其次，将其进行二维Hilbert变换。对Hilbert变换进行二维推广，过程如下：
在二维频域中，频率ω可以表示为(ωx,ωy)，其中，ωx和ωy分别为ω的两个分量(即横坐标和纵坐标)，则可以将二维Hilbert变换的频域表示式-jω/||ω||分解为两个分量，Hx＝-jωx/||ω||和Hy＝-jωy/||ω||。通过对Hx和Hy的逆傅立叶变换，可以得到二维Hilbert变换的时域表示(即hx、hy分别为Hx、Hy的逆傅立叶变换的结果)：
Figure PCTCN2014091046-appb-000026
其中，u和v均取实数，(u,v)用于表示笛卡尔坐标系内的空间位置。需要说明的是，这里的公式(5)也是下面公式(6)中进行卷积的卷积模板。
然后，将人脸图像f(x,y)分别与hx和hy进行卷积运算，可以得到图像二维解析表示的两个虚部fx(x,y)和fy(x,y)，即：
Figure PCTCN2014091046-appb-000027
其中，卷积的过程为：
Figure PCTCN2014091046-appb-000028
这里i，j均为整数；
同理，
Figure PCTCN2014091046-appb-000029
在卷积过程中，u和v表示以(x,y)为原点的笛卡尔坐标系内的空间位置，这里u和v的取值范围均为
Figure PCTCN2014091046-appb-000030
w为正整数，为卷积窗口的大小，即做卷积时，卷积模板的大小为w×w。图7是根据本发明一个优选实施例的基于二维Hilbert变换的卷积模板坐标系的示意图，如图7所示，以w取5为例，该图就是上述的笛卡尔坐标系(平面直角坐标系)，其中，当前点(x,y)做为该坐标系的原点。
可见，二维希尔伯特变换是将公式(5)代入公式(6)后完成的。
因此，对于一幅人脸图像f(x,y)，可以得到其二维解析表示为：
fA(x,y)＝(f(x,y),fx(x,y),fy(x,y))   (7)
(7)式中，f(x,y)是图像二维解析表示的实部，fx(x,y)和fy(x,y)是其两个虚部。基于这些实部和虚部，原始的二维信号f(x,y)可以以被分解为以下三个分量：局部振幅A，局部相位φ和局部方向θ：
Figure PCTCN2014091046-appb-000031
上式中Α描述了信号的能量信息，φ描述了信号的结构信息，θ描述了信号的几何信息。可见，相比于一维的情况，二维位增加了表明局部主方向的了几何信息，不同的分量从不同的角度来描述一个信号，从而可以利用这些分量来表征一个信号，即一个人脸图像信号可以通过局部振幅A，局部相位φ和局部方向θ来表示，而局部振幅A，局部相位φ和局部方向θ又是通过fA(x,y)中的f(x,y)、fx(x,y)和fy(x,y)这三个分量计算出来的。
可见，本实施例将一维Hilbert变换推广到二维，并把其应用到人脸识别的特征提取中，通过2D Hilbert变换，可以提取出人脸图像的A、φ和θ三个分量。
(b)带通滤波器的构造
在实际情况下，信号的长度是有限的，因此，需要在对图像应用二维Hilbert变换之前，对其进行一个带通滤波。从另一个方面来讲，带通滤波有利于维持信号分解中“不变-同变”的性质，这种“不变-同变”性质表示能量(局部振幅)和结构(局部相位和方向)是独立的信息。
Gabor滤波器是获取局部频域信息的一个传统的选择，它们提供了空间域位置和频率域信息的一个较好的融合。然而，它们却有着两个主要的限制：一方面，它的最大带宽被限制在近似一个倍频的范围，另一方面，如果一个人想要获得广泛光谱范围的最大特征定位，Gabor也不是最优的，而且由于Gabor的均值不为零，故容易受到直流分量的影响。作为对Gabor特征的一个改进，Field提出了log-Gabor特征。
log-Gabor滤波器抑制了直流分量，而且能够解决传统Gabor滤波器的带宽限制，同时它在对数频率尺度下的响应仍然是高斯形状的，这样就会使得更多的高频信息被保留下来。
log-Gabor滤波器的频域响应可以表示为：
Figure PCTCN2014091046-appb-000032
其中，ω0是中心频率，σ是带宽的尺度因子，为了保证滤波器组有一个固定的形状，我们保持σ/ω0为一个常数。
由于log-Gabor是带通滤波器，为了更充分地描述一个信号，则需要提取不同的频率分量，从而需要不同尺度的log-Gabor滤波器。在多尺度的log-Gabor滤波器中，参数σ和ω0可以被重写为：
σ＝σratioω0,ω0＝(λminμs-1)-1   (10)
这里λmin的物理意义是最短的波长，μ是关于波长的一个乘子，s是尺度的索引，σratio是σ和ω0的比率σ/ω0。
具体的滤波过程可以通过下式表示：
针对一幅人脸图片f(x,y)，
ffiltered＝F-1(F(f(x,y))*G(ω))   (11)
其中，F(·)和F-1(·)分别表示傅立叶和逆傅立叶变换。
图8是根据本发明一个优选实施例的三个尺度的log-Gabor滤波器频域表示的三维视图，如图8所示，分为上下两行，第一行从左到右依次表示不同尺度的log_Gabor函数在频域表示的俯视图，可以清楚地看到通过不同尺度的log_Gabor函数滤波后，图像保留的是不同频率段的信息；第二行依次表示第一行图像的三维表示，可以明确地表征出log_Gabor函数的带通特性。
图9是根据本发明一个优选实施例的人脸照片在一个尺度下的二维解析表示的示意图，如图9所示，通过以上两步，就可以得到一副人脸的二维解析表示。
(c)稀疏自编码
传统的统计学习方法都是有监督的，即在训练的过程中需要我们人工的干预，一方面训练过程复杂，另一方面需要大量的有标签的数据。而实际上带有标签的数据往往是有限的，这样就限制了算法的效率和准确性。
稀疏自编码学习，能从大量的无标签的样本中通过学习算法寻找这些样本的内在的本质特征，从而缓解了统计学习中的小样本问题，另一方面，稀疏自编码学习能够自动的寻找存在于数据中内在特征，本质上是一个特征学习的过程。
下面对稀疏自编码学习的过程进行一个详细的描述：
针对一个训练样本的集合{x1,x2,x3,…}，其中xi∈Rn表示一个训练样本。自编码神经网络是一个无监督的学习过程，它利用反向传播算法并且使得学习的目标等于输入值。
图10是根据本发明另一优选实施例的一个自编码神经网络的示意图，如图10所示，为自编码神经网络的一个示例：
自编码神经网络的目标是，通过随机梯度下降法得到一个从输入层到输出层的一个假设hw,b(x)≈x，即它尝试逼近一个恒等函数，使得网络的输出
Figure PCTCN2014091046-appb-000033
接近于输入x，其中，W为权重矩阵，b为偏置矩阵。即随机梯度下降法是针对要求解的优化问题，随机地沿着梯度降低的方向移动，最终达到待优化问题的最终的局部或全局最优解。这样，当隐藏层L2的神经元数目m小于训练样本的维数n的时候，仍然可以利用比n小的数据很好的恢复出n维的样本。也就是说，可以通过学习得到n维样本的m个本质特征，从而可以利用这m个特征来表征n维的训练数据。
另一方面，当隐藏层的神经元数目m很大的时候，仍然可以通过增加一些限制，从而使得自编码神经网络能够很好的表征样本的本质特征。这个限制条件就是隐藏层神经元的激活度，可以对这个激活度增加一个稀疏性限制，即保证在一次传播中隐藏层神经元是稀疏地被激活的，从而利用稀疏自编码网络来得到训练数据的结构特征。
图10就是一个典型的神经网络的模型，展示了从输入层(Layer)L1经过隐藏层(Layer)L2到输出层(Layer)L3的一个映射，从输入层L1到隐藏层L2的关系可以通过公式(12)得到，同理隐藏层L2又同时是输出层L3的输入层，故，从隐藏层L2到输出层L3的关系也是通过公式(12)给出的。与一般的神经网络不同的是，一般的神经网络训练过程中需要对输出层的内容进行手工的标定，而本实施例中用到的稀疏自编码神经网络是不需要手工标定的，这里假设输出层和输入层相等的，即，用网络输出来模拟输入的本质特征。
(二)特征比对过程
假设网络的输入为x，则其输出y可以表示为：
Figure PCTCN2014091046-appb-000034
通过稀疏自编码学习算法得到的权重矩阵W和偏置b(即通过随机梯度下降法求解定义的损失函数的最小值)，并分别以公式(8)中的三个分量作为网络的输入，从而会得到各自的输出，即为二维人脸解析表示的编码。
对每一个分量(即公式(8)中的局部振幅、局部方向和局部相位)的编码，分别计算其相似度，这里采用的是向量间夹角的余弦：
Figure PCTCN2014091046-appb-000035
其中，x，y是编码后特征的向量表示，<x,y>表示向量x,y的内积。即利用其夹角的余弦来衡量两个向量的相似程度，然后，总的相似度是通过每一个分量相似度的加权平均得到，其中权重分别为：振幅0.25，相位0.45，方向0.3。
最后，把总的相似度和预先设定的阈值做比较，给出是否是同一个人的判定。
可见，上述实施例利用二维Hilbert变换为载体，得到了人脸图像的解析表示，同时利用稀疏自编码学习算法对得到的解析表示进行编码。既利用了局部纹理信息对光照、姿态、表情等具有鲁棒性的特点，同时算法的时间和空间的复杂度也比较低。在FERET数据库上，四个子库Fb，Fc，DupI，DupII上分别取得了99.7％，99.5％，93.6％，91.5％的首选识别率。
图11是根据本发明一个优选实施例的在DupI数据库上的ROC曲线的示意图，如图11所示，在拍摄时间、光照和表情等变化较大的子库DupI上，误识率为千分之一时，识别率达到了93.21％。
在上述优选实施例中，可以首先对人脸利用不同尺度的带通滤波器log-Gabor进行滤波，对滤波后的图像利用二维Hilbert变换求得其不同分辨率下的二维解析表示；然后，对人脸的解析表示进行稀疏自学习编码，通过学习的方法，自动地寻找数据的本质特征，从而得到人脸的更准确的表示。
综上所述，通过本发明实施例，利用信号的二维Hilbert变换后得到的二维解析信号来表征人脸，还可以利用稀疏自编码算法对变换后的人脸进行编码，作为最终的特征来认证人脸，即利用的是人脸的局部统计特征，不仅能够在一定的程度上解决光照姿态对结果的影响，而且算法的时间和空间复杂度都要比Gabor特征低，但是泛化性和识别结果要优于Gabor特征。
显然，本领域的技术人员应该明白，上述的本发明的各模块或各步骤可以用通用的计算装置来实现，它们可以集中在单个的计算装置上，或者分布在多个计算装置所组成的网络上，可选地，它们可以用计算装置可执行的程序代码来实现，从而可以将它们存储在存储装置中由计算装置来执行，或者将它们分别制作成各个集成电路模块，或者将它们中的多个模块或步骤制作成单个集成电路模块来实现。这样，本发明不限制于任何特定的硬件和软件结合。
以上所述仅为本发明的优选实施例而已，并不用于限制本发明，对于本领域的技术人员来说，本发明可以有各种更改和变化。凡在本发明的精神和原则之内，所作的任何修改、等同替换、改进等，均应包含在本发明的保护范围之内。

Claims (24)

  1. 一种人脸特征的提取方法，包括：
    对获取的人脸图像进行二维希尔伯特变换；以及
    将经过所述二维希尔伯特变换后的所述人脸图像用二维解析信号表示。
  2. 根据权利要求1所述的方法，其中，对获取的所述人脸图像进行所述二维希尔伯特变换的公式为：
    Figure PCTCN2014091046-appb-100001
    其中，f(x,y)＝axy，用于表示M行N列的所述人脸图像的图像矩阵，axy为所述图像矩阵中的元素，x为不大于M的正整数，y为不大于N的正整数，fx(x,y)和fy(x,y)为所述二维解析信号的两个虚部；hx、hy分别为Hx、Hy的逆傅立叶变换的结果，而Hx、Hy分别为所述二维希尔伯特变换的频域公式-jω/||ω||分解出的横坐标分量和纵坐标分量。
  3. 根据权利要求2所述的方法，其中，所述二维希尔伯特变换的时域公式为：
    Figure PCTCN2014091046-appb-100002
    其中，在hx、hy分别与f(x,y)卷积时，(u,v)用于表示以(x,y)为原点的笛卡尔坐标系内的空间位置，u和v的取值范围均为
    Figure PCTCN2014091046-appb-100003
    w为正整数，为卷积窗口的大小。
  4. 根据权利要求1所述的方法，其中，将经过所述二维希尔伯特变换后的所述人脸图像用二维解析信号表示包括：
    通过局部振幅A，局部相位φ和局部方向θ三个分量表示所述人脸图像的二维解析信号，其中，
    Figure PCTCN2014091046-appb-100004
    用于表示所述二维解析信号的能量信息；
    Figure PCTCN2014091046-appb-100005
    用于表示所述二维解析信号的结构信息，sign(·)为符号函数；
    θ＝arctan(fy(x,y)/fx(x,y))，用于表示所述二维解析信号的几何信息；
    且经过所述二维希尔伯特变换后的所述人脸图像的所述二维解析信号为：
    fA(x,y)＝(f(x,y),fx(x,y),fy(x,y))，
    其中，f(x,y)为所述二维解析信号的实部，fx(x,y)和fy(x,y)分别为所述二维解析信号的两个虚部。
  5. 根据权利要求1所述的方法，其中，对获取的人脸图像进行二维希尔伯特变换之前，所述方法还包括：
    分别将获取的所述人脸图像经过带通滤波器进行滤波，其中，所述带通滤波器包括log-Gabor滤波器。
  6. 根据权利要求1所述的方法，其中，将经过所述二维希尔伯特变换后的所述人脸图像用二维解析信号表示之后，所述方法还包括：
    将经过所述二维希尔伯特变换后的所述人脸图像的二维解析信号通过预设的稀疏自编码神经网络，得到相应的最优值，其中，所述预设的稀疏自编码神经网络中的权重矩阵和偏置矩阵由人脸训练样本获得。
  7. 一种采用权利要求1至6中任一项所述人脸特征的提取方法进行人脸特征的认证方法，包括：
    分别使用所述人脸特征的提取方法表示第一人脸图像和第二人脸图像；
    对经过所述人脸特征的提取方法表示后的所述第一人脸图像和所述第二人脸图像进行相似度计算；以及
    根据所述相似度计算的结果完成人脸认证过程。
  8. 根据权利要求7所述的方法，其中，对经过所述人脸特征的提取方法表示后的所述第一人脸图像和所述第二人脸图像进行相似度计算包括：
    对经过所述人脸特征的提取方法表示后的所述第一人脸图像和所述第二人脸图像的每个分量进行相似度计算；
    将每个分量相似度的加权平均得到所述人脸特征的提取方法表示后的所述第一人脸图像和所述第二人脸图像的总的相似度。
  9. 一种人脸特征的提取装置，包括：
    二维希尔伯特变换模块，对获取的人脸图像进行二维希尔伯特变换；以及
    人脸图像表示模块，用于将经过所述二维希尔伯特变换后的所述人脸图像用二维解析信号表示。
  10. 一种采用权利要求9所述人脸特征的提取装置进行人脸特征的认证装置，其中，所述认证装置包括：
    特征提取模块，用于分别使用所述人脸特征的提取装置表示第一人脸图像和第二人脸图像；
    相似度计算模块，对经过所述人脸特征的提取装置表示的所述第一人脸图像和所述第二人脸图像进行相似度计算；以及
    认证模块，用于根据所述相似度计算的结果完成人脸认证过程。
  11. 一种人脸认证方法，包括：
    对第一人脸图像和第二人脸图像进行二维希尔伯特变换，以提取所述第一人脸图像的第一人脸特征和所述第二人脸图像的第二人脸特征；
    计算所述第一人脸特征和所述第二人脸特征的相似度；
    判断所述相似度是否达到预设阈值；以及
    在判断出所述相似度达到所述预设阈值的情况下，确定所述第一人脸图像所表示的人脸和所述第二人脸图像所表示的人脸为同一人脸。
  12. 根据权利要求11所述的人脸认证方法，其中，对第一人脸图像和第二人脸图像进行二维希尔伯特变换，以提取所述第一人脸图像的第一人脸特征和所述第二人脸图像的第二人脸特征包括：
    获取表示目标人脸图像的目标图像矩阵，其中，所述目标人脸图像为所述第一人脸图像或所述第二人脸图像；以及
    对所述目标图像矩阵进行二维希尔伯特变换，得到目标二维解析信号，其中，所述目标人脸图像的目标人脸特征为所述目标二维解析信号。
  13. 根据权利要求12所述的人脸认证方法，其中，对所述目标图像矩阵进行二维希尔伯特变换，得到所述目标二维解析信号包括：
    获取对所述目标图像矩阵进行所述二维希尔伯特变换的卷积模板；
    将所述目标图像矩阵和所述卷积模板进行卷积运算，得到卷积结果；以及
    确定所述目标图像矩阵为所述目标二维解析信号的实部，并确定所述卷积结果为所述目标二维解析信号的虚部。
  14. 根据权利要求13所述的人脸认证方法，其中，通过以下方式获取所述卷积模板：
    将所述二维希尔伯特变换的频域表示式分解为第一分量和第二分量，其中，Hx＝-jωx/||ω||，Hy＝-jωy/||ω||，-jω/||ω||为所述频域表示式，Hx为所述第一分量，Hy为所述第二分量，ωx和ωy分别为ω的横坐标和纵坐标；以及
    对所述第一分量和所述第二分量做逆傅立叶变换，得到所述二维希尔伯特变换的时域表示式，其中，所述卷积模板为所述时域表示式。
  15. 根据权利要求13所述的人脸认证方法，其中，在确定所述目标图像矩阵为所述目标二维解析信号的实部，并确定所述卷积结果为所述目标二维解析信号的虚部之后，所述人脸认证方法还包括：
    确定所述目标二维解析信号的能量信息为局部振幅A；
    确定所述目标二维解析信号的结构信息为局部相位φ；以及
    确定所述目标二维解析信号的几何信息为局部方向θ，
    其中，
    Figure PCTCN2014091046-appb-100006
    Figure PCTCN2014091046-appb-100007
    θ＝arctan(fy(x,y)/fx(x,y))，sign(·)为符号函数；其中，f(x,y)为所述目标二维解析信号的实部，fx(x,y)和fy(x,y)分别为所述目标二维解析信号的两个虚部。
  16. 根据权利要求12所述的人脸认证方法，其中，计算所述第一人脸特征和所述第二人脸特征的相似度包括：
    计算第一二维解析信号的每个分量和第二二维解析信号的对应的每个分量的分量相似度，得到多个分量相似度，其中，所述第一二维解析信号为对所述第一人脸图像的图像矩阵进行二维希尔伯特变换得到的二维解析信号，所述第二二维解析信号为对所述第二人脸图像的图像矩阵进行二维希尔伯特变换得到的二维解析信号；
    按照预设权重值计算所述多个分量相似度的加权平均值；以及
    确定所述加权平均值为所述第一人脸特征和所述第二人脸特征的相似度。
  17. 根据权利要求11所述的人脸认证方法，其中，在对第一人脸图像和第二人脸图像进行二维希尔伯特变换，以提取所述第一人脸图像的第一人脸特征和所述第二人脸图像的第二人脸特征之前，所述人脸认证方法还包括：
    利用带通滤波器对所述第一人脸图像和所述第二人脸图像滤波，其中，所述带通滤波器包括log-Gabor滤波器。
  18. 一种人脸认证装置，包括：
    提取单元，用于对第一人脸图像和第二人脸图像进行二维希尔伯特变换，以提取所述第一人脸图像的第一人脸特征和所述第二人脸图像的第二人脸特征；
    计算单元，用于计算所述第一人脸特征和所述第二人脸特征的相似度；
    判断单元，用于判断所述相似度是否达到预设阈值；以及
    第一确定单元，用于在所述判断单元判断出所述相似度达到所述预设阈值的情况下，确定所述第一人脸图像所表示的人脸和所述第二人脸图像所表示的人脸为同一人脸。
  19. 根据权利要求18所述的人脸认证装置，其中，所述提取单元包括：
    获取子单元，用于获取表示目标人脸图像的目标图像矩阵，其中，所述目标人脸图像为所述第一人脸图像或所述第二人脸图像；以及
    变换子单元，用于对所述目标图像矩阵进行二维希尔伯特变换，得到目标二维解析信号，其中，所述目标人脸图像的目标人脸特征为所述目标二维解析信号。
  20. 根据权利要求19所述的人脸认证装置，其中，所述变换子单元包括：
    获取模块，用于获取对所述目标图像矩阵进行所述二维希尔伯特变换的卷积模板；
    运算模块，用于将所述目标图像矩阵和所述卷积模板进行卷积运算，得到卷积结果；以及
    确定模块，用于确定所述目标图像矩阵为所述目标二维解析信号的实部，并确定所述卷积结果为所述目标二维解析信号的虚部。
  21. 根据权利要求20所述的人脸认证装置，其中，所述获取模块包括：
    分解子模块，用于将所述二维希尔伯特变换的频域表示式分解为第一分量和第二分量，其中，Hx＝-jωx/||ω||，Hy＝-jωy/||ω||，-jω/||ω||为所述频域表示式，Hx为所述第一分量，Hy为所述第二分量，ωx和ωy分别为ω的横坐标和纵坐标；以及
    变化子模块，用于对所述第一分量和所述第二分量做逆傅立叶变换，得到所述二维希尔伯特变换的时域表示式，其中，所述卷积模板为所述时域表示式。
  22. 根据权利要求20所述的人脸认证装置，其中，所述人脸认证装置还包括：
    第二确定单元，用于确定所述目标二维解析信号的能量信息为局部振幅A；
    第三确定单元，用于确定所述目标二维解析信号的结构信息为局部相位φ；以及
    第四确定单元，用于确定所述目标二维解析信号的几何信息为局部方向θ，
    其中，
    Figure PCTCN2014091046-appb-100008
    Figure PCTCN2014091046-appb-100009
    θ＝arctan(fy(x,y)/fx(x,y))，sign(·)为符号函数；其中，f(x,y)为所述目标二维解析信号的实部，fx(x,y)和fy(x,y)分别为所述目标二维解析信号的两个虚部。
  23. 根据权利要求19所述的人脸认证装置，其中，所述计算单元包括：
    第一计算子单元，用于计算第一二维解析信号的每个分量和第二二维解析信号的对应的每个分量的分量相似度，得到多个分量相似度，其中，所述第一二维解析信号为对所述第一人脸图像的图像矩阵进行二维希尔伯特变换得到的二维解析信号，所述第二二维解析信号为对所述第二人脸图像的图像矩阵进行二维希尔伯特变换得到的二维解析信号；
    第二计算子单元，用于按照预设权重值计算所述多个分量相似度的加权平均值；以及
    确定子单元，用于确定所述加权平均值为所述第一人脸特征和所述第二人脸特征的相似度。
  24. 根据权利要求18所述的人脸认证装置，其中，所述人脸认证装置还包括：
    滤波单元，用于利用带通滤波器对所述第一人脸图像和所述第二人脸图像滤波，其中，所述带通滤波器包括log-Gabor滤波器。
PCT/CN2014/091046 2013-12-16 2014-11-13 人脸特征的提取、认证方法及装置 WO2015090126A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201310690584.5A CN103646244B (zh) 2013-12-16 2013-12-16 人脸特征的提取、认证方法及装置
CN201310690584.5 2013-12-16

Publications (1)

Publication Number Publication Date
WO2015090126A1 true WO2015090126A1 (zh) 2015-06-25

Family

ID=50251455

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2014/091046 WO2015090126A1 (zh) 2013-12-16 2014-11-13 人脸特征的提取、认证方法及装置

Country Status (2)

Country Link
CN (1) CN103646244B (zh)
WO (1) WO2015090126A1 (zh)

Cited By (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109427080A (zh) * 2017-08-31 2019-03-05 爱唯秀股份有限公司 快速生成大量复杂光源人脸图像的方法
CN109753942A (zh) * 2019-01-14 2019-05-14 河南理工大学 基于空间金字塔fhog特征的人脸表情识别方法和装置
CN109784325A (zh) * 2017-11-10 2019-05-21 富士通株式会社 开集识别方法和设备及计算机可读存储介质
CN110110734A (zh) * 2018-02-01 2019-08-09 富士通株式会社 开集识别方法、信息处理设备以及存储介质
CN111819568A (zh) * 2018-06-01 2020-10-23 华为技术有限公司 人脸旋转图像的生成方法及装置
CN112732092A (zh) * 2021-01-22 2021-04-30 河北工业大学 基于双视图多尺度卷积神经网络的表面肌电信号识别方法
CN112766215A (zh) * 2021-01-29 2021-05-07 北京字跳网络技术有限公司 人脸融合方法、装置、电子设备及存储介质
CN117218783A (zh) * 2023-09-12 2023-12-12 广东云百科技有限公司 物联网安全管理系统及方法
CN117373100A (zh) * 2023-12-08 2024-01-09 成都乐超人科技有限公司 基于差分量化局部二值模式的人脸识别方法及系统

Families Citing this family (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103646244B (zh) * 2013-12-16 2018-01-09 北京天诚盛业科技有限公司 人脸特征的提取、认证方法及装置
CN104281834B (zh) * 2014-05-16 2017-07-25 华为技术有限公司 一种人脸识别的方法及设备
CN104346607B (zh) * 2014-11-06 2017-12-22 上海电机学院 基于卷积神经网络的人脸识别方法
CN105069400B (zh) * 2015-07-16 2018-05-25 北京工业大学 基于栈式稀疏自编码的人脸图像性别识别系统
CN106934335B (zh) * 2015-12-31 2021-02-02 南通东华软件有限公司 图像识别的方法和装置
CN106203354B (zh) * 2016-07-14 2019-04-12 南京信息工程大学 基于混合深度结构的场景识别方法
CN109189895B (zh) * 2018-09-26 2021-06-04 杭州大拿科技股份有限公司 一种针对口算题的题目批改方法及装置
CN109858463B (zh) * 2019-02-22 2021-03-26 成都云鼎丝路信息技术有限公司 一种双引擎用户识别方法、系统及终端
CN112017156B (zh) * 2020-07-17 2023-02-14 中国科学院西安光学精密机械研究所 基于多光谱视频的空间点目标旋转周期估计方法

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101187975A (zh) * 2007-12-25 2008-05-28 西南交通大学 一种具有光照鲁棒性的人脸特征提取方法
CN101271521A (zh) * 2008-05-13 2008-09-24 清华大学 基于各向异性双树复小波包变换的人脸识别方法
CN101777120A (zh) * 2010-01-28 2010-07-14 山东大学 一种基于序特征的人脸识别图像处理方法
US20130169621A1 (en) * 2011-12-28 2013-07-04 Li Mei Method of creating and transforming a face model and related system
CN103646244A (zh) * 2013-12-16 2014-03-19 北京天诚盛业科技有限公司 人脸特征的提取、认证方法及装置

Family Cites Families (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN100580691C (zh) * 2007-03-16 2010-01-13 上海博康智能信息技术有限公司 综合利用人脸及人体辅助信息的交互式人脸识别系统及方法
CN101079104A (zh) * 2007-06-14 2007-11-28 上海交通大学 基于切信息的人脸识别方法
CN100589118C (zh) * 2007-11-29 2010-02-10 北京中星微电子有限公司 特征提取方法、装置与模式识别方法、装置
US8670560B2 (en) * 2008-10-23 2014-03-11 University Of Ulster Encryption method

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101187975A (zh) * 2007-12-25 2008-05-28 西南交通大学 一种具有光照鲁棒性的人脸特征提取方法
CN101271521A (zh) * 2008-05-13 2008-09-24 清华大学 基于各向异性双树复小波包变换的人脸识别方法
CN101777120A (zh) * 2010-01-28 2010-07-14 山东大学 一种基于序特征的人脸识别图像处理方法
US20130169621A1 (en) * 2011-12-28 2013-07-04 Li Mei Method of creating and transforming a face model and related system
CN103646244A (zh) * 2013-12-16 2014-03-19 北京天诚盛业科技有限公司 人脸特征的提取、认证方法及装置

Cited By (13)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109427080A (zh) * 2017-08-31 2019-03-05 爱唯秀股份有限公司 快速生成大量复杂光源人脸图像的方法
CN109784325A (zh) * 2017-11-10 2019-05-21 富士通株式会社 开集识别方法和设备及计算机可读存储介质
CN110110734B (zh) * 2018-02-01 2023-04-07 富士通株式会社 开集识别方法、信息处理设备以及存储介质
CN110110734A (zh) * 2018-02-01 2019-08-09 富士通株式会社 开集识别方法、信息处理设备以及存储介质
CN111819568A (zh) * 2018-06-01 2020-10-23 华为技术有限公司 人脸旋转图像的生成方法及装置
CN109753942B (zh) * 2019-01-14 2022-11-04 河南理工大学 基于空间金字塔fhog特征的人脸表情识别方法和装置
CN109753942A (zh) * 2019-01-14 2019-05-14 河南理工大学 基于空间金字塔fhog特征的人脸表情识别方法和装置
CN112732092A (zh) * 2021-01-22 2021-04-30 河北工业大学 基于双视图多尺度卷积神经网络的表面肌电信号识别方法
CN112732092B (zh) * 2021-01-22 2023-04-07 河北工业大学 基于双视图多尺度卷积神经网络的表面肌电信号识别方法
CN112766215A (zh) * 2021-01-29 2021-05-07 北京字跳网络技术有限公司 人脸融合方法、装置、电子设备及存储介质
CN117218783A (zh) * 2023-09-12 2023-12-12 广东云百科技有限公司 物联网安全管理系统及方法
CN117373100A (zh) * 2023-12-08 2024-01-09 成都乐超人科技有限公司 基于差分量化局部二值模式的人脸识别方法及系统
CN117373100B (zh) * 2023-12-08 2024-02-23 成都乐超人科技有限公司 基于差分量化局部二值模式的人脸识别方法及系统

Also Published As

Publication number Publication date
CN103646244B (zh) 2018-01-09
CN103646244A (zh) 2014-03-19

Similar Documents

Publication Publication Date Title
WO2015090126A1 (zh) 人脸特征的提取、认证方法及装置
Yuan et al. A convolutional neural network based on TensorFlow for face recognition
Xu et al. Data uncertainty in face recognition
CN102938065B (zh) 基于大规模图像数据的人脸特征提取方法及人脸识别方法
Vageeswaran et al. Blur and illumination robust face recognition via set-theoretic characterization
Li et al. Overview of principal component analysis algorithm
Boutros et al. Iris and periocular biometrics for head mounted displays: Segmentation, recognition, and synthetic data generation
WO2016150240A1 (zh) 身份认证方法和装置
WO2015089949A1 (zh) 一种融合LBP、Gabor特征的人脸聚类方法
El Khiyari et al. Age invariant face recognition using convolutional neural networks and set distances
CN105117708A (zh) 人脸表情识别的方法和装置
CN111639558A (zh) 一种基于ArcFace Loss和改进残差网络的指静脉身份验证方法
Lee et al. Face image retrieval using sparse representation classifier with gabor-lbp histogram
Alaslni et al. Transfer learning with convolutional neural networks for iris recognition
Dong et al. Finger vein verification based on a personalized best patches map
Kerola et al. Cross-view human action recognition from depth maps using spectral graph sequences
Xu et al. An effective recognition approach for contactless palmprint
Tao et al. Finger-vein recognition using bidirectional feature extraction and transfer learning
Hafner et al. Deep iris feature extraction
Mewada et al. Wavelet features embedded convolutional neural network for multiscale ear recognition
Feng et al. Face liveness detection using shearlet-based feature descriptors
Al-Juboori et al. Biometric authentication system based on palm vein
Silva et al. Face sketch recognition from local features
Baruni et al. Age invariant face recognition methods: A review
Satange et al. Composite feature extraction based on Gabor and Zernike moments for face recognition

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: 14870938

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: 14870938

Country of ref document: EP

Kind code of ref document: A1