WO2017088432A1 - 图像识别方法和装置 - Google Patents

图像识别方法和装置 Download PDF

Info

Publication number
WO2017088432A1
WO2017088432A1 PCT/CN2016/084418 CN2016084418W WO2017088432A1 WO 2017088432 A1 WO2017088432 A1 WO 2017088432A1 CN 2016084418 W CN2016084418 W CN 2016084418W WO 2017088432 A1 WO2017088432 A1 WO 2017088432A1
Authority
WO
WIPO (PCT)
Prior art keywords
image
feature vector
matrix
target image
person
Prior art date
Application number
PCT/CN2016/084418
Other languages
English (en)
French (fr)
Inventor
丁守鸿
李季檩
汪铖杰
黄飞跃
吴永坚
谭国富
Original Assignee
腾讯科技(深圳)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 腾讯科技(深圳)有限公司 filed Critical 腾讯科技(深圳)有限公司
Publication of WO2017088432A1 publication Critical patent/WO2017088432A1/zh
Priority to US15/925,028 priority Critical patent/US10713532B2/en

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/10Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
    • G06V40/16Human faces, e.g. facial parts, sketches or expressions
    • G06V40/168Feature extraction; Face representation
    • G06V40/171Local features and components; Facial parts ; Occluding parts, e.g. glasses; Geometrical relationships
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00Pattern recognition
    • G06F18/20Analysing
    • G06F18/21Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/213Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
    • G06F18/2135Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods based on approximation criteria, e.g. principal component analysis
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00Pattern recognition
    • G06F18/20Analysing
    • G06F18/21Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00Pattern recognition
    • G06F18/20Analysing
    • G06F18/24Classification techniques
    • G06F18/241Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2415Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
    • G06F18/24155Bayesian classification
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00Arrangements for image or video recognition or understanding
    • G06V10/70Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/764Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00Arrangements for image or video recognition or understanding
    • G06V10/70Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/77Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
    • G06V10/7715Feature extraction, e.g. by transforming the feature space, e.g. multi-dimensional scaling [MDS]; Mappings, e.g. subspace methods
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/10Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
    • G06V40/16Human faces, e.g. facial parts, sketches or expressions
    • G06V40/161Detection; Localisation; Normalisation
    • G06V40/165Detection; Localisation; Normalisation using facial parts and geometric relationships
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/10Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
    • G06V40/16Human faces, e.g. facial parts, sketches or expressions
    • G06V40/168Feature extraction; Face representation
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/10Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
    • G06V40/16Human faces, e.g. facial parts, sketches or expressions
    • G06V40/172Classification, e.g. identification
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00Arrangements for image or video recognition or understanding
    • G06V10/40Extraction of image or video features
    • G06V10/46Descriptors for shape, contour or point-related descriptors, e.g. scale invariant feature transform [SIFT] or bags of words [BoW]; Salient regional features
    • G06V10/467Encoded features or binary features, e.g. local binary patterns [LBP]

Definitions

  • the present disclosure relates to the field of computer technologies, and in particular, to an image recognition method and apparatus.
  • Face recognition technology usually involves two steps. First, feature extraction is performed on the face image; second, similarity calculation is performed on the extracted feature and the feature in the reference face image.
  • Embodiments of the present invention provide an image recognition method and apparatus, which can solve the above problems.
  • the technical solution is as follows:
  • an image recognition method including:
  • the training matrix is a matrix obtained by training a image in an image library by a joint Bayesian algorithm
  • the target image is identified according to the high dimensional feature vector of the target image and the training matrix.
  • an image recognition apparatus comprising:
  • At least one processor At least one processor
  • the memory stores program instructions that, when executed by the processor, configure the image recognition device to perform the operations of:
  • the training matrix is a matrix obtained by training a image in an image library by a joint Bayesian algorithm
  • the target image is identified according to the high dimensional feature vector of the target image and the training matrix.
  • the LBP feature vector of the target image is extracted by the LBP algorithm, and then the target image is identified according to the LBP feature vector and the training matrix, which is a matrix trained by the joint Bayesian algorithm on the image library; the accuracy of image recognition Higher; the LBP algorithm and the joint Bayesian algorithm can be combined to perform image recognition and improve the accuracy of image recognition.
  • An example of the above image is a face image, and the image in the image library may include images when the user is at different angles, different expressions, or different illuminations, so the above method may still have when the face is occluded or the facial illumination changes strongly. Higher face recognition accuracy.
  • FIG. 1 is a block diagram of a computing device involved in an image recognition method according to various embodiments of the present invention
  • FIG. 2 is a flowchart of an image recognition method according to an embodiment of the present invention.
  • FIG. 3A is a flowchart of an image recognition method according to another embodiment of the present invention.
  • FIG. 3B is a flowchart of a method for extracting an LBP feature vector of an extraction target image according to an embodiment of the present invention
  • FIG. 3C is a schematic diagram of key points of each image obtained by positioning a terminal according to an embodiment of the present invention.
  • FIG. 3D is a schematic diagram of a target image when extracting an LBP feature vector according to an embodiment of the present invention.
  • FIG. 4 is a block diagram showing the structure of an image recognition apparatus according to an embodiment of the present invention.
  • FIG. 5 is a block diagram showing the structure of an image recognition apparatus according to another embodiment of the present invention.
  • the image recognition method may be used in the computing device 100.
  • the computing device may be a variety of terminal devices such as a desktop computer, a mobile phone, a notebook, a tablet, or the like, or the computing device may be a server.
  • the computing device 100 includes a central processing unit (CPU) 101, a system memory 104 including a random access memory (RAM) 102 and a read only memory (ROM) 103, and a connection system memory 104 and System bus 105 of central processing unit 101.
  • CPU central processing unit
  • system memory 104 including a random access memory (RAM) 102 and a read only memory (ROM) 103
  • connection system memory 104 and System bus 105 of central processing unit 101.
  • the computing device 100 also includes a basic input/output system (I/O system) 106 that facilitates transfer of information between various devices within the computer, and a large capacity for storing the operating system 113, applications 112, and other program modules 115.
  • Storage device 107 stores the operating system 113, applications 112, and other program modules 115.
  • the basic input/output system 106 includes a display 108 for displaying information and an input device 109 such as a mouse or keyboard for user input of information.
  • the display 108 and input device 109 are both connected to the central processing unit 101 by an input and output controller 110 connected to the system bus 105.
  • the basic input/output system 106 can also include an input output controller 110 for receiving and processing input from a plurality of other devices, such as a keyboard, mouse, or electronic stylus.
  • input and output controller 110 also provides output to a display screen, printer, or other type of output device.
  • the mass storage device 107 is connected to the central processing unit 101 by a mass storage controller (not shown) connected to the system bus 105.
  • the mass storage device 107 and its associated computer readable medium provide non-volatile storage for the computing device 100. That is, the mass storage device 107 can include a computer readable medium (not shown) such as a hard disk or a CD-ROM drive.
  • the computer readable medium can include computer storage media and communication media.
  • Computer storage media includes volatile and nonvolatile, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data.
  • Computer storage media include RAM, ROM, EPROM, EEPROM, flash memory or other solid state storage technologies, CD-ROM, DVD or other optical storage, tape cartridges, magnetic tape, magnetic disk storage or other magnetic storage devices.
  • RAM random access memory
  • ROM read only memory
  • EPROM Erasable programmable read-only memory
  • EEPROM electrically erasable programmable read-only memory
  • the computing device 100 may also be operated by a remote computer connected to the network via a network such as the Internet. That is, the computing device 100 can be connected to the network 112 through a network interface unit 111 connected to the system bus 105, or can also use a network connection.
  • the port unit 111 is connected to other types of networks or remote computer systems (not shown).
  • the memory also includes one or more programs, the one or more programs being stored in a memory, the one or more programs for performing the image recognition methods provided by the embodiments described below.
  • FIG. 2 shows a flowchart of an image recognition method provided by an embodiment of the present invention, which may be used in the computing device shown in FIG. 1.
  • the image recognition method may include:
  • Step 201 Extract an LBP feature vector of the target image.
  • LBP Local Binary Pattern
  • Uniform coding can be used for coding, which has the characteristics of low algorithm complexity, low memory consumption, and fast calculation speed.
  • Step 202 Calculate a high-dimensional feature vector of the target image according to the LBP feature vector.
  • Step 203 Acquire a training matrix, and the training matrix is a matrix obtained by training the images in the image library by the joint Bayesian algorithm.
  • Step 204 Identify the target image according to the high-dimensional feature vector of the target image and the training matrix.
  • the image recognition method provided by the embodiment extracts the LBP feature vector of the target image by using the LBP algorithm, and then performs image recognition on the target image according to the LBP feature vector and the training matrix, and the training matrix is
  • the matrix obtained by the algorithm is trained in the image library; the accuracy of image recognition is high; the LBP algorithm and the joint Bayesian algorithm can be combined to perform image recognition and improve the accuracy of image recognition.
  • An example of the above image is a face image, and the image in the image library includes images when the user is at different angles, different expressions, or different illuminations, so the above method can still be higher when the face is occluded or the facial illumination changes strongly.
  • the accuracy of face recognition is a face image, and the image in the image library includes images when the user is at different angles, different expressions, or different illuminations, so the above method can still be higher when the face is occluded or the facial illumination changes strongly. The accuracy of face recognition.
  • FIG. 3A is a flowchart of an image recognition method according to an embodiment of the present invention.
  • the embodiment is illustrated by a human face image, and those skilled in the art can understand that the method of the embodiment can also be used for Identification of other images.
  • the method can be implemented by the computing device shown in FIG.
  • the computing device is a terminal (such as a computer, a mobile phone, a tablet computer, etc.) as an example, and the skill is described.
  • the image recognition method may include:
  • Step 301 Extract an LBP feature vector of the target face image.
  • the target face image is the input image to be recognized.
  • This step may include:
  • Step 301a Acquiring k scaled target face images, wherein the k zoomed target face images are images obtained by respectively scaling the target face images according to k preset multiples.
  • the terminal may separately scale the target face image according to each preset multiple of the k preset multiples, thereby obtaining k scaled target face images.
  • k is a positive integer
  • k preset multiples can be evenly distributed.
  • an example of the scaling referred to in this embodiment means that the target face image is reduced.
  • the terminal can reduce the target face image by a multiple of 1.25, 1.5, 1.75, and 2 to obtain 4 reduced target face images.
  • This embodiment does not limit the number of k and the specific value of each preset multiple.
  • Step 301b Determine, for each target face image in the target face image and the k scaled target face images, a face key point in the face image, and extract an LBP feature vector at the face key point.
  • This step can include:
  • the key points of the face may include: left and right eyebrows, left and right eyes, nose and left and right mouth corners, and the like.
  • the terminal can locate each face key point shown in the figure.
  • the LBP feature in the preset area is extracted in a preset area centered on the face key point.
  • the terminal may extract the LBP feature in the preset area by using Uniform coding to obtain an LBP feature histogram.
  • the preset area can be an area of a*a, and a is a positive integer.
  • the terminal may extract the LBP feature in each cell in the preset area by using Uniform coding.
  • the preset area may be a 4*4 area centered on the face key point, and for each cell in the figure, the terminal may calculate a 59-dimensional LBP feature distribution histogram.
  • Step 301c Determine the target person according to the extracted LBP feature vector of each face key point.
  • the LBP feature vector of the face image is
  • the terminal can extract and obtain u LBP features. For example, for the face image shown in FIG. 3D, the terminal can extract 59*27*16 LBP features.
  • the terminal extracts each LBP feature as a parameter in the LBP feature vector, and then combines to obtain an LBP feature vector including each LBP feature.
  • the terminal may use the LBP feature extracted for each facial image as a row or a column in the LBP feature vector, thereby obtaining an LBP feature vector including k+1 rows or k+1 columns.
  • Step 302 Calculate a high-dimensional feature vector of the target face image according to the LBP feature vector.
  • N is greater than the first threshold by less than the second threshold, and, in general, N may be 1440.
  • PCA dimensionality reduction is a commonly used dimension reduction method in image processing. It maps the original feature vector into low-dimensional space through linear transformation.
  • LDA Topic Dirichlet Allocation
  • Step 303 Acquire a training matrix, which is a matrix obtained by training a face image in a face image database by using a joint Bayesian algorithm.
  • the training matrix may be a locally pre-stored matrix or a matrix obtained by the terminal from the server.
  • Step 304 Acquire a high-dimensional feature vector of the reference face image.
  • the high dimensional feature vector of the reference face image may be a vector stored locally. Moreover, the high-dimensional feature vector may be pre-computed and stored locally by the terminal, or may be obtained by the terminal from the server and stored locally, which is not limited in this embodiment.
  • step 301 and step 302. This embodiment does not limit this.
  • Step 305 Calculate the similarity between the target face image and the reference face image according to the high-dimensional feature vector of the target face image, the high-dimensional feature vector of the reference face image, and the training matrix.
  • the target face image and the reference face image are similar.
  • Degree is:
  • transpose vector for x 1 The transpose vector for x 2 .
  • the terminal may perform the following steps:
  • the specific step may include: acquiring an LBP feature vector of each face image in the face image database, and calculating an average value M of all the features in the acquired LBP feature vector.
  • the method for obtaining the LBP feature vector of the face image is similar to the method for obtaining the LBP feature vector of the target image in the embodiment, which is not described herein.
  • x 1 and x 2 can be distributed around 0, which improves the calculation efficiency of similarity.
  • the present embodiment is exemplified by performing face recognition on a target face image by similarity.
  • the terminal may also calculate a variance between the high-dimensional feature vector of the target face image and the high-dimensional feature vector of the reference face image, and then perform face recognition by using the calculated variance.
  • the step of calculating the variance may include:
  • the mean r1 the variance s 1 of the high-dimensional feature vector of the same person in the face image library is obtained.
  • the variance is calculated according to the calculated mean r1, the variance s 1 and the similarity, and the variance s is:
  • the terminal can normalize the calculated s to the interval of 0-100.
  • the size of s indicates the magnitude of the probability that the target face image and the reference face image are the same person.
  • s is a confidence that the value is between 0-100. The larger the s, the higher the probability that the two face images are a person.
  • a threshold may be set. When s is greater than the set threshold, the target face image is determined to be the same person as the reference face image; and when s is less than the set threshold, the target face image and the reference are determined. The face image is not the same person.
  • the set threshold can be set to not less than 50 and not more than 100, and the designer can set the threshold according to the required recognition accuracy. If the required accuracy is high, the threshold is set to a higher value, such as 90; and if the required accuracy is lower, the threshold can be set to a smaller value, such as 60.
  • the face recognition method extracts the LBP feature vector of the target face image by using the LBP algorithm, and then performs face recognition on the target face image according to the LBP feature vector and the training matrix.
  • the matrix is a matrix obtained by training the face image database by the joint Bayesian algorithm; the accuracy of face recognition is high; it is achieved that the LBP algorithm and the joint Bayesian algorithm can be combined to perform face recognition and improve the face.
  • the image in the face image library may include images when the user is at different angles, different expressions or different illuminations, so the above method can still have higher face recognition accuracy when the face is occluded or the facial illumination changes strongly. rate.
  • the face image and the image scaled by the face image are calculated according to the preset multiple, wherein an example of the preset multiple is used to reduce the multiple of the image. , enhances the robustness of scale transformation.
  • the variance can be calculated, and different thresholds can be set for the variance according to the different recognition precisions required, which is better adapted to the actual use requirements.
  • the terminal may obtain the training matrix pre-trained by the server from the server, and may also acquire the training matrix that the terminal itself pre-trains and stores in the local, which is not limited in this embodiment. .
  • the following is an example of calculating and storing the training matrix by the terminal, and the specific steps are as follows:
  • the face image library includes m*n face images, m is the number of people corresponding to the images in the face image library, and n is the number of faces of each person's face image. Wherein, m and n are both positive integers, and each of the n face images of the person may include a face image of the user at different angles, different lighting conditions, and different expressions.
  • This step is similar to the step 301 and the step 302 in the foregoing embodiment.
  • This step is similar to the step 301 and the step 302 in the foregoing embodiment.
  • S ⁇ is the m-dimensional matrix, the covariance matrix is the n face image is a face image of each person in the database;
  • S ⁇ m * n is a square, a face image The covariance matrix between different people in the library.
  • the Gaussian distribution mean ⁇ i of the i-th person in the face image database is calculated, And the joint distribution covariance matrix ⁇ ij of the i-th person and the j-th person, Where x i is the high-dimensional feature vector of the i-th person, and x j is the high-dimensional feature vector of the j-th person.
  • transpose vector for ⁇ i where Is the transpose vector of ⁇ ij .
  • the training matrix A (S ⁇ + S ⁇ ) -1 -(F+G),
  • the face recognition method provided in the above embodiment can be applied to various application scenarios. For example, it can be applied to Internet financial check identity, face recognition, face attribute recognition, face beautification, face cartoon drawing, face tracking, lip language recognition, living body recognition, etc., which can be applied to instant messaging applications, socializing Application sharing platform, watermark camera, creative camera, daily P picture and other products.
  • the most direct application of the technical solution of the embodiment of the present invention is to input two face images and calculate two faces. Similarity can be applied to multiple business scenarios such as face verification and face similarity calculation.
  • FIG. 4 is a structural block diagram of an image recognition apparatus according to an embodiment of the present invention.
  • the image recognition apparatus may include a feature extraction module 401, a first calculation module 402, a matrix acquisition module 403, and an identification module 404.
  • a feature extraction module 401 configured to extract a local binary mode LBP feature vector of the target image
  • a first calculation module 402 configured to calculate a high-dimensional feature vector of the target image according to the LBP feature vector extracted by the feature extraction module 401;
  • a matrix obtaining module 403 configured to acquire a training matrix, where the training matrix is a matrix obtained by training an image in an image library by using a joint Bayesian algorithm;
  • the identification module 404 is configured to identify the target image according to the high-dimensional feature vector of the target image calculated by the first calculating module 402 and the training matrix acquired by the matrix acquiring module 403.
  • the image recognition apparatus extracts an LBP feature vector of a target image by using an LBP algorithm, and further performs image recognition on the target image according to the LBP feature vector and the training matrix, and the training matrix is a joint Bayer
  • the matrix obtained by the algorithm is trained in the image library; the accuracy of image recognition is high; the LBP algorithm and the joint Bayesian algorithm can be combined to perform image recognition and improve the accuracy of image recognition.
  • An example of the above image is a face image, and the image in the image library may include images when the user is at different angles, different expressions, or different illuminations, so the above method may still be more when the face is occluded or the facial illumination changes strongly. High face recognition accuracy.
  • FIG. 5 is a structural block diagram of an image recognition apparatus according to an embodiment of the present invention.
  • the image recognition apparatus may include a feature extraction module 501, a first calculation module 502, a matrix acquisition module 503, and an identification module 504.
  • a feature extraction module 501 configured to extract a local binary mode LBP feature vector of the target image
  • a first calculation module 502 configured to calculate a high-dimensional feature vector of the target image according to the LBP feature vector extracted by the feature extraction module 501;
  • a matrix obtaining module 503 configured to acquire a training matrix, where the training matrix is a matrix obtained by training an image in an image library by using a joint Bayesian algorithm;
  • the identification module 504 is configured to calculate the target image according to the first calculation module 502
  • the high-dimensional feature vector and the training matrix acquired by the matrix obtaining module 503 identify the target image.
  • the identification module 504 includes:
  • a feature acquiring unit 504a configured to acquire a high-dimensional feature vector of the reference image
  • a similarity calculation unit 504b configured to: in the high-dimensional feature vector of the target image is x 1 , a high-dimensional feature vector of the reference image is x 2 , and when the training matrix includes A and G, according to the target image a high-dimensional feature vector, a high-dimensional feature vector of the reference image, and the training matrix, and calculating a similarity between the target image and the reference image, the similarity is:
  • transpose vector for x 1 The transpose vector for x 2 .
  • the image library includes the target image and the reference image.
  • the feature extraction module 501 includes:
  • the image obtaining unit 501a is configured to acquire k scaled target images, and the k target images are obtained by scaling the target image according to k preset multiples, and k is a positive integer;
  • a feature extraction unit 501b configured to determine a key point in the image for each of the target image and the k target images, and extract an LBP feature vector at the key point;
  • the feature determining unit 501c is configured to combine and determine the LBP feature vector of the target image according to the extracted LBP feature vector of the key point.
  • the first calculating module 502 includes:
  • a first calculating unit 502a configured to: when the LBP feature vector of the target image is x r Perform principal component analysis (PCA) dimension reduction, retain the first N-dimensional features, obtain the dimensionality reduction matrix P, N is a positive integer, and x r is the LBP feature vector of the target image. a transpose vector for x r ;
  • PCA principal component analysis
  • the third calculation unit 502c a Bayesian probability model for three-dimensional reduction of the LDA x p, to give dimensionality reduction matrix L;
  • the image is a face image
  • the image library is a face image library
  • the face image library includes m*n face images
  • m is a person corresponding to the image in the face image library.
  • the device also includes:
  • the feature obtaining module 505 is configured to acquire, before the matrix acquiring module 503 acquires the training matrix, a high-dimensional feature vector of each person corresponding to the image in the face image library;
  • the initialization module 506 is configured to initialize S ⁇ and S ⁇ ;
  • S ⁇ is an m-dimensional square matrix, which is a covariance matrix of n face images of each person in the face image library;
  • S ⁇ is an m*n square matrix Is a covariance matrix between different people in the face image library;
  • the second calculation module 507 is configured to calculate F and G:
  • a third calculating module 508 configured to calculate a Gaussian distribution mean ⁇ i of the i-th person in the face image database according to the F and the G, And a joint distribution covariance matrix ⁇ ij of the i-th person and the j-th person, Where x i is a high-dimensional feature vector of the i-th person, and x j is a high-dimensional feature vector of the j-th person;
  • An update module 509 configured to update the S ⁇ and the S ⁇ : among them a transpose vector for ⁇ i , a transpose vector of ⁇ ij ;
  • the second calculating module 507 is further configured to perform the steps of calculating the F and the G again when the S ⁇ and the S ⁇ do not converge;
  • the feature obtaining module 505 includes:
  • the feature extraction unit 505a is configured to extract an LBP feature vector of the face image for each face image in the face image library
  • a fifth calculating unit 505b configured to calculate an average value of all the features in the LBP feature vector of the face image
  • the ninth calculating unit 505f is configured to perform LDA dimension reduction on the x p for each face image of each person to obtain a dimensionality reduction matrix L;
  • the image recognition apparatus extracts an LBP feature vector of a target image by using an LBP algorithm, and further performs image recognition on the target image according to the LBP feature vector and the training matrix, and the training matrix is a joint Bayer
  • the matrix obtained by the algorithm is trained in the image library; the accuracy of image recognition is high; the LBP algorithm and the joint Bayesian algorithm can be combined to perform image recognition and improve the accuracy of recognition.
  • An example of the above image is a face image, and the image in the image library may include images when the user is at different angles, different expressions, or different illuminations, so the above method may still be more when the face is occluded or the facial illumination changes strongly. High face recognition accuracy.
  • the image when calculating the high-dimensional feature vector of the image, the image is scaled according to the image and the preset multiple, wherein the preset multiple may be used to reduce the multiple of the image, and the robustness of the scale transformation is enhanced. Sex.
  • the variance in the embodiment, in the image recognition, the variance can be calculated, and different thresholds can be set for the variance according to the different recognition precisions required, which is better adapted to the actual use requirements.
  • a person skilled in the art may understand that all or part of the steps of implementing the above embodiments may be completed by hardware, or may be instructed by a program to execute related hardware, and the program may be stored in a computer readable storage medium.
  • the storage medium mentioned may be a read only memory, a magnetic disk or an optical disk or the like.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • Physics & Mathematics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • General Physics & Mathematics (AREA)
  • Oral & Maxillofacial Surgery (AREA)
  • General Health & Medical Sciences (AREA)
  • Multimedia (AREA)
  • Evolutionary Computation (AREA)
  • Data Mining & Analysis (AREA)
  • Artificial Intelligence (AREA)
  • Human Computer Interaction (AREA)
  • Evolutionary Biology (AREA)
  • Databases & Information Systems (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Software Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Computing Systems (AREA)
  • Medical Informatics (AREA)
  • Probability & Statistics with Applications (AREA)
  • Geometry (AREA)
  • Image Analysis (AREA)
  • Collating Specific Patterns (AREA)
  • Image Processing (AREA)

Abstract

一种图像识别方法和装置,所述方法包括:提取目标图像的局部二值模式LBP特征向量(201);根据LBP特征向量计算目标图像的高维特征向量(202);获取训练矩阵,训练矩阵为通过联合贝叶斯算法对图像库中的图像训练得到的矩阵(203);根据目标图像的高维特征向量以及训练矩阵,对目标图像进行识别(204)。上述图像识别方法和装置可以将LBP算法和联合贝叶斯算法进行结合来进行识别,提高了图像识别的准确度。

Description

图像识别方法和装置
本申请要求于2015年11月26日提交中国专利局,申请号为201510843973.6,发明名称为“人脸识别方法和装置”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
本公开涉及计算机技术领域,特别涉及一种图像识别方法和装置。
背景技术
人脸识别技术通常包括两个步骤。第一,对人脸图像进行特征提取;第二,对提取到的特征与参考人脸图像中的特征进行相似度计算。
在计算特征的相似度时,通常采用基于马氏距离的计算方式来计算。然而受基于马氏距离的计算算法的限定,上述方案计算得到的相似度的准确度较低。
发明内容
本发明实施例提供了一种图像识别方法和装置,可以解决上述问题。所述技术方案如下:
一方面,提供了一种图像识别方法,包括:
提取目标图像的局部二值模式LBP特征向量;
根据所述LBP特征向量计算所述目标图像的高维特征向量;
获取训练矩阵,所述训练矩阵为通过联合贝叶斯算法对图像库中的图像训练得到的矩阵;
根据所述目标图像的高维特征向量以及所述训练矩阵,对所述目标图像进行识别。
另一方面,提供了一种图像识别装置,包括:
至少一个处理器;和
存储器,其中所述存储器存储有程序指令,所述程序指令当由所述处理器执行时,配置所述图像识别装置执行下述操作:
提取目标图像的局部二值模式LBP特征向量;
根据所述LBP特征向量计算所述目标图像的高维特征向量;
获取训练矩阵,所述训练矩阵为通过联合贝叶斯算法对图像库中的图像训练得到的矩阵;
根据所述目标图像的高维特征向量以及所述训练矩阵,对所述目标图像进行识别。
本发明实施例提供的技术方案的有益效果是:
通过LBP算法提取目标图像的LBP特征向量,进而根据该LBP特征向量以及训练矩阵来对目标图像进行识别,该训练矩阵为通过联合贝叶斯算法对图像库训练得到的矩阵;图像识别的准确度较高;达到了可以将LBP算法和联合贝叶斯算法进行结合来进行图像识别,提高图像识别的准确度的效果。上述图像的一个例子是人脸图像,图像库中的图像,可以包括用户处于不同角度、不同表情或者不同光照时的图像,所以上述方法在人脸被遮挡或者面部光照变化强烈时,仍然可以有较高的人脸识别准确率。
附图说明
为了更清楚地说明本发明实施例中的技术方案,下面将对实施例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发明的一些实施例,对于本领域普通技术人员来讲,在不付出创造性劳动的前提下,还可以根据这些附图获得其他的附图。
图1是本发明各个实施例提供的图像识别方法所涉及的计算装置的框图;
图2是本发明一个实施例提供的图像识别方法的流程图;
图3A是本发明另一个实施例提供的图像识别方法的流程图;
图3B是本发明一个实施例提供的提取目标图像的LBP特征向量的提取方法的流程图;
图3C是本发明一实施例提供的终端定位得到的各个图像关键点的示意图;
图3D是本发明一实施例提供的提取LBP特征向量时的目标图像的示意图;
图4是本发明一个实施例提供的图像识别装置的结构方框图;
图5是本发明另一个实施例提供的图像识别装置的结构方框图。
具体实施方式
为了使本发明的目的、技术方案和优点更加清楚,下面将结合附图对本发明作进一步地详细描述,显然,所描述的实施例仅仅是本发明一部份实施例, 而不是全部的实施例。基于本发明中的实施例,本领域普通技术人员在没有做出创造性劳动前提下所获得的所有其它实施例,都属于本发明保护的范围。
本发明各个实施例所涉及的图像识别方法可以用于计算装置100中。所述计算装置可以是台式计算机、手机、笔记本、平板电脑等各种终端设备,或者,所述计算装置可以是服务器。具体的,请参考图1,所述计算装置100包括中央处理单元(CPU)101、包括随机存取存储器(RAM)102和只读存储器(ROM)103的系统存储器104,以及连接系统存储器104和中央处理单元101的系统总线105。所述计算装置100还包括帮助计算机内的各个器件之间传输信息的基本输入/输出系统(I/O系统)106,和用于存储操作系统113、应用程序112和其他程序模块115的大容量存储设备107。
所述基本输入/输出系统106包括有用于显示信息的显示器108和用于用户输入信息的诸如鼠标、键盘之类的输入设备109。其中所述显示器108和输入设备109都通过连接到系统总线105的输入输出控制器110连接到中央处理单元101。所述基本输入/输出系统106还可以包括输入输出控制器110以用于接收和处理来自键盘、鼠标、或电子触控笔等多个其他设备的输入。类似地,输入输出控制器110还提供输出到显示屏、打印机或其他类型的输出设备。
所述大容量存储设备107通过连接到系统总线105的大容量存储控制器(未示出)连接到中央处理单元101。所述大容量存储设备107及其相关联的计算机可读介质为计算装置100提供非易失性存储。也就是说,所述大容量存储设备107可以包括诸如硬盘或者CD-ROM驱动器之类的计算机可读介质(未示出)。
不失一般性,所述计算机可读介质可以包括计算机存储介质和通信介质。计算机存储介质包括以用于存储诸如计算机可读指令、数据结构、程序模块或其他数据等信息的任何方法或技术实现的易失性和非易失性、可移动和不可移动介质。计算机存储介质包括RAM、ROM、EPROM、EEPROM、闪存或其他固态存储其技术,CD-ROM、DVD或其他光学存储、磁带盒、磁带、磁盘存储或其他磁性存储设备。当然,本领域技术人员可知所述计算机存储介质不局限于上述几种。上述的系统存储器104和大容量存储设备107可以统称为存储器。
根据本发明的各种实施例,所述计算装置100还可以通过诸如因特网等网络连接到网络上的远程计算机运行。也即计算装置100可以通过连接在所述系统总线105上的网络接口单元111连接到网络112,或者说,也可以使用网络接 口单元111来连接到其他类型的网络或远程计算机系统(未示出)。
所述存储器还包括一个或者一个以上的程序,所述一个或者一个以上程序存储于存储器中,所述一个或者一个以上程序用于执行下述实施例提供的图像识别方法。
请参考图2,其示出了本发明一个实施例提供的图像识别方法的流程图,该图像识别方法可以用于图1所示的计算装置。如图2所示,该图像识别方法可以包括:
步骤201,提取目标图像的LBP特征向量。
LBP(Local Binary Pattern,局部二值模式)是一种图像纹理描述算子,其可以表达图像的局部纹理信息,对光照具有不变性。实际实现时,可以使用Uniform编码方式进行编码,具备算法复杂度较低,消耗内存小,计算速度快等特性。
步骤202,根据LBP特征向量计算目标图像的高维特征向量。
步骤203,获取训练矩阵,训练矩阵为通过联合贝叶斯算法对图像库中的图像训练得到的矩阵。
步骤204,根据目标图像的高维特征向量以及训练矩阵,对目标图像进行识别。
综上所述,本实施例提供的图像识别方法,通过LBP算法提取目标图像的LBP特征向量,进而根据该LBP特征向量以及训练矩阵来对目标图像进行图像识别,该训练矩阵为通过联合贝叶斯算法对图像库训练得到的矩阵;图像识别的准确度较高;达到了可以将LBP算法和联合贝叶斯算法进行结合来进行图像识别,提高图像识别的准确度的效果。上述图像的一个例子是人脸图像,图像库中的图像包括用户处于不同角度、不同表情或者不同光照时的图像,所以上述方法在人脸被遮挡或者面部光照变化强烈时,仍然可以有较高的人脸识别准确率。
请参考图3A,其示出了本发明一个实施例提供的图像识别方法的流程图,本实施例以人脸图像来举例说明,本领域的技术人员可以理解本实施例的方法同样可以用于其它图像的识别。该方法可以由图1所示的计算装置来实施。下文以计算装置是终端(例如计算机,手机,平板电脑等)为例进行说明,本领 域的技术人员可以理解,本发明不限于此。如图3A所示,该图像识别方法可以包括:
步骤301,提取目标人脸图像的LBP特征向量。
目标人脸图像为输入的待识别的图像。
可选的,请参考图3B,本步骤可以包括:
步骤301a,获取k个缩放后的目标人脸图像,k个缩放后的目标人脸图像为按照k个预设倍数对目标人脸图像分别进行缩放后得到的图像。
终端可以按照k个预设倍数中的每个预设倍数对目标人脸图像分别进行缩放,进而得到k个缩放后的目标人脸图像。其中,k为正整数,且k个预设倍数可以均匀分布。另外,本实施例所说的缩放的一个例子是指对目标人脸图像进行缩小。
比如,以k为4来举例,终端可以对目标人脸图像按照1.25、1.5、1.75和2的倍数进行缩小,得到4张缩小后的目标人脸图像。本实施例对k的个数以及每个预设倍数的具体数值并不做限定。
步骤301b,对于目标人脸图像以及k个缩放后的目标人脸图像中的每一张人脸图像,确定人脸图像中的人脸关键点,提取人脸关键点处的LBP特征向量。
本步骤可以包括:
(1)、识别目标人脸图像中的人脸框。
(2)、对人脸框中的人脸进行五官定位,得到各个人脸关键点。其中,人脸关键点可以包括:左右眉毛、左右眼睛、鼻子和左右嘴角等等。
比如,请参考图3C,终端可以定位得到图中所示的各个人脸关键点。
(3)、对于每个人脸关键点,在以人脸关键点为中心的预设区域内,提取预设区域内的LBP特征。
可选的,终端可以通过Uniform编码来提取预设区域中的LBP特征,得到LBP特征直方图。预设区域可以为a*a的区域,a为正整数。可选的,终端可以通过Uniform编码来提取预设区域中的每个单元格中的LBP特征。
比如,请参考图3D,预设区域可以为以人脸关键点为中心的4*4的区域,且对于图中的每个单元格,终端可以计算得到59维的LBP特征分布直方图。
步骤301c,根据提取到的各个人脸关键点的LBP特征向量,组合确定目标人 脸图像的LBP特征向量。
在终端对每一张人脸图像进行提取之后,终端可以提取得到u个LBP特征。比如,对于图3D所示的人脸图像,终端可以提取得到59*27*16个LBP特征。
而在终端对目标人脸图像以及k个缩放后的目标人脸图像分别进行提取之后,终端可以得到Y个LBP特征。其中Y=u*(k+1)。比如,以k为4来举例,对于图3D所示的人脸图像,终端可以提取得到5*59*27*16个LBP特征。
终端将提取到的各个LBP特征作为LBP特征向量中的一个参数,进而组合得到包含各个LBP特征的LBP特征向量。可选的,终端可以将对每一张人脸图像提取到的LBP特征作为LBP特征向量中的一行或者一列,进而得到包含k+1行或者k+1列的LBP特征向量。
步骤302,根据LBP特征向量计算目标人脸图像的高维特征向量。
设目标人脸图像的LBP特征向量为xr
第一,对
Figure PCTCN2016084418-appb-000001
进行PCA(Principal Component Analysis,主成分分析)降维,保留前N维特征,得到降维矩阵P,N为正整数,
Figure PCTCN2016084418-appb-000002
为xr的转置向量。
N大于第一阈值小于第二阈值,并且,通常情况下,N可以为1440。PCA降维是图像处理中常用的一种降维方法,其通过线性变换,将原特征向量映射到低维度空间中。
第二,对xr,计算:xp=Pxr
第三,对xp进行LDA(Latent Dirichlet Allocation,三层贝叶斯概率模型)降维,得到降维矩阵L。
第四,对xp计算:x=Lxp,x即为目标人脸图像的高维特征向量。
步骤303,获取训练矩阵,训练矩阵为通过联合贝叶斯算法对人脸图像库中的人脸图像训练得到的矩阵。
该训练矩阵可以为本地预先存储的矩阵,也可以为终端从服务器中获取到的矩阵。
步骤304,获取参考人脸图像的高维特征向量。
该参考人脸图像的高维特征向量可以是存储在本地的向量。并且,该高维特征向量可以为终端预先计算并存储在本地的,也可以是终端从服务器中获取并存储在本地的,本实施例对此并不做限定。
另外,参考人脸图像的高维特征向量的计算方法与目标人脸图像的高维特 征向量的计算方法类似,详细技术细节请参考步骤301和步骤302,本实施例对此并不做限定。
步骤305,根据目标人脸图像的高维特征向量、参考人脸图像的高维特征向量以及训练矩阵,计算目标人脸图像和参考人脸图像的相似度。
可选的,若目标人脸图像的高维特征向量为x1,参考人脸图像的高维特征向量为x2,训练矩阵包括A和G,则目标人脸图像和参考人脸图像的相似度为:
Figure PCTCN2016084418-appb-000003
其中,
Figure PCTCN2016084418-appb-000004
为x1的转置向量,
Figure PCTCN2016084418-appb-000005
为x2的转置向量。
可选的,在计算相似度之前,终端还可以执行如下步骤:
第一,获取人脸图像库中的所有特征的均值M。
具体的本步骤可以包括:获取人脸图像库中的每张人脸图像的LBP特征向量,计算获取到的LBP特征向量中所有特征的均值M。其中,终端获取人脸图像的LBP特征向量的获取方式与本实施例中的获取目标人脸图像的LBP特征向量的获取方式类似,本实施例在此不再赘述。
第二,对x1和x2进行均值化处理。例如,令x1=x1-M;x2=x2-M。
通过对x1和x2进行均值化处理,使得x1和x2可以以0为中心分布,提高了相似度的计算效率。
此外,本实施例以通过相似度来对目标人脸图像进行人脸识别来举例说明。在终端计算得到相似度之后,终端还可以计算目标人脸图像的高维特征向量与参考人脸图像的高维特征向量之间的方差,进而通过计算得到的方差来进行人脸识别。
具体的,计算方差的步骤可以包括:
第一,获取人脸图像库中的同一个人的高维特征向量的均值r1,方差s1
对于人脸图像库中的m*n张人脸图像,在同一个人的n张人脸图像中,计算任意两张人脸图像的相似度r(x1,x2),得到(n-1)个r(x1,x2);对于人脸图像库中的m个人,则一共得到(n-1)*m个r(x1,x2);计算(n-1)*m个r(x1,x2)的均值r1以及方差s1
第二,根据计算得到的均值r1,方差s1以及相似度计算方差,方差s为:
Figure PCTCN2016084418-appb-000006
终端计算得到方差s之后,终端可以将计算得到的s归一化到0-100的区间。 其中,s的大小表示目标人脸图像与参考人脸图像是同一个人的概率的大小。换句话说,s是取值介于0-100之间的置信度。s越大,表示两张人脸图像是一个人的概率越高。
可选的,可以设定一个阈值,当s大于设定阈值时,判定目标人脸图像与参考人脸图像是同一个人的;而当s小于该设定阈值时,判定目标人脸图像与参考人脸图像不是同一个人的。该设定阈值可以设置为不小于50且不大于100,设计人员可以依据所需的识别准确度来设置该阈值。若所需的准确度较高,则该阈值设置为较高的值,如为90;而若所需的准确度较低,则该阈值可以设置为较小的值,如60。
综上所述,本实施例提供的人脸识别方法,通过LBP算法提取目标人脸图像的LBP特征向量,进而根据该LBP特征向量以及训练矩阵来对目标人脸图像进行人脸识别,该训练矩阵为通过联合贝叶斯算法对人脸图像库训练得到的矩阵;人脸识别的准确度较高;达到了可以将LBP算法和联合贝叶斯算法进行结合来进行人脸识别,提高人脸识别的准确度的效果。同时,人脸图像库中的图像可以包括用户处于不同角度、不同表情或者不同光照时的图像,所以上述方法在人脸被遮挡或者面部光照变化强烈时,仍然可以有较高的人脸识别准确率。
本实施例在计算人脸图像的高维特征向量时,根据人脸图像以及按照预设倍数对人脸图像缩放后的图像进行计算,其中,预设倍数的一个例子是用于缩小图像的倍数,增强了尺度变换的鲁棒性。
另外,在本实施例在人脸识别时,可以计算方差,并且可以根据所需的不同识别精度来为方差设置不同的阈值,更好的适应了实际的使用需求。
需要补充说明的一点是,在上述实施例中,终端可以从服务器中获取服务器预先训练的训练矩阵,也可以获取终端自身预先训练并存储在本地的训练矩阵,本实施例对此并不做限定。并且,以下以终端计算并存储训练矩阵来举例说明,其具体步骤如下:
第一,获取人脸图像库中的图像对应的每个人的高维特征向量;
人脸图像库包括m*n张人脸图像,m为人脸图像库中的图像对应的人的个数,n为每个人的人脸图像的张数。其中,m和n均为正整数,且每个人的n张人脸图像可以包括用户在不同角度、不同光照条件下和不同表情时的人脸图像。
(1)、对于人脸图像库中的每张人脸图像,提取人脸图像的LBP特征向量。
本步骤与上述实施例中的步骤301和步骤302类似,详细技术细节请参考上述实施例,本实施例在此不再赘述。
(2)、计算人脸图像的LBP特征向量中所有特征的均值。
(3)、将人脸图像的LBP特征向量中的每个特征减去均值,得到特征向量xr={xri,0<i<m*n},其中,xri表示人脸图像均值处理后的LBP特征向量。
(4)、对
Figure PCTCN2016084418-appb-000007
进行PCA降维,保留前N维特征,得到降维矩阵P,其中,N为正整数,
Figure PCTCN2016084418-appb-000008
为xr的转置向量。
(5)、对每个xr,计算:xp=Pxr
(6)、对于每个人的各张人脸图像,对xp进行LDA降维,得到降维矩阵L。
(7)、对每个xp计算:x=Lxp,x即为一个人的高维特征向量。
第二,初始化Sμ和Sε;Sμ为m维方阵,是人脸图像库中每个人的n张人脸图像的协方差矩阵;Sε为m*n方阵,是人脸图像库中不同人之间的协方差矩阵。
第三,计算F和G:
Figure PCTCN2016084418-appb-000009
第四,根据F和G,计算人脸图像库中第i个人的高斯分布均值μi以及第i个人和第j个人的联合分布协方差矩阵εij
Figure PCTCN2016084418-appb-000011
其中,xi为第i个人的高维特征向量,xj为第j个人的高维特征向量。
第五,根据下式更新Sμ和Sε
Figure PCTCN2016084418-appb-000012
其中
Figure PCTCN2016084418-appb-000013
为μi的转置向量,其中
Figure PCTCN2016084418-appb-000014
为εij的转置向量。
第六,若Sμ和Sε不收敛,则再次执行计算F和G的步骤。
第七,若Sμ和Sε收敛,则根据Sμ和Sε收敛时的F和G,以及收敛时的Sμ和Sε,计算并存储训练矩阵,训练矩阵A=(Sμ+Sε)-1-(F+G),
Figure PCTCN2016084418-appb-000015
需要补充说明的另一点是,上述实施例提供的人脸识别方法可以适用于多种应用场景。例如,可应用于互联网金融核对身份,人脸识别,人脸属性识别,人脸美化,人脸卡通画,人脸跟踪,唇语识别,活体识别等场景,可应用于即时通讯应用程序,社交应用分享平台,水印相机,创意相机,天天P图等产品。本发明实施例技术方案的最直接应用是,输入两张人脸图像,计算两个人脸相 似度,可以应用在人脸验证,人脸相似度计算等多个业务场景中。
请参考图4,其示出了本发明一个实施例提供的图像识别装置的结构方框图,该图像识别装置可以包括:特征提取模块401、第一计算模块402、矩阵获取模块403和识别模块404。
特征提取模块401,用于提取目标图像的局部二值模式LBP特征向量;
第一计算模块402,用于根据所述特征提取模块401提取到的所述LBP特征向量计算所述目标图像的高维特征向量;
矩阵获取模块403,用于获取训练矩阵,所述训练矩阵为通过联合贝叶斯算法对图像库中的图像训练得到的矩阵;
识别模块404,用于根据所述第一计算模块402计算得到的所述目标图像的高维特征向量以及所述矩阵获取模块403获取到的所述训练矩阵,对所述目标图像进行识别。
综上所述,本实施例提供的图像识别装置,通过LBP算法提取目标图像的LBP特征向量,进而根据该LBP特征向量以及训练矩阵来对目标图像进行图像识别,该训练矩阵为通过联合贝叶斯算法对图像库训练得到的矩阵;图像识别的准确度较高;达到了可以将LBP算法和联合贝叶斯算法进行结合来进行图像识别,提高图像识别的准确度的效果。上述图像的一个例子是人脸图像,图像库中的图像可以包括用户处于不同角度、不同表情或者不同光照时的图像,所以上述方法在人脸被遮挡或者面部光照变化强烈时,仍然可以有较高的人脸识别准确率。
请参考图5,其示出了本发明一个实施例提供的图像识别装置的结构方框图,该图像识别装置可以包括:特征提取模块501、第一计算模块502、矩阵获取模块503和识别模块504。
特征提取模块501,用于提取目标图像的局部二值模式LBP特征向量;
第一计算模块502,用于根据所述特征提取模块501提取到的所述LBP特征向量计算所述目标图像的高维特征向量;
矩阵获取模块503,用于获取训练矩阵,所述训练矩阵为通过联合贝叶斯算法对图像库中的图像训练得到的矩阵;
识别模块504,用于根据所述第一计算模块502计算得到的所述目标图像的 高维特征向量以及所述矩阵获取模块503获取到的所述训练矩阵,对所述目标图像进行识别。
可选的,所述识别模块504,包括:
特征获取单元504a,用于获取参考图像的高维特征向量;
相似度计算单元504b,用于在所述目标图像的高维特征向量为x1,所述参考图像的高维特征向量为x2,所述训练矩阵包括A和G时,根据所述目标图像的高维特征向量、所述参考图像的高维特征向量以及所述训练矩阵,计算所述目标图像和所述参考图像的相似度,所述相似度为:
Figure PCTCN2016084418-appb-000016
其中,
Figure PCTCN2016084418-appb-000017
为x1的转置向量,
Figure PCTCN2016084418-appb-000018
为x2的转置向量。
可选的,所述图像库包括所述目标图像和所述参考图像。
可选的,所述特征提取模块501,包括:
图像获取单元501a,用于获取k个缩放后的目标图像,k个所述缩放后的目标图像为按照k个预设倍数对所述目标图像分别进行缩放后得到的图像,k为正整数;
特征提取单元501b,用于对于所述目标图像以及k个所述缩放后的目标图像中的每一张图像,确定所述图像中的关键点,提取所述关键点处的LBP特征向量;
特征确定单元501c,用于根据提取到的所述关键点的LBP特征向量,组合确定所述目标图像的LBP特征向量。
可选的,所述第一计算模块502,包括:
第一计算单元502a,用于在所述目标图像的LBP特征向量为xr时,对
Figure PCTCN2016084418-appb-000019
进行主成分分析PCA降维,保留前N维特征,得到降维矩阵P,N为正整数,xr为所述目标图像的LBP特征向量,
Figure PCTCN2016084418-appb-000020
为xr的转置向量;
第二计算单元502b,用于对所述xr,计算:xp=Pxr
第三计算单元502c,用于对所述xp进行三层贝叶斯概率模型LDA降维,得到降维矩阵L;
第四计算单元502d,用于对所述xp计算:x=Lxp,x即为所述目标图像的高维特征向量。
可选的,所述图像为人脸图像,所述图像库是人脸图像库,人脸图像库包括m*n张人脸图像,m为所述人脸图像库中的图像对应的人的个数,n为每个 人的人脸图像的张数;
所述装置还包括:
特征获取模块505,用于在所述矩阵获取模块503获取所述训练矩阵之前,获取所述人脸图像库中的图像对应的每个人的高维特征向量;
初始化模块506,用于初始化Sμ和Sε;Sμ为m维方阵,是所述人脸图像库中每个人的n张人脸图像的协方差矩阵;Sε为m*n方阵,是所述人脸图像库中不同人之间的协方差矩阵;
第二计算模块507,用于计算F和G:
Figure PCTCN2016084418-appb-000021
第三计算模块508,用于根据所述F和所述G,计算所述人脸图像库中第i个人的高斯分布均值μi
Figure PCTCN2016084418-appb-000022
以及所述第i个人和第j个人的联合分布协方差矩阵εij
Figure PCTCN2016084418-appb-000023
其中,xi为所述第i个人的高维特征向量,xj为所述第j个人的高维特征向量;
更新模块509,用于更新所述Sμ和所述Sε
Figure PCTCN2016084418-appb-000024
Figure PCTCN2016084418-appb-000025
其中
Figure PCTCN2016084418-appb-000026
为μi的转置向量,
Figure PCTCN2016084418-appb-000027
为εij的转置向量;
所述第二计算模块507还用于,在所述Sμ和所述Sε不收敛时,再次执行所述计算F和所述G的步骤;
矩阵存储模块510,用于在所述Sμ和所述Sε收敛时,根据所述Sμ和所述Sε收敛时的所述F和所述G,以及收敛时的所述Sμ和所述Sε,计算并存储所述训练矩阵,所述训练矩阵A=(Sμ+Sε)-1-(F+G),
Figure PCTCN2016084418-appb-000028
可选的,所述特征获取模块505,包括:
特征提取单元505a,用于对于所述人脸图像库中的每张人脸图像,提取所述人脸图像的LBP特征向量;
第五计算单元505b,用于计算所述人脸图像的所述LBP特征向量中所有特征的均值;
第六计算单元505c,用于将所述人脸图像的LBP特征向量中的每个特征减去所述均值,得到特征向量xr={xri,0<i<m*n},其中,xri表示所述人脸图像均值处理后的LBP特征向量;
第七计算单元505d,用于对
Figure PCTCN2016084418-appb-000029
进行PCA降维,保留前N维特征,得到降维矩阵P,其中,N为正整数,
Figure PCTCN2016084418-appb-000030
为xr的转置向量;
第八计算单元505e,用于对每个所述xr,计算:xp=Pxr
第九计算单元505f,用于对于每个人的各张人脸图像,对所述xp进行LDA降维,得到降维矩阵L;
第十计算单元505g,用于对每个所述xp计算:x=Lxp,x即为一个人的高维特征向量。
综上所述,本实施例提供的图像识别装置,通过LBP算法提取目标图像的LBP特征向量,进而根据该LBP特征向量以及训练矩阵来对目标图像进行图像识别,该训练矩阵为通过联合贝叶斯算法对图像库训练得到的矩阵;图像识别的准确度较高;达到了可以将LBP算法和联合贝叶斯算法进行结合来进行图像识别,提高识别的准确度的效果。上述图像的一个例子是人脸图像,图像库中的图像可以包括用户处于不同角度、不同表情或者不同光照时的图像,所以上述方法在人脸被遮挡或者面部光照变化强烈时,仍然可以有较高的人脸识别准确率。
本实施例在计算图像的高维特征向量时,根据图像以及按照预设倍数对图像缩放后的图像进行计算,其中,预设倍数可以是用于缩小图像的倍数,增强了尺度变换的鲁棒性。
另外,在本实施例在图像识别时,可以计算方差,并且可以根据所需的不同识别精度来为方差设置不同的阈值,更好的适应了实际的使用需求。
需要说明的是:上述实施例提供的图像识别装置在进行图像识别时,仅以上述各功能模块的划分进行举例说明,实际应用中,可以根据需要而将上述功能分配由不同的功能模块完成,即将设备的内部结构划分成不同的功能模块,以完成以上描述的全部或者部分功能。另外,上述实施例提供的图像识别装置与图像识别方法实施例属于同一构思,其具体实现过程详见方法实施例,这里不再赘述。
上述本发明实施例序号仅仅为了描述,不代表实施例的优劣。
本领域普通技术人员可以理解实现上述实施例的全部或部分步骤可以通过硬件来完成,也可以通过程序来指令相关的硬件完成,所述的程序可以存储于一种计算机可读存储介质中,上述提到的存储介质可以是只读存储器,磁盘或光盘等。
以上所述仅为本发明的较佳实施例,并不用以限制本发明,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 (19)

  1. 一种图像识别方法,包括:
    提取目标图像的局部二值模式LBP特征向量;
    根据所述LBP特征向量计算所述目标图像的高维特征向量;
    获取训练矩阵,所述训练矩阵为通过联合贝叶斯算法对图像库中的图像训练得到的矩阵;
    根据所述目标图像的高维特征向量以及所述训练矩阵,对所述目标图像进行识别。
  2. 根据权利要求1所述的方法,其中,所述根据所述目标图像的高维特征向量以及所述训练矩阵,对所述目标图像进行识别,包括:
    获取参考图像的高维特征向量;
    根据所述目标图像的高维特征向量、所述参考图像的高维特征向量以及所述训练矩阵,计算所述目标图像和所述参考图像的相似度:
    Figure PCTCN2016084418-appb-100001
    其中,x1为所述目标图像的高维特征向量,x2为所述参考图像的高维特征向量,
    Figure PCTCN2016084418-appb-100002
    为x1的转置向量,
    Figure PCTCN2016084418-appb-100003
    为x2的转置向量,A和G为训练矩阵。
  3. 根据权利要求2所述的方法,其中,所述图像库包括所述目标图像和所述参考图像。
  4. 根据权利要求1所述的方法,其中,所述提取目标图像的局部二值模式LBP特征向量,包括:
    按照k个预设倍数对所述目标图像分别进行缩放以得到k个缩放后的图像,k为正整数;
    对于所述目标图像以及k个所述缩放后的图像中的每一张图像,确定所述图像中的关键点,提取所述关键点处的LBP特征向量;
    根据提取到的所述关键点的LBP特征向量,确定所述目标图像的LBP特征向量。
  5. 根据权利要求1所述的方法,其中,所述根据所述LBP特征向量计算所述目标图像的高维特征向量,包括:
    Figure PCTCN2016084418-appb-100004
    进行主成分分析PCA降维,保留前N维特征,得到降维矩阵P,其中,N为正整数,xr为所述目标图像的LBP特征向量,
    Figure PCTCN2016084418-appb-100005
    为xr的转置向量;
    对所述xr,计算:xp=Pxr
    对所述xp进行三层贝叶斯概率模型LDA降维,得到降维矩阵L;
    计算所述目标图像的高维特征向量x=Lxp
  6. 根据权利要求1所述的方法,其中所述图像为人脸图像。
  7. 根据权利要求6所述的方法,其中,所述图像库包括m*n张人脸图像,m为所述图像库中的图像对应的人的个数,n为每个人的人脸图像的张数;
    所述获取训练矩阵之前,所述方法还包括:
    获取所述图像库中的图像对应的每个人的高维特征向量;
    初始化所述图像库中每个人的n张人脸图像的协方差矩阵Sμ和所述图像库中不同人的图像之间的协方差矩阵Sε,其中Sμ为m维方阵,Sε为m*n方阵;
    计算F和G:
    Figure PCTCN2016084418-appb-100006
    根据所述F和所述G,计算所述图像库中第i个人的高斯分布均值μi
    Figure PCTCN2016084418-appb-100007
    以及所述第i个人和第j个人的联合分布协方差矩阵εij
    Figure PCTCN2016084418-appb-100008
    其中,xi为所述第i个人的高维特征向量,xj为所述第j个人的高维特征向量;
    根据下式更新所述Sμ和所述Sε
    Figure PCTCN2016084418-appb-100009
    其中
    Figure PCTCN2016084418-appb-100010
    为μi的转置向量,
    Figure PCTCN2016084418-appb-100011
    为εij的转置向量;
    根据更新后的所述Sμ和所述Sε以及所述F和所述G,计算训练矩阵
    Figure PCTCN2016084418-appb-100012
  8. 根据权利要求7所述的方法,还包括:
    判断所述更新后的Sμ和所述Sε是否收敛;
    若所述更新后的Sμ和所述Sε不收敛,则再次执行所述计算F和G的步骤;
    若所述更新后的Sμ和所述Sε收敛,则根据Sμ和Sε收敛时的所述F和所述G,以及收敛时的所述Sμ和所述Sε,执行计算训练矩阵的步骤。
  9. 根据权利要求7所述的方法,其中,所述获取图像库中的图像对应的每个人的高维特征向量,包括:
    对于所述图像库中的每张人脸图像,提取所述人脸图像的LBP特征向量;
    计算所述人脸图像的所述LBP特征向量中所有特征的均值;
    将所述人脸图像的LBP特征向量中的每个特征减去所述均值,得到特征向量xr={xri,0<i<m*n},其中,xri表示所述人脸图像均值处理后的LBP特征向量;
    Figure PCTCN2016084418-appb-100013
    进行PCA降维,保留前N维特征,得到降维矩阵P,其中,N为正整数,
    Figure PCTCN2016084418-appb-100014
    为xr的转置向量;
    对每个所述xr,计算:xp=Pxr
    对于每个人的各张人脸图像,对所述xp进行LDA降维,得到降维矩阵L;
    计算每个人的高维特征向量x=Lxp
  10. 一种图像识别装置,包括:
    至少一个处理器;和
    存储器,其中所述存储器存储有程序指令,所述程序指令当由所述处理器执行时,配置所述图像识别装置执行下述操作:
    提取目标图像的局部二值模式LBP特征向量;
    根据所述LBP特征向量计算所述目标图像的高维特征向量;
    获取训练矩阵,所述训练矩阵为通过联合贝叶斯算法对图像库中的图像训练得到的矩阵;
    根据所述目标图像的高维特征向量以及所述训练矩阵,对所述目标图像进行识别。
  11. 根据权利要求10所述的装置,其中,所述根据所述目标图像的高维特征向量以及所述训练矩阵,对所述目标图像进行识别,包括:
    获取参考图像的高维特征向量;
    根据所述目标图像的高维特征向量、所述参考图像的高维特征向量以及所述训练矩阵,计算所述目标图像和所述参考图像的相似度:
    Figure PCTCN2016084418-appb-100015
    其中,x1为所述目标图像的高维特征向量,x2为所述参考图像的高维特征向量,
    Figure PCTCN2016084418-appb-100016
    为x1的转置向量,
    Figure PCTCN2016084418-appb-100017
    为x2的转置向量,A和G为训练矩阵。
  12. 根据权利要求11所述的装置,其中,所述图像库包括所述目标图像和所述参考图像。
  13. 根据权利要求10所述的装置,其中,所述提取目标图像的局部二值模式LBP特征向量,包括:
    按照k个预设倍数对所述目标图像分别进行缩放以得到k个缩放后的图像,k为正整数;
    对于所述目标图像以及k个所述缩放后的图像中的每一张图像,确定所述图像中的关键点,提取所述关键点处的LBP特征向量;
    根据提取到的所述关键点的LBP特征向量,确定所述目标图像的LBP特征向量。
  14. 根据权利要求10所述的装置,其中,所述根据所述LBP特征向量计算所述目标图像的高维特征向量,包括:
    Figure PCTCN2016084418-appb-100018
    进行主成分分析PCA降维,保留前N维特征,得到降维矩阵P,其中,N为正整数,xr为所述目标图像的LBP特征向量,
    Figure PCTCN2016084418-appb-100019
    为xr的转置向量;
    对所述xr,计算:xp=Pxr
    对所述xp进行三层贝叶斯概率模型LDA降维,得到降维矩阵L;
    计算所述目标图像的高维特征向量x=Lxp
  15. 根据权利要求10所述的装置,其中所述图像为人脸图像。
  16. 根据权利要求15所述的装置,其中,所述图像库包括m*n张人脸图像, m为所述图像库中的图像对应的人的个数,n为每个人的人脸图像的张数;
    所述操作还包括:获取所述训练矩阵之前,获取所述图像库中的图像对应的每个人的高维特征向量;
    初始化所述图像库中每个人的n张人脸图像的协方差矩阵Sμ和所述图像库中不同人的图像之间的协方差矩阵Sε,其中Sμ为m维方阵,Sε为m*n方阵;
    计算F和G:
    Figure PCTCN2016084418-appb-100020
    根据所述F和所述G,计算所述图像库中第i个人的高斯分布均值μi
    Figure PCTCN2016084418-appb-100021
    以及所述第i个人和第j个人的联合分布协方差矩阵εij
    Figure PCTCN2016084418-appb-100022
    其中,xi为所述第i个人的高维特征向量,xj为所述第j个人的高维特征向量;
    根据下式更新所述Sμ和所述Sε
    Figure PCTCN2016084418-appb-100023
    其中
    Figure PCTCN2016084418-appb-100024
    为μi的转置向量,其中
    Figure PCTCN2016084418-appb-100025
    为εij的转置向量;
    根据更新后的所述Sμ和所述Sε以及所述F和所述G,计算所述训练矩阵
    Figure PCTCN2016084418-appb-100026
  17. 根据权利要求16所述的装置,其中所述操作还包括:
    判断所述更新后的Sμ和所述Sε是否收敛;
    若所述更新后的Sμ和所述Sε不收敛,则再次执行所述计算F和G的步骤;
    若所述更新后的Sμ和所述Sε收敛,则根据Sμ和Sε收敛时的所述F和所述G,以及收敛时的所述Sμ和所述Sε,执行计算训练矩阵的步骤。
  18. 根据权利要求16所述的装置,其中,所述获取图像库中的图像对应的每个人的高维特征向量,包括:
    对于所述图像库中的每张人脸图像,提取所述人脸图像的LBP特征向量;
    计算所述人脸图像的所述LBP特征向量中所有特征的均值;
    将所述人脸图像的LBP特征向量中的每个特征减去所述均值,得到特征向量xr={xri,0<i<m*n},其中,xri表示所述人脸图像均值处理后的LBP特征向量;
    Figure PCTCN2016084418-appb-100027
    进行PCA降维,保留前N维特征,得到降维矩阵P,其中,N为正整数,
    Figure PCTCN2016084418-appb-100028
    为xr的转置向量;
    对每个所述xr,计算:xp=Pxr
    对于每个人的各张人脸图像,对所述xp进行LDA降维,得到降维矩阵L;
    计算每个人的高维特征向量x=Lxp
  19. 一种计算机可读存储介质,所述存储介质存储有程序指令,所述程序指令当由计算装置的处理器执行时,配置所述装置执行根据权利要求1-9中任一项所述的方法。
PCT/CN2016/084418 2015-11-26 2016-06-02 图像识别方法和装置 WO2017088432A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US15/925,028 US10713532B2 (en) 2015-11-26 2018-03-19 Image recognition method and apparatus

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201510843973.6A CN106803055B (zh) 2015-11-26 2015-11-26 人脸识别方法和装置
CN201510843973.6 2015-11-26

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US15/925,028 Continuation US10713532B2 (en) 2015-11-26 2018-03-19 Image recognition method and apparatus

Publications (1)

Publication Number Publication Date
WO2017088432A1 true WO2017088432A1 (zh) 2017-06-01

Family

ID=58762964

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2016/084418 WO2017088432A1 (zh) 2015-11-26 2016-06-02 图像识别方法和装置

Country Status (3)

Country Link
US (1) US10713532B2 (zh)
CN (1) CN106803055B (zh)
WO (1) WO2017088432A1 (zh)

Cited By (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108765264A (zh) * 2018-05-21 2018-11-06 深圳市梦网科技发展有限公司 图像美颜方法、装置、设备及存储介质
CN109509140A (zh) * 2017-09-15 2019-03-22 阿里巴巴集团控股有限公司 显示方法及装置
CN109754059A (zh) * 2018-12-21 2019-05-14 平安科技(深圳)有限公司 翻拍图像识别方法、装置、计算机设备和存储介质
AU2019200336A1 (en) * 2018-04-02 2019-10-17 Pond5, Inc. Method and system for image searching
CN111260763A (zh) * 2020-01-21 2020-06-09 厦门美图之家科技有限公司 基于人像的卡通形象生成方法、装置、设备及存储介质
CN112330713A (zh) * 2020-11-26 2021-02-05 南京工程学院 基于唇语识别的重度听障患者言语理解度的改进方法
CN112529888A (zh) * 2020-12-18 2021-03-19 平安科技(深圳)有限公司 基于深度学习的人脸图像评估方法、装置、设备及介质

Families Citing this family (14)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US10346461B1 (en) * 2018-04-02 2019-07-09 Pond5 Inc. Method and system for image searching by color
CN108304833A (zh) * 2018-04-17 2018-07-20 哈尔滨师范大学 基于mblbp和dct-bm2dpca的人脸识别方法
CN108416336B (zh) * 2018-04-18 2019-01-18 特斯联(北京)科技有限公司 一种智慧社区人脸识别的方法和系统
CN108960124B (zh) * 2018-06-28 2021-10-01 北京陌上花科技有限公司 用于行人再识别的图像处理方法及装置
CN109242018A (zh) * 2018-08-31 2019-01-18 平安科技(深圳)有限公司 图像验证方法、装置、计算机设备及存储介质
CN109615614B (zh) * 2018-11-26 2020-08-18 北京工业大学 基于多特征融合的眼底图像中血管的提取方法与电子设备
CN109977860B (zh) * 2019-03-25 2021-07-16 上海商汤智能科技有限公司 图像处理方法及装置、电子设备和存储介质
CN110084216B (zh) * 2019-05-06 2021-11-09 苏州科达科技股份有限公司 人脸识别模型训练和人脸识别方法、系统、设备及介质
CN112541564B (zh) * 2019-09-20 2024-02-20 腾讯科技(深圳)有限公司 降低贝叶斯深度神经网络计算复杂度的方法和装置
CN110795592B (zh) * 2019-10-28 2023-01-31 深圳市商汤科技有限公司 图片处理方法、装置及设备
CN111709344B (zh) * 2020-06-09 2023-10-17 上海海事大学 一种基于高斯混合模型的epll图像去光照识别处理方法
CN111738194B (zh) * 2020-06-29 2024-02-02 深圳力维智联技术有限公司 一种用于人脸图像相似性的评价方法和装置
CN114399058B (zh) * 2022-03-25 2022-06-10 腾讯科技(深圳)有限公司 一种模型更新的方法、相关装置、设备以及存储介质
CN115810108B (zh) * 2022-09-23 2023-08-08 南京审计大学 一种基于regnmf的大数据审计中图像特征提取方法

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20100329517A1 (en) * 2009-06-26 2010-12-30 Microsoft Corporation Boosted face verification
CN102004924A (zh) * 2010-11-03 2011-04-06 无锡中星微电子有限公司 一种人头检测系统及方法
CN104573652A (zh) * 2015-01-04 2015-04-29 华为技术有限公司 确定人脸图像中人脸的身份标识的方法、装置和终端
CN204791050U (zh) * 2015-04-01 2015-11-18 北京市商汤科技开发有限公司 身份认证设备
CN105138968A (zh) * 2015-08-05 2015-12-09 北京天诚盛业科技有限公司 人脸认证方法和装置
CN105550658A (zh) * 2015-12-24 2016-05-04 蔡叶荷 一种基于高维lbp与卷积神经网络特征融合的人脸比对方法

Family Cites Families (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US7848566B2 (en) * 2004-10-22 2010-12-07 Carnegie Mellon University Object recognizer and detector for two-dimensional images using bayesian network based classifier
US20150235073A1 (en) * 2014-01-28 2015-08-20 The Trustees Of The Stevens Institute Of Technology Flexible part-based representation for real-world face recognition apparatus and methods
KR102010378B1 (ko) * 2014-09-24 2019-08-13 삼성전자주식회사 객체를 포함하는 영상의 특징을 추출하는 방법 및 장치
US20160173560A1 (en) * 2014-12-12 2016-06-16 Genesis Media Llc Digital Content Delivery Based on Measures of Content Appeal and User Motivation
CN104751143B (zh) * 2015-04-02 2018-05-11 北京中盾安全技术开发公司 一种基于深度学习的人证核验系统及方法

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20100329517A1 (en) * 2009-06-26 2010-12-30 Microsoft Corporation Boosted face verification
CN102004924A (zh) * 2010-11-03 2011-04-06 无锡中星微电子有限公司 一种人头检测系统及方法
CN104573652A (zh) * 2015-01-04 2015-04-29 华为技术有限公司 确定人脸图像中人脸的身份标识的方法、装置和终端
CN204791050U (zh) * 2015-04-01 2015-11-18 北京市商汤科技开发有限公司 身份认证设备
CN105138968A (zh) * 2015-08-05 2015-12-09 北京天诚盛业科技有限公司 人脸认证方法和装置
CN105550658A (zh) * 2015-12-24 2016-05-04 蔡叶荷 一种基于高维lbp与卷积神经网络特征融合的人脸比对方法

Cited By (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109509140A (zh) * 2017-09-15 2019-03-22 阿里巴巴集团控股有限公司 显示方法及装置
AU2019200336A1 (en) * 2018-04-02 2019-10-17 Pond5, Inc. Method and system for image searching
AU2019200336B2 (en) * 2018-04-02 2020-02-20 Pond5, Inc. Method and system for image searching
CN108765264A (zh) * 2018-05-21 2018-11-06 深圳市梦网科技发展有限公司 图像美颜方法、装置、设备及存储介质
CN108765264B (zh) * 2018-05-21 2022-05-20 深圳市梦网科技发展有限公司 图像美颜方法、装置、设备及存储介质
CN109754059A (zh) * 2018-12-21 2019-05-14 平安科技(深圳)有限公司 翻拍图像识别方法、装置、计算机设备和存储介质
CN111260763A (zh) * 2020-01-21 2020-06-09 厦门美图之家科技有限公司 基于人像的卡通形象生成方法、装置、设备及存储介质
CN112330713A (zh) * 2020-11-26 2021-02-05 南京工程学院 基于唇语识别的重度听障患者言语理解度的改进方法
CN112330713B (zh) * 2020-11-26 2023-12-19 南京工程学院 基于唇语识别的重度听障患者言语理解度的改进方法
CN112529888A (zh) * 2020-12-18 2021-03-19 平安科技(深圳)有限公司 基于深度学习的人脸图像评估方法、装置、设备及介质
CN112529888B (zh) * 2020-12-18 2024-04-30 平安科技(深圳)有限公司 基于深度学习的人脸图像评估方法、装置、设备及介质

Also Published As

Publication number Publication date
US10713532B2 (en) 2020-07-14
CN106803055A (zh) 2017-06-06
US20180204094A1 (en) 2018-07-19
CN106803055B (zh) 2019-10-25

Similar Documents

Publication Publication Date Title
WO2017088432A1 (zh) 图像识别方法和装置
US11138413B2 (en) Fast, embedded, hybrid video face recognition system
US11915514B2 (en) Method and apparatus for detecting facial key points, computer device, and storage medium
EP3084682B1 (en) System and method for identifying faces in unconstrained media
US9928405B2 (en) System and method for detecting and tracking facial features in images
WO2021051545A1 (zh) 基于行为识别模型的摔倒动作判定方法、装置、计算机设备及存储介质
US9626552B2 (en) Calculating facial image similarity
US9443325B2 (en) Image processing apparatus, image processing method, and computer program
Li et al. A comprehensive survey on 3D face recognition methods
Mehta et al. Face recognition using scale-adaptive directional and textural features
CN109271930B (zh) 微表情识别方法、装置与存储介质
WO2019033569A1 (zh) 眼球动作分析方法、装置及存储介质
WO2018082308A1 (zh) 一种图像处理方法及终端
Liao et al. Discriminant Analysis via Joint Euler Transform and $\ell_ {2, 1} $-Norm
CN109087240B (zh) 图像处理方法、图像处理装置及存储介质
Biswas et al. Face recognition in low-resolution videos using learning-based likelihood measurement model
Bąk et al. Deep deformable patch metric learning for person re-identification
US9659210B1 (en) System and method for detecting and tracking facial features in images
CN109241942B (zh) 图像处理方法、装置、人脸识别设备及存储介质
Abavisani et al. A robust sparse representation based face recognition system for smartphones
Cui et al. Improving the face recognition system by hybrid image preprocessing
Cui et al. Robust facial landmark localization using classified random ferns and pose-based initialization
Kim et al. Feature scalability for a low complexity face recognition with unconstrained spatial resolution
CN114627560A (zh) 一种动作识别方法、动作识别模型训练方法及相关装置
Tomaselli et al. Low complexity smile detection technique for mobile devices

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: 16867660

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 06/11/2018)

122 Ep: pct application non-entry in european phase

Ref document number: 16867660

Country of ref document: EP

Kind code of ref document: A1