CN107958245A

CN107958245A - A kind of gender classification method and device based on face characteristic

Info

Publication number: CN107958245A
Application number: CN201810031037.9A
Authority: CN
Inventors: 纪政
Original assignee: Shanghai Zheng Peng Mdt Infotech Ltd
Current assignee: Shenzhen Car Top Media Co Ltd
Priority date: 2018-01-12
Filing date: 2018-01-12
Publication date: 2018-04-24

Abstract

The invention discloses a kind of gender classification method and device based on face characteristic, by being pre-processed according to the eyes elements of a fix to facial image；Graphical rule is carried out to the facial image to scale to form multi-resolution Gaussian pyramid；Gabor wavelet conversion is carried out to different scale and direction；Local binary patterns conversion is carried out to the Gabor facial images, and by being both horizontally and vertically evenly dividing the facial image；Extract the histogram data in each region of Gabor local binary patterns facial image；The sample data that masculinity and femininity label is carried in the multiresolution part Gabor binary pattern MLGBP features is gathered, and central junction line dimensionality reduction is carried out to each division provincial characteristics vector；It is trained in Gender Classification model based on support vector machines, obtains gender disaggregated model；Input the feature into the Gender Classification model and obtain analysis result.

Description

Gender classification method and device based on face features

Technical Field

The invention relates to the technical field of face recognition, in particular to a gender classification method and a gender classification device based on face features.

Background

Image-based gender classification is a complex and challenging two-class pattern classification problem because the mechanisms of human brain to identify human gender in images are very poorly studied, and image-feature-based gender classification has attracted the attention of many researchers due to potential applications in human-computer interaction, such as in demographic, video surveillance, and machine vision. For this problem, it is the most challenging to design an excellent image feature extraction algorithm for gender classification. An important problem in the field of pattern classification is how to find an excellent feature extraction method. In general, different feature extraction methods are often suitable for different applications. For example, for the gender classification problem, feature extraction algorithms suitable for gender classification applications need to be sought.

In the process of implementing the technical scheme of the present invention, the inventor of the present application finds that the above prior art has at least the following technical problems:

in the prior art, the method for describing the gender characteristics is difficult to ensure the robustness under the changing conditions of illumination, expression, posture and the like.

Disclosure of Invention

The invention provides a gender classification method and a gender classification device based on human face features, which are used for solving the technical problem that the method for describing gender features in the prior art is difficult to ensure robustness under the changing conditions of illumination, expression, posture and the like.

In one aspect, the invention provides a gender classification method based on human face features, which comprises the following steps: preprocessing the face image according to the eye positioning coordinates; scaling the image scale of the face image according to the original size, one fourth and one sixteenth original sizes to form a multi-resolution Gaussian pyramid, and respectively performing Gaussian filtering; carrying out Gabor wavelet transformation on each face image in the resolution Gaussian pyramid according to different scales and directions to obtain a Gabor face image; carrying out local binary pattern conversion on the Gabor face image, and uniformly dividing the face image according to the horizontal and vertical directions to obtain a Gabor local binary pattern face image; extracting histogram data of each region of the Gabor local binary pattern human face image, and obtaining multi-resolution local Gabor binary pattern MLGBP characteristic vectors according to the sequence from top to bottom and from left to right; collecting sample data with male and female labels in the multiresolution local Gabor binary pattern MLGBP characteristics, and performing central connecting line dimensionality reduction on each divided region characteristic vector to obtain a multiresolution local Gabor binary pattern characteristic vector MLGBP-CCL based on the central connecting line CCL dimensionality reduction; inputting characteristic data with male and female labels in the MLGBP-CCL into a gender classification model based on a support vector machine for training to obtain a gender classification model; and inputting the MLGBP-CCL characteristics into the gender classification model to obtain an analysis result.

Preferably, the inputting the MLGBP-CCL features into the gender classification model to obtain an analysis result further includes: detecting a face image from a video image; extracting multi-resolution local Gabor binary pattern face features MLGBP; obtaining the MLGBP-CCL characteristic vector by utilizing the human face characteristic MLGBP and the central connecting line CCL; and inputting the MLGBP-CCL characteristic vector into the gender classification model to obtain an analysis result.

Preferably, the pretreatment comprises: face detection, face image graying and face image standardization, wherein the face image graying is to convert an RGB image into a YUV format, then directly obtain a Y value, and the formula is that Gray is 0.299 x R +0.587 x G +0.114 x B, and Y is image brightness information; the face image is standardized by converting the input face image into a uniform structure and correcting the face into a uniform size.

Preferably, the multi-resolution gaussian pyramid is a pyramid hierarchy structure formed from bottom to top according to the resolution by performing image scaling and gaussian filtering on the standardized face image.

Preferably, the method further comprises: the Gabor filter is generated by Gabor transformation, a multi-resolution image is decomposed by using Gabor filters with multiple scales and multiple directions, and each block of average gray value in the Gabor image is connected to form a feature vector, wherein the Gabor filter is defined by the following formula:where u and v are defined as the orientation and scale of the Gabor filter, respectively, and z is (x, y)^TAnd | · | | represents norm operation, and σ is a measure of the width and wave of the Gaussian windowThe long ratio.

Preferably, the MLGBP-CCL feature vector is a feature vector obtained by reducing the dimension of the MLGBP feature vector for the center connecting line CCL.

Preferably, the method further comprises: the support vector machine is a machine learning method, and an optimal hyperplane is constructed in a sample input space or a feature space, so that the distance between the hyperplane and two types of sample sets is maximized.

Preferably, the method further comprises: the face coordinate transformation formula is as follows:

wherein, (LX, LY) left eye coordinates, (RX, RY) right eye coordinates, (x, y) original image coordinates, (x ', y') transformed coordinates, (LX, LY) standard face left eye coordinates, (RX, RY) standard face right eye coordinates,

on the other hand, the embodiment of the invention also provides a gender classification device based on the human face characteristics, which comprises: the first preprocessing unit is used for preprocessing the face image according to the eye positioning coordinates; the first filtering unit is used for carrying out image scale scaling on the face image according to the original size, the quarter original size and the sixteenth original size to form a multi-resolution Gaussian pyramid and respectively carrying out Gaussian filtering; the first obtaining unit is used for carrying out Gabor wavelet transformation on each face image in the resolution Gaussian pyramid according to different scales and directions to obtain a Gabor face image; the second obtaining unit is used for carrying out local binary pattern conversion on the Gabor face image, uniformly dividing the face image in the horizontal and vertical directions and obtaining a Gabor local binary pattern face image; a third obtaining unit, configured to extract histogram data of each region of the Gabor local binary pattern face image, and obtain a multiresolution local Gabor binary pattern MLGBP feature vector according to a sequence from top to bottom and from left to right; a fourth obtaining unit, configured to acquire sample data with male and female labels in the multiresolution local Gabor binary pattern MLGBP feature, and perform central connecting line dimensionality reduction on each divided region feature vector to obtain a multiresolution local Gabor binary pattern feature vector MLGBP-CCL based on the central connecting line CCL dimensionality reduction; a fifth obtaining unit, configured to input feature data with male and female labels in the MLGBP-CCL into a gender classification model based on a support vector machine for training, so as to obtain a gender classification model; a sixth obtaining unit, configured to input the MLGBP-CCL features into the gender classification model to obtain an analysis result.

In another aspect, an embodiment of the present invention further provides a gender classification apparatus based on human face features, where the apparatus includes: a memory, a processor and a computer program stored on the memory and executable on the processor, the processor when executing the program implementing the steps of: preprocessing the face image according to the eye positioning coordinates; scaling the image scale of the face image according to the original size, one fourth and one sixteenth original sizes to form a multi-resolution Gaussian pyramid, and respectively performing Gaussian filtering; carrying out Gabor wavelet transformation on each face image in the resolution Gaussian pyramid according to different scales and directions to obtain a Gabor face image; carrying out local binary pattern conversion on the Gabor face image, and uniformly dividing the face image according to the horizontal and vertical directions to obtain a Gabor local binary pattern face image; extracting histogram data of each region of the Gabor local binary pattern human face image, and obtaining multi-resolution local Gabor binary pattern MLGBP characteristic vectors according to the sequence from top to bottom and from left to right; collecting sample data with male and female labels in the multiresolution local Gabor binary pattern MLGBP characteristics, and performing central connecting line dimensionality reduction on each divided region characteristic vector to obtain a multiresolution local Gabor binary pattern characteristic vector MLGBP-CCL based on the central connecting line CCL dimensionality reduction; inputting characteristic data with male and female labels in the MLGBP-CCL into a gender classification model based on a support vector machine for training to obtain a gender classification model; and inputting the MLGBP-CCL characteristics into the gender classification model to obtain an analysis result.

One or more technical solutions in the embodiments of the present invention at least have one or more of the following technical effects:

in the gender classification method and device based on the human face features, the human face images are preprocessed according to the eye positioning coordinates; scaling the image scale of the face image according to the original size, one fourth and one sixteenth original sizes to form a multi-resolution Gaussian pyramid, and respectively performing Gaussian filtering; carrying out Gabor wavelet transformation on each face image in the resolution Gaussian pyramid according to different scales and directions to obtain a Gabor face image; carrying out local binary pattern conversion on the Gabor face image, and uniformly dividing the face image according to the horizontal and vertical directions to obtain a Gabor local binary pattern face image; extracting histogram data of each region of the Gabor local binary pattern human face image, and obtaining multi-resolution local Gabor binary pattern MLGBP characteristic vectors according to the sequence from top to bottom and from left to right; collecting sample data with male and female labels in the multiresolution local Gabor binary pattern MLGBP characteristics, and performing central connecting line dimensionality reduction on each divided region characteristic vector to obtain a multiresolution local Gabor binary pattern characteristic vector MLGBP-CCL based on the central connecting line CCL dimensionality reduction; inputting characteristic data with male and female labels in the MLGBP-CCL into a gender classification model based on a support vector machine for training to obtain a gender classification model; and inputting the MLGBP-CCL characteristics into the gender classification model to obtain an analysis result. The technical problem that the method for describing the gender characteristics in the prior art is difficult to ensure robustness under the changing conditions of illumination, expression, posture and the like is solved, the technical effects of adapting to illumination change and describing better shape texture information to distinguish gender of men and women are achieved, and the method is suitable for different application man-machine interaction and has outstanding progress.

The foregoing description is only an overview of the technical solutions of the present invention, and the embodiments of the present invention are described below in order to make the technical means of the present invention more clearly understood and to make the above and other objects, features, and advantages of the present invention more clearly understandable.

Drawings

Fig. 1 is a schematic flow chart of a gender classification method based on human face features according to an embodiment of the present invention;

FIG. 2 is a schematic diagram of a Gabor wavelet transform process of a face image according to an embodiment of the present invention;

fig. 3 is a schematic structural diagram of a gender classification apparatus based on human face features according to an embodiment of the present invention;

fig. 4 is a schematic structural diagram of another gender classification device based on human face features according to an embodiment of the present invention.

Detailed Description

The embodiment of the invention provides a gender classification method and a gender classification device based on facial features, which are used for solving the technical problem that the method for describing gender features in the prior art is difficult to ensure robustness under the changing conditions of illumination, expression, posture and the like.

The technical method and the device in the embodiment of the invention have the following general idea: preprocessing the face image according to the eye positioning coordinates; scaling the image scale of the face image according to the original size, one fourth and one sixteenth original sizes to form a multi-resolution Gaussian pyramid, and respectively performing Gaussian filtering; carrying out Gabor wavelet transformation on each face image in the resolution Gaussian pyramid according to different scales and directions to obtain a Gabor face image; carrying out local binary pattern conversion on the Gabor face image, and uniformly dividing the face image according to the horizontal and vertical directions to obtain a Gabor local binary pattern face image; extracting histogram data of each region of the Gabor local binary pattern human face image, and obtaining multi-resolution local Gabor binary pattern MLGBP characteristic vectors according to the sequence from top to bottom and from left to right; collecting sample data with male and female labels in the multiresolution local Gabor binary pattern MLGBP characteristics, and performing central connecting line dimensionality reduction on each divided region characteristic vector to obtain a multiresolution local Gabor binary pattern characteristic vector MLGBP-CCL based on the central connecting line CCL dimensionality reduction; inputting characteristic data with male and female labels in the MLGBP-CCL into a gender classification model based on a support vector machine for training to obtain a gender classification model; and inputting the MLGBP-CCL characteristics into the gender classification model to obtain an analysis result. The method achieves the technical effects of adapting to illumination change and describing better shape texture information to distinguish gender of men and women, is suitable for different application human-computer interaction and has outstanding progress.

In order to make the objects, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are some, but not all, embodiments of the present invention. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

Example one

Fig. 1 is a schematic flow chart of a gender classification method based on human face features in an embodiment of the present invention. As shown in fig. 1, the method includes:

step 10: preprocessing the face image according to the eye positioning coordinates;

further, the preprocessing comprises: face detection, face image graying and face image standardization, wherein the face image graying is to convert an RGB image into a YUV format, then directly obtain a Y value, and the formula is that Gray is 0.299 x R +0.587 x G +0.114 x B, and Y is image brightness information; the face image is standardized by converting the input face image into a uniform structure and correcting the face into a uniform size.

Further, the method further comprises: the face coordinate transformation formula isWherein, (LX, LY) left eye coordinates, (RX, RY) right eye coordinates, (x, y) original image coordinates, (x ', y') transformed coordinates, (LX, LY) standard face left eye coordinates, (RX, RY) standard face right eye coordinates,

specifically, the face image is preprocessed according to the eye coordinate positioning of the face image, the preprocessing mainly comprises face detection, face image graying and face image standardization, the face detection mainly comprises face detection and eye positioning, firstly, the main task of the face detection is to judge whether a face exists in the image, and if the face exists, the coordinate position, the face area size and other information of the face in the image need to be given; secondly, extracting a face image area and carrying out graying on the face image area; finally, the eye position information can be manually calibrated for the training sample through detection, so that the standardization of the face image is realized; the image graying refers to a process of converting a color image into a grayscale image, and the image graying is to convert an RGB image into a YUV format and then directly take a Y value, because Y represents luminance information of the image, that is, Gray is 0.299 × R +0.587 × G +0.114 × B; the human face image is standardized, also called geometric normalization, and the purpose is to transform the input human face image into a uniform structure. The method realizes the standardization of the face image by using a geometric correction method, so that the relative positions of key parts of the face in the image are the same, the resolution of the face image is unified into W multiplied by H, and the method takes W as 130 and H as 150; correcting the face contained in the original image to a uniform size, mainly realizing scale correction according to the position information of two eyes of the face, translating and rotating the image, assuming the left and right eye coordinates (LX, LY) and (RX, RY) of the face image, assuming the left and right eye coordinates of the standard face image as (LX, LY) and (RX, RY), respectively, and the width and height as W and H, respectively, the parameters of the standard face image of the invention are LX-30, LY-45, RX-100, RY-45, W-130, H-150, and then the face area coordinate transformation formula is as

wherein α ═ f · cos θ, β ═ f · sin θ, b₀＝f·(W-LX-LY)/2，And (x, y) is the original image coordinate, and (x ', y') is the transformed coordinate, so that a standard face image is obtained.

Step 20: scaling the image scale of the face image according to the original size, one fourth and one sixteenth original sizes to form a multi-resolution Gaussian pyramid, and respectively performing Gaussian filtering;

furthermore, the multi-resolution Gaussian pyramid is used for carrying out image scaling and Gaussian filtering on the standardized face image, and a pyramid hierarchical structure is formed from bottom to top according to the resolution.

Specifically, a pyramid hierarchy is formed on the standardized face image from bottom to top according to the original size, the quarter original size and the one sixteenth original size, image scale scaling is performed, gaussian filtering is performed on the three face images respectively, and the multi-resolution gaussian pyramid is formed. Generally, when an image is compressed, fine information of the image is easily captured; conversely, a coarse redundancy feature can be manifested. That is, towards the top of the pyramid, the fine features of the image are retained, whereas the coarse feature information is presented. Thus, a series of fine and coarse characterizations is much more powerful than a single image analysis. Therefore, the present invention introduces a gaussian pyramid model based on multi-resolution analysis.

First, the gaussian filter is used for image smoothing or sharpening, denoising and has a set of simpler and more standard subsequent processing procedures. The image features are extracted by adjusting the number of filters on different spatial scales, and the Gaussian multiresolution analysis is represented in a pyramid form in the invention. Let g₀(i, j) is the original image, g_k(i, j) (1. ltoreq. k. ltoreq.n) is a Gaussian filter applied to g_k-1(i, j) the resulting image obtained after (i, j). Then, given the sequence: g₀(i,j)→g₁(i,j)→…→g_n(i,j)，Wherein, W_kAnd H_kRespectively represent images g_k(i, j) width and height. By reducing the sampling density in this way, each image in the sequence (except the original image) is 1/2 less wide and high in density than the previous image. If the images are arranged from bottom to top, a pyramid structure with thin tips is formed. The value at each node in the pyramid represents the difference between the two co-gaussian distributions or correlation functions after convolution with the original image. Original image g₀(i, j) each pixel value representing the illumination of a respective image pointThe value is 0 to 255. The original image is located at the bottom or level 0 of the gaussian pyramid. Pyramid level 1 is image g₁(i, j) from the Gaussian filtered g₀(i, j) is obtained. Each value at level 1 is calculated by a 5 x 5 window weighted average at level 0. Similarly, each value in level 2 is also derived by applying the same weighting pattern on level 1. The 5 x 5 weighting mode is chosen by the present invention because it can provide sufficient filtering effect with low computational cost. Function f for layer-to-layer Gaussian filtering process_RIs represented by the formula g_k(i,j)＝f_R(g_k-1(i, j)), i.e., for any level k (0 < k < n) and image node i, j (0 ≦ i < W)_k,0≤j＜H_k) Comprises the following steps:wherein w (r, s) is δ (r) δ(s), andn represents the number of layers of the Gaussian pyramid, and W_kAnd H_kRespectively representing the width and height of the k-th layer image. The pixel density between image layers is reduced by half in width and height, respectively, in the present invention for a 1 and n 2, 3 or 4 multiresolution analysis processes.

Step 30: carrying out Gabor wavelet transformation on each face image in the resolution Gaussian pyramid according to different scales and directions to obtain a Gabor face image;

further, the method further comprises: the Gabor filter is generated by Gabor transformation, a multi-resolution image is decomposed by using Gabor filters with multiple scales and multiple directions, and each block of average gray value in the Gabor image is connected to form a feature vector, wherein the Gabor filter is defined by the following formula:where u and v are defined as the orientation and scale of the Gabor filter, respectively, and z is (x, y)^TAnd | · | | represents norm operation, σ is a metric gaussianWindow width to wavelength ratio.

Specifically, Gabor wavelet transformation is carried out on each face image in the resolution Gaussian pyramid according to different scales and directions to obtain a Gabor face image, the Gabor filter can well describe the ripple shape information of the face, and multi-resolution images are decomposed by using Gabor filters with p scales and q directions. In order to extract the features of the Gabor human face, the invention connects the average gray values of each block in the Gabor image to form a feature vector. The multi-scale, multi-directional Gabor filter is produced by a Gabor transform, i.e., a fourier transform with a gaussian window function. Since windowed fourier transform transduction exhibits local performance in the time-frequency domain, subtle facial features can be described by Gabor filters.

The Fourier transform of the absolute integrable function f (x) can be expressed as

Wherein R represents a real number domain. The essence of the fourier transform is to change the representation domain and the representation values of the function so that some phenomena and laws that cannot be observed in the time domain can be clearly demonstrated in the frequency domain. By fourier transforming the image, we can analyze the spectrum of the image and use it to process and analyze the image. Since non-stationary signals based on time domain are widely existed in nature and engineering fields, it is necessary to know the waveform variation of the corresponding frequency segment. The fourier transform is the integral of a function over the entire time domain, which implies a local variation process of the non-stationary signal. A non-stationary process may be treated as a set of temporally stationary processes. Thus, based on such analysis, the extraction of local processes can be achieved by multiplying f (x) by a moving window function g (x). The windowed Fourier transform can be expressed as

G(ω,a,b)＝∫_Rf(x)Ψ_a,b(x)dx，

Wherein,

ψ(x)＝g(x)e^-iω(ax+b).

if G (x) is a Gaussian function, G (ω, a, b) is called Gabor Transform (GT). Since the Gaussian function G (x) is an even function, G (ω, a, b) can be rewritten as

G(ω,a,b)＝∫_Rf(x)Ψ_a,b(x)dx＝f(x)*Ψ_a,b(x)，

Where a and b are scale and directional exposition, respectively, and denote convolution operations. In addition, in order to describe the moire shape information of the human face well, the multi-resolution image is decomposed using Gabor filters of p scales and q directions. Such a discrete Gabor filter is defined as follows:

where u and v are defined as the orientation and scale of the Gabor filter, respectively, and z is (x, y)^TWhere | l | · | represents a norm operation, σ is a measure of the ratio of the width of the gaussian window to the wavelength, and the wavelet vector is given by,wherein, where k is_maxRepresenting the maximum frequency value, and λ is the spatial factor between the small waves in the frequency domain.

Let I (x, y) be a multi-resolution image whose Gabor image is defined as follows:

G(x,y,u,v)＝I(x,y)*Ψ_u,v(z), where denotes a convolution operation. Gabor wavelets are obtained for p scales (v ∈ {0,1, …, p-1}) and q directions (u ∈ {0,1, …, q-1}), so that p × q Gabor images can be obtained from each multi-resolution image.

Step 40: carrying out local binary pattern conversion on the Gabor face image, and uniformly dividing the face image according to the horizontal and vertical directions to obtain a Gabor local binary pattern face image;

specifically, the Gabor face image is subjected to local binary pattern conversion, the face image is uniformly divided in the horizontal and vertical directions to obtain a Gabor local binary pattern face image, local binary pattern LGBP conversion is performed on the Gabor face image with different scales and different directions, and the face image is uniformly divided in the horizontal and vertical directions, for example, the face image is uniformly divided into 5 × 5 binary pattern face images.

Step 50: extracting histogram data of each region of the Gabor local binary pattern human face image, and obtaining multi-resolution local Gabor binary pattern MLGBP characteristic vectors according to the sequence from top to bottom and from left to right;

step 60: collecting sample data with male and female labels in the multiresolution local Gabor binary pattern MLGBP characteristics, and performing central connecting line dimensionality reduction on each divided region characteristic vector to obtain a multiresolution local Gabor binary pattern characteristic vector MLGBP-CCL based on the central connecting line CCL dimensionality reduction;

specifically, extracting histogram data of each divided region of each Gabor local binary pattern face image in each multi-resolution Gaussian pyramid layer, forming multi-resolution local Gabor binary pattern MLGBP feature vectors from top to bottom and from left to right, collecting multi-resolution local Gabor binary pattern MLGBP feature sample data with male and female labels, and performing center connecting line CCL reduction on each divided region feature vectorAnd dimension is carried out, so that the dimension reduction mapping direction of the central connecting line CCL and the multiresolution local Gabor binary pattern characteristic vector MLGBP-CCL based on the dimension reduction of the central connecting line CCL are obtained. Since the local binary pattern feature LBP has many advantages in describing image texture and constraining illumination transformation, we regard it as a method that can effectively describe the face texture feature. The method operates by comparing the center pixel in a 3 x 3 neighborhood with each of its surrounding pixels to obtain an 8-bit binary string, i.e., a decimal number between 0 and 255. Then, the histogram of the operated image is counted and used as a texture descriptor. To achieve rotation invariance, we use a uniform pattern that extends the original LBP operation. If an 8-bit binary string contains at most two bit transitions, i.e. from 0 to 1 or from 1 to 0, then this pattern is called the LBP uniform pattern, e.g. 00000000, 00011110 and 10000011 are uniform patterns, and to obtain a series of L capsule histograms, the LBP operation is denoted by the following notation, i.e. the LBP operation is represented by the following symbolMeaning that the LBP operation is performed in a neighborhood of P sample points with radius R, the superscript u2 indicates that uniform pattern is used, and non-uniform patterns are collectively labeled as a pattern. The present invention employs a uniform mode strategy, i.e., operates asTo quantify 256 of the LBP values into a histogram with 59 bins, including 58 uniform patterns and one non-uniform pattern, thus totaling 59 patterns, i.e., L-59.

The multi-resolution local Gabor binary pattern characteristic method is used for fusing a multi-resolution Gaussian pyramid, a multi-scale and multidirectional Gabor filter and a local binary pattern. It mainly includes the following several aspects: generating a series of facial images with different sizes by utilizing multi-resolution analysis with a Gaussian pyramid, wherein the facial images imply fine to coarse feature description; carrying out Gabor transformation on each multi-resolution analysis image by using a multi-scale and multi-direction Gabor filter; uniformly dividing each Gabor image into different rectangular areas which are not overlapped with each other; and performing local binary pattern operation on each Gabor image to obtain a series of local binary pattern face images, then performing histogram data extraction on each region of each local binary pattern face image, and arranging the histogram data into a feature vector from top to bottom and from left to right.

Suppose that a multiresolution algorithm is used to obtain n images I of different sizes₀,I₁,…,I_n-1In which I₀Is the original face image. Then, the k-th (0 < k < n) multi-resolution image is converted into p × q sub-Gabor pictures by using p-scale, q-direction Gabor filters. Finally, each Gabor image from the kth multi-resolution picture is divided into m_kRectangular areas not overlapping each otherAnd a histogram with L labels is obtained on each region using the LBP algorithm. Thus, the j (th) dimension in the v (th) scale and u (th) direction Gabor image from the k (th) sub multi-resolution image_kL number of label histograms of each region areWherein,here, f (x, y) denotes a Gabor image from a multi-resolution picture, i is 0,1, …, L-1, j_k＝0,1,…,m_k-1 and the function I { A } isFinally, the histograms of all these MLGBP images are concatenated to form the following feature vector:

wherein j is_k＝0,1,…,m_k-1. It is obvious thatIf m is ordered_kThe MLGBP feature vector may be written as V ═ H_0,0,0,0,…,H_k,u,v,j,…,H_{n-1,q-1,p-1,m-1}Where j is 0,1, …, m-1.

In order to suppress overfitting and improve the gender classification efficiency, the invention adopts a central connecting line algorithm to reduce the dimension of the MLGBP feature vector, and for the MLGBP feature vector, n is 3, q is 8, p is 5, m is 10 × 10, and L is 59, the total dimension of the MLGBP is 708,000, so that the high dimension is very unfavorable for the classifier. Therefore, the MLGBP feature vector needs to be subjected to dimensionality reduction. In order to consider both the spatial information and the class of the samples, the invention performs dimension reduction processing on each region separately. This method is specifically as follows: suppose in the training set, N are present respectively_mMale image sample and N_fIndividual female image samples. Setting a characteristic vector set omega ═ { V ═ V^(t)|t＝1,2,…,(N_m+N_f) Therein of

V^(t)Represents the t (1 ≦ t ≦ N) in the training set_m+N_f) A feature vector. Meanwhile, ω is defined in consideration of category information^(t)For the tag value of the t-th sample, order

Then assume that

Obviously, based on the class information, Γ can be adjusted_k,u,v,jFurther divided into two subsets:

wherein,obviously, setAndare respectively N_mAnd N_f. Thus aggregateAndor the center of these two categories can be calculated using the following equation:

according to the formula, a CCL (central connecting line) dimension reduction method is utilized to find out the corresponding mapping direction, and D is used_k,u,v,jI.e. the mapping direction may be located briefly on the center-connecting line CCL. Thus, the mapping direction of the MLGBP-CCL can be expressed as

Therefore, here, the set Γ_k,u,v,jEach L (L ═ 59) label histogram in (a) is obtainedCan be mapped to a value, i.e.Is represented as follows:as shown in fig. 2, the Gabor transform process is shown for an original image, and in summary, connecting all these mapped individual values forms the following feature vector:

thus, the total dimension of the MLGBP-CCL eigenvectors can be reduced to n × q × p × m.

Step 70: inputting characteristic data with male and female labels in the MLGBP-CCL into a gender classification model based on a support vector machine for training to obtain a gender classification model;

further, the method further comprises: the support vector machine is a machine learning method, and an optimal hyperplane is constructed in a sample input space or a feature space, so that the distance between the hyperplane and two types of sample sets is maximized.

Furthermore, the MLGBP-CCL eigenvector is obtained by reducing the dimension of the MLGBP eigenvector for the central connecting line CCL.

Specifically, feature data with male and female labels in the MLGBP-CCL are input into a gender classification model based on a support vector machine for training to obtain the gender classification model; the MLGBP-CCL eigenvector is obtained by reducing the dimension of the MLGBP eigenvector for the central connecting line CCL. The Support Vector Machine (SVM) is a machine learning method. The basic idea is to construct an optimal hyperplane in a sample input space or a feature space, so that the distance between the hyperplane and two types of sample sets is maximized, and the best generalization capability is obtained. Let the training sample set for a given problem be (x)₁,y₁)，(x₂,y₂)，...，(x_m,y_m) Wherein x is_i∈Rⁿ，y_iE { -1, +1}, i { -1, 2, …, m. It is assumed that the positive and negative samples of the training set can beDivided by a hyperplane, i.e. there is a hyperplane wx + b ═ 0, so that wx_iWhen + b is larger than 0, the output is positive class, otherwise, the output is negative class. For one problem, there may be many hyperplanes that satisfy the condition, but one is called the optimal hyperplane. The optimal hyperplane means that the distance between the point closest to the hyperplane and the hyperplane is maximized, p^*Is the optimal hyperplane. The support vector machine is used for searching the optimal hyperplane, and points which are very close to the optimal hyperplane are the support vector machine.

The distance between the support vector machine and the optimal hyperplane can be obtained as follows:the distance between the two samples in the training set isThus, the problem of constructing an optimal hyperplane is equivalent to that described in

y_i(w^Tx_i+ b) equal to or greater than 1, i equal to 1,2, …, m minimum constraint

This is a quadratic programming problem, whose lagrangian function is constructed:

wherein alpha is_iIs the lagrange multiplier. The optimal solution for quadratic programming is at the saddle point of this lagrangian function. At the saddle point, since the gradient of w and b is zero, it can be obtainedthe theorem of Kuhn-Tucker optimization theory shows that the optimal solution also satisfies the condition that alpha is_i[y_i(w^Tx_i+b)-1]0, obviously only those support directionsLagrange multiplier α of the samples of the measuring machine_iIs non-zero, i.e., only the support vector machine determines the optimal hyperplane, then w can be expressed as,thus, the problem of constructing a hyperplane further translates into a quadratic programming problem,

if a solution to the problem is found, the Lagrangian multiplier is α_o, ⁱThen we can get the following optimal hyperplane w_o，Using already calculated w_oWith a support vector machine x of positive samples, one can obtain,after the support vector machine obtains an optimal hyperplane through training, when a given sample x is predicted, only calculation is neededIt can be decided to which category x belongs.

For the linear inseparable case, a solution called soft-interval support vector machine is proposed, and a penalty factor C is introduced to control the degree of training sample error. The smaller C is, the larger the number of wrongly divided training samples is, but the shorter the training time is; and the larger C is, the closer the obtained hyperplane is to the linear support vector machine, namely, the less the number of the training samples which are wrongly divided is, but the longer the training time is. The most important characteristic of the support vector machine is that the dot product between training samples is K (x) by using a kernel function_i,x_j) Instead of, soThe decision function is obtained as a function of,kernel function K (x)_i,x_j) The selection of (A) can be selected according to actual conditions, and the most common kernel functions include three types: linear, polynomial, and gaussian kernel functions, which are used in the present invention.

In the present invention, for example, n is 3, q is 8, p is 5, m is 7 × 7, and L is 59. Next, gender classification example analysis is performed on 9 different-angle face images, the feature data used for behavior classification is an MLGBP feature vector subjected to CCL dimension reduction, the dimension of the MLGBP feature vector is 5880, and the following table 1 is an MLGBP-CCL feature vector sample.

TABLE 1

The MLGBP-CCL feature vector is obtained by subjecting MLGBP features to CCL, and then carrying out SVM training on a sample of the feature vector so as to obtain a gender classification model.

Step 80: and inputting the MLGBP-CCL characteristics into the gender classification model to obtain an analysis result.

Further, the inputting the MLGBP-CCL features into the gender classification model to obtain an analysis result further includes: detecting a face image from a video image; extracting multi-resolution local Gabor binary pattern face features MLGBP; obtaining the MLGBP-CCL characteristic vector by utilizing the human face characteristic MLGBP and the central connecting line CCL; and inputting the MLGBP-CCL characteristic vector into the gender classification model to obtain an analysis result.

Specifically, in real-time gender classification, firstly, a face image is detected from a video image, secondly, multi-resolution local Gabor binary pattern face features MLGBP are extracted, secondly, the dimension of the MLGBP-CCL features is reduced to form MLGBP-CCL feature vectors by using the dimension reduction mapping direction of the central connecting line CCL obtained in the step S60, and finally, the features are input into a gender classification model in the step S70 to obtain an analysis result. First, left and right eye coordinates (LX, LY) and (RX, RY) of a face image are artificially labeled, assuming that the left and right eye coordinates of a standard face image are (LX, LY) and (RX, RY), respectively, and the width and height are W and H, respectively, the standard face image parameters of the present invention are LX-30, LY-45, RX-100, RY-45, W-130, and H-150, and the face region coordinate transformation formula is

Then, performing multi-resolution Gaussian pyramid, Gabor transformation and LBP operation on the standard face image, then uniformly dividing the feature image according to m which is 7 multiplied by 7, and extracting a histogram of each region to form an MLGBP feature vector; collecting multi-resolution local Gabor binary pattern MLGBP feature sample data with male and female labels, and performing central connecting line CCL dimension reduction on each divided region feature vector to obtain a central connecting line CCL dimension reduction mapping direction D and a multi-resolution local Gabor binary pattern feature vector MLGBP-CCL based on the central connecting line CCL dimension reduction; secondly, inputting MLGBP-CCL characteristic data with male and female labels into a gender classification model based on a Support Vector Machine (SVM) for training so as to obtain a gender classification model; then, the MLGBP feature vector of the face image detected from the video is assumed to be expressed as

V_t＝{H_0,0,0,0,…,H_k,u,v,j,…,H_{n-1,q-1,p-1,m-1}And (c) the step of (c) in which,

and L59;

then, the vector H is mapped in the CCL dimension reduction mapping direction D_k,u,v,jIs mapped to a feature value such that the feature vector V_tIs reduced to nxqxpxm, where n is 3, q is 8, p is 5, and m is 7 × 7, thereby V_tDimension is 5880, i.e., MLGBP-CCL face feature vector.

Finally, the real-time face feature vector V is processed_tAnd inputting the data into the SVM gender classification model, and outputting 0 or 1, wherein 0 represents that the gender characteristic is male, and 1 represents that the gender characteristic is female. The embodiment of the invention solves the technical problem that the method for describing the gender characteristics in the prior art is difficult to ensure the robustness under the changing conditions of illumination, expression, posture and the like, achieves the technical effects of adapting to illumination change and describing better shape texture information to distinguish gender of men and women, is suitable for different application man-machine interaction, has outstanding progress, and simultaneously has the following advantages: the gender classification method based on the human face features can provide real-time people flow statistical information for users, commercial squares or shops; the intelligent advertising machine utilizes the gender classification method based on the human face characteristics to analyze the result and puts in advertising product information suitable for males or females; the robot can embody some personalized services aiming at the processing of males and females through the analysis result of the gender classification method based on the human face characteristics.

Example two

Based on the same inventive concept, as shown in fig. 3, an embodiment of the present application further provides a gender classification device based on human face features, the device includes:

the first preprocessing unit 11 is used for preprocessing the face image according to the eye positioning coordinates;

the first filtering unit 12 is configured to perform image scale scaling on the face image according to an original size, a quarter original size and a sixteenth original size to form a multi-resolution gaussian pyramid, and perform gaussian filtering respectively;

a first obtaining unit 13, where the first obtaining unit 13 is configured to perform Gabor wavelet transform on each face image in the resolution gaussian pyramid according to different scales and directions to obtain a Gabor face image;

a second obtaining unit 14, where the second obtaining unit 14 is configured to perform local binary pattern conversion on the Gabor face image, and uniformly divide the face image in the horizontal and vertical directions to obtain a Gabor local binary pattern face image;

a third obtaining unit 15, where the third obtaining unit 15 is configured to extract histogram data of each region of the Gabor local binary pattern human face image, and obtain a multiresolution local Gabor binary pattern MLGBP feature vector according to an order from top to bottom and from left to right;

a fourth obtaining unit 16, where the fourth obtaining unit 16 is configured to acquire sample data with male and female labels in the multi-resolution local Gabor binary pattern MLGBP feature, and perform central connecting line dimensionality reduction on each divided region feature vector to obtain a multi-resolution local Gabor binary pattern feature vector MLGBP-CCL based on the central connecting line CCL dimensionality reduction;

a fifth obtaining unit 17, where the fifth obtaining unit 17 is configured to input feature data with male and female labels in the MLGBP-CCL into a gender classification model based on a support vector machine for training, so as to obtain a gender classification model;

a sixth obtaining unit 18, where the sixth obtaining unit 18 is configured to input the MLGBP-CCL features into the gender classification model to obtain an analysis result.

Further, the apparatus further comprises: the Gabor filter is generated by Gabor transformation, a multi-resolution image is decomposed by using Gabor filters with multiple scales and multiple directions, and each block of average gray value in the Gabor image is connected to form a feature vector, wherein the Gabor filter is defined by the following formula:where u and v are defined as the orientation and scale of the Gabor filter, respectively, and z is (x, y)^TAnd | · | | represents norm operation, and σ is a measure of the ratio of the gaussian window width to the wavelength.

Further, the apparatus further comprises: the MLGBP-CCL eigenvector is obtained by reducing the dimension of the MLGBP eigenvector for the central connecting line CCL.

Further, the apparatus further comprises: the support vector machine is a machine learning method, and an optimal hyperplane is constructed in a sample input space or a feature space, so that the distance between the hyperplane and two types of sample sets is maximized.

Further, the apparatus further comprises: the face coordinate transformation formula is as follows:

various changes and specific examples of the gender classification method based on human face features in the first embodiment of fig. 1 and 2 are also applicable to the gender classification device based on human face features of the present embodiment, and those skilled in the art can clearly know the implementation method of the gender classification device based on human face features in the present embodiment through the foregoing detailed description of the gender classification method based on human face features, so for the brevity of the description, detailed descriptions are omitted here.

EXAMPLE III

Based on the same inventive concept as the gender identification and classification method based on human face features in the previous embodiment, the invention further provides a computer-readable storage medium, on which a computer program is stored, which when executed by a processor implements the steps of any one of the above-mentioned gender identification and classification methods based on human face features.

Where in fig. 4 a bus architecture (represented by bus 300), bus 300 may include any number of interconnected buses and bridges, bus 300 linking together various circuits including one or more processors, represented by processor 302, and memory, represented by memory 304. The bus 300 may also link together various other circuits such as peripherals, voltage regulators, power management circuits, and the like, which are well known in the art, and therefore, will not be described any further herein. A bus interface 306 provides an interface between the bus 300 and the receiver 301 and transmitter 303. The receiver 301 and the transmitter 303 may be the same element, i.e., a transceiver, providing a means for communicating with various other apparatus over a transmission medium.

The processor 302 is responsible for managing the bus 300 and general processing, and the memory 304 may be used for storing data used by the processor 302 in performing operations.

The technical scheme provided in the embodiment of the application at least has the following technical effects or advantages:

As will be appreciated by one skilled in the art, embodiments of the present invention may be provided as a method, system, or computer program product. Accordingly, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, and the like) having computer-usable program code embodied therein.

The present invention is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each flow and/or block of the flow diagrams and/or block diagrams, and combinations of flows and/or blocks in the flow diagrams and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable information processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable information processing apparatus, create means for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.

These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable information processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means which implement the function specified in the flowchart flow or flows and/or block diagram block or blocks.

These computer program instructions may also be loaded onto a computer or other programmable information processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.

It will be apparent to those skilled in the art that various changes and modifications may be made in the present invention without departing from the spirit and scope of the invention. Thus, if such modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include such modifications and variations.

Claims

1. A gender classification method based on human face features is characterized by comprising the following steps:

preprocessing the face image according to the eye positioning coordinates;

scaling the image scale of the face image according to the original size, one fourth and one sixteenth original sizes to form a multi-resolution Gaussian pyramid, and respectively performing Gaussian filtering;

carrying out Gabor wavelet transformation on each face image in the resolution Gaussian pyramid according to different scales and directions to obtain a Gabor face image;

carrying out local binary pattern conversion on the Gabor face image, and uniformly dividing the face image according to the horizontal and vertical directions to obtain a Gabor local binary pattern face image;

extracting histogram data of each region of the Gabor local binary pattern human face image, and obtaining multi-resolution local Gabor binary pattern MLGBP characteristic vectors according to the sequence from top to bottom and from left to right;

collecting sample data with male and female labels in the multiresolution local Gabor binary pattern MLGBP characteristics, and performing central connecting line dimensionality reduction on each divided region characteristic vector to obtain a multiresolution local Gabor binary pattern characteristic vector MLGBP-CCL based on the central connecting line CCL dimensionality reduction;

inputting characteristic data with male and female labels in the MLGBP-CCL into a gender classification model based on a support vector machine for training to obtain a gender classification model;

and inputting the MLGBP-CCL characteristics into the gender classification model to obtain an analysis result.

2. The method of claim 1, wherein the inputting the MLGBP-CCL features into the gender classification model to yield analysis results further comprises:

detecting a face image from a video image;

extracting multi-resolution local Gabor binary pattern face features MLGBP;

obtaining the MLGBP-CCL characteristic vector by utilizing the human face characteristic MLGBP and the central connecting line CCL;

and inputting the MLGBP-CCL characteristic vector into the gender classification model to obtain an analysis result.

3. The method of claim 1, wherein the pre-processing comprises: face detection, face image graying and face image standardization,

the face image graying is to convert an RGB image into a YUV format, then directly obtain a Y value, and the formula is Gray-Y-0.299 xR +0.587 xG +0.114 xB, wherein Y is image brightness information;

the face image is standardized by converting the input face image into a uniform structure and correcting the face into a uniform size.

4. The method of claim 3, wherein the multi-resolution Gaussian pyramid is a pyramid hierarchy formed from bottom to top in resolution for image scaling and Gaussian filtering of a normalized face image.

5. The method of claim 4, wherein the method further comprises:

the Gabor filter is generated by Gabor transformation, a multi-resolution image is decomposed by using Gabor filters with multiple scales and multiple directions, and each block of average gray value in the Gabor image is connected to form a feature vector, wherein the Gabor filter is defined by the following formula:

where u and v are defined as the orientation and scale of the Gabor filter, respectively, and z is (x, y)^TAnd | · | | represents norm operation, and σ is a measure of the ratio of the gaussian window width to the wavelength.

6. The method of claim 2, wherein the MLGBP-CCL feature vector is a feature vector obtained by reducing a dimension of the MLGBP feature vector for a center-connected line CCL.

7. The method of claim 1, wherein the method further comprises: the support vector machine is a machine learning method, and an optimal hyperplane is constructed in a sample input space or a feature space, so that the distance between the hyperplane and two types of sample sets is maximized.

8. The method of claim 1, wherein the method further comprises:

the face coordinate transformation formula is as follows:

<mrow> <mfenced open = "[" close = "]"> <mtable> <mtr> <mtd> <msup> <mi>x</mi> <mo>&prime;</mo> </msup> </mtd> </mtr> <mtr> <mtd> <msup> <mi>y</mi> <mo>&prime;</mo> </msup> </mtd> </mtr> <mtr> <mtd> <mn>1</mn> </mtd> </mtr> </mtable> </mfenced> <mo>=</mo> <mfenced open = "[" close = "]"> <mtable> <mtr> <mtd> <mi>&alpha;</mi> </mtd> <mtd> <mi>&beta;</mi> </mtd> <mtd> <mrow> <mo>(</mo> <mi>l</mi> <mi>x</mi> <mo>+</mo> <mi>r</mi> <mi>x</mi> <mo>)</mo> <mo>/</mo> <mn>2</mn> </mrow> </mtd> </mtr> <mtr> <mtd> <mrow> <mo>-</mo> <mi>&beta;</mi> </mrow> </mtd> <mtd> <mi>&alpha;</mi> </mtd> <mtd> <mrow> <mo>(</mo> <mi>l</mi> <mi>y</mi> <mo>+</mo> <mi>r</mi> <mi>y</mi> <mo>)</mo> <mo>/</mo> <mn>2</mn> <mo>+</mo> <msub> <mi>b</mi> <mn>0</mn> </msub> </mrow> </mtd> </mtr> <mtr> <mtd> <mn>0</mn> </mtd> <mtd> <mn>0</mn> </mtd> <mtd> <mn>1</mn> </mtd> </mtr> </mtable> </mfenced> <mo>&CenterDot;</mo> <mfenced open = "[" close = "]"> <mtable> <mtr> <mtd> <mi>x</mi> </mtd> </mtr> <mtr> <mtd> <mi>y</mi> </mtd> </mtr> <mtr> <mtd> <mn>1</mn> </mtd> </mtr> </mtable> </mfenced> <mo>,</mo> </mrow>

9. a gender classification device based on human face features, the device comprising:

the first preprocessing unit is used for preprocessing the face image according to the eye positioning coordinates;

the first filtering unit is used for carrying out image scale scaling on the face image according to the original size, the quarter original size and the sixteenth original size to form a multi-resolution Gaussian pyramid and respectively carrying out Gaussian filtering;

the first obtaining unit is used for carrying out Gabor wavelet transformation on each face image in the resolution Gaussian pyramid according to different scales and directions to obtain a Gabor face image;

the second obtaining unit is used for carrying out local binary pattern conversion on the Gabor face image, uniformly dividing the face image in the horizontal and vertical directions and obtaining a Gabor local binary pattern face image;

a third obtaining unit, configured to extract histogram data of each region of the Gabor local binary pattern face image, and obtain a multiresolution local Gabor binary pattern MLGBP feature vector according to a sequence from top to bottom and from left to right;

a fourth obtaining unit, configured to acquire sample data with male and female labels in the multiresolution local Gabor binary pattern MLGBP feature, and perform central connecting line dimensionality reduction on each divided region feature vector to obtain a multiresolution local Gabor binary pattern feature vector MLGBP-CCL based on the central connecting line CCL dimensionality reduction;

a fifth obtaining unit, configured to input feature data with male and female labels in the MLGBP-CCL into a gender classification model based on a support vector machine for training, so as to obtain a gender classification model;

a sixth obtaining unit, configured to input the MLGBP-CCL features into the gender classification model to obtain an analysis result.

10. A gender classification device based on human face features, the device comprising: memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor when executing the program implements the steps of:

preprocessing the face image according to the eye positioning coordinates;