CN106960176A - A kind of pedestrian's gender identification method based on transfinite learning machine and color characteristic fusion - Google Patents

Abstract

A kind of pedestrian's gender identification method based on transfinite learning machine and color characteristic fusion of the present invention, including：Extract the learning machine feature that transfinites of the training image of unmarked gender attribute；The input training image hsv color feature of unmarked gender attribute is extracted, the learning machine feature that will transfinite is combined with color characteristic, obtain fusion feature, pedestrian's gender sorter is trained using linear SVM SVM according to fusion feature and training image label；Utilize training gained model extraction test image feature, extract its hsv color feature simultaneously, then two category features are merged, obtains the fusion feature of test image, fusion feature is classified with linear SVM SVM pedestrian's gender sorter obtained by training process.The present invention is extracted to input picture to transfinite learning characteristic and color characteristic and carries out effective integration, realizes the complementation of two kinds of features, pedestrian's gender attribute is more effectively caught, so as to improve pedestrian's sex discrimination.

Description

A kind of pedestrian's gender identification method based on transfinite learning machine and color characteristic fusion

Technical field

It is more particularly to a kind of based on transfiniting learning machine and color characteristic melts the present invention relates to computer vision and pattern-recognition Pedestrian's gender identification method of conjunction.

Background technology

With actively promoting for " smart city ", thousands of video monitoring cameras gradually cover various public arenas, Basic guarantee facility is provided for municipal public safety management.For the consideration of many factors such as anti-terrorism, national public safety, depending on Frequency monitoring system is in the urgent need to the quick identity recognizing technology under a kind of remote, non-mated condition of target, so as to remote When can quickly confirm pedestrian's identity, realize intelligent early-warning.And it is used as the important supplementary means of remote identity recognizing technology, pedestrian Techniques of Gender Recognition is an indispensable part in intelligent video monitoring system.

Pedestrian's Techniques of Gender Recognition has extensive market application foreground, such as Intelligent Human-Machine Interface field, and pedestrian's sex is known Other quantity statistics system and automatic passenger number statistical system etc., its foundation that can be determined as business and other important decisions.

It is expert in people's gender identification method, relatively common method is based on traditional manual feature (such as HOG, SIFT Deng) extract, then carry out feature learning (as with SVM SVMs) realization identification.Either based on popular at present The method of deep learning carries out depth characteristic extraction, study and identification.But due to single traditional-handwork feature extracting method Accuracy of identification is not universal high, although and deep learning can improve sex accuracy of identification, but its precision is dependent on a large amount of instructions Practice data, adjust ginseng skill and high-capability computing device.

The content of the invention

It is an object of the invention to overcome the shortcomings of existing method, the present invention proposes a kind of based on the learning machine that transfinites Pedestrian's gender identification method of (Extreme Learning Machine) and color characteristic fusion, solves traditional-handwork extraction Characterization method accuracy of identification is not high and deep learning relies on big data, adjusts the problem of joining skill and high-performance calculation hardware.

The present invention includes training stage and test phase, in the training stage, to one point of input training pedestrian's picture training Layer transfinites learning machine, obtains model and extracts the feature of training image, while extract input picture hsv color feature, by this two Category feature carries out effective integration and obtains a kind of fusion feature with high distinction, then utilizes linear branch to gained fusion feature Hold vector machine training grader.In test phase, the study spy of the learning machine model extraction test image that transfinites trained is utilized Levy, and merged with the hsv color feature of test image, fusion feature is entered with training gained linear SVM SVM Row classification, exports recognition result.

A kind of pedestrian's gender identification method based on transfinite learning machine and color characteristic fusion includes training process and identification Process, it is specific as follows：

Training process S1：The training image of unmarked gender attribute is inputted, is then fed into the learning machine that transfinites of layering and instructs Practice, obtain model M (M is to solve the output weight beta transfinited in hierarchical network obtained by the object function of learning machine autocoder), And the feature a of training image is extracted with training gained model M；

Training process S2：The input training image hsv color feature b of unmarked gender attribute is extracted, by training process S1 Obtain learning machine feature a and the color characteristic b that transfinites be combined, obtain fusion feature V=[a, b], according to fusion feature V with (training and test image select gained to training image label from PETA pedestrian images storehouse, and the image library provides every image Pedestrian's sex label) utilize linear SVM SVM training pedestrian's gender sorter；

Identification process S3：Gained model M is trained to extract test image feature a' with training process S1, test image feature is carried Take process identical with training image characteristic extraction procedure.While extraction hsv color feature b ', then two category features are melted Close, obtain the fusion feature V' of test image, gained linear SVM SVM is trained to fusion feature V' with training process S2 Classified.

It is preferred that, training process S1 also includes, and carries out the super of pretreatment retraining layering to the training image of input first Learning machine is limited, and is transfinited learning machine feature a with model extraction obtained by training；Identification process S3 also includes, first the survey to input Attempt extract again test image feature a' as to be pre-processed.

It is preferred that, described pretreatment, including：

3.1) gray processing processing is carried out to input color pedestrian image, the coloured image of input is converted into gray level image, Only retain monochrome information；

3.2) specific quantity m sub-block { x is extracted to pedestrian's gray level image of acquisition⁽¹⁾,...,x^(m), sub-block x⁽ⁱ⁾Name For local receptor field (Local receptive field), m=8；

3.3) local contrast normalized is performed to each local receptor field respectively；

3.4) whitening processing is performed to each local receptor field respectively.

It is preferred that, it is described to each local receptor field execution local contrast normalized, it is specific as follows：

Wherein：For the pixel value of i-th of sub-block (local receptor field) (j, k) position,To return by local contrast The pixel value of i-th of sub-block (local receptor field) (j, k) position, k after one change₁For the length of local receptor field, k₂For local experiences Wild width, C is integer constant, and value is 10.

It is preferred that, it is described to each local receptor field execution whitening processing, it is specific as follows：

[D, U]=Eig (cov (Y)),

z⁽ⁱ⁾=U (D+diag (ε))^-1/2U^Ty⁽ⁱ⁾, i=1 ..., m

Wherein：z⁽ⁱ⁾For the sub-block (local receptor field) after whitening processing, size is k₁×k₂, Y is by m (m=8) height Block y⁽ⁱ⁾Constitute, cov () is covariance function, Eig () is characterized value analytic function, and D is characterized value, and U is characterized vector, diag () represents diagonal matrix, and ε is 0.1.

It is preferred that, the learning machine feature a that transfinites for extracting training image, including：

6.1) by k₁×k₂Local receptor field z⁽ⁱ⁾Column vector is converted to, dimension is k₁k₂, it is expressed as Z⁽ⁱ⁾；Then by m (m =8) individual local receptor field be combined into the form of matrix：

The transfinite object function of learning machine autocoder of first layer is：

Wherein, W is random orthogonal weight,It is L for size₁×L₁Unit matrix, b to standardize random column vector,

Dimension is L₁.I is the row vector that all elements are 1, and dimension is m (m=8), and σ () is sigmoid functions；α₁, α₂

To change the factor of influence that concealed nodes are distributed with output node；K is regularization factors.

Make H=σ (α₁(WZ₁+ bi)), then the output weight of first layer is：

6.2) the Feature Mapping figure of first layer is calculated according to 6.1) gained output weight betaIts In：I is input picture,Represent convolution operation；

6.3) in the layer network of learning machine second that transfinites, to the output characteristic mapping graph of first layerExtract specific quantity m (m =local receptor field 8), and local contrast normalized and whitening processing are performed to the local receptor field of acquisition；

6.4) to 6.3) obtaining local receptor field repeat step 6.1) and 6.2) operation, the learning machine that the transfinites acquisition second layer Feature Mapping figureWherein：L₂For the number of second layer output characteristic mapping graph；

6.5) L for obtaining first layer₁The L that individual Feature Mapping figure is obtained with the second layer₂Individual Feature Mapping figure is cascaded, Obtain L₁×(L₂+ 1) individual Feature Mapping figure；

6.6) in the learning machine third layer network that transfinites, to the L for learning machine acquisition of transfiniting₁×(L₂+ 1) individual Feature Mapping figure enters Row binary quantization processing so that Feature Mapping figure only has 0 and 1 two kind of value；

6.7) to every L₁Individual binary features mapping graph is quantified, will be per L₁Individual binary features mapping graph is compressed into one Individual Feature Mapping figure, obtains L₂Feature Mapping figure after+1 compressionFeature Mapping after compression The pixel value span of figure is 0 to 255；

6.8) block division is carried out to the characteristic pattern after compression, can obtain block is：

Wherein, B_iFor to characteristic pattern obtain after block division i-th piece, x × y represents the Feature Mapping figure after compression Size, w₁×w₂Represent to divide block size, s₁×s₂Represent step-length；

6.9) using histc to each block B_iBuild its histogram；

6.10) histogram of all pieces of cascade, obtains the feature of input pedestrian image：

Wherein：Wherein, N_B=(L₂+1)×[(x-w₁)/s₁+1]×[(y-w₂)/s₂+1]；hist(B_i) represent to block B_iStructure The histogram built；Represent that dimension is N in real number field_BVector；Represent the number of bits in histogram；X is compression The length of Feature Mapping figure afterwards, y is the width of Feature Mapping figure after compression, w₁For the length of divided block, w₂For the width of divided block, s₁For water Square to division step-length, s₂For the division step-length of vertical direction.

It is preferred that, the input training image hsv color feature b for extracting unmarked gender attribute, including：

7.1) input pedestrian image is converted into HSV color spaces by rgb color space；

7.2) series is quantified to different colours (H, S, V) path setting difference for inputting pedestrian image：

H_QL=8, S_QL=2, V_QL=2；

7.3) the respective max pixel value H in H in input pedestrian image, S, V passages is found out_max, S_max, V_max；

7.4) H, S, the quantized value of V passages of each pixel in input pedestrian image are calculated：

H_QV=H_QL×h(x,y)/H_max,

S_QV=S_QL×s(x,y)/S_max,

V_QV=V_QL×v(x,y)/V_max；

Wherein, h (x, y) represents the pixel value of input pedestrian image H passages (x, y) point, and s (x, y) represents input pedestrian's figure As the pixel value of channel S (x, y) point, v (x, y) represents the pixel value of input pedestrian image V passages (x, y) point；

7.5) the hsv color histogram matrix A that size is 8 × 2 × 2 is created, 0 is initialized with；According to step 7.4) The quantized value of three passages of each pixel obtained by calculating counts each matrix coordinate [H as matrix coordinate_QV,S_QV,V_QV] The number of times of appearance, is recorded as Q [H_QV,S_QV,V_QV]；

7.6) 8 × 2 × 2HSV color histogram matrix As are converted to 1 × 32 characteristic vector b.

It is preferred that, the obtaining step of the fusion feature V includes：

8.1) 2 kinds of different classes of features of n pedestrian images are obtained：a₁, b₁, a₂, b₂..., a_n, b_nCarry out Fusion Features, Obtain fusion feature V：

Wherein, often row represents the fusion feature of single sample to matrix V；

8.2) normalized matrix V* is obtained by row standardization to matrix V：

Wherein, v_minRepresent the minimum value of each row, v_maxThe maximum of each row is represented, v is the element of original matrix, and v* is Element obtained by after normalizing operation in matrix.

The present invention constructs a kind of pedestrian's sex identification model based on transfinite learning machine and color characteristic fusion, the model Transfinited learning machine using pedestrian image training layering, and optimize relevant parameter.Transfinited learning machine mould using the obtained layering of training Type extracts the feature of training image, then merges the hsv color feature of manual extraction, is instructed using Linear SVM SVMs Practice grader.In identification process, obtained layering is trained to transfinite the feature simultaneously combined test of learning machine model extraction test image The hsv color feature of image is then carried out with obtaining fusion feature with training gained linear SVM SVM to fusion feature Classification, exports recognition result.The present invention can be widely used in intelligent video monitoring system, Intelligent Human-Machine Interface field, OK People's sex is recognized among quantity statistics and automatic passenger number statistical system, is used as business and the foundation of other important decisions decision.

The present invention is described in further detail below in conjunction with drawings and Examples, but one kind of the present invention is based on transfinite Habit machine and pedestrian's gender identification method of color characteristic fusion are not limited to embodiment.

Brief description of the drawings

Fig. 1 is pedestrian sex recognizer flow chart of the present invention based on transfinite learning machine and color characteristic fusion；

Fig. 2 carries out block to the characteristic pattern after compression for the present invention and divides figure.

Embodiment

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

Pedestrian sex identification model of the present embodiment based on transfinite learning machine and color characteristic fusion, the model includes feature Extraction module and perceptron functional module, the embodiment specifically includes training process and identification process, shown in Figure 1, this hair A kind of bright pedestrian's sex in transfinite learning machine and color characteristic fusion knows method, and step is as follows：

1) training image of unmarked gender attribute is inputted, is then fed into the learning machine that transfinites of layering and trains, mould is obtained Type M (M for solve the output weight beta that transfinites obtained by the object function of learning machine autocoder in hierarchical network), and with training Gained model M extracts the feature a of training image；

2) the input training image hsv color feature b of unmarked gender attribute is extracted, by transfiniting that training process S1 is obtained Learning machine feature a is combined with color characteristic b, fusion feature V=[a, b] is obtained, according to fusion feature V and training image mark (training and test image select gained to label from PETA pedestrian images storehouse, and the image library provides pedestrian's sex of every image Label) pedestrian's gender sorter is trained (it should be noted that in machine learning field, supporting using linear SVM SVM Vector machine SVM, which is one, the learning model of supervision, is generally used for two classification problems, the present invention relates to support vector machines Place utilize prior art)；

3) use training process 1) training gained model M extract test image feature a', test image characteristic extraction procedure with Training image characteristic extraction procedure is identical.While extraction hsv color feature b ', then two category features are merged, surveyed The fusion feature V' of picture is attempted, with training process 2) train gained linear SVM SVM to classify fusion feature V'.

Further, step 1) and step 3) in, input training to the coloured image of input with after test pedestrian image, entering The processing of row gray processing, gray level image is converted into by the coloured image of input, only retains monochrome information.

Further, the learning machine feature a that transfinites of training image is extracted, process is as follows：

In the first layer of learning machine that transfinites, specific quantity m sub-block { x is extracted to pedestrian's gray level image of acquisition⁽¹⁾,..., x^(m), sub-block x⁽ⁱ⁾It is named as local receptor field (Local receptive field), m=8.

Then local contrast normalized is performed to each local receptor field of acquisition, it is specific as follows：

Wherein：For the pixel value of i-th of sub-block (local receptor field) (j, k) position,To pass through local contrast The pixel value of i-th of sub-block (local receptor field) (j, k) position, k after normalization₁For the length of local receptor field, k₂For local sense By wild width, C is integer constant, and value is 10.

Whitening processing is performed to each local receptor field of acquisition, it is specific as follows：

[D, U]=Eig (cov (Y)),

z⁽ⁱ⁾=U (D+diag (ε))^-1/2U^Ty⁽ⁱ⁾, i=1 ..., m

Wherein：z⁽ⁱ⁾For the sub-block (local receptor field) after whitening processing, size is k₁×k₂, Y is by m (m=8) height Block y⁽ⁱ⁾Constitute, cov () is covariance function, Eig () is characterized value analytic function, and D is characterized value, and U is characterized vector, diag () represents diagonal matrix, and ε is 0.1.

It is each after local contrast normalization and whitening processing followed by learning machine autocoder study of transfiniting Local receptor field, comprise the following steps that：

By k₁×k₂Local receptor field z⁽ⁱ⁾Column vector is converted to, dimension is k₁k₂, it is expressed as Z⁽ⁱ⁾.Then by m (m=8) Individual local receptor field is combined into the form of matrix：

The transfinite object function of learning machine autocoder of first layer is：

Wherein：W is random orthogonal weight.It is L for size₁×L₁Unit matrix.B ties up to standardize random column vector Spend for L₁.I is the row vector that all elements are 1, and dimension is m (m=8).σ () is sigmoid functions.α₁, α₂Hidden to change The factor of influence that node is distributed with output node.K is regularization factors.

Make H=σ (α₁(WZ₁+ bi)), then the output weight of first layer is：

The Feature Mapping figure of first layer is calculated according to the output weight beta of acquisitionWherein：I is Input picture,Represent convolution operation.

In the layer network of learning machine second that transfinites, to the output characteristic mapping graph of first layerExtract specific quantity m (m=8) Local receptor field, and local contrast normalized and whitening processing are performed to each local receptor field of acquisition.

Then each local receptor field obtained also with the learning machine autocoder (ELM-AE) that transfinites to the second layer Learnt, obtain the Feature Mapping figure of the second layerWherein：L₂For of second layer output characteristic mapping graph Number.Then L first layer obtained₁The L that individual Feature Mapping figure is obtained with the second layer₂Individual Feature Mapping figure is cascaded, and obtains L₁ ×(L₂+ 1) individual Feature Mapping figure.

In the learning machine third layer network that transfinites, to the L for learning machine acquisition of transfiniting₁×(L₂+ 1) individual Feature Mapping figure carries out two System quantification treatment so that Feature Mapping figure only has two kinds of values, 0 and 1.

Then to every L₁Individual binary features mapping graph is quantified, will be per L₁Individual binary features mapping graph is compressed into one Individual Feature Mapping figure, obtains L₂Feature Mapping figure after+1 compressionFeature Mapping figure after compression Pixel value span be 0 to 255.

Described in reference picture 2, block division is carried out to the characteristic pattern after compression, can obtain block is：

Wherein, B_iFor to characteristic pattern obtain after block division i-th piece, x × y represents the Feature Mapping figure after compression Size, w₁×w₂Represent to divide block size, s₁×s₂Step-length is represented, x is the length of Feature Mapping figure after compression, and y is feature after compression The width of mapping graph.w₁For the length of divided block, w₂For the width of divided block.s₁For the division step-length of horizontal direction, s₂For vertical direction Divide step-length.

Using histc to each block B_iBuild its histogram.

All pieces of histogram is finally cascaded, input picture feature is obtained：

Wherein, N_B=(L₂+1)×[(x-w₁)/s₁+1]×[(y-w₂)/s₂+1]；hist(B_i) represent to block B_iWhat is built is straight Fang Tu；Represent that dimension is N in real number field_BVector；Represent the number of bits in histogram.

It should be noted that test image feature a' extraction process is identical with training image characteristic extraction procedure, this reality Apply in example and be not repeated.

Further, step 2) and step 3) in, hsv color feature extraction is extracted, extraction process is following (not to be marked to extract The input training image hsv color feature b of note gender attribute is illustrated, the hsv color feature b of test image ' extraction Journey is the same, and the present embodiment is not repeated)：

Input pedestrian image is converted into HSV color spaces by rgb color space.

Path setting differences different to image quantify series：

H_QL=8, S_QL=2, V_QL=2.

Find out the respective max pixel value in H in input pedestrian image, S, V passages：H_max, S_max, V_max。

Calculate the H, S, the quantized value of V passages of each pixel in input pedestrian image：

H_QV=H_QL×h(x,y)/H_max,

S_QV=S_QL×s(x,y)/S_max,

V_QV=V_QL×v(x,y)/V_max。

Wherein h (x, y) represents the pixel value of input pedestrian image H passages (x, y) point, and s (x, y) represents input pedestrian image The pixel value of channel S (x, y) point, v (x, y) represents the pixel value of input pedestrian image V passages (x, y) point.

The hsv color histogram matrix A that size is 8 × 2 × 2 is created, 0 is initialized with.Obtained by calculating The quantized value of three passages of each pixel counts each matrix coordinate [H as matrix coordinate_QV,S_QV,V_QV] occur number of times, note Record as Q [H_QV,S_QV,V_QV].If for example, [H_QV,S_QV,V_QV]=[7,2,1] number of times that occurs is 3, then Q [H_QV,S_QV,V_QV]= 3。

8 × 2 × 2HSV color histogram matrix As are converted to 1 × 32 characteristic vector b.

Further, the learning machine feature a that will transfinite is merged with color characteristic b, and fusion steps are as follows：

Obtain 2 kinds of different classes of features of n pedestrian images：a₁, b₁, a₂, b₂..., a_n, b_nFusion Features are carried out, are obtained Fusion feature V：

Wherein：Often row represents the fusion feature of single sample to matrix V.

Normalized matrix V is obtained by row standardization to matrix V^*：

Wherein：v_minRepresent the minimum value of each row, v_maxThe maximum of each row is represented, v is the element of original matrix, v^*For Element obtained by after normalizing operation in matrix.

Above-described embodiment is intended merely to the explanation present invention, and is not used as limitation of the invention.As long as according to this hair Bright technical spirit, is changed, modification etc. will all fall in the range of the claim of the present invention to above-described embodiment.

Claims (8)

1. a kind of pedestrian's gender identification method based on transfinite learning machine and color characteristic fusion, it is characterised in that methods described Including training process and identification process, step is as follows：

Training process S1：The training image of unmarked gender attribute is inputted, being then fed into layering transfinites training in learning machine, obtains Model M, and with training gained model M extract training image the learning machine feature a that transfinites；

Training process S2：The input training image hsv color feature b of unmarked gender attribute is extracted, training process S1 is obtained The learning machine feature a and color characteristic b that transfinites be combined, obtain fusion feature V=[a, b], according to fusion feature V and training Image tag trains pedestrian's gender sorter using linear SVM SVM；

Identification process S3：Gained model M is trained to extract test image feature a' with training process S1, while extracting its hsv color Feature b ', then two category features are merged, the fusion feature V' of test image is obtained, with linear branch obtained by training process S2 Vector machine SVM pedestrian gender sorter is held to classify to fusion feature V'.

2. pedestrian's gender identification method according to claim 1 based on transfinite learning machine and color characteristic fusion, it is special Levy and be, training process S1 also includes, the training image of input is pre-processed transfinite again learning machine feature a first；Identification Process S3 also includes, and the test image of input is pre-processed extract again test image feature a' first.

3. pedestrian's gender identification method according to claim 2 based on transfinite learning machine and color characteristic fusion, it is special Levy and be, the pretreatment, including：

3.1) gray processing processing is carried out to the colorful pedestrian image of input, the coloured image of input is converted into gray level image, only Retain monochrome information；

3.2) specific quantity m sub-block { x is extracted to pedestrian's gray level image of acquisition⁽¹⁾,...,x^(m)}；Wherein, sub-block x⁽ⁱ⁾Name For local receptor field, m=8；

3.3) local contrast normalized is performed to each local receptor field respectively；

3.4) whitening processing is performed to each local receptor field respectively.

4. pedestrian's gender identification method according to claim 3 based on transfinite learning machine and color characteristic fusion, it is special Levy and be, it is described to each local receptor field execution local contrast normalized, it is specific as follows：

$y_{j,k}^{\left(i\right)}=\frac{x_{j,k}^{\left(i\right)}-\frac{1}{k_1{k}_2}{\mathrm{\Σ}}_{j=1}^{k_1}{\mathrm{\Σ}}_{k=1}^{k_2}{x}_{j,k}^{\left(i\right)}}{(\sqrt{\frac{1}{k_1{k}_2}{\mathrm{\Σ}}_{j=1}^{k_1}{\mathrm{\Σ}}_{k=1}^{k_2}{(x_{j,k}^{\left(i\right)}-\frac{1}{k_1{k}_2}{\mathrm{\Σ}}_{j=1}^{k_1}{\mathrm{\Σ}}_{k=1}^{k_2}{x}_{j,k}^{\left(i\right)})}^2}+C)},$

Wherein：For the pixel value of i-th of sub-block (j, k) position,For i-th of sub-block after local contrast is normalized The pixel value of (j, k) position, k₁For the length of local receptor field, k₂For the width of local receptor field, C is integer constant, and value is 10。

5. pedestrian's gender identification method according to claim 3 based on transfinite learning machine and color characteristic fusion, it is special Levy and be, it is described to each local receptor field execution whitening processing, it is specific as follows：

[D, U]=Eig (cov (Y)),

z⁽ⁱ⁾=U (D+diag (ε))^-1/2U^Ty⁽ⁱ⁾, i=1 ..., m

Wherein：z⁽ⁱ⁾For the sub-block after whitening processing, size is k₁×k₂, Y is by m sub-block y⁽ⁱ⁾Constitute, cov () is association side Difference function, Eig () is characterized value analytic function, and D is characterized value, and U is characterized vector, and diag () represents diagonal matrix, and ε is 0.1。

6. pedestrian's gender identification method according to claim 2 based on transfinite learning machine and color characteristic fusion, it is special Levy and be, the learning machine feature a that transfinites for extracting training image, including：

6.1) by k₁×k₂Local receptor field z⁽ⁱ⁾Column vector is converted to, dimension is k₁k₂, it is expressed as Z⁽ⁱ⁾；By m local experiences Open country is combined into the form of matrix：

$Z_1=\[Z^{\left(1\right)},Z^{\left(2\right)},...,Z^{\left(m\right)}\]\∈R^{k_1{k}_2\×m}$

The transfinite object function of learning machine autocoder of first layer is：

$\underset{W,b}{minK}\left|\right|Z_1-{\mathrm{\α}}_2\mathrm{\β}\mathrm{\σ}\left({\mathrm{\α}}_1\right({\mathrm{WZ}}_1+bi\left)\right)|{|}_{L2}^2+|\left|\mathrm{\β}\right|{|}_{L2}^2$

$\begin{array}{cc}s.t& {\mathrm{WW}}^T=I_{L_1},b^{T}b=1\end{array}$

Wherein：W is random orthogonal weight,It is L for size₁×L₁Unit matrix, b is standardizes random column vector, and dimension is L₁, i is the row vector that all elements are 1, and dimension is m, and σ () is sigmoid functions；α₁、α₂To change concealed nodes and output The factor of influence of Node distribution；K is regularization factors；

Make H=σ (α₁(WZ₁+ bi)), then the output weight of first layer is：

$\mathrm{\β}=({\mathrm{\β}}_1,...,{\mathrm{\β}}_{L_1})=\frac{1}{{\mathrm{\α}}_2}{Z}_1{H}^T{(\frac{I_{L_1}}{K}+{\mathrm{HH}}^T)}^{-1}$

6.2) the Feature Mapping figure of first layer is calculated according to output weight betaWherein, I schemes for input Picture,Represent convolution operation；

6.3) in the layer network of learning machine second that transfinites, to the output characteristic mapping graph of first layerExtract specific quantity m part Receptive field, and local contrast normalized and whitening processing are performed to the local receptor field of acquisition；

6.4) to 6.3) obtaining local receptor field repeat step 6.1) and 6.2) operation, the spy of the learning machine that the transfinites acquisition second layer Levy mapping graphWherein, L₂For the number of second layer output characteristic mapping graph；

6.5) L for obtaining first layer₁The L that individual Feature Mapping figure is obtained with the second layer₂Individual Feature Mapping figure is cascaded, and is obtained L₁×(L₂+ 1) individual Feature Mapping figure；

6.6) in the learning machine third layer network that transfinites, to the L for learning machine acquisition of transfiniting₁×(L₂+ 1) individual Feature Mapping figure carries out two System quantification treatment so that Feature Mapping figure only has 0 and 1 two kind of value；

6.7) to every L₁Individual binary features mapping graph is quantified, will be per L₁Individual binary features mapping graph is compressed into a spy Mapping graph is levied, L is obtained₂Feature Mapping figure after+1 compressionFeature Mapping figure after compression Pixel value span is 0 to 255；

6.8) block division is carried out to the characteristic pattern after compression, obtaining block is：

$\{B_1{B}_2,...,B_{(L_2+1)\×\[(x-w_1)/s_1+1\]\×\[(y-w_2)/s_2+1\]}\}$

Wherein, B_iFor to characteristic pattern obtain after block division i-th piece, x × y represents the Feature Mapping figure size after compression, w₁×w₂Represent to divide block size, s₁×s₂Represent step-length；

6.9) using histc to each block B_iBuild its histogram；

6.10) histogram of all pieces of cascade, obtains the learning machine feature that transfinites of input pedestrian image：

$a=f\left(\mathrm{Im}age\right)={\[hist{\left(B_1\right)}^{T}hist{\left(B_2\right)}^T,...,hist{\left(B_{N_B}\right)}^T\]}^T\∈R^{N_B\left(2^{L_1}\right)}$

Wherein, N_B=(L₂+1)×[(x-w₁)/s₁+1]×[(y-w₂)/s₂+1]；hist(B_i) represent to block B_iThe Nogata of structure Figure；Represent that dimension is N in real number field_BVector；Represent the number of bits in histogram；X reflects for feature after compression The length of figure is penetrated, y is the width of Feature Mapping figure after compression, w₁For the length of divided block, w₂For the width of divided block, s₁For horizontal direction Divide step-length, s₂For the division step-length of vertical direction.

7. pedestrian's gender identification method according to claim 1 based on transfinite learning machine and color characteristic fusion, it is special Levy and be, the input training image hsv color feature b for extracting unmarked gender attribute, including：

7.1) input pedestrian image is converted into HSV color spaces by rgb color space；

7.2) series is quantified to different colours (H, S, V) path setting difference for inputting pedestrian image：

H_QL=8, S_QL=2, V_QL=2；

7.3) the respective max pixel value H in H in input pedestrian image, S, V passages is found out_max, S_max, V_max；

7.4) H, S, the quantized value of V passages of each pixel in input pedestrian image are calculated：

H_QV=H_QL×h(x,y)/H_max,

S_QV=S_QL×s(x,y)/S_max,

V_QV=V_QL×v(x,y)/V_max；

Wherein, h (x, y) represents the pixel value of input pedestrian image H passages (x, y) point, and s (x, y) represents that input pedestrian image S leads to The pixel value of road (x, y) point, v (x, y) represents the pixel value of input pedestrian image V passages (x, y) point；

7.5) the hsv color histogram matrix A that size is 8 × 2 × 2 is created, 0 is initialized with；According to step 7.4) calculate The quantized value of resulting three passages of each pixel counts each matrix coordinate [H as matrix coordinate_QV,S_QV,V_QV] occur Number of times, be recorded as Q [H_QV,S_QV,V_QV]；

7.6) 8 × 2 × 2HSV color histogram matrix As are converted to 1 × 32 characteristic vector b.

8. pedestrian's gender identification method according to claim 1 based on transfinite learning machine and color characteristic fusion, it is special Levy and be, the obtaining step of the fusion feature V includes：

8.1) 2 kinds of different classes of features of n pedestrian images are obtained：a₁, b₁, a₂, b₂..., a_n, b_nFusion Features are carried out, are obtained Fusion feature V：

$V=\left[\begin{array}{cc}{a}_1& b_1\\ a_2& b_2\\ .& .\\ .& .\\ .& .\\ a_n& b_n\end{array}\right]$

Wherein, often row represents the fusion feature of single sample to matrix V；

8.2) normalized matrix V is obtained by row standardization to matrix V^*：

$v^{*}=\frac{v-v_{min}}{v_{max}-v_{min}},$

Wherein, v_minRepresent the minimum value of each row, v_maxThe maximum of each row is represented, v is the element of original matrix, v^*For standard Element obtained by changing after operation in matrix.