CN105930812A - Vehicle brand type identification method based on fusion feature sparse coding model - Google Patents

Abstract

The invention discloses a vehicle brand type identification method based on a fusion feature sparse coding model. The method comprises the following steps: 1) carrying out locating extraction on a front face region of a vehicle and pretreatment on a front face image of the vehicle; 2) extracting vehicle front face features and constructing a fusion feature; 3) constructing the sparse coding model based on the fusion feature; 4) constructing a non-negativity constraint sparse coding model; and 5) carrying out vehicle brand type identification through a reconstruction error minimum method. Classification of different vehicle brands is realized by effectively extracting the features of the front face of the vehicle; and the method is used for automatically extracting vehicle brand information in shot traffic gate videos and carrying out classification, thereby realizing intelligent management of gate video data.

Description

A kind of vehicle brand kind identification method based on fusion feature sparse coding model

Technical field

Patent of the present invention relates to intelligent transportation research field, the mainly research of vehicle brand sorting technique.

Background technology

Weigh sensor system application demand for vehicle brand is extensive, such as inspection of solving a case, the statistics of public security deparment and traffic police department Investigation, parking lot, the occasion such as vehicle management of residential quarters.Vehicle identification method based on computer vision is typical Pattern recognition is that equipment is easy and simple to handle in the applied research of the people-Che-road-environment of intelligent transportation field, advantage, failure rate is low, Can the use of round-the-clock all the period of time, the information in abundant digging vehicle image, real-time intelligent efficiency can greatly by Vehicle management personnel free from uninteresting complicated artificial cognition work, save great amount of cost and human and material resources. Shortcoming is to extract reliable feature description vehicle brand identifying accurately the most fast and effectively still to require study with classifying.

For research and development a new generation vehicle brand identification system, this system by the bayonet socket camera (video sensor) at crossing, front end, Video transmission system, vehicle brand information processing system form, and can monitor the traffic of traffic block port in real time, again may be used To realize identification and the investigation of vacation (overlapping) board car of vehicle brand.At present, the feature extraction that vehicle brand identification is used Method has Curvelet conversion, HOG feature, PHOG feature, Harr feature, EOH feature, Gabor wavelet etc.. But, above-mentioned recognition methods is all to extract single feature to carry out the identification of vehicle brand, so the technical program research Vehicle brand kind identification method based on fusion feature sparse coding model.

Summary of the invention

It is an object of the invention to solve existing recognition methods and extract the shortcoming that feature is single, it is provided that be a kind of based on fusion feature The vehicle brand kind identification method of sparse coding model.

The technical solution used in the present invention is: a kind of vehicle brand type identification side based on fusion feature sparse coding model Method, comprises the following steps:

1) before vehicle, the location in face region is extracted and the pretreatment of face image before vehicle；

2) extract vehicle front face feature and build fusion feature；

3) sparse coding model based on fusion feature is built；

4) nonnegativity restriction sparse coding model is built；

5) reconstructed error minimum method is used to carry out vehicle brand type identification.

As preferably, described step 1) in front face zone location be according to relative position between face with car plate before vehicle Relation, so needing to carry out the location to car plate before car face location, first finding according to template matching and having right angle characteristic Point, extracts the coordinate position of 4 angle points of car plate, to obtain the center-of-mass coordinate of car plate.Assume car plate in bayonet socket view data Width and height be respectively w and h pixel, the height of car face and width are respectively W and H pixel, by car plate 4 The coordinate position of individual angle point determines that the center-of-mass coordinate of car plate is for (x, y), then learns according to the statistics of great amount of images data, car The distance of the left margin of face, right margin and car plate barycenter is the twice of car plate width, i.e. 2w, the coboundary of car face with The distance of car plate barycenter is five times of car plate height, i.e. 5h, and the distance of the lower boundary of car face and car plate center of mass point is two Car plate height again, i.e. 2h.

The pretreatment work of image includes the histogram equalization of image, size normalization etc..The gatherer process of vehicle image In affected by illumination, shooting distance and focal length, image present different bright-dark degree, image contrast the highest and Car face position in entire image and size are uncertain, thus cause the car face size detected inconsistent.In order to reduce The impact of illumination, strengthens the contrast of image, preferably extracts the feature of vehicle brand, to image in this Image semantic classification Carrying out equalization processing, cardinal principle is to utilize certain Function Mapping to convert the pixel grey scale in original image so that The probability density of the gradation of image of equalization is uniformly distributed, and the dynamic range of images after equalization is increased, and contrast increases By force.The inconsistent coupling that can affect between car face data of car face size, normalization size will not be to feature extraction and classifying Calculating cause the biggest pressure, also can be effectively maintained the information of car face simultaneously, use bilinear interpolation in the method The process being normalized, normalization a size of 512 × 256, and for vehicle brand identification from the point of view of, mainly find car Diversity between type and vehicle, it should as far as possible reduce the interference information between classification, so herein by car car plate district on the face Territory uses Lycoperdon polymorphum Vitt RGB (128,128,128) to be filled with.

As preferably, described step 2) build fusion feature by selecting multiple features superposition extraction method, will two dimensional image Carry out converting the one-level characteristic vector extracted, construct rational dictionary as rarefaction representation data input carry out sample The sparse coding of notebook data extracts sparse coefficient as secondary characteristics.Based on different feature extraction principles, feature has folded Additivity.The local feature description that can higher level be extracted merges as the input of subordinate's feature extraction.Definition conversion Function: γ=f_α(β), f_α(β) represent β feature in the enterprising line translation of α feature space, then the conversion after a combination thereof is special Levying space is d=n dimension.

As preferably, step 3) in build fusion feature sparse coding model:

If original image is I, then (x, y) represents the gray-scale pixel values of piece image to I, and (x y) represents the space coordinates of pixel； Characteristic vector after one-level feature extraction is T, and intrinsic dimensionality is t.Definition transforming function transformation function is: T_t=Γ₁(I (x, y)), Wherein T_tRepresent the one-level characteristic vector after conversion.Then visually-perceptible system is by stimulating the receptive field of generation to external world Feature, is expressed as the active state of optic cell, and the model of this process information coding is described as formula (1), i.e.

$T_t=\underset{i}{\mathrm{\Σ}}{\mathrm{\α}}_i{t}_i+\mathrm{\ϵ}---\left(1\right)$

Wherein, t_iRepresent the feature bases of simulation primary visual system main visual cortex V1 district receptive field；α_iIt is the dilutest Sparse coefficient vector, represents the response to each basic function, the moving type of corresponding main visual cortex V1 district simple cell neuron State；ε usually assumes that as white Gaussian noise.For the T after conversion_tSignal, if test sample is that y, A are for believing after conversion Number composition training sample space, x is sparse vector.When sparse vector t_iL₀Time the most sparse, there is formula (2), i.e.

$\begin{array}{cccc}{\hat{x}}_0=\min imize& \left|\right|x|{|}_0& s.t.& y=Ax\end{array}---\left(2\right)$

Above formula equation and l₁The solution of minimization problem is with solving, i.e.

$\begin{array}{cccc}{\hat{x}}_1=\min imize& \left|\right|x|{|}_1& s.t.& y=Ax\end{array}---\left(3\right)$

If but when the linear restriction in above formula is false, following unconstrained optimization problem formula (4) can be translated into, i.e.

$\begin{array}{cc}C\left({\hat{x}}_p\right)=\min imize& \left|\right|y-Ax|{|}_p^p+\mathrm{\λ}|\left|x\right|{|}_1\end{array}---\left(4\right)$

By above statement it can be seen that first have to the training of feature after structure converts when the sparse vector of solving equation Sample space, according to after extracting and the characteristic vector that converts, uses K-singular value decomposition method to set up sample characteristics space, Assuming that the feature inputting training sample is Y, sample characteristics space is A, and the algorithm flow of foundation is as follows:

The first step: initialize the dictionary D of random distribution₀∈R^n×K, initially enter first step for each training sample profit Solving of sparse vector is carried out with orthogonal matching pursuit algorithm (Orthogonal Matching Pursuit, OMP) algorithm, Constantly iteration makes formulas.t.||x||₀≤T₀Set up, solve acquisition sparse vector x_i。

Second step: to dictionary D^(J-1)In every string k=1,2 ..., K carries out dictionary and updates by column: assume that sample group isCalculate E_kError matrix isThen E is made_kCan only be from correspondence Row in ω_kChoose and getDecompose with singular value decomposition method againWherein can make rowFor square The result of first row of battle array U is selected, and updates dictionary, with the first row of V Yu Δ (1,1) product be updated middle coefficient to Amount

3rd step: the new dictionary of formation after having updated by columnMaking Its Sparse Decomposition, ceaselessly iteration is until meeting end condition Till error minimum.

The dictionary D of final foundation inputs the feature samples space A of Sample Establishing according to being.Setting up the mistake of feature space Needing in journey constantly to utilize OMP Algorithm for Solving sparse vector, the solution procedure of OMP algorithm is as follows:

The first step: initial conditions was complete dictionary D=[d₁,d₂,...,d_L], primary signal y and degree of rarefication M, output It is to support indexed set Λ_m=Λ_m-1, sparse coefficient

Second step: first to redundancy r₀=y, supports indexed set Λ₀=Φ, primary iteration m=1, initialize；

3rd step: assume that signal supports in the m time iteration and integrate as Λ_m=Λ_m-1∪(λ_m), the most ceaselessly iterative computation Draw support indexUpdate residual errorUntil reaching iteration End condition m=M.

As preferably, described step 4) structure nonnegativity restriction sparse coding model:

From the angle of neuro physiology, more weak background is stimulated more sensitive by V1 district neuronal cell, and stings Swashing can not be negative value.Feature according to the abstract extraction of one-level is appreciated that these eigenvalues are all nonnegativities, therefore, is subject to The inspiration of neuro physiology, and the value combining each element of the characteristic vector after one-stage transfor-mation is non-negative, then can be right Characteristic signal carries out nonnegativity rarefaction representation.By Algorithms of Non-Negative Matrix Factorization (the Non-negative Matrix of Lee et al. Factorization, NMF) and the canonical algorithm that proposes of Olshausen et al. combine, then will be formed a kind of new Sparse coding algorithm, referred to as non-negative sparse coding algorithm, its object function is defined as:

$C\left(x\right)=\frac{1}{2}\left|\right|y-Ax|{|}^2+\mathrm{\λ}\underset{i,j}{\mathrm{\Σ}}f\left(x_{ij}\right)---\left(5\right)$

Wherein, constraints is λ ＞ 0；Y represents the column vector of test image shifting combination, and element therein is equal More than or equal to zero；A represents that basic function, element therein are all higher than or equal to zero；X represents openness coefficient, wherein Element x_ijIt is all higher than or equal to zero.The openness of sparse vector x is determined by the concrete form of penalty, is defined as f(·).Then taking its function expression is f (x)=| x |=x (x >=0), and therefore target function type (5) is of equal value under nonnegativity condition In formula (6), i.e.

$C\left(x\right)=\frac{1}{2}\left|\right|y-Ax|{|}^2+\mathrm{\λ}\underset{i,j}{\mathrm{\Σ}}{x}_{ij}---\left(6\right)$

So function f () is a strict increase function.As f (x)=x, | x₁| ＞ | x₂| ＞ ... ＞ | x_m| time, then have f(x₁) ＞ f (x₂) ＞ .. ＞ f (x_m), f (x) is absolute value strict increase function.So, as long as y-Ax value is constant, make When obtaining f (x) reduction, then object function is always reducing.

Main thought based on nonnegativity sparse coding is: orderτ ＞ 1, A=τ A.I.e. A is multiplied by one and puts Big coefficient, x is multiplied by a coefficient of reduction, | | y-Ax | |²Being worth constant, in x, all elements reduces, and f (x) reduces, then may be used To ensure that object function C (x) always successively decreases.So for given basic function, the optimization to sparse coefficient x can use formula (7) calculate, i.e.

xⁱ⁺¹=xⁱ.*{(A^Ty)./(A^TAy+λ)} (7)

Wherein, " .* " and " ./" represents the dot product of matrix and point removes respectively.Employing formula (7) more new regulation realizes the iteration of x Process, then the x after updating still meets nonnegativity, because its update method is by being multiplied by a non-negative factor (A^Ty)./(A^TAy+ λ) realize.As long as when the initial value of sparse vector x is set to positive number, then in the iteration of x During any permissible accuracy can be converged to global minimum.Given x is constant, it is considered to the optimization problem of A. The gradient descent algorithm of employing standard, the more new regulation obtaining A is:

$A=\mathrm{\β}\[A^t+\frac{\mathrm{\μ}}{\mathrm{\β}}(y-A^{t}x)x^T\]---\left(8\right)$

Wherein, μ is Learning Step, and β is learning rate.As long as step size mu is more than zero and sufficiently small, Projected Gradient Ensure that reduction target function value.

As preferably, described step 5) use reconstructed error minimum method to carry out vehicle brand type identification.

For given k class vehicle brand image, by all kinds of training samples are carried out dictionary learning, m can be obtained It is suitable for the basic function set of reconstruct original vehicle brand image.For arbitrary test sample, calculate it with every kind of basic function Set carries out the reconstructed error of rarefaction representation, corresponding to the generic that classification is this sample that reconstructed error is minimum.That is: If total k class, the column vector in each sample v under each class describesIf the i-th class comprises n_iIndividual sample This, then have

$A_i=[v_{{\mathrm{in}}_1},v_{{\mathrm{in}}_2},v_{{\mathrm{in}}_3},...,v_{{\mathrm{in}}_i}]\∈R^{m\×n_i}---\left(9\right)$

If y belongs to the i-th class, thenI.e. y can be carried out linearly by the sample of the i-th class Combination approaches.How to determine that vehicle brand based on rarefaction representation is classified, i.e. then can be weighed by the sparse solution tried to achieve Structure goes out all kinds of image, then by carrying out asking residual error with original test sample, classifying rules be reconstructed residual minimum be just That class.

Beneficial effect: before the present invention efficiently extracts vehicle, the feature of face realizes the classification to different vehicle brand, is used for Automatically extract in the traffic block port video photographed vehicle brand information and classify, it is achieved the intelligence to bayonet socket video data Energyization manages.

Detailed description of the invention

Below in conjunction with specific embodiments, the technical program is further illustrated:

A kind of vehicle brand kind identification method based on fusion feature sparse coding model, comprises the following steps:

The first step: use template matching method detection car plate position, extract the coordinate of car plate, according to the phase of car plate with car face To relation extract car face picture, the pretreatment work then the car face picture extracted being correlated with, equal including rectangular histogram The work such as weighing apparatusization, size normalization, normalize to 512 × 256 by car face image data；

Second step: by selecting multiple features superposition extraction method to build fusion feature, conversion will be carried out extract by two dimensional image One-level characteristic vector, construct rational dictionary and input as the data of rarefaction representation and carry out the sparse volume of sample data Code extracts sparse coefficient as secondary characteristics.Based on different feature extraction principles, feature has additivity.Can be by upper The local feature description that level is extracted merges as the input of subordinate's feature extraction.Definition transforming function transformation function: γ=f_α(β), f_α(β) represent β feature in the enterprising line translation of α feature space, then the transform characteristics space after a combination thereof is d=n dimension. 4th step: the foundation of fusion feature sparse coding model.If original image for I, then (x y) represents piece image to I Gray-scale pixel values, (x y) represents the space coordinates of pixel；Characteristic vector after one-level feature extraction is T, feature dimensions Number is t.

3rd step: set original image as I, then (x, y) represents the gray-scale pixel values of piece image to I, and (x y) represents the sky of pixel Between coordinate；Characteristic vector after one-level feature extraction is T, and intrinsic dimensionality is t.Definition transforming function transformation function is: T_t=Γ₁(I (x, y)), wherein T_tRepresent the one-level characteristic vector after conversion.Then visually-perceptible system is by stinging to external world Swashing the receptive field feature produced, be expressed as the active state of optic cell, the model of this process information coding is retouched State as formula (1), i.e.

$T_t=\underset{i}{\mathrm{\Σ}}{\mathrm{\α}}_i{t}_i+\mathrm{\ϵ}---\left(1\right)$

Wherein, t_iRepresent the feature bases of simulation primary visual system main visual cortex V1 district receptive field；α_iIt is the dilutest Sparse coefficient vector, represents the response to each basic function, the moving type of corresponding main visual cortex V1 district simple cell neuron State；ε usually assumes that as white Gaussian noise.For the T after conversion_tSignal, if test sample is that y, A are for believing after conversion Number composition training sample space, x is sparse vector.When sparse vector t_iL₀Time the most sparse, sparse vector and l₁ The solution of minimization problem, with solving, when difference solution, can be translated into following unconstrained optimization problem.By above State it can be seen that first have to the training sample space of feature after structure converts when the sparse vector of solving equation, according to After extraction and the characteristic vector that converts, K-singular value decomposition method is used to set up sample characteristics space, empty setting up feature Need during between constantly to utilize OMP Algorithm for Solving sparse vector.

4th step: the foundation of nonnegativity restriction sparse coding model.Object function is defined as:

$C\left(x\right)=\frac{1}{2}\left|\right|y-Ax|{|}^2+\mathrm{\λ}\underset{i,j}{\mathrm{\Σ}}f\left(x_{ij}\right)---\left(2\right)$

Wherein, constraints is λ ＞ 0；Y represents the column vector of test image shifting combination, and element therein is equal More than or equal to zero；A represents that basic function, element therein are all higher than or equal to zero；X represents openness coefficient, wherein Element x_ijIt is all higher than or equal to zero.The openness of sparse vector x is determined by the concrete form of penalty, is defined as f(·).Then taking its function expression is f (x)=| x |=x (x >=0), and therefore target function type (2) is of equal value under nonnegativity condition In formula (3), i.e.

$C\left(x\right)=\frac{1}{2}\left|\right|y-Ax|{|}^2+\mathrm{\λ}\underset{i,j}{\mathrm{\Σ}}{x}_{ij}---\left(3\right)$

So function f () is a strict increase function.As f (x)=x, | x₁| ＞ | x₂| ＞ ... ＞ | x_m| time, then have f(x₁) ＞ f (x₂) ＞ .. ＞ f (x_m), f (x) is absolute value strict increase function.So, as long as y-Ax value is constant, make When obtaining f (x) reduction, then object function is always reducing.

Main thought based on nonnegativity sparse coding is: orderτ ＞ 1, A=τ A.I.e. A is multiplied by one and puts Big coefficient, x is multiplied by a coefficient of reduction, | | y-Ax | |²Being worth constant, in x, all elements reduces, and f (x) reduces, then may be used To ensure that object function C (x) always successively decreases.So for given basic function, the optimization to sparse coefficient x can use formula (4) calculate, i.e.

xⁱ⁺¹=xⁱ.*{(A^Ty)./(A^TAy+λ)} (4)

Wherein, " .* " and " ./" represents the dot product of matrix and point removes respectively.Employing formula (4) more new regulation realizes the iteration of x Process, then the x after updating still meets nonnegativity, because its update method is by being multiplied by a non-negative factor (A^Ty)./(A^TAy+ λ) realize.As long as when the initial value of sparse vector x is set to positive number, then in the iteration of x During any permissible accuracy can be converged to global minimum.Given x is constant, it is considered to the optimization problem of A. The gradient descent algorithm of employing standard, the more new regulation obtaining A is:

$A=\mathrm{\β}\[A^t+\frac{\mathrm{\μ}}{\mathrm{\β}}(y-A^{t}x)x^T\]---\left(5\right)$

Wherein, μ is Learning Step, and β is learning rate.As long as step size mu is more than zero and sufficiently small, Projected Gradient Ensure that reduction target function value.

5th step: use reconstructed error minimum method to carry out vehicle brand type identification based on sparse coding model.For given K class vehicle brand image, by all kinds of training samples are carried out dictionary learning, m can be obtained and be best suitable for reconstructing original The basic function set of vehicle brand image.For arbitrary test sample, calculate it and carry out sparse with every kind of basic function set The reconstructed error represented, corresponding to the generic that classification is this sample that reconstructed error is minimum.That is: total k class is set, A column vector in each sample v under each class describesIf the i-th class comprises n_iIndividual sample, then have

$A_i=[v_{{\mathrm{in}}_1},v_{{\mathrm{in}}_2},v_{{\mathrm{in}}_3},...,v_{{\mathrm{in}}_i}]\∈R^{m\×n_i}---\left(6\right)$

If y belongs to the i-th class, thenI.e. y can be carried out linearly by the sample of the i-th class Combination approaches.How to determine that vehicle brand based on rarefaction representation is classified, i.e. then can be weighed by the sparse solution tried to achieve Structure goes out all kinds of image, and then by carrying out asking residual error with original test sample, classifying rules determines according to reconstructed error minimum Affiliated classification.

It should be pointed out that, for those skilled in the art, under the premise without departing from the principles of the invention, Can also make some improvements and modifications, these improvements and modifications also should be regarded as protection scope of the present invention.In the present embodiment The clearest and the most definite each ingredient all can use prior art to be realized.

Claims (6)

1. a vehicle brand kind identification method based on fusion feature sparse coding model, it is characterised in that: include with Lower step:

1) before vehicle, the location in face region is extracted and the pretreatment of face image before vehicle；

2) extract vehicle front face feature and build fusion feature；

3) sparse coding model based on fusion feature is built；

4) nonnegativity restriction sparse coding model is built；

5) reconstructed error minimum method is used to carry out vehicle brand type identification.

A kind of vehicle brand type identification side based on fusion feature sparse coding model the most according to claim 1 Method, it is characterised in that: described step 1) in vehicle before face zone location be according to phase between face with car plate before vehicle To position relationship, so needing to carry out the location to car plate before car face location, first finding according to template matching and having directly Angle characteristic point, extracts the coordinate position of 4 angle points of car plate, to obtain the center-of-mass coordinate of car plate；Assume bayonet socket view data The width of middle car plate and height are respectively w and h pixel, the width of car face and height and are respectively W and H pixel, by The coordinate position of 4 angle points of car plate determine the center-of-mass coordinate of car plate for (x, y), the left margin of car face, right margin and car plate The distance of barycenter is the twice of car plate width, i.e. 2w, and the coboundary of car face is car plate height with the distance of car plate barycenter Five times, i.e. 5h, and the car plate height that distance is twice of the lower boundary of car face and car plate center of mass point, i.e. 2h；

Before vehicle, the pretreatment work of face image includes histogram equalization and the size normalization of image；Described image straight Side's figure equalization processing is to utilize certain Function Mapping to convert the pixel grey scale in original image so that the figure of equalization As the probability density of gray scale is uniformly distributed, the dynamic range of images after equalization is increased, and contrast strengthens；Described figure The size normalization of picture uses the process that bilinear interpolation is normalized, normalization a size of 512 × 256, and car is on the face License plate area uses Lycoperdon polymorphum Vitt RGB (128,128,128) to be filled with.

A kind of vehicle brand type identification side based on fusion feature sparse coding model the most according to claim 1 Method, it is characterised in that: described step 2) in build fusion feature by select multiple features superposition extraction method, will two dimension Image carries out converting the one-level characteristic vector extracted, construct rational dictionary as rarefaction representation data input enter The sparse coding of row sample data extracts sparse coefficient as secondary characteristics；Based on different feature extraction principles, feature has There is additivity；The local feature description that can higher level be extracted merges as the input of subordinate's feature extraction；Definition Transforming function transformation function: γ=f_α(β), f_α(β) represent β feature in the enterprising line translation of α feature space, then the change after a combination thereof Changing feature space is d=n dimension.

A kind of vehicle brand type identification side based on fusion feature sparse coding model the most according to claim 1 Method, it is characterised in that: described step 3) the middle fusion feature sparse coding model that builds:

If original image is I, then (x, y) represents the gray-scale pixel values of piece image to I, and (x y) represents the space coordinates of pixel； Characteristic vector after one-level feature extraction is T, and intrinsic dimensionality is t；Definition transforming function transformation function is: T_t=Γ₁(I (x, y)), Wherein T_tRepresent the one-level characteristic vector after conversion；Then visually-perceptible system is by stimulating the receptive field of generation to external world Feature, is expressed as the active state of optic cell, and the model of this process information coding is described as formula (1), i.e.

$T_t=\underset{i}{\Σ}{\mathrm{\α}}_i{t}_i+\mathrm{\ϵ}---\left(1\right)$

Wherein, t_iRepresent the feature bases of simulation primary visual system main visual cortex V1 district receptive field；α_iIt is the dilutest Sparse coefficient vector, represents the response to each basic function, the moving type of corresponding main visual cortex V1 district simple cell neuron State；ε usually assumes that as white Gaussian noise；For the T after conversion_tSignal, if test sample is that y, A are for believing after conversion Number composition training sample space, x is sparse vector；When sparse vector t_iL₀Time the most sparse, there is formula (2), i.e.

$\begin{array}{cccc}{\hat{x}}_0=\min imize& \left|\right|x|{|}_0& s.t.& y=Ax\end{array}---\left(2\right)$

Above formula equation and l₁The solution of minimization problem is with solving, i.e.

$\begin{array}{cccc}{\hat{x}}_1=\min imize& \left|\right|x|{|}_1& s.t.& y=Ax\end{array}---\left(3\right)$

If but when the linear restriction in above formula is false, following unconstrained optimization problem formula (4) can be translated into, i.e.

$\begin{array}{cc}C\left({\hat{x}}_p\right)=\min imize& \left|\right|y-Ax|{|}_p^p+\mathrm{\λ}|\left|x\right|{|}_1\end{array}---\left(4\right)$

By above statement it can be seen that first have to the training of feature after structure converts when the sparse vector of solving equation Sample space, according to after extracting and the characteristic vector that converts, uses K-singular value decomposition method to set up sample characteristics space, Assuming that the feature inputting training sample is Y, sample characteristics space is A, and the algorithm flow of foundation is as follows:

The first step: initialize the dictionary D of random distribution₀∈R^n×K, initially enter first step for each training sample profit Carrying out solving of sparse vector with OMP algorithm, continuous iteration makes formulaSet up, Solve acquisition sparse vector x_i；

Second step: to dictionary D^(J-1)In every string k=1,2 ..., K carries out dictionary and updates by column: assume that sample group isCalculate E_kError matrix isThen E is made_kCan only be from correspondence Row in ω_kChoose and getDecompose with singular value decomposition method againWherein can make rowFor square The result of first row of battle array U is selected, and updates dictionary, with the first row of V Yu Δ (1,1) product be updated middle coefficient to Amount

3rd step: the new dictionary of formation after having updated by columnMaking Its Sparse Decomposition, ceaselessly iteration is until meeting end condition Till error minimum；

The dictionary D of final foundation inputs the feature samples space A of Sample Establishing according to being；Setting up the mistake of feature space Needing in journey constantly to utilize OMP Algorithm for Solving sparse vector, the solution procedure of OMP algorithm is as follows:

The first step: initial conditions was complete dictionary D=[d₁,d₂,...,d_L], primary signal y and degree of rarefication M, output It is to support indexed set Λ_m=Λ_m-1, sparse coefficient

Second step: first to redundancy r₀=y, supports indexed set Λ₀=Φ, primary iteration m=1, initialize；

3rd step: assume that signal supports in the m time iteration and integrate as Λ_m=Λ_m-1∪(λ_m), the most ceaselessly iterative computation Draw support indexUpdate residual errorUntil reaching iteration End condition m=M.

A kind of vehicle brand type identification side based on fusion feature sparse coding model the most according to claim 1 Method, it is characterised in that: described step 4) in structure nonnegativity restriction sparse coding model:

The object function of non-negative sparse coding algorithm is defined as:

$C\left(x\right)=\frac{1}{2}\left|\right|y-Ax|{|}^2+\mathrm{\λ}\underset{i,j}{\Σ}f\left(x_{ij}\right)---\left(5\right)$

Wherein, constraints is λ ＞ 0；Y represents the column vector of test image shifting combination, and element therein is equal More than or equal to zero；A represents that basic function, element therein are all higher than or equal to zero；X represents openness coefficient, wherein Element x_ijIt is all higher than or equal to zero；The openness of sparse vector x is determined by the concrete form of penalty, is defined as f(·)；Then taking its function expression is f (x)=| x |=x (x >=0), and therefore target function type (5) is of equal value under nonnegativity condition In formula (6), i.e.

$C\left(x\right)=\frac{1}{2}\left|\right|y-Ax|{|}^2+\mathrm{\λ}\underset{i,j}{\Σ}{x}_{ij}---\left(6\right)$

So function f () is a strict increase function；As f (x)=x, | x₁| ＞ | x₂| ＞ ... ＞ | x_m| time, then have f(x₁) ＞ f (x₂) ＞ ... ＞ f (x_m), f (x) is absolute value strict increase function；So, as long as y-Ax value is constant, make When obtaining f (x) reduction, then object function is always reducing；

Main thought based on nonnegativity sparse coding is: orderτ ＞ 1, A=τ A；I.e. A is multiplied by one and puts Big coefficient, x is multiplied by a coefficient of reduction, | | y-Ax | |²Being worth constant, in x, all elements reduces, and f (x) reduces, then may be used To ensure that object function C (x) always successively decreases；So for given basic function, the optimization to sparse coefficient x can use formula (7) calculate, i.e.

xⁱ⁺¹=xⁱ.*{(A^Ty)./(A^TAy+λ)} (7) Wherein, " .* " and " ./" represents the dot product of matrix and point removes respectively；Employing formula (7) more new regulation realizes the iteration of x Process, then the x after updating still meets nonnegativity, because its update method is by being multiplied by a non-negative factor (A^Ty)./(A^TAy+ λ) realize；As long as when the initial value of sparse vector x is set to positive number, then in the iteration of x During any permissible accuracy can be converged to global minimum；Given x is constant, it is considered to the optimization problem of A； The gradient descent algorithm of employing standard, the more new regulation obtaining A is:

$A=\mathrm{\β}\[A^t+\frac{\mathrm{\μ}}{\mathrm{\β}}(y-A^{t}x)x^T\]---\left(8\right)$

Wherein, μ is Learning Step, and β is learning rate；As long as step size mu is more than zero and sufficiently small, Projected Gradient Ensure that reduction target function value.

A kind of vehicle brand type identification side based on fusion feature sparse coding model the most according to claim 1 Method, it is characterised in that: described step 5) in use reconstructed error minimum method carry out vehicle brand type identification；

For given k class vehicle brand image, by all kinds of training samples are carried out dictionary learning, m can be obtained It is suitable for the basic function set of reconstruct original vehicle brand image；For arbitrary test sample, calculate it with every kind of basic function Set carries out the reconstructed error of rarefaction representation, corresponding to the generic that classification is this sample that reconstructed error is minimum；That is: If total k class, the column vector in each sample v under each class describesIf the i-th class comprises n_iIndividual sample This, then have

If y belongs to the i-th class, thenI.e. y can be carried out linearly by the sample of the i-th class Combination approaches；How to determine that vehicle brand based on rarefaction representation is classified, i.e. then can be weighed by the sparse solution tried to achieve Structure goes out all kinds of image, then by carrying out asking residual error with original test sample, classifying rules be reconstructed residual minimum be just That class.