CN105930812A

Movatterモバイル変換

Info

Publication number: CN105930812A
Application number: CN201610268208.0A
Authority: CN
Inventors: 赵池航; 陈爱伟; 张小琴
Original assignee: Southeast University
Current assignee: Southeast University
Priority date: 2016-04-27
Filing date: 2016-04-27
Publication date: 2016-09-07

Abstract

Translated fromChinese

本发明公开了一种基于融合特征稀疏编码模型的车辆品牌类型识别方法，包括以下步骤：1)车辆前脸区域的定位提取及车辆前脸图像的预处理；2)提取车辆前脸特征并构建融合特征；3)构建基于融合特征的稀疏编码模型；4)构建非负性约束稀疏编码模型；5)采用重构误差最小法进行车辆品牌类型识别。本发明有效地提取车辆前脸的特征来实现对不同车辆品牌的分类，用于自动提取拍摄到的交通卡口视频中车辆品牌信息并进行分类，实现对卡口视频数据的智能化管理。The invention discloses a vehicle brand type recognition method based on a fusion feature sparse coding model, comprising the following steps: 1) positioning and extraction of the vehicle front face area and preprocessing of the vehicle front face image; 2) extracting vehicle front face features and constructing Fusion features; 3) Construct a sparse coding model based on fusion features; 4) Construct a non-negativity constrained sparse coding model; 5) Use the method of minimum reconstruction error to identify vehicle brand types. The invention effectively extracts the features of the front face of the vehicle to realize the classification of different vehicle brands, is used to automatically extract and classify the vehicle brand information in the captured traffic checkpoint video, and realizes the intelligent management of the checkpoint video data.

Description

Translated fromChinese

一种基于融合特征稀疏编码模型的车辆品牌类型识别方法A Vehicle Brand Type Recognition Method Based on Fusion Feature Sparse Coding Model

技术领域technical field

本发明专利涉及智能交通研究领域，主要是车辆品牌分类方法的研究。The invention patent relates to the field of intelligent transportation research, mainly the research on vehicle brand classification methods.

背景技术Background technique

对于车辆品牌的智能化识别系统应用需求广泛，如公安交警部门的破案稽查、统计调研、停车场、住宅小区的车辆管理等场合。基于计算机视觉的车辆识别方法是典型的模式识别在智能交通领域的人-车-路-环境的应用研究，优点是设备操作简便，故障率低，可以全天候全时段的使用，充分挖掘车辆图像中的信息，实时智能化效率可以极大的将车辆管理人员从枯燥繁复的人工判别工作中解放出来，节约大量成本和人力物力资源。缺点是如何快速有效的提取可靠的特征描述车辆品牌并准确的识别与分类仍有待研究。There is a wide range of application requirements for the intelligent identification system of vehicle brands, such as investigations by public security and traffic police departments, statistical investigations, parking lots, and vehicle management in residential quarters. The vehicle recognition method based on computer vision is a typical application research of pattern recognition in the field of intelligent transportation of people-vehicle-road-environment. Real-time intelligent efficiency can greatly liberate vehicle management personnel from the boring and complicated manual identification work, saving a lot of cost and human and material resources. The disadvantage is that how to quickly and effectively extract reliable features to describe vehicle brands and accurately identify and classify them remains to be studied.

为研发新一代车辆品牌识别系统，该系统由前端路口的卡口相机(视频传感器)、视频传输系统、车辆品牌信息处理系统组成，既能实时监控交通卡口的交通状况，又可以实现车辆品牌的识别以及假(套)牌车的查处。目前，车辆品牌识别运用的特征提取方法有Curvelet变换、HOG特征、PHOG特征、Harr特征、EOH特征、Gabor小波等。但是，上述的识别方法都是提取单一的特征进行车辆品牌的识别，所以本技术方案研究基于融合特征稀疏编码模型的车辆品牌类型识别方法。In order to develop a new generation of vehicle brand recognition system, the system is composed of a checkpoint camera (video sensor) at the front intersection, a video transmission system, and a vehicle brand information processing system. identification and investigation of fake (set) license plates. At present, the feature extraction methods used in vehicle brand recognition include Curvelet transform, HOG feature, PHOG feature, Harr feature, EOH feature, Gabor wavelet, etc. However, the above identification methods all extract a single feature to identify the vehicle brand, so this technical solution studies the vehicle brand type identification method based on the fusion feature sparse coding model.

发明内容Contents of the invention

本发明的目的在于解决现有识别方法提取特征单一的缺点，提供一种基于融合特征稀疏编码模型的车辆品牌类型识别方法。The purpose of the present invention is to solve the shortcoming of single feature extraction in existing recognition methods, and provide a vehicle brand type recognition method based on a fusion feature sparse coding model.

本发明采用的技术方案为：一种基于融合特征稀疏编码模型的车辆品牌类型识别方法，包括以下步骤：The technical solution adopted in the present invention is: a vehicle brand type recognition method based on fusion feature sparse coding model, comprising the following steps:

1)车辆前脸区域的定位提取及车辆前脸图像的预处理；1) Positioning and extraction of the front face area of the vehicle and preprocessing of the front face image of the vehicle;

2)提取车辆前脸特征并构建融合特征；2) Extract vehicle front face features and construct fusion features;

3)构建基于融合特征的稀疏编码模型；3) Build a sparse coding model based on fusion features;

4)构建非负性约束稀疏编码模型；4) Construct a non-negativity constrained sparse coding model;

5)采用重构误差最小法进行车辆品牌类型识别。5) Using the method of minimum reconstruction error to identify the vehicle brand type.

作为优选，所述步骤1)中的前脸区域定位是根据车辆前脸与车牌之间的相对位置关系，所以车脸定位之前需要进行对车牌的定位，首先根据模板匹配寻找具有直角特性点，提取车牌4个角点的坐标位置，以得到车牌的质心坐标。假设卡口图像数据中车牌的宽度和高度分别为w和h像素，车脸的高度和宽度分别为W和H个像素点，由车牌4个角点的坐标位置确定车牌的质心坐标为(x,y)，则根据大量图像数据的统计得知，车脸的左边界、右边界与车牌质心的距离均为车牌宽度的两倍，即2w，车脸的上边界与车牌质心的距离为车牌高度的五倍，即5h，而车脸的下边界与车牌质心点的距离为两倍的车牌高度，即2h。As preferably, the location of the front face area in the step 1) is based on the relative positional relationship between the front face of the vehicle and the license plate, so before the location of the car face, the location of the license plate needs to be carried out. At first, according to the template matching, a right-angle characteristic point is searched for, Extract the coordinates of the four corners of the license plate to obtain the coordinates of the center of mass of the license plate. Assuming that the width and height of the license plate in the bayonet image data are w and h pixels respectively, and the height and width of the car face are respectively W and H pixels, the coordinates of the center of mass of the license plate are determined by the coordinates of the four corner points of the license plate as (x ,y), according to the statistics of a large amount of image data, the distance between the left boundary and the right boundary of the car face and the center of mass of the license plate is twice the width of the license plate, that is, 2w, and the distance between the upper boundary of the car face and the center of mass of the license plate is Five times the height, that is, 5h, and the distance between the lower boundary of the car face and the center of mass of the license plate is twice the height of the license plate, that is, 2h.

图像的预处理工作包括图像的直方图均衡化、尺寸归一化等。车辆图像的采集过程中受光照、拍摄距离和焦距的影响，图像呈现出不同的明暗程度、图像的对比度不高且车脸在整幅图像中的位置和大小不确定，从而导致检测出的车脸尺寸不一致。为了减少光照的影响，增强图像的对比度，更好地提取车辆品牌的特征，本图像预处理中对图像进行均衡化处理，主要原理是利用某种函数映射对原始图像中的像素灰度做变换，使得均衡化的图像灰度的概率密度均匀分布，均衡化后的图像动态范围得到增大，对比度增强。车脸尺寸的不一致会影响车脸数据之间的匹配，归一化尺寸不会对特征提取与分类的计算造成很大的压力，同时也会很好的保留车脸的信息，在本方法中采用双线性插值进行归一化的处理，归一化尺寸为512×256，而对于车辆品牌识别来讲，主要是寻找车型与车型之间的差异性，应该尽量减少类别间的干扰信息，所以本文将车脸上的车牌区域使用灰色RGB(128,128,128)进行填充。Image preprocessing includes image histogram equalization, size normalization, etc. In the process of vehicle image collection, affected by the illumination, shooting distance and focal length, the image presents different levels of light and shade, the contrast of the image is not high, and the position and size of the vehicle face in the entire image are uncertain, resulting in the detected vehicle The face size is inconsistent. In order to reduce the influence of light, enhance the contrast of the image, and better extract the characteristics of the vehicle brand, the image is equalized in this image preprocessing. The main principle is to use a certain function mapping to transform the grayscale of the pixel in the original image. , so that the probability density of the equalized image gray level is uniformly distributed, the dynamic range of the equalized image is increased, and the contrast is enhanced. The inconsistency of the size of the car face will affect the matching between the car face data. The normalized size will not cause great pressure on the calculation of feature extraction and classification, and it will also preserve the information of the car face well. In this method Bilinear interpolation is used for normalization processing, and the normalized size is 512×256. For vehicle brand recognition, it is mainly to find the difference between models and models, and the interference information between categories should be minimized. Therefore, in this paper, the license plate area on the car face is filled with gray RGB (128, 128, 128).

作为优选，所述步骤2)构建融合特征通过选择多特征叠加提取法，即将二维图像进行变换提取出的一级特征向量，构建出合理的字典作为稀疏表示的数据输入来进行样本数据的稀疏编码提取稀疏系数作为二级特征。基于不同的特征提取原理，特征具有叠加性。可以将上级提取到的局部特征描述作为下级特征提取的输入进行融合。定义变换函数：γ＝f_α(β)，f_α(β)表示将β特征在α特征空间上进行变换，则其组合后的变换特征空间为d＝n维。As a preference, said step 2) constructing the fusion feature by selecting the multi-feature superposition extraction method, that is, transforming the two-dimensional image to extract the first-level feature vector, constructing a reasonable dictionary as the data input of the sparse representation to perform the sparseness of the sample data Encoding extracts sparse coefficients as secondary features. Based on different feature extraction principles, features are superimposed. The local feature description extracted by the upper level can be used as the input of the lower level feature extraction for fusion. Define the transformation function: γ=f_α (β), f_α (β) means to transform the β feature on the α feature space, then the transformed feature space after its combination is d=n dimension.

作为优选，步骤3)中构建融合特征稀疏编码模型：As preferably, in step 3), construct fusion feature sparse coding model:

设原始图像为I，则I(x,y)表示一幅图像的灰度像素值，(x,y)表示像素的空间坐标；经过一级特征提取后的特征向量为T，特征维数为t。定义变换函数为：T_t＝Γ₁(I(x,y))，其中T_t代表经过变换后的一级特征向量。则视觉感知系统通过对外界刺激产生的感受野特征，将其表达为视觉细胞的活动状态，这一过程用信息编码的模型描述为式(1)，即Suppose the original image is I, then I(x, y) represents the gray pixel value of an image, and (x, y) represents the spatial coordinates of the pixel; the feature vector after the first-level feature extraction is T, and the feature dimension is t. The transformation function is defined as: T_t =Γ₁ (I(x,y)), where T_t represents the transformed first-level feature vector. Then the visual perception system expresses the receptive field characteristics of external stimuli as the activity state of visual cells. This process is described by the information coding model as formula (1), namely

${T T}_{t t} = = \underset{i i}{Σ Σ} {α α}_{i i} {t t}_{i i} + + ϵ ϵ - - - - - - ((11))$

其中，t_i表示模拟初级视觉系统主视皮层V1区感受野的特征基向量；α_i是随机稀疏系数矢量，表示对各个基函数的响应，对应主视皮层V1区简单细胞神经元的活动状态；ε通常假设为高斯白噪声。对于变换后的T_t信号，设测试样本为y，A为变换后信号组成的训练样本空间，x为稀疏向量。当稀疏向量t_i的l₀足够稀疏时，有式(2)，即Among them, t_i represents the characteristic basis vector simulating the receptive field of the V1 area of the primary visual system; α_i is a random sparse coefficient vector, which represents the response to each basis function, corresponding to the activity state of simple cell neurons in the V1 area of the main visual cortex ; ε is usually assumed to be Gaussian white noise. For the transformed T_t signal, let the test sample be y, A be the training sample space composed of the transformed signal, and x be the sparse vector. When the l₀ of the sparse vector t_i is sufficiently sparse, there is formula (2), namely

$\begin{matrix} {\overset{^^}{x x}}_{00} = = min min i i m m i i z z e e & | | | | x x | | {| |}_{00} & s the s . . t t . . & y the y = = A A x x \end{matrix} - - - - - - ((22))$

上式方程与l₁最小化问题的解同解，即The above equation has the same solution as the solution of the l₁ minimization problem, namely

$\begin{matrix} {\overset{^^}{x x}}_{11} = = min min i i m m i i z z e e & | | | | x x | | {| |}_{11} & s the s . . t t . . & y the y = = A A x x \end{matrix} - - - - - - ((33))$

但如果上式中的线性约束不成立时，可以将其转化为下列无约束优化问题式(4)，即However, if the linear constraint in the above formula does not hold, it can be transformed into the following unconstrained optimization problem formula (4), namely

$\begin{matrix} C C (({\overset{^^}{x x}}_{p p})) = = min min i i m m i i z z e e & | | | | y the y - - A A x x | | {| |}_{p p}^{p p} + + λ λ | | | | x x | | {| |}_{11} \end{matrix} - - - - - - ((44))$

通过以上的表述可以看出，在求解方程的稀疏向量时首先要构建变换后特征的训练样本空间，根据提取后并变换的特征向量，运用K-奇异值分解方法建立样本特征空间，假设输入训练样本的特征为Y，样本特征空间为A，建立的算法流程如下所示：From the above statement, it can be seen that when solving the sparse vector of the equation, the training sample space of the transformed features must first be constructed. According to the extracted and transformed feature vectors, the K-singular value decomposition method is used to establish the sample feature space. Assuming that the input training The feature of the sample is Y, and the feature space of the sample is A. The established algorithm flow is as follows:

第一步：初始化随机分布的字典D₀∈R^n×K，首先进入第一步骤对于每个训练样本利用正交匹配追踪算法(Orthogonal Matching Pursuit，OMP)算法进行稀疏向量的求解，不断迭代使得式s.t.||x||₀≤T₀成立，求解获得稀疏向量x_i。Step 1: Initialize the randomly distributed dictionary D₀ ∈ R^n×K , first enter the first step and use the Orthogonal Matching Pursuit (OMP) algorithm to solve the sparse vector for each training sample, and continuously iterate so that Mode st||x||₀ ≤ T₀ is established, and the sparse vector x_i is obtained by solving.

第二步：对字典D^(J-1)中的每一列k＝1,2,...,K进行字典逐列更新：假设样本组为计算E_k误差矩阵即为然后令E_k只能从对应的列中ω_k选取获取到再用奇异值分解法来分解其中可以令列为矩阵U的第一列的结果来选则更新字典，用V与Δ(1,1)乘积的第一列进行更新中间系数向量The second step: update each column k=1,2,...,K in the dictionary D^(J-1) column by column: suppose the sample group is The calculation of the E_k error matrix is Then let E_k can only be obtained by selecting ω_k from the corresponding column Then use the singular value decomposition method to decompose which can be ordered Select the update dictionary for the result of the first column of the matrix U, and use the first column of the product of V and Δ(1,1) to update the intermediate coefficient vector

第三步：逐列更新完后形成的新字典作稀疏分解，不停的迭代直到满足终止条件误差最小为止。Step 3: A new dictionary formed after updating column by column Do sparse decomposition and iterate continuously until the termination condition is satisfied and the error is minimized.

最终建立的字典D即为根据输入样本建立的特征样本空间A。在建立特征空间的过程中需要不断的利用OMP算法求解稀疏向量，OMP算法的求解过程如下所示：The finally established dictionary D is the feature sample space A established based on the input samples. In the process of establishing the feature space, it is necessary to continuously use the OMP algorithm to solve the sparse vector. The solution process of the OMP algorithm is as follows:

第一步：输入条件为过完备字典D＝[d₁,d₂,...,d_L]，原始信号y和稀疏度M,输出的是支撑索引集Λ_m＝Λ_m-1，稀疏系数Step 1: The input condition is an overcomplete dictionary D=[d₁ ,d₂ ,...,d_L ], the original signal y and the degree of sparsity M, and the output is the support index set Λ_m = Λ_m-1 , sparse coefficient

第二步：首先对冗余r₀＝y，支撑索引集Λ₀＝Φ，初始迭代m＝1，进行初始化；Step 2: Initialize redundancy r₀ =y, support index set Λ₀ =Φ, initial iteration m=1;

第三步：假设在第m次迭代中信号支撑集为Λ_m＝Λ_m-1∪(λ_m)，然后不停的迭代计算得出支撑索引更新残差直到达到迭代终止条件m＝M。Step 3: Assume that the signal support set in the mth iteration is Λ_m = Λ_m-1 ∪(λ_m ), and then iteratively calculate the support index update residuals Until the iteration termination condition m=M is reached.

作为优选，所述步骤4)的构建非负性约束稀疏编码模型：As preferably, the construction non-negativity constraint sparse coding model of described step 4):

从神经生理学的角度出发，V1区神经元细胞对较弱的背景刺激较为敏感，而且刺激不能为负值。根据一级抽象提取的特征可以得知这些特征值都是非负性的，因此，受神经生理学的启发，并结合一级变换后的特征向量的每个元素的值均为非负，则可以对特征信号进行非负性稀疏表示。将Lee等人的非负矩阵分解算法(Non-negative MatrixFactorization，NMF)和Olshausen等人提出的标准算法相结合，那么就会形成一种新的稀疏编码算法，称为非负稀疏编码算法，其目标函数定义为：From the perspective of neurophysiology, neurons in the V1 area are more sensitive to weaker background stimuli, and the stimuli cannot be negative. According to the features extracted by the first-level abstraction, it can be known that these eigenvalues are non-negative. Therefore, inspired by neurophysiology, combined with the value of each element of the eigenvector after the first-level transformation is non-negative, then it can be used for The feature signal is non-negative sparse representation. Combining the non-negative matrix factorization algorithm (Non-negative Matrix Factorization, NMF) of Lee et al. and the standard algorithm proposed by Olshausen et al., a new sparse coding algorithm will be formed, called non-negative sparse coding algorithm, whose The objective function is defined as:

$C C ((x x)) = = \frac{11}{22} | | | | y the y - - A A x x | | {| |}^{22} + + λ λ \underset{i i,, j j}{Σ Σ} f f (({x x}_{i i j j})) - - - - - - ((55))$

其中，约束条件为λ＞0；y表示测试图像转换组合而成的列向量，其中的元素均大于或等于零；A表示基函数，其中的元素均大于或等于零；x表示稀疏性系数，其中的元素x_ij均大于或等于零。稀疏向量x的稀疏性由惩罚函数的具体形式所决定，定义为f(·)。则取其函数表达式为f(x)＝|x|＝x(x≥0)，因此目标函数式(5)在非负性条件下等价于式(6)，即Among them, the constraint condition is λ>0; y represents the column vector formed by the test image conversion and combination, and the elements in it are all greater than or equal to zero; A represents the basis function, and the elements in it are all greater than or equal to zero; x represents the sparsity coefficient, where The elements x_ij are all greater than or equal to zero. The sparsity of the sparse vector x is determined by the specific form of the penalty function, which is defined as f(·). Then take its function expression as f(x)=|x|=x(x≥0), so the objective function formula (5) is equivalent to formula (6) under the condition of non-negativity, that is

$C C ((x x)) = = \frac{11}{22} | | | | y the y - - A A x x | | {| |}^{22} + + λ λ \underset{i i,, j j}{Σ Σ} {x x}_{i i j j} - - - - - - ((66))$

这样函数f(·)是一个严格增函数。当f(x)＝x，|x₁|＞|x₂|＞...＞|x_m|时，则有f(x₁)＞f(x₂)＞..＞f(x_m)，f(x)为绝对值严格增函数。所以，只要y-Ax值不变，使得f(x)减小时，那么目标函数总是在减小。Thus the function f(·) is a strictly increasing function. When f(x)=x, |x₁ |＞|x₂ |＞...＞|x_m |, then f(x₁ )＞f(x₂ )＞..＞f(x_m ) , f(x) is a strictly increasing function of absolute value. Therefore, as long as the value of y-Ax remains unchanged, so that f(x) decreases, the objective function is always decreasing.

基于非负性稀疏编码的主要思想是：令τ＞1，A＝τA。即A乘上一个放大系数，x乘上一个缩小系数，||y-Ax||²值不变，x中所有元素缩小，f(x)减小，则可以保证目标函数C(x)总是递减。那么对于给定的基函数，对稀疏系数x的优化可以用式(7)进行计算，即The main idea based on non-negative sparse coding is: Let τ>1, A=τA. That is, A is multiplied by an enlargement factor, x is multiplied by a reduction coefficient, the value of ||y-Ax||² remains unchanged, all elements in x are reduced, f(x) is reduced, and the total value of the objective function C(x) can be guaranteed is decreasing. Then for a given basis function, the optimization of the sparse coefficient x can be calculated by formula (7), namely

xⁱ⁺¹＝xⁱ.*{(A^Ty)./(A^TAy+λ)} (7)xⁱ⁺¹ = xⁱ .*{(A^T y)./(A^T Ay+λ)} (7)

其中，“.*”和“./”分别代表矩阵的点乘和点除。采用式(7)更新规则来实现x的迭代过程，则更新后的x仍然满足非负性，因为它的更新方法是通过乘以一个非负因子(A^Ty)./(A^TAy+λ)来实现的。只要当稀疏向量x的初始值设置为正数，那么在x的迭代过程中对任意要求的精度都能够收敛到全局最小值。给定x不变，考虑A的优化问题。采用标准的梯度下降算法，得到A的更新规则为：Among them, ".*" and "./" represent the dot multiplication and division of the matrix respectively. Using the update rule of formula (7) to realize the iterative process of x, the updated x still satisfies the non-negativity, because its update method is by multiplying a non-negative factor (^AT y)./(^AT Ay+ λ) to achieve. As long as the initial value of the sparse vector x is set to a positive number, then any required accuracy can converge to the global minimum during the iterative process of x. Given that x is constant, consider the optimization problem of A. Using the standard gradient descent algorithm, the update rule of A is obtained as:

$A A = = β β [[{A A}^{t t} + + \frac{μ μ}{β β} ((y the y - - {A A}^{t t} x x)) {x x}^{T T}]] - - - - - - ((88))$

其中，μ是学习步长，β为学习速率。只要步长μ大于零且足够小，投影梯度算法就可以保证减小目标函数值。Among them, μ is the learning step size and β is the learning rate. As long as the step size μ is greater than zero and small enough, the projected gradient algorithm is guaranteed to reduce the objective function value.

作为优选，所述步骤5)采用重构误差最小法进行车辆品牌类型识别。Preferably, said step 5) adopts the method of minimum reconstruction error to identify the vehicle brand type.

对于给定的k类车辆品牌图像，通过对各类训练样本进行字典学习，可获得m个最适合重构原始车辆品牌图像的基函数集合。对于任意的测试样本，计算其用每种基函数集合进行稀疏表示的重构误差，对应于重构误差最小的类别为该样本的所属类别。即：设共有k类，每个类下的每个样本用v中的一个列向量来描述，若第i类包含n_i个样本，则有For a given k types of vehicle brand images, by performing dictionary learning on various training samples, m basis function sets that are most suitable for reconstructing the original vehicle brand images can be obtained. For any test sample, calculate the reconstruction error of its sparse representation with each basis function set, and the category corresponding to the smallest reconstruction error is the category of the sample. That is: suppose there are k classes in total, and each sample under each class is described by a column vector in v , if the i-th class contains n_i samples, then

${A A}_{i i} = = [[{v v}_{{in in}_{11}},, {v v}_{{in in}_{22}},, {v v}_{{in in}_{33}},, . . . . . .,, {v v}_{{in in}_{i i}}]] &Element; &Element; {R R}^{m m \times \times {n no}_{i i}} - - - - - - ((99))$

若y属于第i类，则即y可以由第i类的样本进行线性组合来逼近。如何确定基于稀疏表示的车辆品牌分类，即可以通过求得的稀疏解然后重构出各类图像，然后通过与原始测试样本进行求残差，分类规则为重构残差最小的就为那一类。If y belongs to the i-th class, then That is, y can be approximated by a linear combination of samples of the i-th class. How to determine the vehicle brand classification based on sparse representation, that is, we can reconstruct various images through the obtained sparse solution, and then calculate the residual with the original test sample. The classification rule is that the reconstruction residual is the smallest. kind.

有益效果：本发明有效地提取车辆前脸的特征来实现对不同车辆品牌的分类，用于自动提取拍摄到的交通卡口视频中车辆品牌信息并进行分类，实现对卡口视频数据的智能化管理。Beneficial effects: the present invention effectively extracts the features of the front face of the vehicle to realize the classification of different vehicle brands, and is used to automatically extract and classify the vehicle brand information in the captured traffic checkpoint video, and realize the intelligence of the checkpoint video data manage.

具体实施方式detailed description

下面结合具体实施方案对本技术方案进一步说明：Below in conjunction with specific implementation scheme, this technical scheme is further described:

一种基于融合特征稀疏编码模型的车辆品牌类型识别方法，包括以下步骤：A vehicle brand type recognition method based on a fusion feature sparse coding model, comprising the following steps:

第一步：使用模板匹配方法检测车牌位置，提取车牌的坐标，根据车牌与车脸的相对关系提取车脸图片，接着对提取到的车脸图片进行相关的预处理工作，包括直方图均衡化、尺寸归一化等工作，将车脸图片数据归一化到512×256；Step 1: Use the template matching method to detect the position of the license plate, extract the coordinates of the license plate, extract the car face picture according to the relative relationship between the license plate and the car face, and then perform related preprocessing work on the extracted car face picture, including histogram equalization , size normalization, etc., normalize the car face image data to 512×256;

第二步：通过选择多特征叠加提取法构建融合特征，即将二维图像进行变换提取出的一级特征向量，构建出合理的字典作为稀疏表示的数据输入来进行样本数据的稀疏编码提取稀疏系数作为二级特征。基于不同的特征提取原理，特征具有叠加性。可以将上级提取到的局部特征描述作为下级特征提取的输入进行融合。定义变换函数：γ＝f_α(β)，f_α(β)表示将β特征在α特征空间上进行变换，则其组合后的变换特征空间为d＝n维。第四步：融合特征稀疏编码模型的建立。设原始图像的为I，则I(x,y)表示一幅图像的灰度像素值，(x,y)表示像素的空间坐标；经过一级特征提取后的特征向量为T，特征维数为t。Step 2: Construct the fusion feature by selecting the multi-feature superposition extraction method, that is, convert the first-level feature vector extracted from the two-dimensional image, and construct a reasonable dictionary as the data input of the sparse representation to perform sparse coding of the sample data and extract the sparse coefficient as a secondary feature. Based on different feature extraction principles, features are superimposed. The local feature description extracted by the upper level can be used as the input of the lower level feature extraction for fusion. Define the transformation function: γ=f_α (β), f_α (β) means to transform the β feature on the α feature space, then the transformed feature space after its combination is d=n dimension. The fourth step: the establishment of fusion feature sparse coding model. Assuming that the original image is I, then I(x, y) represents the grayscale pixel value of an image, and (x, y) represents the spatial coordinates of the pixel; the feature vector after the first-level feature extraction is T, and the feature dimension for t.

第三步：设原始图像为I，则I(x,y)表示一幅图像的灰度像素值，(x,y)表示像素的空间坐标；经过一级特征提取后的特征向量为T，特征维数为t。定义变换函数为：T_t＝Γ₁(I(x,y))，其中T_t代表经过变换后的一级特征向量。则视觉感知系统通过对外界刺激产生的感受野特征，将其表达为视觉细胞的活动状态，这一过程用信息编码的模型描述为式(1)，即Step 3: Let the original image be I, then I(x, y) represents the grayscale pixel value of an image, and (x, y) represents the spatial coordinates of the pixel; the feature vector after the first-level feature extraction is T, The feature dimension is t. The transformation function is defined as: T_t =Γ₁ (I(x,y)), where T_t represents the transformed first-level feature vector. Then the visual perception system expresses the receptive field characteristics of external stimuli as the activity state of visual cells. This process is described by the information coding model as formula (1), namely

其中，t_i表示模拟初级视觉系统主视皮层V1区感受野的特征基向量；α_i是随机稀疏系数矢量，表示对各个基函数的响应，对应主视皮层V1区简单细胞神经元的活动状态；ε通常假设为高斯白噪声。对于变换后的T_t信号，设测试样本为y，A为变换后信号组成的训练样本空间，x为稀疏向量。当稀疏向量t_i的l₀足够稀疏时，稀疏向量的与l₁最小化问题的解同解，当不同解时，可以将其转化为下列无约束优化问题。通过以上的表述可以看出，在求解方程的稀疏向量时首先要构建变换后特征的训练样本空间，根据提取后并变换的特征向量，运用K-奇异值分解方法建立样本特征空间，在建立特征空间的过程中需要不断的利用OMP算法求解稀疏向量。Among them, t_i represents the characteristic basis vector simulating the receptive field of the V1 area of the primary visual system; α_i is a random sparse coefficient vector, which represents the response to each basis function, corresponding to the activity state of simple cell neurons in the V1 area of the main visual cortex ; ε is usually assumed to be Gaussian white noise. For the transformed T_t signal, let the test sample be y, A be the training sample space composed of the transformed signal, and x be the sparse vector. When the l₀ of the sparse vector t_i is sufficiently sparse, the solution of the sparse vector and l₁ minimization problem is the same solution, and when the solution is different, it can be transformed into the following unconstrained optimization problem. From the above description, it can be seen that when solving the sparse vector of the equation, the training sample space of the transformed features must first be constructed. According to the extracted and transformed feature vectors, the K-singular value decomposition method is used to establish the sample feature space. In the process of space, it is necessary to continuously use the OMP algorithm to solve the sparse vector.

第四步：非负性约束稀疏编码模型的建立。目标函数定义为：The fourth step: the establishment of non-negativity constrained sparse coding model. The objective function is defined as:

$C C ((x x)) = = \frac{11}{22} | | | | y the y - - A A x x | | {| |}^{22} + + λ λ \underset{i i,, j j}{Σ Σ} f f (({x x}_{i i j j})) - - - - - - ((22))$

其中，约束条件为λ＞0；y表示测试图像转换组合而成的列向量，其中的元素均大于或等于零；A表示基函数，其中的元素均大于或等于零；x表示稀疏性系数，其中的元素x_ij均大于或等于零。稀疏向量x的稀疏性由惩罚函数的具体形式所决定，定义为f(·)。则取其函数表达式为f(x)＝|x|＝x(x≥0)，因此目标函数式(2)在非负性条件下等价于式(3)，即Among them, the constraint condition is λ>0; y represents the column vector formed by the test image conversion and combination, and the elements in it are all greater than or equal to zero; A represents the basis function, and the elements in it are all greater than or equal to zero; x represents the sparsity coefficient, where The elements x_ij are all greater than or equal to zero. The sparsity of the sparse vector x is determined by the specific form of the penalty function, which is defined as f(·). Then take its function expression as f(x)=|x|=x(x≥0), so the objective function formula (2) is equivalent to formula (3) under the condition of non-negativity, namely

$C C ((x x)) = = \frac{11}{22} | | | | y the y - - A A x x | | {| |}^{22} + + λ λ \underset{i i,, j j}{Σ Σ} {x x}_{i i j j} - - - - - - ((33))$

基于非负性稀疏编码的主要思想是：令τ＞1，A＝τA。即A乘上一个放大系数，x乘上一个缩小系数，||y-Ax||²值不变，x中所有元素缩小，f(x)减小，则可以保证目标函数C(x)总是递减。那么对于给定的基函数，对稀疏系数x的优化可以用式(4)进行计算，即The main idea based on non-negative sparse coding is: Let τ>1, A=τA. That is, A is multiplied by an enlargement factor, x is multiplied by a reduction coefficient, the value of ||y-Ax||² remains unchanged, all elements in x are reduced, f(x) is reduced, and the total value of the objective function C(x) can be guaranteed is decreasing. Then for a given basis function, the optimization of the sparse coefficient x can be calculated by formula (4), namely

xⁱ⁺¹＝xⁱ.*{(A^Ty)./(A^TAy+λ)} (4)xⁱ⁺¹ = xⁱ .*{(A^T y)./(A^T Ay+λ)} (4)

其中，“.*”和“./”分别代表矩阵的点乘和点除。采用式(4)更新规则来实现x的迭代过程，则更新后的x仍然满足非负性，因为它的更新方法是通过乘以一个非负因子(A^Ty)./(A^TAy+λ)来实现的。只要当稀疏向量x的初始值设置为正数，那么在x的迭代过程中对任意要求的精度都能够收敛到全局最小值。给定x不变，考虑A的优化问题。采用标准的梯度下降算法，得到A的更新规则为：Among them, ".*" and "./" represent the dot multiplication and division of the matrix respectively. Using the update rule of formula (4) to realize the iterative process of x, the updated x still satisfies the non-negativity, because its update method is by multiplying a non-negative factor (^AT y)./(^AT Ay+ λ) to achieve. As long as the initial value of the sparse vector x is set to a positive number, then any required accuracy can converge to the global minimum during the iterative process of x. Given that x is constant, consider the optimization problem of A. Using the standard gradient descent algorithm, the update rule of A is obtained as:

$A A = = β β [[{A A}^{t t} + + \frac{μ μ}{β β} ((y the y - - {A A}^{t t} x x)) {x x}^{T T}]] - - - - - - ((55))$

第五步：基于稀疏编码模型采用重构误差最小法进行车辆品牌类型识别。对于给定的k类车辆品牌图像，通过对各类训练样本进行字典学习，可获得m个最适合重构原始车辆品牌图像的基函数集合。对于任意的测试样本，计算其用每种基函数集合进行稀疏表示的重构误差，对应于重构误差最小的类别为该样本的所属类别。即：设共有k类，每个类下的每个样本用v中的一个列向量来描述，若第i类包含n_i个样本，则有Step 5: Based on the sparse coding model, the method of minimum reconstruction error is used to identify the vehicle brand type. For a given k types of vehicle brand images, by performing dictionary learning on various training samples, m basis function sets that are most suitable for reconstructing the original vehicle brand images can be obtained. For any test sample, calculate the reconstruction error of its sparse representation with each basis function set, and the category corresponding to the smallest reconstruction error is the category of the sample. That is: suppose there are k classes in total, and each sample under each class is described by a column vector in v , if the i-th class contains n_i samples, then

${A A}_{i i} = = [[{v v}_{{in in}_{11}},, {v v}_{{in in}_{22}},, {v v}_{{in in}_{33}},, . . . . . .,, {v v}_{{in in}_{i i}}]] &Element; &Element; {R R}^{m m \times \times {n no}_{i i}} - - - - - - ((66))$

若y属于第i类，则即y可以由第i类的样本进行线性组合来逼近。如何确定基于稀疏表示的车辆品牌分类，即可以通过求得的稀疏解然后重构出各类图像，然后通过与原始测试样本进行求残差，分类规则根据重构误差最小确定所属的类别。If y belongs to the i-th class, then That is, y can be approximated by a linear combination of samples of the i-th class. How to determine the vehicle brand classification based on sparse representation, that is, various images can be reconstructed through the obtained sparse solution, and then the residual error with the original test sample is calculated, and the classification rule determines the category according to the minimum reconstruction error.

应当指出，对于本技术领域的普通技术人员来说，在不脱离本发明原理的前提下，还可以做出若干改进和润饰，这些改进和润饰也应视为本发明的保护范围。本实施例中未明确的各组成部分均可用现有技术加以实现。It should be pointed out that those skilled in the art can make some improvements and modifications without departing from the principle of the present invention, and these improvements and modifications should also be regarded as the protection scope of the present invention. All components that are not specified in this embodiment can be realized by existing technologies.