CN108021947A - Visual sense-based hierarchical extreme learning machine target identification method - Google Patents

Visual sense-based hierarchical extreme learning machine target identification method Download PDF

Info

Publication number
CN108021947A
CN108021947A CN 201711419430 CN201711419430A CN108021947A CN 108021947 A CN108021947 A CN 108021947A CN 201711419430 CN201711419430 CN 201711419430 CN 201711419430 A CN201711419430 A CN 201711419430A CN 108021947 A CN108021947 A CN 108021947A
Authority
CN
Grant status
Application
Patent type
Prior art keywords
elm
network
training
layer
feature extraction
Prior art date
Application number
CN 201711419430
Other languages
Chinese (zh)
Other versions
CN108021947B (en )
Inventor
张弘
罗昭慧
李军伟
Original Assignee
北京航空航天大学
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date

Links

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING; COUNTING
    • G06KRECOGNITION OF DATA; PRESENTATION OF DATA; RECORD CARRIERS; HANDLING RECORD CARRIERS
    • G06K9/00Methods or arrangements for reading or recognising printed or written characters or for recognising patterns, e.g. fingerprints
    • G06K9/62Methods or arrangements for recognition using electronic means
    • G06K9/6217Design or setup of recognition systems and techniques; Extraction of features in feature space; Clustering techniques; Blind source separation
    • G06K9/6256Obtaining sets of training patterns; Bootstrap methods, e.g. bagging, boosting
    • GPHYSICS
    • G06COMPUTING; CALCULATING; COUNTING
    • G06KRECOGNITION OF DATA; PRESENTATION OF DATA; RECORD CARRIERS; HANDLING RECORD CARRIERS
    • G06K9/00Methods or arrangements for reading or recognising printed or written characters or for recognising patterns, e.g. fingerprints
    • G06K9/62Methods or arrangements for recognition using electronic means
    • G06K9/6267Classification techniques
    • GPHYSICS
    • G06COMPUTING; CALCULATING; COUNTING
    • G06NCOMPUTER SYSTEMS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N99/00Subject matter not provided for in other groups of this subclass
    • G06N99/005Learning machines, i.e. computer in which a programme is changed according to experience gained by the machine itself during a complete run

Abstract

The present invention relates to a visual sense-based hierarchical extreme learning machine target identification method. The method comprises the following steps of (1) calibrating an image data set,dividing the image data set into a training set and a measurement set according to a certain proportion, processing into the forms of the unified sizes, and expanding into the column vectors; (2) designing a hierarchical extreme learning machine network model, according to the size of the data set, setting the number of the sparse coding layers of a feature extraction network and the number of the nodes of each network layer; (3) training the feature extraction network, namely training to obtain the weights of the sparse coding layers, connecting the input data with a hidden layer via randomweight, and then iterating and solving the weight of the hidden layer via a fast iterative shrinkage threshold algorithm (FISTA); (4) taking the feature vectors extracted by the feature extraction network as the input of an extreme learning machine (ELM), training an ELM network to obtain the parameters of the hidden layer of the ELM; (5) inputting the data of the test set in the feature extraction network to extract the features, and then inputting in the ELM to obtain the classification results.

Description

一种基于视觉的分层极限学习机目标识别方法 A learning machine target identification method based on hierarchical limit vision

技术领域 FIELD

[0001] 本发明涉及一种基于视觉的分层极限学习机目标识别方法,属于计算机视觉信息领域。 [0001] The present invention relates to a layered ELM target recognition method based on vision, computer vision information belongs.

背景技术 Background technique

[0002] 目标识别是通过特定的算法分析图像中是否存在特定的目标,主要提取最能代表目标的特征数据,通过一定方法把这些数据训练出数学模型,最后根据此模型对未知的目标进行分类与识别。 [0002] The object recognition is whether a particular object is present by certain algorithms analyze the image, the primary extracts feature data most representative of the target by a certain method the data to train the mathematical model, and finally according to the model to classify unknown targets and recognition. 基于机器学习的目标识别方法中的一个关键步骤就是设计一个泛化性能好并且输出稳定的分类器,当前运用较为广泛的分类器是支持向量机(Support Vector Machine,SVM)和BP 神经网络(Back Propagation Neural Network)。 Target Recognition based on machine learning is a key step is to design a good generalization performance and stable output sorter, more extensive use of current classifier is SVM (Support Vector Machine, SVM) and BP neural network (Back Propagation Neural Network).

[0003] 虽然这些方法作为目标识别模型的分类器都有着不错的泛化性能,但是它们依然存在着不可忽视的缺陷,对于BP神经网络来说,由于其参数多且复杂,训练极为耗时,同时需求大量的训练数据,而SVM是基于小样本设计的,所以很好的解决了训练数据这个问题, 但同样存在训练耗时,参数寻优慢等实时性问题,特别是针对一些高维数的复杂数据,计算复杂度会非常高。 [0003] Although these methods as a classifier object recognition model has a good generalization performance, but they still can not be ignored defects for BP neural network, because of its parameters are many and complex, time-consuming training, At the same time demand a large amount of training data, while SVM is designed based on a small sample, it is a good solution to the problem of training data, but there are also real-time problem consuming training, parameter optimization and slow, especially for some high-dimensionality complex data, computational complexity is very high.

[0004] 稀疏编码算法(Sparse Coding)是一种无监督学习算法,它用来寻找一组完备的基向量来更高效地表示样本数据。 [0004] Sparse coding algorithm (Sparse Coding) is an unsupervised learning algorithm, which is used to find a complete set of basis vectors to represent the sample data more efficiently. 形如主成分分析(PCA),但是比PCA灵活,可以设置隐含层节点数,来适应不同的数据,从而提取高效的特征表达。 Shaped as principal component analysis (PCA), but flexible than PCA, hidden layer nodes may be provided, to accommodate different data to extract features and efficient expression.

[0005] 极限学习机(Extream Leanring Machine,ELM)是一种有监督的学习算法,输入向量与的ELM的隐含层通过随机权重连接,然后通过有监督的训练隐含层与输出层之间的权重参数。 [0005] ELM (Extream Leanring Machine, ELM) is a supervised learning algorithm, hidden layer input vector ELM reconnection with random weights, and then through between the hidden layer and the output with a layer Training Supervision the weight parameter. 没有反向传播的过程,极大的减少了网络训练的耗时。 No back-propagation process, greatly reducing the time-consuming network training. 同时ELM网络参数少,相比于一般的卷积神经网络(CNN),计算复杂度较低。 Meanwhile ELM less network parameters, as compared to the general convolutional neural network (CNN), a lower computational complexity.

[0006] 因此,本发明结合稀疏编码算法与极限学习机算法的优点,提出了一种分层极限学习机的目标识别算法。 [0006] Accordingly, the present invention combines the advantages of sparse coding algorithm and machine learning algorithms limit, target identification algorithm layered limit learning machine. 分层极限学习机能够很好利用稀疏编码学习到的特征,对比一般的极限学习机算法,它能够提高算法识别性能的同时把训练的耗时控制在合理范围之内。 ELM well hierarchical sparse coding learned features, in general ELM comparison algorithms, the algorithm which can improve the recognition performance while consuming training is controlled within reasonable limits.

发明内容 SUMMARY

[0007] 本发明技术解决问题:克服现有技术的不足,提供一种基于视觉的分层极限学习机目标识别方法,满足识别准确率要求的同时把训练耗时控制在合理范围之内。 Technical invention [0007] The present problem: to overcome the deficiencies of the prior art, provide a layered ELM target recognition method based on visual, recognition accuracy while satisfying the required training time consuming controlled within reasonable limits.

[0008] 本发明技术解决方案:一种基于视觉的分层极限学习机目标识别方法,包括以下步骤: [0008] The technical solution of the invention: A hierarchical ELM vision-based object recognition method, comprising the steps of:

[0009] ⑴将图像数据集进行标定,并分为训练集与测试集; [0009] ⑴ the calibrated image data set, and divided into a training set and test set;

[0010] (2)根据训练集的大小,设置特征提取网络的稀疏编码层数及每个网络层节点数; 利用训练集训练所述特征提取网络,得到特征提取网络的稀疏编码层权重,从而获得已训练的特征提取网络及训练集的有效的特征向量; [0010] (2) extracting the training set according to the size of the network is provided wherein the number of layers and sparse coding each network layer nodes; sparse coding layer weights using the training set of training the network feature extraction, feature extraction network weights obtained, whereby feature extraction to obtain a trained network training set and effective eigenvectors;

[0011] (3)将步骤(2)得到的特征向量作为极限学习机网络ELM的输入,训练ELM并得到ELM的隐含层参数,从而获得已训练的ELM; Obtained [0011] (3) The step (2) as a limiting feature vectors ELM network learning machine input, and to give training ELM ELM hidden layer parameters, to obtain a trained ELM;

[0012] (4)由步骤(2)得到的已训练的特征提取网络与步骤⑶得到的已训练的极限学习机组合得到分层极限学习机网络模型,将步骤(1)得到的测试集输入到所述分层极限学习机网络模型并得到目标识别结果。 [0012] (4) from step (2) obtained in the network trained with the feature extraction step ⑶ obtained ELM composition has been trained network hierarchical ELM model, the step (1) to give the test set input ELM to the layered network model and target recognition results obtained.

[0013] 所述步骤⑵中,得到特征提取网络的稀疏编码层权重的具体过程包括: [0013] The step ⑵ obtain network feature extraction heavy weight concrete layer sparse coding process comprises:

[0014] 所述特征提取网络由多层的稀疏编码器构成,对于每一层稀疏编码器,将输入通过随机权重与隐含层相连,隐含层到输出层之间为稀疏编码层权重,然后采用阈值收缩算法(FISTA)迭代使输出与输入保持相同,从而求得稀疏编码层的权重。 [0014] The feature extraction network constructed of multiple layers sparse encoder, sparse coding for each layer, a hidden layer is connected to an input random weights to the hidden layer is a layer of sparse coding weights between the output layer, - threshold algorithm is then employed (FISTA) iterative input output remains the same, thereby to obtain the right weight sparse coding layer.

[0015] 所述步骤⑶中,得到ELM的隐含层参数的具体过程包括: [0015] The step ⑶ give hidden layer of ELM specific process parameters comprising:

[0016] 将步骤(2)中得到训练集的有效的特征向量输入ELM,有效的特征向量与ELM的随机权重相连得到输出H,通过公式: [0016] The effective eigenvectors in step (2) obtained in the training set input ELM, Random Weight ELM effective eigenvectors and weights to obtain an output coupled to H, by the equation:

[0017] [0017]

Figure CN108021947AD00041

[0018]得到极限学习机隐含层参数『,其中T为训练集的标签,其中C是规范化系数,C的取值为2_3()。 [0018] get ELM hidden layer parameters ", where T is the training set of tags, where C is a normalization factor, the value of C 2_3 ().

[0019] 本发明与现有技术相比的优点在于: [0019] The advantages of the present invention compared to the prior art in that:

[0020] (1)本发明的特征提取网络通过多层稀疏编码器的叠加能够从复杂的数据中提取有效的特征信息,因而对于一般的极限学习机(ELM)方法,提取特征的鲁棒性更好,识别率更尚; [0020] (1) extracting features of the present invention by a multilayer overlay network is sparse coding can be extracted from the feature information effectively complex data, and thus to limit learning machine in general (the ELM) method, the robustness of feature extraction better recognition rate is still more;

[0021] (2)本发明利用有监督的极限学习机方法对特征进行分类识别,对比于支持向量机(SVM),可以适应更大的样本集,泛化能力也更好,对比于卷积神经网络(CNN)方法,没有反向传播过程训练时间短,计算速度更快; [0021] (2) The present invention utilizes a supervised learning machine limits features classification method, in contrast to the support vector machine (the SVM), can accommodate a larger sample set, but also better generalization ability, in contrast to the convolution neural network (CNN) method, there is no back-propagation procedure short training time, faster calculation;

[0022] ⑶本发明将特征提取网络与极限学习机结合(ELM),可以根据不同大小的数据集来设置特征提取网络的稀疏编码层数,自适应程度高,可以应用到不同场景条件下的目标识别任务当中。 [0022] ⑶ the present invention is a network feature extraction ELM binding (the ELM) may be set according to the size of the data sets of different feature extraction layers sparse coding network, a high degree of adaptation, can be applied to different scenarios under conditions object recognition task them.

附图说明 BRIEF DESCRIPTION

[0023] 图1基于视觉的分层极限学习机目标识别方法流程方框示意图; [0023] FIG 1 is a hierarchical vision-based object recognition ELM schematic process flow block;

[0024] 图2分层极限学习机网络模型示意图; [0024] FIG. 2 ELM hierarchical network model schematic;

[0025] 图3手写数字目标识别效果图。 [0025] FIG Target Recognition of handwritten numeral 3 in FIG.

具体实施方式 detailed description

[0026] 要分为两部分:无监督的特征提取网络以及有监督的极限学习机分类网络,该方法充分利用了分层极限学习机在目标识别中的准确性,以及训练过程中的低耗时特点。 [0026] should be divided into two parts: the network unsupervised feature extraction and supervised ELM classified network, which takes full advantage of the hierarchical ELM accuracy in target recognition, as well as the training process and low consumption when characteristics.

[0027] 如图1所示,一种基于视觉的分层极限学习机目标识别方法,该方法的具体步骤如下: [0027] 1 A layered visual object recognition method based ELM, specific steps of the method are as follows:

[0028] (1)将图像数据集进行标定,并按照一定比例分为训练集与测集,并处理为统一大小的格式,然后将图像数据展开为列向量。 [0028] (1) the calibrated image data set, and divided into a training set and test set according to a certain proportion, and processed into a format uniform size, and then expands the image data column vector.

[0029] ⑵训练无监督特征提取网络: [0029] ⑵ training unsupervised feature extraction network:

[0030] 如图2所示,无监督特征提取网络主要由多层稀疏编码器叠加构成。 [0030] As shown in FIG 2, unsupervised feature extraction of a multilayer network is mainly composed of a sparse overlay encoder. 根据数据大小,设定稀疏编码层数,若数据集较大,可以适当增加层数,数据集较小,可以适当减少层数,一般可以设置为三层; The size of the data set sparse coding layers, if the data set is large, the number of layers can be increased, the data set is small, the number of layers may be appropriately reduced, generally can be set to three;

[0031] 对于一个稀疏编码层,主要过程描述如下: [0031] For a sparse coding layer, the main process is described as follows:

[0032] 设输入数据为X,k个η维原始特征样本 [0032] The set of input data X, k th sample of the original feature dimensions η

Figure CN108021947AD00051

经过随机权重与隐含层连接,设隐含层节点数量为L.则隐含层第i个节点输入为: After random hidden layer connection weights and, provided the number of nodes of the hidden layer nodes L. input layer i is implied as:

[0033] [0033]

Figure CN108021947AD00052

[0034] 其中Wi,bi为随机参数,得到一个L维随机特征向量,记为H: [0034] where Wi, bi random parameter, to obtain a random L-dimensional feature vector, referred to as H:

[0035] h (x) = [g (x;wi,bi),——,g (x;WL,bL)] [0035] h (x) = [g (x; wi, bi), -, g (x; WL, bL)]

[0036] 隐含层输出H通过隐含层权重炉与输出层相连,期望得到一个与输入相同的输出, 即从随机特征中恢复原始数据,因此目标的损失&函数可以表示为: [0036] H hidden layer output by the hidden layer and the output layer weights connected to the furnace, it is desirable to obtain the same output with one input, i.e. restore the original data from the random characteristics, so the loss of the target & amp; function can be expressed as:

[0037] [0037]

Figure CN108021947AD00053

[0038] 可以定义为 [0038] can be defined as

Figure CN108021947AD00054

[0039] 通过阈值收缩算法(shrinkage-thresholding algorithm,FISTA)反复迭代求解出隐含层参数妒。 [0039] Algorithm shrinkage (shrinkage-thresholding algorithm, FISTA) by iteration threshold parameter jealous solved the hidden layer. 设j代表迭代次数,Vp为P ®h)梯度函数,γ为Vp的Lipschitz常数,迭代开始,设 J represents the number of iterations provided, Vp is P ®h) gradient function, gamma] is a Lipschitz constant Vp of iteration start, provided

Figure CN108021947AD00055

对于j (j多1),按照下面步骤: For j (j over), the following steps:

Figure CN108021947AD00056

[0043] 求解出隐含层参数之后,使用隐含层参数妒作为稀疏编码层的权重,对输入数据提取有效的特征。 After [0043] parameters solved the hidden layer, the hidden layer used as weights sparse coding parameters Jealous layer weight, effective feature extraction on the input data. 稀疏编码器学习到的特征可以简洁准确反应出原始数据的特征,这种无监督的稀疏编码过程提取到的特征应用于有监督的分类层,可以提高目标识别的准确率。 Learning sparse coding features may be simple to accurately reflect the characteristics of the original data, this unsupervised sparse coding process is applied to the extracted feature with a supervised classification layer, can improve the accuracy of target recognition.

[0044] (3)训练有监督的分类网络: [0044] (3) training supervised classification network:

[0045] 有监督的分类网络主要由单层极限学习机(ELM)构成,设输入层节点数量为Ρ,隐含层节点数量为L。 [0045] The supervised classification network mainly composed of a single ELM (ELM), provided the number of input layer node Ρ, the number of hidden layer nodes is L. 与稀疏编码层类似,ELM隐含层的第i个节点的输出为: Similarly sparse coding layer, the i-th output of the hidden layer nodes as ELM:

[0046] [0046]

Figure CN108021947AD00057

[0047] 贝Ij相当于将P维向量映射至IjL维向量:h (X) = [g (x;wi,bi),. . .,g (x;WL,bL)] [0047] Ij shell P-dimensional vector corresponds to a mapping IjL dimensional vector: h (X) = [.. G (x; wi, bi) ,., g (x; WL, bL)]

[0048] 其中Wl为第输入层节点与隐含层节点之间的第i个连接A1为偏置,一般为随机赋值,bi为偏置,g (X)为激活函数,这里使用sigmoid函数: [0048] wherein Wl is connected to the i-th input layer A1 between the first node and hidden layer node is biased, typically randomly assigned, bias BI is, g (X) is the activation function, a sigmoid function used here:

[0049] [0049]

Figure CN108021947AD00058

[0050] 对应L维输出向量,记为Η: [0050] the corresponding L-dimensional output vector, referred to as Η:

[0051] [0051]

Figure CN108021947AD00059

[0052] 使用ELM训练分类器,在ELM算法中,输入权重和偏置是随机分配的因此只有妒需要训练。 [0052] The use of ELM trained classifier algorithm in the ELM, the weights and bias inputs are assigned randomly so that only the required training jealous. 令yk表示输入xk对应的实际的输出向量,则将所有的训练样本带入公式中可以得到: So that the actual output yk represents the input vector xk corresponding, then all the training samples into the formula can be obtained:

[0053] [0053]

Figure CN108021947AD00061

[0054] 其中: [0054] wherein:

Figure CN108021947AD00062

[0058] 设T为训练样本标签,训练的目标是使得训练误差|T-ffie| I2和输出权重I lrl I的范数最小。 [0058] Let T be the training sample labels, training goal is to make the training error | T-ffie | I2 and the minimum output weight LRL I I norm. 因此训练过程可以表示为一个有约束最优化问题: Therefore, the training process can be represented as a constrained optimization problem:

[0059] [0059]

Figure CN108021947AD00063

[0060] 其中C是规范化系数,用来平衡拟合函数的平滑度和函数拟合值与真实数据距离差距这两者之间的关系。 [0060] where C is a normalization factor, used to balance the relationship between the two fitting smoothness value function, and function fitting the data from the real gap. 可以使用拉格朗日法解决此问题,如果矩阵I Lagrangian method can be used to solve this problem, if the matrix I

Figure CN108021947AD00064

为非奇异矩阵,则 Non-singular matrix, then

Figure CN108021947AD00065

,若为奇异矩阵,则有: If the matrix is ​​singular, there are:

[0061] [0061]

Figure CN108021947AD00066

. 求解得到的π作为ELM隐含层权重,与L维输出向量H连接,可以得到最终的分类结果。 Solving obtained as π ELM weight hidden layer weights, connected to the output of the L-dimensional vector H, the result of classification can be obtained.

[0062] ⑷测试训练得到的分层极限学习机模型: [0062] layered ELM model ⑷ test training obtained:

[0063] 利用(2)迭代得到的多层稀疏编码层权重βΗ,对输入图像列向量提取有效的特征, 然后将此特征向量输入到极限学习机(ELM)当中,进行分类识别,过程如下: [0063] using (2) a multilayer iterative sparse coding layer weights obtained heavy βΗ, extraction of the input image sequence effective feature vectors, the feature vector is then input this ELM (the ELM) which, for classification, as follows:

[0064] 设最终输出层的节点数目记作Μ,第i个隐含层节点和第j个输出层节点之间的权重为妒,对于输入向量X,节点j的输出为: [0064] The number of nodes of the output layer finally disposed referred to as [mu], a weight between the i-th hidden layer node and a j-th node of the output layer weight jealous, the input vector X, the output node j is:

[0065] [0065]

Figure CN108021947AD00067

Figure CN108021947AD00068

[0066] 其中随机连接权重与偏置wi,bi应与ELM训练过程中保持一致,K为ELM有监督训练得到。 [0066] wherein the random bias connection weights wi, bi ELM should be consistent with the training process, K is obtained ELM supervised training.

[0067] 因此输入样本X对应的输出为: 在识别阶段,给定一个样本X, 则该样本所属类别为: [0067] Thus the output of the input sample X corresponds to: the identification phase, a given sample X, the sample belongs to category:

[0068] [0068]

Figure CN108021947AD00071

[0069] 通过上述方法得到样本的测试结果,将测试集的测试分类结果与测试机的真实结果相比较,得到此分层极限学习机的准确率; [0069] The test results obtained by the above method the sample, the real results of the test results of the classification test set is compared with the test machine, to obtain accuracy limit of this hierarchy learning machine;

[0070] (5)若准确率满足实际要求,可将此模型应用于实际工程,若不达到要求,则扩大数据集,改变各个网络层的节点数目以及特征提取层层数,重复进行步骤二、三,提高准确率。 [0070] (5) When the accuracy meet the actual requirements, it can be practical engineering application of this method, if not meet the requirements, the expansion of data sets, changing the number of nodes, and wherein each network layer extracted several layers, two steps are repeated and third, to improve accuracy.

[0071] 下面通过一个具体实施例来验证本发明所提出的基于视觉的分层极限学习机目标识别方法的性能,所采用的是手写数字图像作为验证对象。 [0071] By following a specific embodiment of the present invention is proposed to verify the performance of the vision-based object recognition method ELM layered, used as a digital image of the handwritten verification target.

[0072] 基于视觉的分层极限学习机目标识别方法,其具体实现步骤如下: [0072] Target Identification ELM based layered visual, specific implementation steps are as follows:

[0073] (1)将图像数据集进行标定,并按照一定比例,本发明经过试验确定为4比1分为训练集与测集,并处理为统一大小的格式,然后将图像数据展开为列向量; [0073] (1) the calibrated image data set, and a certain percentage, the present inventors have experimentally determined ratio of 1 to 4 is divided into a training set and test set, and the processing is unified format size, and then expands the image data column vector;

[0074] (2)设计分层极限学习机网络模型,根据数据集大小,在这里设置为三层; [0074] (2) Design of a hierarchical network model ELM, the data set size, set here to three;

[0075] (3)训练特征提取网络,即训练得到多层稀疏编码层权重,输入数据通过随机权重与隐含层相连,然后通过FISTA算法迭代求解出隐含层权重炉,即为稀疏编码层的权重,最后通过三次稀疏编码层提取有效的特征输出; [0075] (3) Training feature extraction network, i.e., the trained weights multilayer sparse coding layer, the hidden layer is connected to the input data by the random weights, and then the hidden layer weights iterative solver algorithm by weight FISTA furnace, i.e. sparse coding layer weight, and finally outputs the extracted valid sparse coding features by three layers;

[0076] ⑷将特征提取网络提取到的特征向量作为极限学习机(ELM)的输入,训练ELM网络。 [0076] ⑷ network feature extraction to extract feature vectors as input ELM (ELM), the ELM network training. 输入向量与随机权重相连得到输出H,训练集的标签为T,则极限学习机隐含层参数K根据奇异性可以表示为: Random input vector and the weight to obtain an output coupled to H, a training set of tags T, the ELM The hidden layer parameter K can be expressed as singularity:

Figure CN108021947AD00072

从而训练得到ELM隐含层参数妒; Training hidden layer thus obtained ELM Jealous parameters;

[0077] (5)将测试集样本输入(3)训练得到的特征提取网络提取特征,然后输入极限学习机(ELM)得到测试集数据的测试分类结果,如图3所示,为手写数字的识别结果,说明本发明方法均能识别正确; [0077] (5) a test sample input collector (3) of the trained network feature extraction feature extraction, and then enter the ELM (the ELM) to give the test data set classification test results, shown in Figure 3, digital handwritten recognition result, the method of the present invention is able to identify the correct;

[0078] (7)将测试集的测试分类结果与测试机的真实结果相比较,在minst数据集上得到此分层极限学习机的准确率为95%。 [0078] (7) The classification of the test results of the real set of test results comparing the test machine, to obtain this hierarchy ELM accuracy rate of 95% on minst dataset.

[0079] 提供以上实施例仅仅是为了描述本发明的目的,而并非要限制本发明的范围。 [0079] The above embodiments are merely provided for the purpose of describing the present invention, and are not intended to limit the scope of the invention. 本发明的范围由所附权利要求限定。 Scope of the invention defined by the appended claims. 不脱离本发明的精神和原理而做出的各种等同替换和修改,均应涵盖在本发明的范围之内。 Without departing from the spirit and principles of the present invention made various equivalents and modifications should fall within the scope of the present invention.

Claims (3)

  1. 1. 一种基于视觉的分层极限学习机目标识别方法,其特征在于,包括以下步骤: (1) 将图像数据集进行标定,并分为训练集与测试集; (2) 根据训练集的大小,设置特征提取网络的稀疏编码层数及每个网络层节点数;利用训练集训练所述特征提取网络,得到特征提取网络的稀疏编码层权重,从而获得已训练的特征提取网络及训练集的有效的特征向量; (3) 将步骤⑵得到的特征向量作为极限学习机网络ELM的输入,训练ELM并得到ELM的隐含层参数,从而获得已训练的ELM; ⑷由步骤⑵得到的已训练的特征提取网络与步骤(3)得到的已训练的极限学习机组合得到分层极限学习机网络模型,将步骤(1)得到的测试集输入到所述分层极限学习机网络模型并得到目标识别结果。 A vision-based object recognition ELM hierarchical method, comprising the steps of: (1) the calibrated image data set, and divided into a training set and test set; (2) according to the training set size, provided the network sparse coding feature extraction layers and the number of nodes in each network layer; using a training set of training the network feature extraction, feature extraction to obtain sparse coding layer network weights, trained to obtain a feature extraction network training set and effective eigenvectors; (3) the step of feature vectors obtained as ⑵ ELM ELM network input, and to give training ELM ELM hidden layer parameters, to obtain a trained ELM; ⑵ ⑷ step is obtained from training feature extraction network and step (3) to give the ELM composition has been trained network hierarchical ELM model, the step (1) obtained in the test set is input to the hierarchical network model and ELM give target recognition results.
  2. 2. 根据权利要求1所述的基于视觉的分层极限学习机目标识别方法,其特征在于:所述步骤(2)中,得到特征提取网络的稀疏编码层权重的具体过程包括: 所述特征提取网络由多层的稀疏编码器构成,对于每一层稀疏编码器,将输入通过随机权重与隐含层相连,隐含层与输出层之间为稀疏编码层权重,然后采用阈值收缩算法(FISTA)迭代使输出与输入保持相同,从而求得稀疏编码层的权重。 2. The method according to claim visual object recognition learning machine based on hierarchical limit, characterized in that said 1: the step (2) to give a heavy weight concrete sparse coding layer network feature extraction process comprises: the features extracting a plurality of layers of a sparse network encoder, sparse coding for each layer, a hidden layer is connected to an input random weights between the hidden layer and the output layer is a layer weight sparse coding weights, and scaling algorithm using the threshold value ( FISTA) iterative input output remains the same, thereby to obtain the right weight sparse coding layer.
  3. 3. 根据权利要求1所述的基于视觉的分层极限学习机目标识别方法,其特征在于:所述步骤(3)中,得到ELM的隐含层参数的具体过程包括: 将步骤(2)中得到训练集的有效的特征向量输入ELM,有效的特征向量与ELM的随机权重相连得到输出H,通过公式: 3. A vision-based object recognition ELM hierarchical method of claim 1, wherein: said step (3), to give the specific process parameters ELM hidden layer comprises: Step (2) obtained in the training set of input feature vector ELM effective, efficient random weights feature vectors to obtain an output connected to a weight ELM H, by the equation:
    Figure CN108021947AC00021
    得到极限学习机隐含层参数f,其中T为训练集的标签,其中C是规范化系数。 Obtained ELM hidden layer parameter f, where T is the tag of the training set, where C is a normalization factor.
CN 201711419430 2017-12-25 2017-12-25 A learning machine target identification method based on hierarchical limit vision CN108021947B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN 201711419430 CN108021947B (en) 2017-12-25 2017-12-25 A learning machine target identification method based on hierarchical limit vision

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN 201711419430 CN108021947B (en) 2017-12-25 2017-12-25 A learning machine target identification method based on hierarchical limit vision

Publications (2)

Publication Number Publication Date
CN108021947A true true CN108021947A (en) 2018-05-11
CN108021947B CN108021947B (en) 2018-10-26

Family

ID=62074646

Family Applications (1)

Application Number Title Priority Date Filing Date
CN 201711419430 CN108021947B (en) 2017-12-25 2017-12-25 A learning machine target identification method based on hierarchical limit vision

Country Status (1)

Country Link
CN (1) CN108021947B (en)

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104899549A (en) * 2015-04-17 2015-09-09 重庆大学 SAR target recognition method based on range profile time-frequency image identification dictionary learning
EP2325294B1 (en) * 2000-06-08 2016-08-24 The Regents of The University of California Visual-servoing optical microscopy
CN106096543A (en) * 2016-06-08 2016-11-09 东华大学 Hand-written digital identification method based on improved extreme learning machine (ELM)
CN107085704A (en) * 2017-03-27 2017-08-22 杭州电子科技大学 Rapid facial expression recognition method based on ELM self-encoding algorithm

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP2325294B1 (en) * 2000-06-08 2016-08-24 The Regents of The University of California Visual-servoing optical microscopy
CN104899549A (en) * 2015-04-17 2015-09-09 重庆大学 SAR target recognition method based on range profile time-frequency image identification dictionary learning
CN106096543A (en) * 2016-06-08 2016-11-09 东华大学 Hand-written digital identification method based on improved extreme learning machine (ELM)
CN107085704A (en) * 2017-03-27 2017-08-22 杭州电子科技大学 Rapid facial expression recognition method based on ELM self-encoding algorithm

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
林雨: ""极限学习机与自动编码器的融合算法研究"", 《中国优秀硕士学位论文全文数据库 信息科技辑》 *
沈力波,邵枫等: ""基于极限学习机的全参考立体图像质量评价"", 《计算机辅助设计与图形学学报》 *

Also Published As

Publication number Publication date Type
CN108021947B (en) 2018-10-26 grant

Similar Documents

Publication Publication Date Title
Ou et al. Asymmetric transitivity preserving graph embedding
Quattoni et al. An efficient projection for l 1,∞ regularization
CN103793718A (en) Deep study-based facial expression recognition method
CN102208038A (en) Image classification method based on visual dictionary
CN103605972A (en) Non-restricted environment face verification method based on block depth neural network
Chao Introduction to pattern recognition
CN101833667A (en) Pattern recognition classification method expressed based on grouping sparsity
CN102156871A (en) Image classification method based on category correlated codebook and classifier voting strategy
CN1952953A (en) Posture recognition method of human's face based on limited Boltzmann machine neural network
Xu et al. Video anomaly detection based on a hierarchical activity discovery within spatio-temporal contexts
Dong et al. Automatic age estimation based on deep learning algorithm
Leng et al. A 3D model recognition mechanism based on deep Boltzmann machines
Li et al. Learning balanced and unbalanced graphs via low-rank coding
CN103530689A (en) Deep learning-based clustering method
CN102831446A (en) Image appearance based loop closure detecting method in monocular vision SLAM (simultaneous localization and mapping)
Andrés et al. Efficient aerodynamic design through evolutionary programming and support vector regression algorithms
CN103020642A (en) Water environment monitoring and quality-control data analysis method
Luo et al. Robust discrete code modeling for supervised hashing
CN102324047A (en) Hyper-spectral image ground object recognition method based on sparse kernel representation (SKR)
Langone et al. Incremental kernel spectral clustering for online learning of non-stationary data
CN101140216A (en) Gas-liquid two-phase flow type recognition method based on digital graphic processing technique
Wang et al. Feature selection and multi-kernel learning for adaptive graph regularized nonnegative matrix factorization
Xie et al. Deepshape: Deep-learned shape descriptor for 3D shape retrieval
Wu et al. Optimization and improvement based on K-Means Cluster algorithm
CN103488869A (en) Wind power generation short-term load forecast method of least squares support vector machine

Legal Events

Date Code Title Description
PB01
SE01
GR01