Image classification method based on extreme learning machine trained by OLQR algorithm

Abstract

The application discloses an extreme learning machine image classification method based on OLQR algorithm training, and combines the idea of optimal control with machine learning model training organically, reserves the powerful feature extraction capability of the extreme learning machine, converts the solution of the ELM output weight into a series of state feedback control problems of control systems by using the OLQR algorithm, fully utilizes the fast and stable characteristics of the optimal control technology to make the ELM training error quickly decay, and provides a new idea for the improvement and optimization of the machine learning algorithm and the image classification.

Description

Technical Field

[0001] This invention belongs to the field of image classification technology in computer vision, specifically relating to an image classification method for extreme learning machines trained based on the OLQR algorithm. Background Technology

[0002] Image recognition and classification are key and hot topics in computer vision research, with wide applications across various industries. For example, in the medical field, they assist in disease diagnosis, including but not limited to tumor detection and lesion analysis; in the transportation field, they are used for license plate recognition and traffic sign recognition; and in the security field, they are used for facial recognition and behavior analysis. In general, the development of image recognition and classification technology has brought tremendous changes and opportunities to many fields. With the continuous advancement of algorithms and models, image recognition and classification will play an even more important role in more areas, bringing greater convenience and innovation to people's lives.

[0003] The core of image recognition and classification technology lies in deep learning and neural networks. By constructing deep neural network models and training them with large amounts of labeled data, computers can extract features from images, learn patterns, and make accurate predictions. Commonly used image recognition and classification algorithms include Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs). However, the training process of neural networks uses the backpropagation algorithm, which leverages the principle of gradient descent to continuously reduce the error between the model's output and the actual labels through multiple iterations of learning, thus completing the model training. As the number of network layers in a neural network model increases, the parameters that need to be adjusted also grow exponentially. This makes parameter training not only time-consuming and laborious but also poses significant challenges to researchers' model validation and debugging processes. Even minor changes to the model design require extensive training to see the final effect, rather than immediate results. Therefore, some scholars are considering improvements to the model structure and training methods of deep learning to address the current predicament.

[0004] Extreme Learning Machine (ELM) is one such solution. ELM is a single-hidden-layer feedforward neural network algorithm. Due to its simple structure, random generation of weights and biases of hidden layer nodes during training, and determination of output weights by solving the least-norm solution of a system of linear equations, ELM can achieve training speed improvements of hundreds or thousands of times compared to many traditional machine learning algorithms. Moreover, it can achieve good generalization performance while maintaining high learning accuracy, and is increasingly being used in the fields of image recognition and classification.

[0005] However, since ELM training requires all training data to be known, in practical applications, the training dataset is acquired incrementally. When the training dataset changes, ELM must be retrained, which significantly limits its efficiency and increases resource consumption. Therefore, it is necessary to consider using online learning methods to improve the model training process. Online learning is an efficient model training mechanism that allows the model to overcome the limitation of needing all data to be known during training. Sample data can be imported into the model one by one or in batches for training. When new data is generated after the model has been trained, the model only needs to be updated based on the current data without retraining. Summary of the Invention

[0006] In view of this, the present invention provides an image classification method based on the OLQR algorithm for Extreme Learning Machine training, which can achieve faster convergence speed, less training time and stronger generalization performance than the existing ELM image classification method.

[0007] This invention provides an image classification method for extreme learning machines trained using the OLQR algorithm, which specifically includes the following steps:

[0008] Step 1: Collect image data and label the categories to obtain training and test sets. Standardize the feature vectors of the images. The sample set is... ,in, Let k be the feature vector of the k-th image. Let N be the label of the k-th image, and N be the total number of samples.

[0009] Step 2: The weight training process of the Extreme Learning Machine (ELM) model is transformed into an online learning task using an online quadratic programming controller algorithm. An online learning ELM image recognition model based on this algorithm is established, as shown in the following equation:

[0010] ,

[0011] Among them, w i Let β be the input weight vector connecting the i-th hidden layer node and the input layer node, i=1,2,…,L, where β is the weight vector. i The output weight matrix b is used to connect the i-th hidden layer node and the output layer node. i Let be the bias of the i-th hidden layer node. For activation function, For the output matrix, w represents the number of hidden layer nodes. i and b i All are randomly generated;

[0012] Step 3: Using the training set generated in Step 1, train the online learning extreme learning machine image recognition model based on the OLQR algorithm using the online quadratic programming controller algorithm;

[0013] Step 4: In actual use, after performing the same standardization process as in Step 1 on the image to be classified, input it into the online learning extreme learning machine image recognition model based on the OLQR algorithm trained in Step 3 for calculation and inference, and output the category of the image to be classified to complete the image classification.

[0014] Step 3, which involves using an online quadratic programming controller algorithm to train an online learning extreme learning machine image recognition model based on the OLQR algorithm, includes:

[0015] Step 3.1: Construct an online learning framework using an online quadratic programming controller algorithm to compute the hidden layer output matrix of the extreme learning machine. The objective function is designed as follows:

[0016] ,

[0017] in, Used to measure the magnitude of error. Used to penalize the weight vector. For balance coefficient, and All are weighted control matrices;

[0018] Step 3.2: Initialize the online learning extreme learning machine image recognition model based on the online quadratic programming controller algorithm, and set the model parameters: balance coefficient. Each time, import the number of training data points M and the number of hidden layer nodes L of the model, and set the initial values ​​of the weight matrix. Randomly generate the weights of the hidden layer and bias Determine the activation function;

[0019] Step 3.3: Import M training samples As the training dataset for this round, t is a random number less than (NM). This training dataset is input into an online learning extreme learning machine image recognition model based on an online quadratic programming controller algorithm. The hidden layer output matrix is ​​calculated. The hidden layer output matrix The input matrix x′ of the online quadratic programming controller algorithm is then used to fit the data flow using a linear regression model, transforming the training process of the Extreme Learning Machine into an online learning task.

[0020] ,

[0021] in, , Let be the connection weight from the i-th hidden layer node to the output node, and d be the dimension of the sample features. The noise in the k-th training data sample;

[0022] Step 3.4: Using the input matrix x′ as input, calculate the output layer weight coefficients using an online quadratic programming controller algorithm. The training of an online learning extreme learning machine image recognition model based on an online quadratic programming controller algorithm was completed.

[0023] The online quadratic programming controller algorithm employs singular value decomposition (SVD) on the matrix. Perform eigenvalue decomposition to complete the output layer weight coefficients The solution is to find the solution, where, Let be the input matrix at a certain time. The process of inputting the product of a matrix and its transpose at a certain time includes:

[0024] Step 5.1: Set the loop variable The initial value is M;

[0025] Step 5.2: Let the current time index r = s, and obtain the current input matrix. and output matrix ;

[0026] Step 5.3: Calculate the product of the input matrix and its transpose at time r. ;

[0027] Step 5.4: Using the unitary matrix , Singular value decomposition is used to calculate and You can get and ,in, For matrix equations The solution, These are weighting coefficients;

[0028] Step 5.5 and They are respectively and The eigenvalues, will and Substitution achievable , equivalent to ,calculate get ;

[0029] Step 5.6, according to The feedback gain matrix is ​​calculated. ;

[0030] Step 5.7: Calculate the current error. ;

[0031] Step 5.8, Update and ;

[0032] Step 5.9, if Exit the loop and complete the output layer weight coefficients. The calculation; if Then, increment s by M and execute step 5.2; after completing the above calculations, the output layer weight coefficients can be obtained. .

[0033] Furthermore, the activation function sigmoid function .

[0034] Furthermore, the aforementioned and All are unit arrays.

[0035] Beneficial effects:

[0036] This invention organically combines the concept of optimal control with machine learning model training. While retaining the powerful feature extraction capabilities of Extreme Learning Machines (ELM), it uses the OLQR algorithm to transform the solution of ELM output weights into a series of state feedback control problems of a control system. It fully leverages the fast and stable characteristics of optimal control techniques to rapidly decay ELM training errors. Furthermore, by adding a model weight regularization term to the quadratic objective function, it effectively suppresses overfitting while ensuring model accuracy. Compared to the standard Extreme Learning Machine, the OLQR algorithm ensures that each iteration is optimal through optimal control, significantly shortening the convergence process of ELM and thus effectively improving training speed. Furthermore, the application of regularization techniques improves model generalization performance while suppressing overfitting. This provides a new approach for improving and optimizing machine learning algorithms and image classification. Attached Figure Description

[0037] Figure 1 This is a schematic diagram of the network structure of the Extreme Learning Machine.

[0038] Figure 2 This is a schematic diagram of the closed-loop system control structure for an online quadratic programming controller algorithm.

[0039] Figure 3 This is a flowchart illustrating an image classification method for an extreme learning machine trained using the OLQR algorithm, as provided by the present invention. Detailed Implementation

[0040] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0041] To facilitate a better understanding of the Extreme Learning Machine (ELM) image classification method provided in this invention, the basic principles of the ELM and the Online Linear Quadratic Regulator (OLQR) algorithm are explained as follows:

[0042] The symbols and their meanings are shown in Table 1:

[0043] Table 1. Symbols and their meanings

[0044]

[0045] The single hidden layer neural network structure of Extreme Learning Machine (ELM) is as follows: Figure 1 As shown, for a given A dataset of samples Assuming the number of hidden layer nodes is Then the mathematical model of the extreme learning machine can be expressed as:

[0046]

[0047] To further simplify the problem, arranging the above variables in a certain order yields the following form:

[0048] , ,

[0049]

[0050] Then formula (1) can be rewritten as formula (2):

[0051]

[0052] Where H is the hidden layer output matrix, Let be the output weight matrix to be determined. This is the output matrix. Therefore, the training process of ELM can be summarized into three steps: ① Initialization, randomly generating hidden layer parameters. and ② Based on the selected activation function The ELM hidden layer output matrix is ​​obtained by calculating the weights and biases of the initialized hidden layer. ③ Calculate the output layer weight matrix .

[0053] To solve for the output layer weight matrix, it is necessary to use the output matrix of the ELM. and label matrix To construct the loss function, the standard extreme learning machine uses the root mean squared error (MSE) as the loss function, as shown in formula (3), where Representing the square of the Frobenius norm:

[0054]

[0055] Using the Moore-Penrose generalized inverse method, the solution to problem (3) can be obtained as follows:

[0056]

[0057] in, for The Moore-Penrose generalized inverse matrix. There are many methods for calculating the Moore-Penrose generalized inverse. If the orthogonal projection method is used, the calculation formula is as follows:

[0058]

[0059] Solve to complete the output weight matrix This completes the training of the Extreme Learning Machine, allowing the output weights obtained from the training to predict new data samples.

[0060] The Online Quadratic Programming Controller (OLQR) algorithm uses the following linear regression model to fit the data flow. , The input feature vector for the online quadratic programming controller algorithm:

[0061]

[0062] in, , d is the dimension of the sample features. Let be the noise in the k-th training data sample. Assume the weight vector obtained at time point r is... Then the weight vector at time point r+1 is Two time points for the same input sample The residuals are recorded as follows:

[0063]

[0064]

[0065] in, This represents the error of the k-th sample at time r. The error of the k-th sample at time r+1 is represented by equation (8). Subtracting equation (7) from equation (8) gives the change in the residual of sample k:

[0066]

[0067] Performing the same operation on M input samples yields:

[0068]

[0069]

[0070] Therefore, the residual changes of the M samples can be written in the following compact form:

[0071]

[0072] The symbols and their meanings in formula (10) are shown in Table 2:

[0073] Table 2 Symbols and their meanings

[0074]

[0075] To facilitate explaining online learning from the perspective of feedback control, Treat it as a state vector and use symbols Replace it. Then use By substitution, the following discrete linear time-varying control system is obtained:

[0076]

[0077] in, and These represent the system input matrix and the control input, respectively. To represent the system's state vector... Decay to 0, design control law , construct as Figure 2 The closed-loop state feedback system shown adjusts the given control input. This allows the state to quickly stabilize to 0.

[0078] As r increases It is also constantly updated, thus obtaining a series of time-invariant control systems. If these systems are made to stabilize and converge in sequence, a control input sequence can be obtained. This makes these systems stable, and they have:

[0079]

[0080] Therefore, for a given r, the online learning problem can be transformed into the control problem corresponding to system (11). To solve for the optimal feedback control input using infinite-time LQR, a virtual time-invariant dynamic system is constructed as follows:

[0081]

[0082] in, Obviously, if If the row order is full rank, then the system has complete controllability and observability, which means that there exists an optimal control that can stabilize system (13). Therefore, the objective function of the optimal control of system (13) is:

[0083]

[0084] in, It is a state weighting matrix. It is a control weighting matrix; to simplify the problem, here... and All are set to identity matrices. The bias in the prediction was measured. Then the strength of the control input is constrained. These are weighting coefficients used to balance the two optimization objectives.

[0085] According to optimal control theory, the optimal solution to the quadratic optimization problem (14) can be obtained as follows:

[0086]

[0087] in, It is a matrix equation:

[0088]

[0089] The solution. Since the time complexity of solving equation (16) is very high, we consider using some linear algebra techniques to reduce the time complexity of solving equation (16). From:

[0090]

[0091] We can obtain:

[0092]

[0093] Then formula (16) can be rewritten as:

[0094]

[0095] Then we have:

[0096]

[0097] make Equation (20) can be written as:

[0098]

[0099] in, and for A symmetric matrix of dimension . From equation (21), we can obtain:

[0100]

[0101] because and If a matrix is ​​commutative, then there exists a unitary matrix. , , so that:

[0102]

[0103] in,

[0104]

[0105]

[0106] and They are respectively and Substituting equation (23) into equation (21) yields the eigenvalues:

[0107]

[0108] Equivalent to:

[0109]

[0110] because If the number is positive, solve for... get:

[0111]

[0112] Complete equation (16) Solve for it.

[0113] This invention provides an image classification method for Extreme Learning Machines (ELM) trained using the OLQR algorithm. The core idea is to use the OLQR algorithm to calculate the output weight matrix based on the feature mapping of traditional ELMs, thereby constructing an ELM image classification model based on the OLQR algorithm. Then, the model is trained using a large number of image samples containing labeled data. After that, the trained model can be used to predict and infer the labels of image samples with unknown labels, thereby achieving image classification.

[0114] This invention provides an image classification method for Extreme Learning Machines trained using the OLQR algorithm, the process of which is as follows: Figure 3 As shown, it includes the following steps:

[0115] Step 1: Collect a certain amount of image data and perform category labeling to obtain a dataset. Divide the dataset into a training set and a test set according to a certain ratio. Standardize the feature vectors of the images. The sample set is... , where x is the feature vector of the image, y is the label of the image, and N is the total number of samples.

[0116] Step 2: By converting the weight training process of the Extreme Learning Machine (ELM) model into an online learning task using the Online Quadratic Programming Controller (OLQR) algorithm, an Online Learning Extreme Learning Machine (OLQR-OS-ELM) image recognition model based on the Online Quadratic Programming Controller (OLQR) algorithm is established. The number of hidden layer nodes in the network structure of the Extreme Learning Machine is... As shown in the following formula:

[0117]

[0118] Among them, w i β is the input weight vector connecting the i-th hidden layer node and the input layer node. i The output weight matrix b is used to connect the i-th hidden layer node and the output layer node. i Let be the bias of the i-th hidden layer node. For activation function, This is the output matrix. The weight vector is w. i and bias b i For random generation, activation function Usually the sigmoid function .

[0119] Step 3: Using the training set generated in Step 1, train the online learning extreme learning machine image recognition model based on the OLQR algorithm using the online quadratic programming controller algorithm.

[0120] ELM randomly generates the weights of the hidden layers during the initialization process. and bias Then sample data The input is passed to the model and activated by the function. The hidden layer output matrix is ​​then obtained:

[0121]

[0122] At this point, the matrix can be... It can be viewed as a dataset after feature mapping, analogous to the original input dataset:

[0123]

[0124] The mapped dataset still contains N samples and N labels, but the feature dimension becomes L. Therefore, the weight training process of ELM can be transformed into an online learning task using the OLQR algorithm, resulting in an online learning extreme learning machine (OLQR-OS-ELM) image recognition model based on the OLQR algorithm. The training process of OLQR-OS-ELM is as follows:

[0125] Step 3.1: Construct an online learning framework using the OLQR algorithm to calculate the hidden layer output matrix of the Extreme Learning Machine. The objective function is designed as follows:

[0126]

[0127] in, Used to measure the magnitude of error. The regularization term is used to penalize the weight vector. Balance coefficient. Used to balance two optimization objectives: when When the value is small, the model's main optimization objective is the training error. In this case, the model will try to minimize the training error to improve model accuracy, but overfitting may occur. When... When the values ​​are large, the main optimization objective of the model becomes the weight vector to improve the model's generalization performance, but the model's accuracy will decrease relatively. This can be achieved by setting the matrix... and More refined weighted control can be applied to the model, and further, to simplify the computation process, [the following can be done]: and All are set as identity matrices.

[0128] The penalty applied to the weight vector described above is the regularization process. Regularization techniques suppress model overfitting by controlling the size of the weight vector. If the weight vector is too large, even small changes in the input variables will be excessively amplified by the weight vector, leading to model instability. If the weight vector can be controlled within a small range, it can ensure that the model maintains good performance when the input changes, that is, when the model is used with new samples, thereby improving the overall performance of the model.

[0129] Step 3.2: Initialize the online learning extreme learning machine (OLQR-OS-ELM) image recognition model based on the OLQR algorithm, and set the model parameters including: balance coefficients. Each time, import the number of training data sets M and the number of hidden layer nodes L of the model, and set the initial value of the OLQR algorithm weight matrix. Randomly generate the weights of the hidden layer and bias Determine the activation function (usually the sigmoid function).

[0130] Step 3.3: Import M training samples As the training dataset for this round, t can be a random number less than (NM). The training dataset for this round is input into the online learning extreme learning machine image recognition model based on the OLQR algorithm as follows: The hidden layer output matrix is ​​calculated. The hidden layer output matrix The input matrix x of the online quadratic programming controller algorithm ′ By using a linear regression model to fit the data flow, the training process of the Extreme Learning Machine is transformed into an online learning task.

[0131]

[0132] in, , Let be the connection weight from the i-th hidden layer node to the output node, and d be the dimension of the sample features. The noise is in the k-th training data sample.

[0133] Step 3.4: Input matrix x ′ The output layer weight coefficients are calculated using an online quadratic programming controller algorithm for the input. An online learning extreme learning machine image recognition model based on the OLQR algorithm was obtained, and the training of the image recognition model was completed.

[0134] To further improve the execution efficiency of image recognition models, this invention improves the OLQR algorithm by employing Singular Value Decomposition (SVD) on the matrix. Eigenvalue decomposition is performed to complete the output layer weight coefficients. The solution is to find the solution, where, This is the product of the input matrix and its transpose at a certain time, and also the product of the input sample and its transpose at a certain time. Given the input matrix at a certain time, the specific process is as follows:

[0135] Set loop variable The initial value is M;

[0136] Let the index at the current time be r=s, and obtain the current input matrix. and output matrix ;

[0137] calculate ;

[0138] Calculation using singular value decomposition get and ;

[0139] use and Calculated ;

[0140] according to Calculated ;

[0141] Calculate the current error ;

[0142] renew ;

[0143] renew ;

[0144] like Exit the loop and complete the output layer weight coefficients. The calculation; if Then let s increment M to continue the loop;

[0145] The output layer weight coefficients can be obtained by completing the above calculations. .

[0146] The algorithm flow for solving linear regression online using the OLQR algorithm is as follows:

[0147]

[0148] Based on this, the present invention modifies the training process of the online learning extreme learning machine image recognition model based on the OLQR algorithm, i.e., step 3.4 above, as follows:

[0149] Step 3.4.1: Set the loop variable s = M (maximum value is N);

[0150] Step 3.4.2, Calculation ,in ;

[0151] Step 3.4.3, Calculation ;

[0152] Step 3.4.4, Calculation ;

[0153] Steps 3, 4, and 5: Calculate using eigenvalue decomposition get and ;

[0154] Step 3.4.6: Calculate using formulas (24), (25), and (30) to obtain... ;

[0155] Step 3.4.7, Calculation ;

[0156] Step 3.4.8, Calculation ;

[0157] Step 3.4.9, Update ;

[0158] Step 3.4.10: Let r = r + M; if the training data has been completely traversed or s is greater than the maximum value, then break out of the loop and output the weight matrix. As the output weight matrix Otherwise, if s is not greater than the maximum value, then increment s by M and return to step 3.4.2 to continue training.

[0159] Step 4: In actual use, after performing the same standardization process as in Step 1 on the image to be classified, input it into the online learning extreme learning machine image recognition model based on the OLQR algorithm trained in Step 3 for computation and inference, and obtain the category output of the image to be classified, thus completing the image classification.

[0160] This invention provides an image classification method based on an Extreme Learning Machine trained using the OLQR algorithm. Experiments were conducted on multiple UCI (University of California, Irvine) datasets, which are standard datasets commonly used in machine learning. The comparison algorithms were traditional Extreme Learning Machine and Online Sequence Extreme Learning Machine (OS-ELM). Compared with these existing algorithms, this invention has the following advantages and positive effects:

[0161] This invention exhibits higher classification accuracy and better generalization performance, effectively suppressing model overfitting. The error convergence during training is faster than OS-ELM based on recursive least squares training, achieving a 3-5 times speed advantage, which becomes more pronounced with increasing dataset and sample dimensionality. Parameter sensitivity experiments demonstrate that this invention performs well on most hyperparameters (balance coefficient / regularization coefficient). Excellent performance can be achieved within a range of combinations of the number of training samples (M) imported each time, which greatly simplifies the parameter adjustment process and improves the efficiency of model optimization.

[0162] The online learning extreme learning machine image recognition model based on the OLQR algorithm provided by this invention applies optimal control theory to the training process of machine learning models and uses state feedback instead of output feedback to train the model, which is the first of its kind in the field of extreme learning machines.

[0163] In summary, the above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An image classification method for an extreme learning machine trained using the OLQR algorithm, characterized in that, Specifically, the following steps are included: Step 1: Collect image data and label the categories to obtain training and test sets. Standardize the feature vectors of the images. The sample set is... ,in, Let k be the feature vector of the k-th image. Let N be the label of the k-th image, and N be the total number of samples. Step 2: The weight training process of the Extreme Learning Machine (ELM) model is transformed into an online learning task using an online quadratic programming controller algorithm. An online learning ELM image recognition model based on this algorithm is established, as shown in the following equation: ， Among them, w i Let β be the input weight vector connecting the i-th hidden layer node and the input layer node, i=1,2,…,L. i The output weight matrix b is used to connect the i-th hidden layer node and the output layer node. i Let be the bias of the i-th hidden layer node. For activation function, For the output matrix, w represents the number of hidden layer nodes. i and b i All are randomly generated; Step 3: Using the training set generated in Step 1, train the online learning extreme learning machine image recognition model based on the OLQR algorithm using the online quadratic programming controller algorithm; Step 4: In actual use, after performing the same standardization process as in Step 1 on the image to be classified, input it into the online learning extreme learning machine image recognition model based on the OLQR algorithm trained in Step 3 for calculation and inference, and output the category of the image to be classified to complete the image classification. Step 3, which involves using an online quadratic programming controller algorithm to train an online learning extreme learning machine image recognition model based on the OLQR algorithm, includes: Step 3.1: Construct an online learning framework using an online quadratic programming controller algorithm to compute the hidden layer output matrix of the extreme learning machine. The objective function is designed as follows: ， in, Used to measure the magnitude of error. Used to penalize the weight vector. For balance coefficient, and All are weighted control matrices; Step 3.2: Initialize the online learning extreme learning machine image recognition model based on the online quadratic programming controller algorithm, and set the model parameters: balance coefficient. Each time, import the number of training data points M and the number of hidden layer nodes L of the model, and set the initial values ​​of the weight matrix. Randomly generate the weights of the hidden layer and bias Determine the activation function; Step 3.3: Import M training samples As the training dataset for this round, t is a random number less than (NM). This training dataset is input into an online learning extreme learning machine image recognition model based on an online quadratic programming controller algorithm. The hidden layer output matrix is ​​calculated. The hidden layer output matrix The input matrix x of the online quadratic programming controller algorithm ′ Then, a linear regression model is used to fit the data flow, transforming the training process of the Extreme Learning Machine into an online learning task: ， in, , Let be the connection weight from the i-th hidden layer node to the output node, and d be the dimension of the sample features. The noise in the k-th training data sample; Step 3.4: Input matrix x ′ The output layer weight coefficients are calculated using an online quadratic programming controller algorithm for the input. The training of an online learning extreme learning machine image recognition model based on an online quadratic programming controller algorithm was completed. The online quadratic programming controller algorithm employs singular value decomposition (SVD) on the matrix. Perform eigenvalue decomposition to complete the output layer weight coefficients The solution is to find, where, Let be the input matrix at a certain time. The process of inputting the product of a matrix and its transpose at a certain time includes: Step 5.1: Set the loop variable The initial value is M; Step 5.2: Let the current time index r = s, and obtain the current input matrix. and output matrix ; Step 5.3: Calculate the product of the input matrix and its transpose at time r. ; Step 5.4: Using the unitary matrix , Singular value decomposition is used to calculate and You can get and ,in, For matrix equations The solution, These are weighting coefficients; Step 5.5 and They are respectively and The eigenvalues, will and Substitution achievable , equivalent to ,calculate get ; Step 5.6, according to The feedback gain matrix is ​​calculated. ; Step 5.7: Calculate the current error. ; Step 5.8, Update and ; Step 5.9, if Exit the loop and complete the output layer weight coefficients. The calculation; if Then, increment s by M and execute step 5.2; after completing the above calculations, the output layer weight coefficients can be obtained. .

2. The image classification method for Extreme Learning Machines according to claim 1, characterized in that, The activation function sigmoid function .

3. The image classification method for Extreme Learning Machines according to claim 1, characterized in that, The and All are unit arrays.

Images