An image recognition method based on limit learning machine of cyclic iteration least square

Abstract

The application discloses an image recognition method based on a limit learning machine of a cyclic iteration least square, randomly extracts image features by using convolution operation, and obtains an output weight matrix of hidden layer nodes by using the cyclic iteration least square, so as to realize training of the limit learning machine model, improve the generalization performance of the algorithm under the condition of reducing the heavy calculation burden and improving the calculation efficiency, and meet the needs of image recognition to a certain extent, and provide a new thought and a new way for more accurate image recognition.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of deep learning, and particularly relates to an image recognition method based on a limit learning machine of cyclic iteration least square. BACKGROUND

[0002] Image recognition is one of key technologies in the information age, and tasks that cannot be realized by traditional sensor technology can be completed by using image recognition technology. Meanwhile, with the aid of image recognition and fusion of multi-source information, monitoring effect can be more accurate. With the continuous development of computer technology, computer computing capacity is further improved, and information processing capacity is greatly improved. Relying on intelligent image recognition algorithm, recognition accuracy can be improved, and time required for recognition can be reduced, which will play an important role in the field of artificial intelligence in the future. In the past few years, convolutional neural network has become a mainstream machine learning method for various tasks in the field of image recognition. However, the traditional convolutional neural network adopts a back propagation method based on gradient descent for training, and along with the increase of the number of convolutional neural network parameters, network training time becomes longer. Therefore, how to shorten the training time of neural network while ensuring the accuracy of image recognition is of great significance. SUMMARY

[0003] Therefore, the application provides an image recognition method based on a limit learning machine of cyclic iteration least square, which realizes accurate and rapid recognition of images.

[0004] The image recognition method based on the limit learning machine of cyclic iteration least square provided by the application comprises the following steps:

[0005] Step 1: sequentially performing convolution, nonlinear mapping and pooling processing on collected image samples to extract features, and constructing a training sample set from the features and labels of all image samples; the image samples are represented as {X i ,Y i |X i ∈R p×q×3 ,Y i ∈R,i=1,2,...,N} wherein X i represents the i-th image as input of the model, Y i represents the label corresponding to the i-th image as output of the model, p is the height of each image, q is the width of each image, 3 represents three color channels of RGB, and N is the total number of samples; the features are represented as wherein n, b, m and s all represent the size of the convolution kernel.

[0006] Step 2, establishing an extreme learning machine model, using the training sample set to solve the output weight matrix of the hidden layer node by using a cyclic iterative least square, completing the training of the extreme learning machine model, the extreme learning machine model is shown in formula (5):

[0007] Y 1×N = β 1×k ·J k×N (5)

[0008]

[0009] Wherein, Y 1×N ∈R 1×N represents the output matrix of the extreme learning machine model, N is the number of samples, k is the number of hidden layer nodes, J k×N is the output matrix of the hidden layer node, And S (p-n-m+1+q-b-s+1)×N ∈R (p-n-m+1+q-b-s+1)×N is the input data of the extreme learning machine model, (p-n-m+1+q-b-s+1) is the feature dimension of each input sample vector, α k×(p-n-m+1+q-b-s+1) is the input weight matrix between the input to the hidden layer node, λ k×N is the bias matrix of the hidden layer node, is a selectable nonlinear activation function; α k×(p-n-m+1+q-b-s+1) and λ k×N are randomly generated and remain unchanged after generation;

[0010] Step 3, in actual use, the features of the image to be identified are extracted in the manner of step 1, and the features are input into the extreme learning machine model trained in step 2 to obtain the type of the image.

[0011] Further, the step 2 of using the training sample set to solve the output weight matrix of the hidden layer node by using a cyclic iterative least square to complete the training of the extreme learning machine model comprises the following steps:

[0012] Step 1.1, the inverse matrix is calculated by formula (6) The value of the element in the ith row and jth column in the tth iteration:

[0013]

[0014] Wherein, U N×k represents the inverse matrix of the hidden layer node output matrix J k×N , and satisfies J k×N *U N×k =I k×k ; I k×k is a unit matrix; is a vector and a vector Perform inner product operations; For vectors The square of the 2-norm; the maximum value of the number of iterations t is T;

[0015] Step 1.2: Calculate the residual at the t-th iteration using formula (7).

[0016]

[0017] Among them, residual The initial value is

[0018] Step 1.3: Increment i by 1. When i ≤ N, execute step 1.1; when i > N, execute step 1.4.

[0019] Step 1.4: Increment j by 1. If j ≤ k, execute step 1.1; if j > k, execute step 1.5.

[0020] Step 1.5: Calculate the t-th iteration error of the inverse matrix using formula (8):

[0021]

[0022] Among them, E k×k The iteration error of the inverse matrix is ​​initially set to E0. k×k =I k×k ;

[0023] Step 1.6: Increment t by 1. When t ≤ T, execute step 1.1; when t > T, execute step 1.7.

[0024] Step 1.7: Calculate and output the output weight matrix of the hidden layer nodes using formula (9). End this process:

[0025]

[0026] Further, in step 1, the acquired image samples are sequentially subjected to convolution, nonlinear mapping, and pooling to extract features. A training sample set is constructed from the features and labels of all image samples, specifically including the following steps:

[0027] Step 1.1: Obtain the i-th image X using formula (1). i Image based on average pixels Image X i Let X be a triple array. i =[X 1i ,X 2i ,X 3i ], where X1i Represents image X i The components of the R channel, X 2i Represents image X i The components of the G channel, X 3i Represents image X i The components of the B channel:

[0028]

[0029] Step 1.2: Apply convolution, nonlinear mapping, and pooling to the image obtained in Step 1.1 using formulas (2), (3), and (4) in sequence:

[0030]

[0031] H i (p-n+1)×(q-b+1) =φ(F i (p-n+1)×(q-b+1) (3)

[0032]

[0033] Among them, F i (p-n+1)×(q-b+1) ∈R (p-n+1)×(q-b+1) W is the output matrix of the convolutional layer. n×b For a two-dimensional convolution kernel, W n×b This is to generate randomly and keep it unchanged after generation; * indicates a convolution operation; The input data for the convolutional layer; H i (p -n+1)×(q-b+1) The output feature mapping matrix is ​​φ; φ is an optional nonlinear activation function. P is the output matrix of the pooling layer. m×s It is a two-dimensional convolution kernel, in which all elements are 1, and satisfy the conditions p > n > m, pnm > 0 and q > b > s, qbs > 0;

[0034] Step 1.3: Convert the output matrix of the pooling layer. Convert to feature vector

[0035] Beneficial effects:

[0036] This invention utilizes convolution operations to randomly extract image features and uses iterative least squares to obtain the output weight matrix of hidden layer nodes, thereby enabling the training of the Extreme Learning Machine model. This can improve the generalization performance of the algorithm while reducing excessive computational burden and increasing computational efficiency, and can meet the needs of image recognition to a certain extent. At the same time, it provides new ideas and approaches for more accurate image recognition. Detailed Implementation

[0037] The following examples illustrate the invention in detail.

[0038] This invention provides an image recognition method based on iterative least squares extreme learning machine. The basic idea is to establish an image dataset, preprocess the dataset, label images of different types to form training samples, and denot these samples as {X}. i ,Y i |X i ∈R p×q×3 ,Y i ∈R,i=1,2,...,N}, where X i Y represents the i-th image as the input to the model. i The output of the model is p, where p is the height of each image, q is the width of each image, 3 represents the three color channels RGB of the image, and N is the total number of samples. Image features are randomly extracted using convolution operations. The dataset formed by the features of the training samples and their corresponding labels is input into the Extreme Learning Machine model to establish an image recognition model. On this basis, the output weight matrix of the hidden layer nodes is obtained by iterative least squares, thereby completing the training of the Extreme Learning Machine model and realizing rapid image recognition.

[0039] This invention provides an image recognition method based on a cyclic iterative least squares extreme learning machine, including the construction of a feature matrix of image data training samples, and the construction and training of an extreme learning machine model based on the calculation of output weights using cyclic iterative least squares. The specific steps are as follows:

[0040] Step 1: Collect image data and corresponding labels to form training samples. The training samples are represented as {X}. i ,Y i |X i ∈R p ×q×3 ,Y i ∈R,i=1,2,...,N}, where X i Y represents the i-th image as the input to the model. i The label corresponding to the i-th image is the output of the model, p is the height of each image, q is the width of each image, 3 represents the three color channels RGB, and N is the total number of samples.

[0041] Step 2: Use formula (1) to obtain the i-th image X. i Image based on average pixels Image X i Let X be a triple array. i =[X 1i ,X 2i ,X 3i ], where X1i Represents image X i The components of the R channel, X 2i Represents image X i The components of the G channel, X 3i Represents image X i The components of the B channel:

[0042]

[0043] Step 3: Apply convolution, nonlinear mapping, and pooling to the image obtained in Step 2 using formulas (2), (3), and (4) in sequence:

[0044]

[0045] H i (p-n+1)×(q-b+1) =φ(F i (p-n+1)×(q-b+1) (3)

[0046]

[0047] Among them, F i (p-n+1)×(q-b+1) ∈R (p-n+1)×(q-b+1) W is the output matrix of the convolutional layer. n×b For a two-dimensional convolution kernel, W n×b To ensure random generation and immutability after generation, n and b represent the kernel size; * indicates convolution operation; The input data for the convolutional layer; H i (p-n+1)×(q-b+1) The output feature mapping matrix is ​​φ; φ is an optional nonlinear activation function. P is the output matrix of the pooling layer. m×s The kernel is a two-dimensional convolution kernel, where m and s represent the size of the kernel. All elements in the kernel are 1 and satisfy the conditions p > n > m, pnm > 0 and q > b > s, qbs > 0.

[0048] Step 4: Output matrix of pooling layer Convert to feature vector Repeat steps 1 to 3 to form a feature matrix containing N image data training samples. A training sample set is established, consisting of the features and corresponding labels of the image data training samples. The training sample set includes two parts: a training sample set and a test sample set. The training sample set is used to train the model, and the test sample set is used to test the trained model.

[0049] Step 5: Establish an Extreme Learning Machine (ELM) model and train it using the training sample set established in Step 4. The ELM model is shown in Equation (5):

[0050] Y 1×N =β 1×k ·J k×N (5)

[0051]

[0052] Among them, Y 1×N ∈R 1×N Let J represent the output matrix of the Extreme Learning Machine model, where N is the number of samples, k is the number of hidden layer nodes, and J is the number of hidden layer nodes. k×N S is the output matrix of the hidden layer nodes. (p-n-m+1+q-b-s+1)×N ∈R (p-n-m+1+q-b-s+1)×N For the input data of the Extreme Learning Machine model, (pn-m+1+qb-s+1) is the feature dimension of each input sample vector, and α is the input data. k×(p-n-m+1+q-b-s+1) Let λ be the input weight matrix between the hidden layer nodes. k×N Let be the bias matrix of the hidden layer nodes. α is a selectable nonlinear activation function. k×(p-n-m+1+q-b-s+1) and λ k×N All are randomly generated and remain unchanged after generation.

[0053] Step 6: Train the Extreme Learning Machine model using a training sample set consisting of the feature matrix and corresponding labels of the image data training samples. This invention utilizes iterative least squares to solve the output weight matrix of the hidden layer nodes, thereby completing the training of the Extreme Learning Machine model.

[0054] Step 7: In practical use, the image type can be obtained by inputting the features of the image to be identified into the Extreme Learning Machine model trained in Step 6.

[0055] Specifically, step 6 above uses iterative least squares to solve for the output weight matrix of the hidden layer nodes to complete the training of the extreme learning machine model, which includes the following steps:

[0056] Define U N×k J represents the output matrix of the hidden layer nodes. k×N The inverse matrix of J, and satisfying J k×N *U N×k =I k×k , among which, I k ×k It is the identity matrix. E k×k The iteration error of the inverse matrix is ​​initially set to E0. k×k =I k×k ; The remainder is the initial value. Set the number of iterations T, the iteration number t, and let t = 1.

[0057] For example, if k = 4 and N = 4, then the identity matrix and inverse matrix can be represented as:

[0058]

[0059] Step 1.1: Calculate the inverse matrix using formula (6). The value of the element in the i-th row and j-th column at the t-th iteration:

[0060]

[0061] in, For vectors sum vector Perform inner product operations. For vectors The square of the 2-norm.

[0062] Step 1.2: Calculate the residual at the t-th iteration using formula (7):

[0063]

[0064] Step 1.3: Increment i by 1. When i ≤ N, execute step 1.1; when i > N, execute step 1.4.

[0065] Step 1.4: Increment j by 1. If j ≤ k, execute step 1.1; if j > k, execute step 1.5.

[0066] Step 1.5: Calculate the t-th iteration error of the inverse matrix using formula (8):

[0067]

[0068] Step 1.6: Increment t by 1. When t ≤ T, execute step 1.1; when t > T, execute step 1.7.

[0069] Step 1.7: Calculate and output the output weight matrix of the hidden layer nodes using formula (9). End this process.

[0070]

[0071] In the above process, the output weight matrix of the hidden layer nodes is solved by a cyclic iterative least squares algorithm, which makes the model easier to converge, thereby improving the learning speed of the extreme learning machine model and shortening the training time of the model.

[0072] 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 recognition method based on a cyclic iterative least squares extreme learning machine, characterized in that, Includes the following steps: Step 1: Perform convolution, nonlinear mapping, and pooling on the acquired image samples sequentially to extract features. Construct a training sample set from the features and labels of all image samples; the image samples are represented as {X}. i ,Y i |X i ∈R p×q×3 ,Y i ∈R,i=1,2,...,N}, where X i Let Y represent the i-th image as the input to the model. i Let p represent the label corresponding to the i-th image as the model output, p be the height of each image, q be the width of each image, 3 represent the three color channels (RGB), and N be the total number of samples; the feature is represented as... Where n, b, m, and s all represent the size of the convolution kernel; Step 2: Establish an Extreme Learning Machine (ELM) model. Using the training sample set, solve for the output weight matrix of the hidden layer nodes using iterative least squares to complete the training of the ELM model. The ELM model is shown in Equation (5): Y 1×N <β 1×k ·J k×N (5) Among them, Y 1×N ∈R 1×N Let J represent the output matrix of the Extreme Learning Machine model, where N is the number of samples, k is the number of hidden layer nodes, and J is the number of hidden layer nodes. k×N Output matrix for hidden layer nodes. And satisfy S (p-n-m+1+q-b-s+1)×N ∈R (p-n-m+1+q-b-s+1)×N For the input data of the Extreme Learning Machine model, (pn-m+1+qb-s+1) is the feature dimension of each input sample vector, and α is the input data. k×(p-n-m+1+q-b-s+1) Let λ be the input weight matrix between the hidden layer nodes. k×N Let be the bias matrix of the hidden layer nodes. α is a selectable nonlinear activation function. k×(p-n-m+1+q-b-s+1) and λ k×N All are randomly generated and remain unchanged after generation; Step 3: In practical use, the features of the image to be identified are extracted using the method in Step 1. The features are then input into the Extreme Learning Machine model trained in Step 2 to obtain the image type. Step 2 uses the training sample set to solve the output weight matrix of the hidden layer nodes using iterative least squares to complete the training of the extreme learning machine model, including the following steps: Step 1.1: Calculate the inverse matrix using formula (6). The value of the element in the i-th row and j-th column at the t-th iteration: Among them, U N×k J represents the output matrix of the hidden layer nodes. k×N The inverse matrix of J, and satisfying J k×N *U N×k =I k×k ;I k×k It is the identity matrix; For vectors sum vector Perform inner product operations; For vector J i k×1 The square of the 2-norm; the maximum value of the number of iterations t is T; Step 1.2: Calculate the residual at the t-th iteration using formula (7). : Among them, residual The initial value is Step 1.3: Increment i by 1. When i ≤ N, execute step 1.1; when i > N, execute step 1.

4. Step 1.4: Increment j by 1. If j ≤ k, execute step 1.1; if j > k, execute step 1.

5. Step 1.5: Calculate the t-th iteration error of the inverse matrix using formula (8): Among them, E k×k The iteration error of the inverse matrix is ​​initially set to E0. k×k =I k×k ; Step 1.6: Increment t by 1. If t ≤ T, execute step 1.1; if t > T, execute step 1.

7. Step 1.7: Calculate and output the output weight matrix of the hidden layer nodes using formula (9). End this process:

2. The image recognition method according to claim 1, characterized in that, In step 1, the acquired image samples are sequentially processed by convolution, nonlinear mapping, and pooling to extract features. A training sample set is constructed from the features and labels of all image samples, specifically including the following steps: Step 1.1: Obtain the i-th image X using formula (1). i Image based on average pixels Image X i Let X be a triple array. i =[X 1i ,X 2i ,X 3i ], where X 1i Represents image X i The components of the R channel, X 2i Represents image X i The components of the G channel, X 3i Represents image X i The components of the B channel: Step 1.2: Apply convolution, nonlinear mapping, and pooling to the image obtained in Step 1.1 using formulas (2), (3), and (4) in sequence: Among them, F i (p-n+1)×(q-b+1) ∈R (p-n+1)×(q-b+1) W is the output matrix of the convolutional layer. n×b For a two-dimensional convolution kernel, W n×b This is to generate randomly and keep it unchanged after generation; * indicates a convolution operation; The input data for the convolutional layer; H i (p -n+1)×(q-b+1) The output feature mapping matrix is ​​φ; φ is an optional nonlinear activation function. P is the output matrix of the pooling layer. m×s It is a two-dimensional convolution kernel, in which all elements are 1, and satisfy the conditions p > n > m, pnm > 0 and q > b > s, qbs > 0; Step 1.3: Convert the output matrix of the pooling layer. Convert to feature vector

Images