A fine image classification method

Abstract

The application relates to a fine image classification method, which is characterized by the following steps: constructing a decomposition dictionary learning model, constructing a self-adaptive local order preserving constraint model, constructing a classifier model, constructing a target function, solving the target function and constructing a classification method model. The application can not only adaptively capture the local geometric structure information of atoms and the high-order rank information features of the neighborhood of each decomposition atom in the decomposition dictionary learning process, but also can measure the contribution of multi-view data to the representation of an object through the distance between atoms, enhance the discrimination performance of the decomposition coefficients and the decomposition dictionary. Furthermore, the decomposition coefficients and the class label matrix of the fine image are used to construct a classifier model, the classifier parameters are constrained by a norm, the robustness of the classifier learning model is enhanced, the time complexity of the algorithm is reduced, and the classification performance of the adaptive local order preserving constraint multi-view decomposition dictionary learning algorithm is improved.

Description

TECHNICAL FIELD

[0001] The application relates to a fine image classification method, in particular to a fine image classification method based on adaptive local order-preserving multi-view decomposition dictionary learning.

[0002] The application relates to a fine image classification method, in particular to a fine image classification method based on adaptive local order-preserving multi-view decomposition dictionary learning. BACKGROUND

[0003] The fine image contains view information of multiple different views, can represent object features from different angles, and the multi-view decomposition dictionary can be used for sparse decomposition of the multi-view fine image, can extract and select multi-view features with certain robustness, can learn a classifier with good classification performance, and is widely applied to multi-view fine image classification. However, the current multi-view decomposition dictionary learning model equally treats each view, ignores the contribution of different view fine images to representing objects and the local position information between different view fine images, and reduces the discrimination performance of the multi-view decomposition dictionary learning model. SUMMARY

[0004] The application provides a fine image classification method, and particularly relates to a fine image classification method based on adaptive local order-preserving multi-view decomposition dictionary learning.

[0005] The application provides a fine image classification method, and particularly relates to a fine image classification method based on adaptive local order-preserving multi-view decomposition dictionary learning.

[0006] In order to achieve the above object, the fine image classification method mainly comprises the following steps:

[0007] Firstly, a decomposition dictionary learning model is constructed, the decomposition dictionary learning model is constructed by inputting multi-view fine image data, the pixel value of the image is converted into an input vector based on a column vector, and the decomposition dictionary learning model is constructed by using different view data; a decomposition dictionary and a corresponding decomposition coefficient are learned for each view data.

[0008] Secondly, an adaptive local order-preserving constraint model is constructed, the one-to-one correspondence relationship between the decomposition atom and the profile learned in the decomposition dictionary learning model is used to construct the adaptive local order-preserving constraint model; in the decomposition dictionary learning process, the neighborhood relationship of the decomposition atom and the high-order ordering information of each decomposition atom neighborhood are maintained, so that the discrimination performance of the decomposition dictionary is enhanced.

[0009] Thirdly, a classifier model is constructed, the decomposition coefficient learned in the decomposition dictionary learning model and the class label matrix of the multi-view data are used to construct the classifier model, and the norm constraint is performed on the classifier parameter; a classifier parameter with certain robustness and discrimination is learned, so that the classification performance of the algorithm is improved.

[0010] Fourthly, a target function is constructed, based on the decomposed dictionary learning model, an adaptive local order preserving constraint model and a classifier model are added, so that the target function of the multi-view decomposed dictionary learning algorithm is constructed; so as to enhance the discriminant performance of the decomposed dictionary and the discriminator, and further improve the classification performance of the algorithm.

[0011] Fifthly, the target function is solved, the gradient descent method is adopted to solve the target function, and the multi-view decomposed dictionary and the decomposed coefficient are obtained.

[0012] Sixthly, the model of the classification method is constructed, and the data of the test sample is inputted, so that the class label of the test sample is outputted.

[0013] As a further improvement of the above method, in the fifth step of solving the target function, the output of each layer is the input of the next layer, until the target function of the output layer of the last layer is solved, so that the multi-view decomposed dictionary and the decomposed coefficient are obtained.

[0014] As a further improvement of the above method, in the sixth step, the model of the classification method is to decompose the test sample by using the classifier and the decomposed dictionary, and to determine the class label of the test sample according to the product of the decomposed coefficient and the classifier.

[0015] The present application can not only adaptively capture the local geometric structure information of atoms and the high-order rank information features of the neighborhood of each decomposed atom in the decomposed dictionary learning process, but also can measure the contribution of multi-view data to the representation of objects through the distance between atoms, enhance the discriminant performance of the decomposed coefficient and the decomposed dictionary. Moreover, the classifier model is constructed by using the decomposed coefficient and the class label matrix of the fine image, the robustness of the classifier learning model is enhanced and the time complexity of the algorithm is reduced by constraining the parameters of the classifier, and the classification performance of the adaptive local order preserving constraint multi-view decomposed dictionary learning algorithm is improved. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 The flowchart of the present application is shown. DETAILED DESCRIPTION

[0017] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0018] Reference Figure 1 The fine image classification method according to an embodiment of the present application mainly includes the following steps:

[0019] Firstly, a decomposition dictionary learning model is constructed, and the pixel values of the image are converted into input vectors based on column vectors by inputting multi-view fine image data to construct the decomposition dictionary learning model, and different view data is used to construct the decomposition dictionary learning model; a decomposition dictionary and a corresponding decomposition coefficient are learned for each view data.

[0020] Secondly, an adaptive local order preserving constraint model is constructed, and the one-to-one correspondence between the decomposition atoms learned in the decomposition dictionary learning model and the profiles is used to construct the adaptive local order preserving constraint model; the neighborhood relationship of the decomposition atoms and the high-order ordering information of each decomposition atom neighborhood are maintained during the decomposition dictionary learning process, thereby enhancing the discriminative performance of the decomposition dictionary.

[0021] Thirdly, a classifier model is constructed, and the decomposition coefficients learned in the decomposition dictionary learning model and the class label matrix of the multi-view data are used to construct the classifier model, and the norm constraint of the classifier parameters is performed; a classifier parameter with high robustness and discriminativeness is learned, thereby improving the classification performance of the algorithm.

[0022] Fourthly, a target function is constructed, and the adaptive local order preserving constraint model and the classifier model are added to the decomposition dictionary learning model, thereby constructing the target function of the multi-view decomposition dictionary learning algorithm; the discriminative performance of the decomposition dictionary and the discriminator is enhanced, thereby improving the classification performance of the algorithm.

[0023] Fifthly, the target function is solved, and the gradient descent method is used to solve the target function to obtain the multi-view decomposition dictionary and the decomposition coefficient.

[0024] Sixthly, a model of the classification method is constructed, and the data of the test sample is inputted, thereby outputting the class label of the test sample.

[0025] In the fifth step of solving the target function, the output of each layer is the input of the next layer, and the target function of the output layer of the last layer is solved, thereby obtaining the multi-view decomposition dictionary and the decomposition coefficient. In the sixth step, the model of the classification method is used to decompose the test sample by using the classifier and the decomposition dictionary, and the class label of the test sample is determined according to the product of the decomposition coefficient and the classifier.

[0026] Reference Figure 1 In terms of algorithm, it is assumed that is the view data of the fine image, wherein is a set of real numbers, is the view data set, is the dimension of the data, is the the number of view samples, . is the decomposition dictionary learned for each view, is the decomposition coefficient matrix corresponding to each decomposition dictionary.

[0027] First, the decomposition dictionary learning model for each view is constructed as follows:

[0028] (1);

[0029] where, denotes the constraint set of the decomposition dictionary, denotes the sparse factor, denotes the norm constraint.

[0030] Second, the adaptive local order preserving constraint model is constructed.

[0031] Suppose and are the neighboring decomposition atoms of decomposition atom , then their corresponding profiles , and can be expressed as , and . If decomposition atoms and are similar, then their corresponding profiles and are also similar. If , then , denotes the distance formula. Therefore, the adaptive local order preserving constraint model is constructed as follows:

[0032] (2);

[0033] where, is the total number of multi-view decomposition atoms, is the weight of decomposition atoms and , denotes the sequence number of the nearest neighbor of decomposition atom , and is a symmetric matrix and . Since profiles are row vectors of the decomposition coefficient matrix, when the distance between decomposition atoms is calculated using the Euler distance formula, the adaptive local order preserving constraint model can be converted to:

[0034] (3);

[0035] wherein, , is a decomposition coefficient matrix, is a weight matrix of decomposition atoms, is a symmetric matrix calculated by elements in .

[0036] Thirdly, constructing a classifier model.

[0037] (4);

[0038] wherein, wherein is a real number set, is the number of classes of multi-view data, is the number of multi-view samples, , is the class label vector of the th sample, and the non-zero position indicates the class label of the sample, is a classifier parameter, is a regularization parameter, denotes norm constraint.

[0039] Fourthly, in order to learn a multi-view decomposition dictionary with robustness, the objective function proposed by the present patent is as follows:

[0040] (5);

[0041] wherein, , , are all parameters.

[0042] Referring to Figure 1 , the decomposition dictionary is updated first in terms of algorithm update. Assuming that other variables in the objective function are constants, the objective function can be converted to:

[0043] (6);

[0044] By using Cholesky decomposition and singular value decomposition, the following can be obtained:

[0045] (7);

[0046] wherein, is obtained by Cholesky decomposition of , and are obtained by singular value decomposition of , is the identity matrix, is a parameter.

[0047] Fifth step, update the classifier .

[0048] Assuming that other variables in the objective function are constants, the objective function is converted to

[0049] (8) ;

[0050] The direct derivative of the above formula can obtain the solution of the classifier as

[0051] (9) ;

[0052] where, is a diagonal matrix, , is the first row of the matrix .

[0053] Finally, update the decomposition coefficient .

[0054] Assuming that other variables in the objective function are constants, the objective function is converted to

[0055] (10) ;

[0056] Assuming , the decomposition coefficient matrix can be constructed by , and the above formula is converted to

[0057] (11) ;

[0058] The direct derivative of the above formula can obtain the solution of the decomposition coefficient as

[0059] (12) ;

[0060] where, is the identity matrix.

[0061] Referring to Figure 1 , in terms of classification method, the classifier and the decomposition dictionary are used, and the classification method for multi-view data is as follows:

[0062] (13) ;

[0063] where, indicates multi-view data Class label.

[0064] The application can not only adaptively capture the local geometric structure information of atoms and the high-order rank information features of each decomposition atom neighborhood in the decomposition dictionary learning process, but also can measure the contribution of multi-view data to the representation of objects through the distance between atoms, enhance the discriminant performance of decomposition coefficients and decomposition dictionary. Moreover, the class label matrix of decomposition coefficients and fine images is used to construct a classifier model, and the classifier parameters are constrained by norm, which enhances the robustness of the classifier learning model, reduces the time complexity of the algorithm, and improves the classification performance of the adaptive local order-preserving constraint multi-view decomposition dictionary learning algorithm.

[0065] Finally, it should be noted that: the above only for the preferred embodiments of the application, and not for limiting the application, although the application has been described in detail with reference to the foregoing embodiments, for those skilled in the art, it still can be modified to the technical solutions recorded in the foregoing embodiments, or equivalent replacement of some technical features. Any modification, equivalent replacement, improvement, etc. within the spirit and principles of the application shall be included in the protection scope of the application.

Claims

1. A refined image classification method, characterized in that, The main steps include: First, a decomposition dictionary learning model is constructed by inputting multi-view fine image data. The pixel values ​​of the image are converted into column vector-based input vectors, and a decomposition dictionary learning model is constructed using data from different viewpoints, so as to learn a decomposition dictionary and corresponding decomposition coefficients for each viewpoint. Secondly, an adaptive local order-preserving constraint model is constructed. This model utilizes the one-to-one correspondence between decomposition atoms and profiles learned in the decomposition dictionary learning model, where profiles are row vectors of the decomposition coefficient matrix. The steps for constructing the adaptive local order-preserving constraint model are as follows: Assume... and It is the decomposition of atoms The nearest neighbor decomposition atoms, then their corresponding profiles , and It can be represented as , and , For fine image data; if decomposed into atoms and If they are similar, then their corresponding profiles and They are also similar; if ,but , The distance formula is used; therefore, the adaptive local order-preserving constraint model is constructed as follows: ,in It is the total number of atoms decomposed from multiple perspectives. To decompose atoms and The weight, Indicates the decomposition of atoms of The nearest neighbor sequence number, It is a symmetric matrix and Since profiles are row vectors of the decomposition coefficient matrix, when calculating the distances between decomposed atoms using the Euler distance formula, the adaptive local order-preserving constraint model can be transformed into... ,in, , The coefficient matrix is ​​decomposed. To decompose the weight matrix of the atoms, To utilize A symmetric matrix calculated from the elements in the matrix; Third, construct a classifier model by using the decomposition coefficients learned in the decomposition dictionary learning model and the class label matrix of multi-view data, and impose norm constraints on the classifier parameters. Fourth, construct the objective function. Based on the dictionary decomposition learning model, add an adaptive local order-preserving constraint model and a classifier model to construct the objective function of the multi-view dictionary decomposition learning algorithm. Fifth, the objective function is solved using the gradient descent method to obtain the multi-view decomposition dictionary and decomposition coefficients; Sixth, construct a model for the classification method, input the data of the test samples, and output the class labels of the test samples.

2. The refined image classification method according to claim 1, characterized in that, In the fifth step of solving the objective function, the output of each layer becomes the input of the next layer, until the objective function of the output layer of the last layer is solved, thereby obtaining the multi-view decomposition dictionary and decomposition coefficients.

3. The refined image classification method according to claim 1, characterized in that, In the sixth step, the classification method model uses a classifier and a decomposition dictionary to decompose the test samples, and determines the class label of the test samples based on the product of the decomposition coefficients and the classifier.

Images