Fine image classification method based on deep transfer dictionary learning

Abstract

The application relates to a fine image classification method based on deep migration dictionary learning, which is characterized by comprising the following steps: constructing a deep dictionary learning model, learning a source field dictionary and a target field dictionary to extract complex nonlinear features of fine images, then solving a target function of an output layer in each layer of the deep dictionary learning model by using a gradient descent method to obtain an encoding dictionary and an encoding coefficient of the output layer, and finally obtaining classifier parameters through the obtained encoding dictionary and the encoding coefficient to construct a classification method model. The application solves the problem that source field fine images and target field fine images belong to different distributions, thereby improving the classification performance of the fine images.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of fine image classification, and particularly relates to a fine image classification method based on deep transfer dictionary learning. BACKGROUND

[0002] Deep dictionary learning algorithm can extract complex nonlinear features of fine images, and is widely used in fine image classification and other fields. However, the current deep dictionary learning algorithm belongs to the mode classification problem under the condition of independent and identically distributed, which is difficult to meet in the application of fine image classification. Generally, fine images are images reflecting the subtle differences of objects taken by different devices in different environments, and they usually have multiple different types of feature distribution, which makes the current deep dictionary learning method difficult to obtain satisfactory performance in fine image classification application. Therefore, the present application proposes a semi-supervised deep transfer dictionary learning model and applies it to a fine image classification system to improve the classification performance of fine images. SUMMARY

[0003] In order to achieve the above purpose, the fine image classification method based on deep transfer dictionary learning of the present application mainly includes the following steps:

[0004] Firstly, a deep dictionary learning model is constructed using a synthetic dictionary, and the gradient descent method is used to update the dictionary and the encoding coefficient of each layer;

[0005] Secondly, the source domain dictionary and the target domain dictionary are learned respectively by using the fine images with class labels in the source domain and the target domain, and the complex nonlinear features of the fine images in the source domain and the target domain are extracted respectively;

[0006] Thirdly, in each layer of the deep dictionary learning model, an adaptive maximum mean difference model is constructed using the row vectors of the atom and the encoding coefficient matrix to reduce the distribution difference between the source domain and the target domain, and the gradient descent method is used to solve the objective function of the output layer to obtain the encoding dictionary and the encoding coefficient of the output layer;

[0007] Fourthly, the target domain encoding coefficient is extracted from the encoding coefficient matrix of the output layer, and the classifier parameter is obtained by combining the class label matrix of the target domain sample;

[0008] Fifthly, the model of the classification method is constructed, the target domain dictionary of the output layer and the test sample are used to calculate the corresponding encoding coefficient, and then the model of the classification method is constructed by combining the classifier parameter;

[0009] Finally, the test sample is input into the model of the classification method, and the class label of the test sample is output.

[0010] As a further improvement of the above scheme, the third step is to solve the objective function of the output layer of the last layer in each layer of the deep dictionary learning model, and the output of each layer is the input of the next layer, until the output layer of the last layer is obtained.

[0011] As a further improvement of the above scheme, in the sixth step, the model of the classification method is the class corresponding to the maximum value of the product of the classifier parameters and the test sample encoding coefficient value.

[0012] The deep migration dictionary learning model is proposed for the deep dictionary learning scheme assuming that fine image classification belongs to a same distribution mode classification problem, and the source field fine image and the target field fine image belong to different distribution problems, and a semi-supervised deep migration dictionary learning model is designed by using part of the classification data in the target field to improve the classification performance of the fine image. BRIEF DESCRIPTION OF DRAWINGS

[0013] Figure 1 The flowchart of the present application is shown. DETAILED DESCRIPTION

[0014] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only 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.

[0015] Referring to Figure 1 The fine image classification method based on deep migration dictionary learning mainly includes the following steps:

[0016] First, a deep dictionary learning model is constructed by using a synthetic dictionary, and the dictionary and the encoding coefficient of each layer are updated by using the gradient descent method.

[0017] Secondly, the fine images with class labels in the source field and the target field are used to learn the source field dictionary and the target field dictionary respectively, and the complex nonlinear features of the fine images in the source field and the target field are extracted respectively.

[0018] Third, in each layer of the deep dictionary learning model, an adaptive maximum mean discrepancy model is constructed using the atomic and row vectors of the encoding coefficient matrix to reduce the distribution difference between the source domain and the target domain, and a gradient descent method is used to solve the objective function of the output layer to obtain the encoding dictionary and the encoding coefficient of the output layer.

[0019] Fourth, the target domain encoding coefficient is extracted from the encoding coefficient matrix of the output layer, and the classifier parameters are obtained by combining the class label matrix of the target domain sample.

[0020] Fifth, the model of the classification method is constructed, and the corresponding encoding coefficient of the test sample is calculated using the target domain dictionary of the output layer and the test sample, and then the model of the classification method is constructed by combining the classifier parameters.

[0021] Finally, the test sample is input into the model of the classification method, and the class label of the test sample is output.

[0022] The third step is to input the test sample into the model of the classification method, and the class label of the test sample is output.

[0023] Referring to Figure 1 In terms of algorithm, it is assumed that is the fine image sample with known class label in the source domain, is the sample with known class label in the target domain, is the dimension of the sample, and are the number of known class label samples in the source domain and the target domain, respectively. is the dictionary of the first layer, wherein and are the dictionaries of the first layer learned using the source domain sample and the target domain sample, respectively. is the encoding coefficient of the first layer, wherein and are the encoding coefficients of the first layer learned using the source domain sample and the target domain sample, respectively.

[0024] The deep transfer dictionary learning model is constructed as follows:

[0025] First step: input the source domain fine image with class label and into the deep transfer dictionary learning model.

[0026] Step Two: The layer depth migration dictionary is constructed as follows:

[0027] (1);

[0028] in These are parameters. The first term is the source domain data reconstruction term, the second term is the target domain data reconstruction term, and the third term is the adaptive maximum average difference term. Their construction method is as follows:

[0029] (2);

[0030] in, Indicates the first character in the source dictionary Row vectors of the coding coefficient matrix corresponding to each atom Indicates the first [item] in the target dictionary Row vectors of the coding coefficient matrix corresponding to each atom It is the transpose of the coding coefficient matrix. It is a matrix, and its elements are calculated as follows:

[0031] (3);

[0032] in and It represents the number of atoms in the source domain dictionary and the target domain, both of which are equal to the number of labeled samples in the corresponding domain.

[0033] Step 3: The layer (output layer) depth transfer dictionary is constructed as follows:

[0034] (4);

[0035] The first term is the source domain data reconstruction term, the second term is the target domain data reconstruction term, and the third term is the maximum average difference term, constructed as shown in formula (2). Additionally, and They are the first Source domain coding coefficients and target domain coding coefficients, and They are the first Source domain dictionary and target domain dictionary It is the first Layer coding coefficient matrix.

[0036] Reference Figure 1 Regarding algorithm updates, since each layer of the deep transfer dictionary learning model has a similar objective function, in order to reduce computational complexity, the 1st layer... The objective function of the layer is optimized as follows:

[0037] (5) ;

[0038] wherein, is the first layer encoding coefficient matrix, when , . is the first dictionary, is the matrix constructed by formula (2). Thus, the update of the algorithm is as follows:

[0039] First, the first layer dictionary is updated as follows:

[0040] (6) ;

[0041] Second, the second layer dictionary is updated as follows:

[0042] (7).

[0043] Referring to Figure 1 , in terms of the classification method, the output layer target domain encoding coefficient and the target domain sample class label matrix are used to construct the classifier parameter as follows:

[0044] (8) ;

[0045] For a test sample , its encoding coefficient can be calculated by the output layer dictionary using the following formula:

[0046] (9) ;

[0047] The class label of the test sample is predicted using the following formula:

[0048] (10).

[0049] ​To address the assumption that fine-grained image classification follows a similar distribution pattern in deep dictionary learning schemes, this invention proposes a deep transfer dictionary learning model. This model resolves the issue of source and target domain fine-grained images belonging to different distributions. A semi-supervised deep transfer dictionary learning model is designed using partial class label data from the target domain to improve the classification performance of fine-grained images. Furthermore, in each layer of the transfer dictionary learning model, an adaptive maximum mean difference model is constructed using transfer dictionary atoms and their corresponding encoding coefficient matrix row vectors to minimize the distributional differences between the source and target domains. This invention utilizes a deep dictionary to extract complex nonlinear features from fine-grained images and reduces domain differences by constructing a layer-by-layer adaptive maximum mean difference model, thereby improving the performance of deep dictionary learning algorithms in multi-domain fine-grained image classification systems.

[0050] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A refined image classification method based on deep transfer dictionary learning, characterized in that, The main steps include: First, a deep dictionary learning model is constructed using a synthetic dictionary, and gradient descent is used to update the dictionary and encoding coefficients of each layer. Secondly, by using the fine images with class labels in the source and target domains respectively, the source domain dictionary and the target domain dictionary are learned respectively, and the complex nonlinear features of the fine images in the source and target domains are extracted respectively. Third, in each layer of the deep dictionary learning model, an adaptive maximum mean difference model is constructed using the row vectors of the atoms and coding coefficient matrices to reduce the distribution difference between the source and target domains. Gradient descent is then used to solve the objective function of the output layer, thereby obtaining the coding dictionary and coding coefficients of the output layer. The deep transfer dictionary learning model is constructed as follows: The source domain image is transformed into a finely labeled image... and As input to the deep transfer dictionary learning model; construct the first Layer depth migration dictionary as shown in the formula As shown, where, These are parameters. The first term is the source domain data reconstruction term, the second term is the target domain data reconstruction term, and the third term is the adaptive maximum average difference term. Their construction method is as shown in the formula. As shown, where, Indicates the first character in the source dictionary Row vectors of the coding coefficient matrix corresponding to each atom Indicates the first [item] in the target dictionary Row vectors of the coding coefficient matrix corresponding to each atom It is the transpose of the coding coefficient matrix. It is a matrix, and its elements are calculated as follows: ,in, and The number of atoms in the source domain dictionary and the target domain are both equal to the number of labeled samples in the corresponding domain; construct the first... The layer, also known as the output layer, has a deep transfer dictionary as shown in the formula. As shown, the first term is the source domain data reconstruction term, the second term is the target domain data reconstruction term, and the third term is the maximum average difference term, which is constructed in a similar way to the adaptive maximum average difference term construction method in the second step. and They are the first Source domain coding coefficients and target domain coding coefficients, and They are the first Source domain dictionary and target domain dictionary It is the first Layer coding coefficient matrix; Fourth, extract the target domain coding coefficients from the coding coefficient matrix of the output layer, and combine them with the class label matrix of the target domain samples to obtain the classifier parameters; Fifth, construct the model of the classification method, use the target domain dictionary of the output layer and test samples to calculate its corresponding encoding coefficients, and then combine the classifier parameters to construct the model of the classification method. Finally, the test samples are input into the model of the classification method, thereby outputting the class labels of the test samples.

2. The refined image classification method based on deep transfer dictionary learning according to claim 1, characterized in that, In the third step, the output of each layer in the deep dictionary learning model becomes the input of the next layer, until the objective function of the output layer of the last layer is solved, thereby obtaining the encoding dictionary and encoding coefficients of the output layer.

3. The refined image classification method based on deep transfer dictionary learning according to claim 1, characterized in that, In the sixth step, the classification method model uses the class corresponding to the largest value of the product of the classifier parameters and the coding coefficients of the test sample as the class label of the test sample.

Images