Width learning image classification method and device based on mixed norm regularization constraint

Abstract

The present invention discloses a width learning image classification method and device based on mixed norm regularization constraint, which relates to the field of image processing technology, wherein the method comprises: obtaining image data to be classified, dividing the preprocessed image data into training samples and test samples, and obtaining class labels corresponding to the training samples; inputting the training samples into a width learning network for processing to generate width conversion features; according to the width conversion features and the class labels, using an ε dragging method and a mixed norm regularization constraint method, constructing a discriminant width learning network classification optimization objective function based on mixed norm regularization constraint; using a two-step iterative method to optimize and solve the objective function to achieve image classification; the present invention takes a width learning network as the main body, requires fewer parameters, has a simpler network structure, enhances feature extraction capabilities, and can achieve performance similar to that of a deep neural network in less time.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and more particularly to a width learning image classification method and device based on mixed norm regularization constraints. Background Art

[0002] At present, as one of the main problems in the field of artificial intelligence, image classification tasks focus on identifying the categories of certain images to solve some specific problems, such as face recognition, image retrieval, target detection, etc., and are an important part of the field of computer vision. With the continuous development of science and technology, human demand for information continues to increase, and the amount of information contained in images is huge. How to process the information contained in images is also one of the research issues in the field of image classification.

[0003] However, the Broad Learning System (BLS) commonly used in image classification is different from deep neural networks. BLS requires fewer parameters and has a simpler network structure. It can achieve similar performance to deep neural networks in less time. However, the above algorithm still has the following shortcomings: (1) The Broad Learning System (BLS) uses strict binary labels for class labels, which severely limits the flexibility of the model; (2) Although BLS has strong feature extraction capabilities and can obtain better features through sparse autoencoders, there is inevitably node redundancy in the wide network framework, which affects the performance of the model.

[0004] Therefore, how to provide an image classification method that can solve the above problems is an issue that those skilled in the art need to solve urgently. Summary of the invention

[0005] In view of this, the present invention provides a width learning image classification method and device based on mixed norm regularization constraints, which takes a width learning network as the main body, requires fewer parameters, has a simpler network structure, enhances feature extraction capabilities, and can achieve performance similar to that of a deep neural network in less time.

[0006] In order to achieve the above object, the present invention adopts the following technical solution:

[0007] A width learning image classification method based on mixed norm regularization constraint comprises the following steps:

[0008] Acquire image data to be classified, preprocess the image data, divide the preprocessed image data into training samples and test samples, and obtain class labels corresponding to the training samples;

[0009] Constructing a width learning network, inputting the training samples into the width learning network for processing, and generating width conversion features;

[0010] According to the width conversion feature and the class label, using the ε drag method and the mixed norm regularization constraint method, a discriminant width learning network classification optimization objective function based on the mixed norm regularization constraint is constructed;

[0011] A two-step iterative method is used to optimize and solve the objective function to achieve image classification.

[0012] Preferably, the specific expression of the classification optimization objective function of the discriminant width learning network based on the mixed norm regularization constraint is:

[0013]

[0014] Where W represents the projection matrix, A represents the training sample width conversion feature matrix, λ1 and λ2 represent regularization parameters, B represents the drag direction of ε, a constant matrix composed of +1 and -1, ⊙ represents the Hadamard product of the matrix; M is a constant matrix representing the distance of label dragging, ||·|| F represents the Frobenius norm of the matrix, ||·|| 2,1 Represents the matrix Norm.

[0015] Preferably, the specific process of solving the objective function using the two-step iterative method includes:

[0016] (1) Solve the projection matrix W:

[0017] make Taking the partial derivative of W and setting it equal to 0, we get:

[0018]

[0019] In the formula, R = Y + B⊙M, D is the diagonal element D ii =1 / ‖W i ‖2 is a diagonal matrix;

[0020] (2) Optimize the label drag distance matrix M:

[0021] Let E = AW-Y, the optimization problem is simplified to Decompose the optimization problem into sub-problems to be solved element by element, and obtain M ij =max(B ij E ij ,0), that is, M=max(B⊙E,0);

[0022] (3) Iterate the above (1) and (2) to obtain the optimal solution of the label drag distance matrix M and the projection matrix W.

[0023] Preferably, the termination condition of the iterative optimization solution is:

[0024] (obj (t-1) -obj (t) )≤ε0

[0025] Where, obj (t-1) 、obj (t) They represent the objective function values ​​of the t-1th and tth iterations respectively, and ε0 is the set value.

[0026] Preferably, the specific steps of obtaining the width conversion feature are:

[0027] (i) Randomly generate weight W ei With bias β ei , using the linear transformation mapping function φ i (·) Generate N from input data X g The i-th group of randomly mapped features is:

[0028]

[0029] Combining all the mapping features, we get:

[0030] (ii) Using the mapping features obtained in (i) Generate l groups of enhanced features, each group of enhanced features generally contains an enhanced node. The network has only three layers, and the input layer is followed by the enhanced layer. The jth group of enhanced features can be expressed as:

[0031]

[0032] Among them, W hj and β hj It is also randomly generated by the network. j =tanh(·) represents the jth activation function, and all enhanced features are combined to obtain:

[0033] (iii) Combine all the mapping features and enhanced features to obtain the final width conversion feature:

[0034]

[0035] Preferably, the steps of image data preprocessing are as follows:

[0036] Unifying the sizes of the training samples and the test samples;

[0037] Divide N images as training samples and the remaining images as test samples; divide the image database four times to obtain four groups of training samples and test samples;

[0038] The training samples and the test samples are standardized.

[0039] Furthermore, the present invention also provides a classification device using any of the above-mentioned width learning image classification methods based on mixed norm regularization constraints, comprising:

[0040] An acquisition module is used to acquire image data to be classified, preprocess the image data, divide the preprocessed image data into training samples and test samples, and acquire class labels corresponding to the training samples;

[0041] A processing module, used for constructing a width learning network, inputting the training samples into the width learning network for processing, and generating width conversion features;

[0042] A calculation module, used to construct a classification optimization objective function of a discriminant width learning network based on a mixed norm regularization constraint according to the width conversion feature and the class label by using an ε drag method and a mixed norm regularization constraint method;

[0043] The solution module is used to solve the objective function by two-step iteration to achieve image classification.

[0044] It can be seen from the above technical solution that, compared with the prior art, the present invention discloses a method, device and storage medium for image classification based on mixed norm regularization constraints, which has the following features:

[0045] Beneficial effects:

[0046] (1) Based on the wide learning network, it requires fewer parameters, has a simpler network structure, and enhances feature extraction capabilities, and can achieve performance similar to that of a deep neural network in less time.

[0047] (2) The ε-drag method adopted in the present invention can, on the one hand, expand the distance between different categories and better fit the regression task. On the other hand, the use of the relaxed regression target strategy can also greatly improve the flexibility and discrimination of the model.

[0048] (3) The hybrid norm regularization constraint adopted by the present invention combines the norm with the norm to regularize the projection matrix trained based on the width learning network model: under the norm constraint, the projection matrix is ​​more compact and can be equipped with a row sparse structure. Redundant features can be removed during the training process. By applying the norm at the same time, the model can be further constrained to prevent overfitting of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0050] Figure 1 The overall flow chart of the width learning image classification method based on mixed norm regularization constraint provided by the present invention;

[0051] Figure 2 A block diagram of the structural principle of a width learning image classification device based on a mixed norm regularization constraint provided by the present invention;

[0052] Figure 3a Some images from the ExtendedYaleB database;

[0053] Figure 3b Some images of the AR database;

[0054] Figure 3c Some images of the LFW database. DETAILED DESCRIPTION

[0055] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0056] See attached Figure 1 As shown, the embodiment of the present invention discloses a width learning image classification method based on mixed norm regularization constraint, comprising the following steps:

[0057] Obtain image data to be classified, preprocess the image data, divide the preprocessed image data into training samples and test samples, and obtain class labels corresponding to the training samples;

[0058] Construct a width learning network, input the training samples into the width learning network for processing, generate width conversion features to enhance feature extraction capabilities;

[0059] According to the width conversion features and class labels, the ε drag method and the mixed norm regularization constraint method are used to construct the classification optimization objective function of the discriminant width learning network based on the mixed norm regularization constraint;

[0060] A two-step iterative method is used to optimize and solve the objective function to achieve image classification.

[0061] In a specific embodiment, the steps of image data preprocessing are as follows:

[0062] Unify the sizes of training samples and test samples;

[0063] Divide N images as training samples and the remaining images as test samples; divide the image database four times to obtain four groups of training samples and test samples;

[0064] The training samples and test samples are standardized.

[0065] In a specific embodiment, the specific steps of obtaining the width conversion feature are:

[0066] (i) Randomly generate weight W ei With bias β ei , using the linear transformation mapping function φ i (·) Generate N from input data X g The i-th group of randomly mapped features is:

[0067]

[0068] Combining all the mapping features, we get:

[0069] (ii) Using the mapping features obtained in (i) Generate l groups of enhanced features, each group of enhanced features generally contains an enhanced node, and the jth group of enhanced features can be expressed as: (there are only three layers, and the enhancement layer is after the input layer)

[0070]

[0071] Among them, W hj and β hj It is also randomly generated by the network. j =tanh(·) represents the jth activation function, and all enhanced features are combined to obtain:

[0072]

[0073] (iii) Combine all the mapping features and enhanced features to obtain the final width conversion feature:

[0074]

[0075] (iv) Connect the width conversion feature A to the network output, and the optimization problem is specifically expressed as:

[0076]

[0077] In a specific embodiment, the ε drag method is used to expand the distance between different categories and better fit the regression task, greatly improving the flexibility of the optimization problem model. The target matrix Y is dragged and relaxed to Y+B⊙M, and the objective function is:

[0078]

[0079] The mixed norm will Norm and The norm is combined with regularization constraints on the projection matrix trained based on the width learning network model. Under the norm constraint, the projection matrix is ​​more compact and can be equipped with a row sparse structure. Redundant features can be removed during training. The norm can further impose constraints on the model to prevent overfitting of the model, and the objective function is further optimized as follows:

[0080]

[0081] Where W represents the projection matrix, A represents the training sample width conversion feature matrix, λ1 and λ2 represent regularization parameters, B represents the drag direction of ε, a constant matrix composed of +1 and -1, ⊙ represents the Hadamard product of the matrix; M is a constant matrix representing the distance of label dragging, ||·|| F represents the Frobenius norm of the matrix, ||·|| 2,1 Represents the matrix Norm.

[0082] In a specific embodiment, the specific process of solving the objective function using the two-step iterative method includes:

[0083] (1) Solve the projection matrix W:

[0084] make Find the partial derivative of W and set it equal to 0. According to the ridge regression of the pseudo-inverse matrix, we can get:

[0085]

[0086] Get the projection matrix

[0087] In the formula, R = Y + B⊙M, D is the diagonal element D ii =1 / ||W i ||2 composed of a diagonal matrix;

[0088]

[0089] (2) Optimize the label drag distance matrix M:

[0090] make Simplified to Decompose the optimization problem into sub-problems to be solved element by element. Since B represents the dragging direction of ε, it is a constant matrix composed of +1 and -1, that is, Then (E ij -B ij M ij ) 2 =(B ij E ij -M ij ) 2 , considering the non-negative constraint of M, we can get the optimal solution of the subproblem: M ij =max(B ij E ij ,0), that is:

[0091] M=max(B⊙E,0) (19)

[0092] (3) Iterate the above (1) and (2) to obtain the optimal solution of the label drag distance matrix M and the projection matrix W.

[0093] In a specific embodiment, the termination condition of the iterative optimization solution is:

[0094] (obj (t-1) -obj (t) )≤ε0

[0095] Where, obj (t-1) 、obj (t) They represent the objective function values ​​of the t-1th and tth iterations respectively, and ε0 is the set value.

[0096] The above-mentioned discriminant width learning network algorithm based on mixed norm regularization constraint is as follows:

[0097] 1: Input training sample {X,Y}, feature mapping function φ i (·), activation function ξ j (·), regularization parameters λ1, λ2, number of feature map node groups N g , the number of nodes in each set of feature map nodes N f , the number of enhanced features N e ;

[0098] 2: Initialize M=0, D=I d ;

[0099] 3: Randomly generate weights and biases W of appropriate dimensions ei , βei ,i=1,2,...,N g ; Through the mapping function φ i (·) Calculate the i-th group of feature nodes Z i =φ i (XW ei +β ei ), i=1,2,...,N g , combining all mapping feature nodes to obtain

[0100] 4: Randomly generate weights and biases W of appropriate dimensions hj and β hj , through the activation function ξ j (·) Generate enhanced features Generating Width Conversion Features by Collecting Feature Nodes and Enhancing Features

[0101] 5: Given M and D, calculate the projection matrix W using formula (17);

[0102] 6: Using W, update M through formula (19) and update D through formula (18);

[0103] 7: Using the updated M, D, update the projection matrix W using formula (17);

[0104] 8: t = t + 1;

[0105] 9: until: (obj (t-1) -obj t )≤ε0;

[0106] Among them, obj (t-1) 、obj t They represent the objective function values ​​of the t-1th and tth iterations respectively, and ε0 is the set value.

[0107] See attached Figure 2 As shown, an embodiment of the present invention further provides a device for a width learning image classification method based on a mixed norm regularization constraint using any one of the above embodiments, comprising:

[0108] An acquisition module is used to acquire image data to be classified, preprocess the image data, divide the preprocessed image data into training samples and test samples, and acquire class labels corresponding to the training samples;

[0109] A processing module is used to construct a width learning network, input the training samples into the width learning network for processing, and generate width conversion features;

[0110] The calculation module is used to construct a classification optimization objective function of a discriminant width learning network based on a mixed norm regularization constraint according to the width conversion features and class labels, using the ε drag method and the mixed norm regularization constraint method;

[0111] The solution module is used to solve the objective function using two-step iterations to achieve image classification.

[0112] An embodiment of the present invention further provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, an image classification method as described in any one of the above embodiments is implemented.

[0113] The method provided by the embodiment of the present invention is compared with several commonly used classification methods, including ELM, SRC, LRC, CRC, SVM, DLSR, LRRR, LRLR, RELSR, IDLSR and BLS. In addition, the method provided by the embodiment of the present invention is compared with the standard width learning network method alone to further evaluate the classification performance of the method of the present invention.

[0114] Figure 3 (a)-(c) are partial image database embodiments. (a) is a partial image of the ExtendedYaleB database, (b) is a partial image of the AR database, and (c) is a partial image of the LFW database.

[0115] (1) The image data in the ExtendedYaleB database are uniformly resized to 32×32 pixels. We randomly select 10, 15, 20, and 25 images from each category as training sets, and the rest as test sets.

[0116] (2) For the AR database, we uniformly resize the images to 50×40 pixels and select 4, 6, 8, and 12 samples from each category as training sets, and the rest are used as test sets.

[0117] (3) For the LFW database, the images therein are uniformly adjusted to a size of 32×32 pixels, 5, 6, 7, and 8 images are randomly selected from each category as training sets, and the rest are used as test sets. The processing of the remaining image database embodiments is similar to the above operation.

[0118] The method provided in the embodiment of the present invention is compared with other methods to analyze performance. The specific analysis results are shown in Tables 1-3.

[0119] Table 1 Comparison of classification accuracy (AVER%) between this method (MDBLS) and other methods on the EXTENDEDYALEB face database

[0120]

[0121] Table 2 Comparison of classification accuracy of this method (MDBLS) and other methods on AR face database (AVER%)

[0122]

[0123] Table 3 Comparison of classification accuracy (AVER%) between this method (MDBLS) and other methods on LFW face database

[0124]

[0125] Through the above comparison, we can find that MDBLS is superior to standard BLS in terms of parameter setting and accuracy. In most cases, the total number of combined feature nodes of MDBLS is smaller. And the two parameters λ1 and λ2 are larger. From the perspective of accuracy, MDBLS can achieve higher accuracy with fewer training samples and lower standard derivation. Using regression target relaxation technology norm regularization term and With the mixed norm regularization constraint of the norm regularization term, the MDBLS method is more discriminative and robust.

[0126] In addition, based on the above-mentioned partial database, the curve of the objective function value and the number of iterations with different training times is found, and the method of the present invention is theoretically guaranteed to converge to a local minimum. In the above-mentioned experiment, the objective value of the algorithm of the present invention decreases monotonically at each iteration and can converge to a stable value of no more than 30 iterations, which once again proves the effectiveness of the method.

[0127] Each embodiment in this specification is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same and similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part. The above description of the disclosed embodiment enables professionals and technicians in this field to implement or use the present invention.

[0128] Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A width learning image classification method based on mixed norm regularization constraint, characterized in that: The following steps are involved: Acquire image data to be classified, preprocess the image data, divide the preprocessed image data into training samples and test samples, and obtain class labels corresponding to the training samples; Constructing a width learning network, inputting the training samples into the width learning network for processing, and generating width conversion features; According to the width conversion feature and the class label, the ε drag method and the mixed norm regularization constraint method are used to construct a classification optimization objective function of the discriminant width learning network based on the mixed norm regularization constraint. The specific expression is: Where W represents the projection matrix, A represents the training sample width conversion feature matrix, λ1 and λ2 represent regularization parameters, B represents the drag direction of ε, a constant matrix composed of +1 and -1, ⊙ represents the Hadamard product of the matrix; M is a constant matrix representing the distance of label dragging, ||·|| F represents the Frobenius norm of the matrix, ||·|| 2,1 l represents the matrix 2,1 norm; The objective function is optimized and solved by a two-step iterative method to achieve image classification. The specific process includes: (1) Solve the projection matrix W: make Taking the partial derivative of W and setting it equal to 0, we get: In the formula, R = Y + B⊙M, D is the diagonal element D ii =1 / ||W i ||2 composed of a diagonal matrix; (2) Optimize the label drag distance matrix M: Let E = AW-Y, the optimization problem is simplified to Decompose the optimization problem into sub-problems to be solved element by element, and obtain M ij =max(B ij E ij ,0), that is: M=max(B⊙E,0); (3) Iterate (1) and (2) to obtain the optimal solution of the label drag distance matrix M and the projection matrix W.

2. According to claim 1, a width learning image classification method based on mixed norm regularization constraint is characterized in that: The termination condition of iterative optimization solution is: (object) (t-1) -obj (t) )≤ε0 Where, obj (t-1) 、obj (t) They represent the objective function values ​​of the t-1th and tth iterations respectively, and ε0 is the set value.

3. The method for image classification based on width learning with mixed norm regularization constraint according to claim 1, characterized in that: The specific steps to obtain the width conversion feature are: (i) Randomly generate weight W ei With bias β ei , using the linear transformation mapping function φ i (·) Generate N from input data X g The i-th group of randomly mapped features is: Combining all the mapping features, we get: (ii) Using the mapping features obtained in (i) Generate l groups of enhanced features, each group of enhanced features contains an enhanced node, and the jth group of enhanced features can be expressed as: Among them, W hj and β hj It is also randomly generated by the network. j =tanh(·) represents the jth activation function, and all enhanced features are combined to obtain: (iii) Combine all the mapping features and enhanced features to obtain the final width conversion feature:

4. The method for image classification based on width learning with mixed norm regularization constraint according to claim 1, characterized in that: The steps of image data preprocessing are as follows: Unifying the sizes of the training samples and the test samples; Divide N images as training samples and the remaining images as test samples; divide the image database four times to obtain four groups of training samples and test samples; The training samples and the test samples are standardized.

5. A classification device using a width learning image classification method based on a mixed norm regularization constraint as described in any one of claims 1 to 4, characterized in that: include: An acquisition module is used to acquire image data to be classified, preprocess the image data, divide the preprocessed image data into training samples and test samples, and acquire class labels corresponding to the training samples; A processing module, used for constructing a width learning network, inputting the training samples into the width learning network for processing, and generating width conversion features; A calculation module, used to construct a classification optimization objective function of a discriminant width learning network based on a mixed norm regularization constraint according to the width conversion feature and the class label by using an ε drag method and a mixed norm regularization constraint method; The solution module is used to solve the objective function by two-step iteration to achieve image classification.

Images