Competition non-negative representation classification method and system based on flexible factor

Abstract

The invention discloses a competition non-negative representation classification method and system based on a flexible factor, and relates to the technical field of image mode classification. The method comprises the following steps: acquiring a test sample and a training sample set for image mode classification, and preprocessing samples in the training sample set; calculating a representation coefficient of the test sample on the training sample set based on the preprocessed training sample set; and classifying the test sample based on the representation coefficient to obtain a category corresponding to the test sample identified as the minimum residual error. According to the method and the system provided by the invention, the subspace competition representation items and the category representation vectors can be constrained at the same time, the contribution of each category representation vector is increased, the test samples are classified based on the representation coefficients, and the category corresponding to the test samples identified as the minimum residual error is obtained. And various image classification tasks are effectively completed, and the image classification accuracy is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of image pattern classification, and in particular to a competitive non-negative representation classification method and system based on flexible factors. Background Art

[0002] The field of pattern classification has always been a hot research topic. In recent years, a large number of mature methods have been widely used in various fields. However, the Convolutional Neural Network (CNN) model has shown excellent performance and has spawned many outstanding applications. Despite its significant advantages, a large number of experiments have shown that a large amount of training data is often required to fully understand the underlying distribution of the data. In the case of limited data set size and insufficient training samples, the Representation Based Classification Method (RBCM) becomes an attractive choice.

[0003] Sparse Representation-based Classification (SRC) treats all training samples as a dictionary and represents the test samples by linear combination of elements extracted from the dictionary. In this process, norm constraints are imposed on the combination coefficients to obtain sparse representation coefficients. Collaborative Representation-based Classification (CRC) points out that the improvement in classification accuracy mainly comes from the collaborative representation formed between the test samples and training samples from different categories, rather than relying solely on sparsity. By adopting regularization, CRC can achieve a performance level similar to SRC while effectively reducing the computational burden.

[0004] In order to resolve the ambiguity about the impact of or regularization on classification performance, the nonnegative representation classification method (NRC) believes that although it can be well explained mathematically to approximate test samples using samples from multiple categories through complex addition and subtraction operations, this contradicts the intuitive physical meaning. In actual scenarios, samples of the same category tend to show positive correlation, while samples of different categories tend to show no correlation rather than negative correlation. Therefore, a non-negative constraint is imposed on the representation coefficient to prevent negative values, which promotes the improvement of the model's classification performance. Through a large number of experiments, NRC has demonstrated excellent performance.

[0005] However, there is no difference between the original NRC and the traditional CRC in the classification process. The representation process is independent of the classification process. NRC ignores the competing representation information. At the same time, NRC does not add additional constraints to the representation vector, which leads to the lower performance of NRC image pattern classification and the lower accuracy of image pattern classification tasks. Summary of the invention

[0006] In order to overcome the problem that the existing image pattern classification technology ignores the competitive representation information, resulting in low performance of NRC image pattern classification and further low accuracy of image pattern classification tasks, the present invention proposes a competitive non-negative representation classification method and system based on flexible factors, which can simultaneously constrain subspace competitive representation items and category representation vectors, increase the contribution of each category representation vector, complete various image classification tasks and effectively improve the accuracy of image classification.

[0007] To achieve the purpose of the present invention, the present invention adopts the following technical solutions:

[0008] A competitive non-negative representation classification method based on flexible factors, the method comprising the following steps:

[0009] Acquire test samples and training sample sets for image pattern classification, and preprocess the samples in the training sample sets;

[0010] Based on the preprocessed training sample set, calculate the representation coefficient of the test sample on the training sample set;

[0011] The test sample is classified based on the representation coefficient to obtain a category corresponding to the minimum residual error identified as the test sample.

[0012] In the above technical scheme, preprocessing the samples in the training sample set can improve the reliability of the sample data in the training sample set and, at the same time, improve the efficiency of subsequent data calculations; calculating the representation coefficient of the test sample on the training sample set, the obtained representation coefficient can simultaneously constrain the subspace competition representation item and the category representation vector, increase the contribution of each category representation vector, and classify the test sample based on the representation coefficient to obtain the category corresponding to the minimum residual identified by the test sample, thereby effectively completing various image classification tasks and improving the accuracy of image classification.

[0013] Furthermore, a training sample set for image pattern classification is obtained, and the expression is:

[0014]

[0015] in, n i Represents the i-th type of training sample X iC represents the total number of sample categories; n represents the total number of training samples; α represents the feature vector of each sample; m represents the feature dimension of each sample.

[0016] Furthermore, the process of preprocessing the samples in the training sample set includes:

[0017] The samples in the training sample set are normalized using the l2 norm.

[0018] In the above technical solution, performing l2 norm normalization processing on the samples in the training sample set can improve the reliability of the sample data in the training sample set and, at the same time, improve the efficiency of subsequent data calculation.

[0019] Furthermore, the process of calculating the representation coefficient of the test sample on the training sample set includes:

[0020] Setting a neural network model to learn sample data changes between test samples and training sample sets, and obtaining a competitive non-negative representation classification model that grasps sample data changes between test samples and training sample sets;

[0021] The competitive non-negative representation classification model constructs a competitive non-negative representation classification objective function based on the test sample and training sample set;

[0022] Solve the competitive non-negative representation classification objective function to obtain the representation coefficient of the test sample on the training sample set.

[0023] Furthermore, the competitive non-negative representation classification model constructs a competitive non-negative representation classification objective function of a non-negative least squares problem based on the test sample and the training sample set, and the expression is:

[0024]

[0025] Use the variable splitting method to solve the representation coefficient α, introduce the auxiliary variable z, and rewrite the competitive non-negative representation classification objective function into the following form:

[0026]

[0027] Among them, the first term represents the overall representation error term; the second term represents the competition representation error term, f i represents the flexibility factor; the third term represents the category coefficient constraint term, y represents the test sample, and α is the representation coefficient. λ>0 and γ>0 both represent weight balance parameters, X i represents the i-th type of training samples, α i represents the coefficient vector corresponding to the i-th class of training samples.

[0028] Furthermore, the process of solving the competitive non-negative representation classification objective function includes:

[0029] The alternating direction multiplier method is used to solve the corresponding augmented Lagrangian function, which is expressed as:

[0030]

[0031] Fixed f i , z and δ, update α, the expression is:

[0032]

[0033] Fix α, z and δ, update f i , the expression is:

[0034]

[0035] Fixed α, f i and δ, update z, the expression is:

[0036]

[0037] Update δ, the expression is:

[0038] δ = δ + μ (z - α);

[0039] Repeat the above iterative steps until the convergence condition of the alternating direction multiplier method is met or the number of iterations exceeds the preset threshold T, and the global optimal solution is obtained;

[0040] Where δ represents the Lagrange multiplier vector, μ>0 represents the penalty parameter, <δ, z-α> represents the inner product of two vectors, and the initialization variable α0=f i0 = z0 = δ0 = 0, using α t , f it , z t and δ t represents the optimization variable and Lagrange multiplier when the iteration is t (t=0, 1, 2, ..., T), and T is a preset threshold.

[0041] Furthermore, the expression of the convergence condition of the alternating direction multiplier method is:

[0042] ||α t -z t ||2≤η,||α t+1 -α t ||2≤η and ||z t+1 -z t ||2≤ηsatisfies at the same time;

[0043] Wherein, η represents a smaller preset threshold.

[0044] In the above technical solution, a flexible factor is introduced to represent the error between the test sample and the competitive representation of each subspace. The flexible factor is used as the weight of each type of representation vector, so that the representation vector of the category with a small competitive representation error has a higher representation weight, thereby improving the representation ability of the model.

[0045] Furthermore, the process of classifying the test samples based on the representation coefficients includes:

[0046] Calculate the residual of the test sample in each category of the training sample set, the expression is:

[0047] r i =||yX i α i ||2;

[0048] According to the residual classification, the test sample is identified as the category corresponding to the minimum residual, and the expression is:

[0049]

[0050] Among them, X i represents the i-th type of training samples, α i represents the coefficient vector corresponding to the i-th class of training samples.

[0051] In the above technical solution, the test samples are classified based on the representation coefficients to obtain the category corresponding to the minimum residual error identified as the test sample, thereby effectively completing various image classification tasks and improving the accuracy of image classification.

[0052] A competitive non-negative representation classification system based on flexible factors, the system comprising:

[0053] An acquisition module, used for acquiring a test sample and a training sample set for image pattern classification;

[0054] A processing module, used for preprocessing samples in the training sample set;

[0055] A calculation module, used for calculating the representation coefficient of the test sample on the training sample set based on the preprocessed training sample set;

[0056] The classification module is used to classify the test sample based on the representation coefficient to obtain the category corresponding to the minimum residual error of the test sample.

[0057] An electronic device comprises a memory, a processor and a computer program stored in the memory and running on the processor, wherein when the processor executes the computer program, the steps of a competitive non-negative representation classification method based on flexible factors are implemented.

[0058] Compared with the prior art, the present invention has the following beneficial effects:

[0059] The present invention proposes a competitive non-negative representation classification method and system based on flexible factors. First, the samples in the training sample set are preprocessed to improve the reliability of the sample data in the training sample set and improve the efficiency of subsequent data calculation. Then, the representation coefficient of the test sample on the training sample set is calculated. The obtained representation coefficient can simultaneously constrain the subspace competitive representation item and the category representation vector, increase the contribution of each category representation vector, and classify the test sample based on the representation coefficient to obtain the category corresponding to the minimum residual error identified as the test sample, thereby effectively completing various image classification tasks and improving the accuracy of image classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 A flowchart of a competitive non-negative representation classification method based on flexible factors provided in an embodiment of the present application;

[0061] Figure 2 A schematic diagram of the structure of a competitive non-negative representation classification model provided in an embodiment of the present application;

[0062] Figure 3 A schematic diagram of the structure of a competitive non-negative representation classification system based on flexible factors provided in an embodiment of the present application. DETAILED DESCRIPTION

[0063] In order to facilitate the understanding of the present invention, the present invention will be described more fully below with reference to the relevant drawings. Preferred embodiments of the present invention are provided in the drawings. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. On the contrary, the purpose of providing these embodiments is to make the understanding of the disclosure of the present invention more thorough and comprehensive.

[0064] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art of the present invention. The terms used herein in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The term "and / or" used herein includes any and all combinations of one or more related listed items.

[0065] Embodiment 1:

[0066] A competitive non-negative representation classification method based on flexible factors, see Figure 1 , the method comprises the following steps:

[0067] Step S1: obtaining a test sample and a training sample set for image pattern classification, and preprocessing the samples in the training sample set;

[0068] Step S2: Based on the preprocessed training sample set, calculating the representation coefficient of the test sample on the training sample set;

[0069] Step S3: classify the test sample based on the representation coefficient to obtain the category corresponding to the minimum residual error of the test sample.

[0070] In step S1, a training sample set for image pattern classification is obtained, and the expression is:

[0071]

[0072] in, n i Represents the i-th type of training sample X i C represents the total number of sample categories; n represents the total number of training samples; α represents the feature vector of each sample; m represents the feature dimension of each sample.

[0073] Furthermore, the process of preprocessing the samples in the training sample set includes:

[0074] The samples in the training sample set are normalized using the l2 norm.

[0075] It can be understood that performing l2 norm normalization processing on the samples in the training sample set can improve the reliability of the sample data in the training sample set and at the same time improve the efficiency of subsequent data calculation.

[0076] As a preferred embodiment, in step S2, see Figure 2 , the process of calculating the representation coefficient of the test sample on the training sample set includes:

[0077] Step S21: Setting a neural network model to learn sample data changes of a test sample and a training sample set, and obtaining a competitive non-negative representation classification model that grasps the sample data changes of the test sample and the training sample set;

[0078] Step S22: The competitive non-negative representation classification model constructs a competitive non-negative representation classification objective function based on the test sample and the training sample set;

[0079] Step S23: Solve the competitive non-negative representation classification objective function to obtain the representation coefficient of the test sample on the training sample set.

[0080] In step S22, the competitive non-negative representation classification model constructs a competitive non-negative representation classification (FCNR) objective function of a non-negative least squares problem based on the test sample and the training sample set, and the expression is:

[0081]

[0082] Among them, the first term represents the overall representation error term; the second term represents the competition representation error term, f i represents the flexibility factor; the third term represents the category coefficient constraint term, y represents the test sample, and α is the representation coefficient. λ>0 and γ>0 both represent weight balance parameters, X i represents the i-th type of training samples, α i represents the coefficient vector corresponding to the i-th class of training samples.

[0083] In the second item, the test sample is reconstructed by each class of samples. The test sample of this class can often better reconstruct the test sample y, so that the flexibility factor f of this class i Smaller.

[0084] Use f i As the constraint coefficient of each category in the third term, the representation coefficient {α i} i≠k The optimization weight is larger, and the representation coefficient of this class {α k} indicates a greater weight, where k is the label of y. This enhances the overall reconstruction capability of the representation coefficient.

[0085] Furthermore, FCNR is a non-negative least squares problem, which requires the use of variable splitting method to solve the representation coefficient α, introduce auxiliary variables z, and rewrite the competitive non-negative representation classification (FCNR) objective function into the following form:

[0086]

[0087] In step S23, the process of solving the competitive non-negative representation classification objective function includes:

[0088] The alternating direction method of multipliers (ADMM) is used to solve the corresponding augmented Lagrange function (ALF), which is expressed as:

[0089]

[0090] In formula (4), δ is the Lagrange multiplier vector, μ>0 is the penalty parameter, and <δ, z-α> is the inner product of two vectors. Initialize variable α0=f i0 = z0 = δ0 = 0, using α t , f it , z t and δ trepresents the optimization variable and Lagrange multiplier when the iteration is t (t = 0, 1, 2, ..., T), and T is the preset threshold. Use the alternating optimization method to update the variables α, f i and z, the detailed solution process is as follows:

[0091] (1) Fixed f i , z and δ, update α, the expression is:

[0092]

[0093] In order to facilitate optimization, the matrix H is introduced i =diag(0, ..., 0, 1, ..., 1, 0, ..., 0), where the element of the i-th category is 1 and the rest are 0. Therefore, the problem Thus there is It can be transformed into the following problem:

[0094]

[0095] By taking the derivative of the objective function of this problem with respect to α and setting it to 0, we can get the analytical solution of α:

[0096]

[0097] in,

[0098] (2) Fix α, z and δ, and update f i , the expression is:

[0099]

[0100] The objective function of this problem is f i Taking the derivative and setting it to 0, we can get f i The analytical solution is:

[0101]

[0102] (3) Fixed α, f i and δ, update z, the expression is:

[0103]

[0104] The analytical solution of z is:

[0105]

[0106] (4) Update δ, the expression is:

[0107] δ = δ + μ (z - α);

[0108] Repeat the above iterative steps until the convergence condition of the alternating direction multiplier method is met or the number of iterations exceeds the preset threshold T, and the global optimal solution is obtained;

[0109] Where δ represents the Lagrange multiplier vector, μ>0 represents the penalty parameter, <δ, z-α> represents the inner product of two vectors, and the initialization variable α0=f i0 = z0 = δ0 = 0, using α t , f it , z t and δ t represents the optimization variable and Lagrange multiplier when the iteration is t (t=0, 1, 2, ..., T), and T is a preset threshold.

[0110] Furthermore, the expression of the convergence condition of the alternating direction multiplier method is:

[0111] ||α t -z t ||2≤η,||α t+1 -α t ||2≤η and ||z t+1 -z t ||2≤ηsatisfies at the same time;

[0112] Wherein, η represents a smaller preset threshold.

[0113] Since the least squares objective function, linear equality, and non-negative constraints are all strict convex optimization problems, solving formula (3) through ADMM can ensure convergence to the global optimal solution.

[0114] It can be understood that the flexible factor is introduced to represent the error between the test sample and the competitive representation of each subspace. The flexible factor is used as the weight of each type of representation vector, so that the representation vector of the category with a small competitive representation error has a higher representation weight, thereby improving the representation ability of the model.

[0115] As a preferred embodiment, in step S3, the process of classifying the test sample based on the representation coefficient includes:

[0116] Calculate the residual of the test sample in each category of the training sample set, the expression is:

[0117] r i =||yX i α i ||2;

[0118] According to the residual classification, the test sample is identified as the category corresponding to the minimum residual, and the expression is:

[0119]

[0120] Among them, Xi represents the i-th type of training samples, α i represents the coefficient vector corresponding to the i-th class of training samples.

[0121] It can be understood that by classifying the test samples based on the representation coefficients, the test samples are identified as the categories corresponding to the minimum residuals, thereby effectively completing various image classification tasks and improving the accuracy of image classification.

[0122] In this embodiment, the samples in the training sample set are preprocessed, which can improve the reliability of the sample data in the training sample set and improve the efficiency of subsequent data calculations. The representation coefficient of the test sample on the training sample set is calculated, and the obtained representation coefficient can simultaneously constrain the subspace competition representation item and the category representation vector, increase the contribution of each category representation vector, and classify the test sample based on the representation coefficient to obtain the category corresponding to the minimum residual identified by the test sample, thereby effectively completing various image classification tasks and improving the accuracy of image classification.

[0123] Embodiment 2:

[0124] This example provides further experimental data proof based on the method proposed in the example, as follows:

[0125] In order to more comprehensively evaluate the performance of FCNR, this paper selects the AR face dataset, the Extended Yale B face dataset and the USPS handwriting dataset as verification platforms, and conducts comparative experiments on FCNR with the current mainstream classification methods.

[0126] The AR Face dataset contains 4,000 frontal color face images of 126 different individuals, 56 of whom are males and 70 of whom are females. The images show changes in lighting conditions, facial expressions, and occlusion. The images of each individual were taken in two independent periods, 14 days apart. A subset of 100 individuals (50 males and 50 females) was selected as training and test data. Each individual has 14 images, 7 for training and 7 for testing. All images were cropped to a size of 60×43 and normalized to unit norm. Finally, PCA was used to project these images into feature subspaces of dimensions 54, 120, and 300. Table 1 lists the classification accuracy of different methods on the AR Face dataset in detail.

[0127] Table 1 Classification accuracy of different methods on AR face dataset

[0128]

[0129] It is obvious that in all three dimensions, the FCNR algorithm surpasses all other methods to achieve the highest classification results. In 54 dimensions, it is 85.6%, which is 0.4% higher than NRC, in 120 dimensions, it is 91.6%, which is 0.3% higher than NRC, and in 300 dimensions, it is 94.3%, which is 1% higher than NRC. The experimental results strongly demonstrate the excellent performance of FCNR on AR datasets.

[0130] The Extended Yale B dataset contains 64 frontal face images taken under different lighting conditions. The original image size is 192×168 pixels. These images are resized to 54×48 pixels and normalized to unit norm. The dataset is randomly divided into two halves, each containing 32 images of each person, one half as the training set and the other half as the test set. Finally, PCA is used to project the original images into subspaces of 84, 150, and 300 dimensions. The classification accuracy of different methods on the Extended Yale B dataset is detailed in Table 2.

[0131] Table 2 Classification accuracy of different methods on the Extended Yale B handwritten digits dataset

[0132]

[0133] Obviously, in all three dimensions, the FCNR algorithm surpasses all other methods to achieve the highest classification results. In 84 dimensions, it is 96.9%, which is 1.7% higher than NRC, in 150 dimensions, it is 97.6%, which is 0.7% higher than NRC, and in 300 dimensions, it is 98.8%, which is 0.4% higher than NRC. The experimental results strongly demonstrate the excellent performance of FCNR on the Extended Yale B dataset.

[0134] USPS is a set of grayscale handwritten digit image datasets with various font styles, containing 9298 images in total, showing all the numbers from 0 to 9. These images are divided into two sets, 7291 images for training and 2007 images for testing. Each image is standardized, with a uniform size of 16×16 pixels and centered. In this experimental setting, 50, 100, and 200 images of each digit are randomly selected for training, and the remaining samples are allocated to testing. The experiment was tested ten times independently, and Table 3 lists in detail the average classification accuracy of different methods on the USPS handwritten digit dataset.

[0135] Table 3 Classification accuracy of different methods on USPS handwritten digits dataset

[0136]

[0137] Obviously, in all experiments with three numbers of training samples, the FCNR algorithm surpassed all other methods to obtain the highest classification results. In the experiment with 50 training samples per class, FCNR improved by 92.5% over NRC by 1.3%, in the experiment with 100 training samples, 93.5% over NRC by 1.1%, and in the experiment with 300 training samples, 94.3% over NRC by 1.1%. The experimental results strongly demonstrate the excellent performance of FCNR on the USPS handwritten digit dataset.

[0138] Embodiment three:

[0139] A competitive non-negative representation classification system based on flexible factors, see Figure 3 , the system comprising:

[0140] An acquisition module, used for acquiring a test sample and a training sample set for image pattern classification;

[0141] A processing module, used for preprocessing samples in the training sample set;

[0142] A calculation module, used for calculating the representation coefficient of the test sample on the training sample set based on the preprocessed training sample set;

[0143] The classification module is used to classify the test sample based on the representation coefficient to obtain the category corresponding to the minimum residual error of the test sample.

[0144] Embodiment 4:

[0145] An electronic device comprises a memory, a processor and a computer program stored in the memory and running on the processor, wherein when the processor executes the computer program, the steps of a competitive non-negative representation classification method based on flexible factors are implemented.

[0146] The above descriptions are merely embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A competitive non-negative representation classification method based on flexible factors, characterized in that: The method comprises the following steps: Acquire test samples and training sample sets for image pattern classification, and preprocess the samples in the training sample sets; Based on the preprocessed training sample set, calculate the representation coefficient of the test sample on the training sample set; The test sample is classified based on the representation coefficient to obtain a category corresponding to the minimum residual error identified as the test sample.

2. The competitive non-negative representation classification method based on flexible factor according to claim 1 is characterized in that: Get the training sample set for image pattern classification. The expression is: in, n i Represents the i-th type of training sample X i C represents the total number of sample categories; n represents the total number of training samples; α represents the feature vector of each sample; m represents the feature dimension of each sample.

3. The competitive non-negative representation classification method based on flexible factor according to claim 1 is characterized in that: The process of preprocessing the samples in the training sample set includes: The samples in the training sample set are normalized using the l2 norm.

4. The competitive non-negative representation classification method based on flexible factor according to claim 2 is characterized in that: The process of calculating the representation coefficient of the test sample on the training sample set includes: Setting a neural network model to learn sample data changes between test samples and training sample sets, and obtaining a competitive non-negative representation classification model that grasps sample data changes between test samples and training sample sets; The competitive non-negative representation classification model constructs a competitive non-negative representation classification objective function based on the test sample and training sample set; Solve the competitive non-negative representation classification objective function to obtain the representation coefficient of the test sample on the training sample set.

5. The competitive non-negative representation classification method based on flexible factor according to claim 4 is characterized in that: The competitive non-negative representation classification model constructs a competitive non-negative representation classification objective function of a non-negative least squares problem based on the test sample and the training sample set. The expression is: Use the variable splitting method to solve the representation coefficient α, introduce the auxiliary variable z, and rewrite the competitive non-negative representation classification objective function into the following form: Among them, the first term represents the overall representation error term; the second term represents the competition representation error term, f i represents the flexibility factor; the third term represents the category coefficient constraint term, y represents the test sample, α is the representation coefficient, λ>0 and γ>0 both represent the weight balance parameter, X i represents the i-th type of training samples, α i represents the coefficient vector corresponding to the i-th class of training samples.

6. The competitive non-negative representation classification method based on flexible factor according to claim 5 is characterized in that: The process of solving the competitive non-negative representation classification objective function includes: The alternating direction multiplier method is used to solve the corresponding augmented Lagrangian function, which is expressed as: Fixed f i , z and δ, update α, the expression is: Fix α, z and δ, update f i , the expression is: Fixed α, f i and δ, update z, the expression is: Update δ, the expression is: δ = δ + μ (z - α); Repeat the above iterative steps until the convergence condition of the alternating direction multiplier method is met or the number of iterations exceeds the preset threshold T, and the global optimal solution is obtained; Where δ represents the Lagrange multiplier vector, μ>0 represents the penalty parameter, <δ, z-α> represents the inner product of two vectors, and the initialization variable α0=f i0 = z0 = δ0 = 0, using α t , f it , z t and δ t represents the optimization variable and Lagrange multiplier when the iteration is t (t=0, 1, 2, ..., T), and T is a preset threshold.

7. The competitive non-negative representation classification method based on flexible factor according to claim 6 is characterized in that: The expression of the convergence condition of the alternating direction multiplier method is: ||α t -z t ||2≤η, ||a t+1 -α t ||2≤η and ||z t+1 -z t ||2≤ηsatisfies at the same time; Wherein, η represents a smaller preset threshold.

8. The competitive non-negative representation classification method based on flexible factor according to claim 7 is characterized in that: The process of classifying the test sample based on the representation coefficient includes: Calculate the residual of the test sample in each category of the training sample set, the expression is: r i =||yX i a i ||2; According to the residual classification, the test sample is identified as the category corresponding to the minimum residual, and the expression is: Among them, X i represents the i-th type of training samples, α i represents the coefficient vector corresponding to the i-th class of training samples.

9. A competitive non-negative representation classification system based on flexible factors, characterized in that: The system comprises: An acquisition module, used for acquiring a test sample and a training sample set for image pattern classification; A processing module, used for preprocessing samples in the training sample set; A calculation module, used for calculating the representation coefficient of the test sample on the training sample set based on the preprocessed training sample set; The classification module is used to classify the test sample based on the representation coefficient to obtain the category corresponding to the minimum residual error of the test sample.

10. An electronic device, characterized in that: The method comprises a memory, a processor, and a computer program stored in the memory and running on the processor, wherein when the processor executes the computer program, the steps of the method according to any one of claims 1 to 8 are implemented.

Images