CN-122023956-A - Self-adaptive dictionary learning method based on L0 regularization and mixed discrimination constraint

CN122023956ACN 122023956 ACN122023956 ACN 122023956ACN-122023956-A

Abstract

The invention discloses an adaptive dictionary learning method based on L0 regularization and mixed discrimination constraint in the technical field of automatic control and adjustment, which comprises the following steps of S1 dividing an original image dataset into a training subset and a verification subset according to a preset proportion, S2 constructing a class sub-dictionary and splicing to generate a global initial dictionary, initializing an all-zero coding coefficient matrix, an SVM discrimination vector and bias parameters, S3 constructing a mixed objective function integrating a reconstruction error term, a Fisher discrimination constraint term, an SVM discrimination loss term and a dissimilarity penalty term, restraining the correlation among atoms of different classes through the dissimilarity penalty term, S4 solving a mixed objective function model by adopting an alternative optimization strategy, S5 solving a test coding coefficient on a test sample to determine a prediction class, improving robustness by utilizing an enhancement and sparsity mechanism, enhancing dictionary discrimination force by mixed discrimination constraint, reducing complexity by combining alternative optimization, and considering efficiency and generalization.

Inventors

YANG BAOQING
YANG ZHOUSHENG

Assignees

扬州大学

Dates

Publication Date: 20260512
Application Date: 20260130

Claims (10)

1. The self-adaptive dictionary learning method based on L0 regularization and mixed discrimination constraint is characterized by comprising the following steps: S1, acquiring an original image data set to be classified, extracting image features, performing dimension reduction processing on the image features by utilizing principal component analysis, and reserving variance information of a preset proportion, performing L2 norm normalization of column vectors on a dimension-reduced sample matrix to obtain a small sample data set, performing illumination disturbance data enhancement on the small sample data set, expanding the number of the small sample data set to a preset multiple of the number of the original image data set, and dividing the small sample data set into a training subset and a verification subset according to the preset proportion; S2, extracting a main component and a class mean vector from the training sample, constructing a class sub-dictionary, and splicing to generate a global initial dictionary; S3, constructing a mixed objective function of fusion reconstruction error items, fisher identification constraint items, SVM identification loss items and dissimilarity penalty items, wherein the mixed objective function aims at minimizing reconstruction errors, enhancing intra-class compactness and inter-class separability through Fisher criteria, guiding coding coefficients to adapt to a classifier through SVM loss, and inhibiting correlation among atoms of different classes through dissimilarity penalty items; s4, solving a mixed objective function model by adopting an alternate optimization strategy; S5, solving the test coding coefficient of the test sample by using a trained mixed objective function model, and calculating a classification score based on the trained SVM discrimination vector and the bias to determine a prediction category.
2. The adaptive dictionary learning method based on L0 regularization and mixed discriminant constraint of claim 1, wherein said mixed objective function model analysis processing steps are as follows: under the condition of fixed dictionary and SVM parameters, constructing an extended signal vector and an extended dictionary through mathematical transformation, and updating coding coefficients by utilizing an improved orthogonal matching pursuit algorithm; Under the condition of fixing coding coefficients and SVM parameters, updating a dictionary through a Lagrangian dual algorithm, and executing soft orthogonalization processing; Under the condition of fixed dictionary and coding coefficient, updating SVM discrimination vector and bias based on LBFGS algorithm; And monitoring and executing early shutdown by verifying the accuracy of the subset, and outputting optimal model parameters.
3. The adaptive dictionary learning method based on L0 regularization and mixed discriminant constraint of claim 2, wherein the dimension reduction process of principal component analysis in S1 comprises preserving variance information of more than 95% of image features, compressing feature dimensions to preset dimensions, normalizing each column vector of a sample matrix Execution of The data enhancement is to execute illumination disturbance on the small sample data set, expand the small sample data set to be 3 times of the number of the original image data set, divide the training subset and the verification subset according to 8:2, and keep the distribution of various samples consistent.
4. The adaptive dictionary learning method based on L0 regularization and mixed discriminant constraint of claim 3, wherein the step of constructing the class sub-dictionary in S2 comprises extracting training samples by principal component analysis The principal components are taken as atoms, and the mean value vector of the training sample is calculated and normalized to be taken as the first From a total of atoms Sub-dictionary of individual atom constituent classes , wherein, Expressed as a sequence of training samples, all class sub-dictionaries are concatenated to form a global initial dictionary 。
5. The adaptive dictionary learning method based on the L0 regularization and mixed discrimination constraint of claim 4, wherein the mathematical expression of the mixed objective function model in S3 is as follows: Wherein, the In order to train the sample, In order to be a dictionary of words, For the coding coefficients of the code, Is a discriminant vector for the SVM, Is biased; is the Frobenius norm of the matrix; Reconstructing an error term; is Fisher constraint term; the sum of matrix main diagonal elements; Is an intra-class divergence matrix; Is an inter-class divergence matrix; is a smooth regularization term coefficient; Is the total number of training samples; is an SVM discrimination item; in order to train the sample tag, Is a weight coefficient; a dissimilarity penalty term; is a penalty coefficient; And Coding coefficients of different categories are respectively represented; Index for training samples; indexing for all categories; for carrying out L2 norm normalization constraint on dictionary atoms, preventing the dictionary atom scale from scaling arbitrarily; for all atoms in the dictionary All are applicable; Is vector quantity I.e. the number of non-zero elements in the vector; for sparsity threshold, i.e. limiting at most retention in the encoded vector of each training sample A number of non-zero coefficients; To all training samples All are applicable.
6. The adaptive dictionary learning method based on L0 regularization and hybrid discrimination constraints of claim 5, wherein the step of updating the coding coefficients using the improved orthogonal matching pursuit algorithm in S4 is as follows: Constructing an extended signal vector Expanding dictionary Wherein, the signal vector is expanded From a mathematical derivation of the SVM penalty term, Mathematical derivation from Fisher constraint terms, expansion dictionary For a mathematically transformed SVM parameter matrix, For a Fisher constrained diagonal matrix, For the dissimilarity punishment of the related terms, The mixed objective function is converted into a standard sparse coding form Iterative selection of atomic update coefficients most relevant to residuals using orthogonal matching pursuit algorithm Until the number of non-zero coefficients reaches Or residual variation is less than a threshold; the specific construction parameters of the expansion signal vector and the expansion dictionary are as follows: signal component corresponding to SVM loss term ; Dictionary components ; Signal components corresponding to Fisher constraint terms ; Dictionary components , wherein, And solving by constructing a least square problem which converts the complex mixed constraint optimization problem into a standard.
7. The adaptive dictionary learning method based on the L0 regularization and hybrid discriminant constraint of claim 6, wherein the updating and soft orthogonalization of the dictionary in S4 comprises: Solving for Obtaining a new dictionary Applying an update amplitude limitation formula Preventing iteration vibration, presetting round number at each interval, and checking class sub-dictionary Performing QR decomposition to obtain an orthogonal matrix And utilize the fusion coefficient Update atoms: To enhance the linear independence of atoms.
8. The adaptive dictionary learning method based on L0 regularization and mixed discriminant constraint of claim 7, wherein the SVM parameter updating step in S4 comprises the steps of using current coding coefficient As input feature to train sample tags To achieve this, the minimum of the following SVM loss function is solved for updating by LBFGS algorithm And : ; Wherein, the Is a regularization parameter.
9. The adaptive dictionary learning method based on L0 regularization and mixed discrimination constraint of claim 8, wherein the early stop system comprises calculating classification accuracy of the current mixed objective function model on the training subset and the verification subset after each iteration is finished, determining convergence or fitting risk if the accuracy of the verification subset is not improved continuously for 5 iterations, immediately stopping the iteration, and using the dictionary corresponding to the highest verification accuracy Coding Discrimination vector Bias and method of making same Saved as final model parameters.
10. The adaptive dictionary learning method based on the L0 regularization and hybrid discrimination constraint of claim 9, wherein the step of classifying the test sample in S5 comprises the steps of: For an input test sample Dictionary completed by training And regularization coefficient The test coding coefficients are solved by an analytic solution or a least square method: ; Calculating a classification score vector Selecting And the class index corresponding to the element with the largest medium value is used as a prediction class, and the prediction class is compared with the real label to calculate the test accuracy.

Description

Self-adaptive dictionary learning method based on L0 regularization and mixed discrimination constraint Technical Field The invention relates to the technical field of computer vision and image processing, in particular to an adaptive dictionary learning method based on L0 regularization and mixed discrimination constraint. Background Dictionary learning is used as a core technology of sparse representation, the development of the dictionary learning always surrounds three targets of improving discriminant, reducing complexity and adapting to multiple scenes, an evolution path of unsupervised-supervised-multi-constraint collaboration is experienced, early dictionary learning (such as KSVD and LC-KSVD) takes reconstruction error minimization as a core, atoms are learned from data in an unsupervised mode, but obvious limitations exist, such as insufficient discriminant and atoms lack of category semantic information, the dictionary learning is only suitable for signal processing tasks such as denoising, compression and the like, the generalization capability is weak, and the dictionary learning is sensitive to small samples and high noise data. In the prior art, a structured dictionary learning method applying Fisher discriminant is provided, discriminant is improved through double constraint, reconstruction errors of similar samples by constraint class sub-dictionaries are minimized, intra-class aggregation is guaranteed, and then the ratio of inter-class divergence to intra-class divergence is maximized through Fisher criteria, so that inter-class separation is enhanced. In the method, in the tasks of face recognition (Extended Yaleb), target detection and the like, the classification accuracy is improved by 10-15% compared with KSVD, and a foundation for supervising dictionary learning is laid. However, FDDL relies on class labels to construct Fisher constraint during training, labels cannot be obtained during testing, sparse coding is performed by removing the constraint, so that model generalization capability is reduced, and for example, in polarized SAR ship detection, the accuracy is suddenly reduced by 8% -12% when clutter of a test sample overlaps with a target subspace. In SVGDl learning method, classification loss is fused into dictionary learning objective function for the first time, coding coefficient is guided to optimize to the direction of 'intra-class aggregation and inter-class separation' by means of finger loss, so that the dictionary has reconstruction capability and can be directly adapted to SVM classifier. On the Extended Yaleb dataset, the classification accuracy of SVGDL is improved by 3% -5% compared with FDDL%, and the reasoning speed is faster. However, the SVGDL only fuses SVM loss and reconstruction error, and does not consider constraints such as Fisher class divergence, non-target class dissimilarity and the like, so that class confusion is serious under complex scenes (such as 15 class Scene classification of Scene 15). In order to solve the problems, the invention provides an adaptive dictionary learning method based on L0 regularization and mixed discrimination constraint, which combines a range loss term and Fisher discrimination criterion and adds coefficient dissimilarity constraint terms. Not only is the inter-class separation strengthening of FDDL considered, but also the reconstruction capability of SVGDl on the dictionary is considered, and coefficient dissimilarity constraint items can be used for associating sub-dictionaries belonging to the same class, so that the discriminant of the dictionary is further enhanced. Finally, the L0 norm constraint is employed to encourage sample encoding to select mainly atoms of the relevant class. Disclosure of Invention Aiming at the defects existing in the prior art, the invention provides an adaptive dictionary learning method based on L0 regularization and mixed discrimination constraint. The invention aims to realize the self-adaptive dictionary learning method based on L0 regularization and mixed discrimination constraint, which comprises the following steps: S1, acquiring an original image data set to be classified, extracting image features, performing dimension reduction processing on the image features by utilizing principal component analysis, and reserving variance information of a preset proportion, performing L2 norm normalization of column vectors on a dimension-reduced sample matrix to obtain a small sample data set, performing illumination disturbance data enhancement on the small sample data set, expanding the number of the small sample data set to a preset multiple of the number of the original image data set, and dividing the small sample data set into a training subset and a verification subset according to the preset proportion; S2, extracting a main component and a class mean vector from the training sample, constructing a class sub-dictionary, and splicing to generate a global initial dictionary; S3, con