Image cross-domain recognition method based on relaxed domain adaptive width learning

Abstract

This invention proposes an image cross-domain recognition method based on relaxed-domain adaptive width learning, belonging to the technical fields of image recognition and transfer learning. The method includes: preprocessing and feature extraction of source and target domain image data; extracting width features of the source and target domains respectively through a width learning network and concatenating them to construct a joint width feature matrix; constructing an inter-domain difference metric matrix and transforming the joint width feature matrix to obtain a domain alignment term; constructing an orientation matrix and a non-negative drag matrix to establish a label enhancement matrix; constructing a relaxed-domain adaptive width learning model containing a domain alignment term, a source domain recognition reconstruction term, and a structural consistency term; and employing an alternating optimization strategy to achieve joint optimization, obtaining a domain alignment transformation matrix and a recognition projection matrix, and predicting the category of image samples in the target domain. This invention effectively alleviates the distribution offset problem in image cross-domain recognition and significantly improves recognition accuracy and the model's generalization ability.

Description

Technical Field

[0001] This invention relates to the technical field of image recognition and transfer learning, and in particular to an image cross-domain recognition method based on relaxed domain adaptive width learning. Background Technology

[0002] With the widespread application of artificial intelligence technology in fields such as medical imaging, industrial vision, and remote sensing analysis, cross-domain image recognition has gradually become a key technology for improving the generalization ability of models. In the field of cross-domain image recognition, due to the influence of different acquisition devices, imaging conditions, and environmental factors, image data from the source and target domains often have different feature distributions, causing traditional image recognition models based on the independent and identically distributed assumption to experience a significant performance degradation in the target domain. Therefore, how to effectively reduce the distributional differences between the source and target domains has become a core issue in cross-domain learning.

[0003] Broad Learning Systems (BLS) are widely used in various image recognition tasks due to their simple structure, high training efficiency, and ease of incremental updates. BLS achieves strong representational capabilities by constructing random feature nodes and augmentation nodes, and completes recognition modeling using analytical solutions. However, traditional BLS is often trained primarily on source domain image data, and its output weights typically handle both feature mapping and discriminative functions simultaneously. When there are distributional differences between the source and target domains, the model struggles to reduce inter-domain discrepancies while maintaining discriminative power, leading to a decline in target domain performance.

[0004] Furthermore, traditional BLS often employs fixed hard label encoding, resulting in a lack of flexibility in label representation. This makes it difficult to adapt to inter-class structural changes caused by cross-domain distribution shifts, further affecting the stability of image recognition boundaries. While existing deep domain adaptive methods can achieve cross-domain alignment, they typically rely on complex networks and iterative training, leading to high computational costs and making it difficult to simultaneously meet the structural controllability and efficient solution requirements emphasized by BLS. Therefore, it is necessary to introduce domain alignment and label relaxation mechanisms for cross-domain image recognition tasks while maintaining the simplicity and efficiency advantages of BLS. This would reduce inter-domain distribution differences while enhancing inter-class separability, thereby improving the model's generalization ability in the target domain.

[0005] Patent application number 202110743638.4 discloses a cross-domain image classification method based on unsupervised domain adaptation, comprising: S1: constructing source domain image sample data and target domain image sample data as training data; S2: normalizing the images in the training data; S3: inputting the images in the normalized training data into a domain adaptive network for training, wherein the domain adaptive network is used for cross-domain image classification; S4: during the training of the domain adaptive network, if the total loss value does not reach a preset convergence condition, iteratively updating the parameters of the domain adaptive network until the total loss value reaches the preset convergence condition, and recording the converged domain adaptive network as the trained domain adaptive classification model; S5: using the trained domain adaptive classification model for cross-domain image classification. This invention reduces the loss of important information during feature extraction, obtains more common features, and improves the accuracy of cross-domain image classification and the generalization ability of the image classification network model. However, the above methods still suffer from reduced recognition performance when the image distributions of the source and target domains differ significantly; at the same time, the fixed source domain labels limit the inter-class spacing; and the training process also suffers from high computational cost and low efficiency. Summary of the Invention

[0006] To address the performance degradation of traditional width learning methods when significant distribution differences exist between the source and target domains, this invention proposes a cross-domain image recognition method based on Relaxed Domain Adaptation Broad Learning System (RDABLS). By constructing a domain alignment mechanism based on Maximum Mean Discrepancy (MMD) and a label relaxation mechanism based on direction and drag matrices, cross-domain feature distribution alignment, flexible expansion of the label space, and enhanced discriminative ability are achieved, thereby effectively improving cross-domain image recognition performance and model generalization ability. This application solves the technical problems of cross-domain distribution bias, label limitations, and low training efficiency through domain alignment mechanisms, label relaxation techniques, and efficient solution strategies.

[0007] To achieve the above objectives, the technical solution of the present invention is implemented as follows: an image cross-domain recognition method based on relaxed domain adaptive width learning, the steps of which include:

[0008] Step 1: Preprocess and extract features from the source and target domain image data to obtain the training sample set for the source domain and the test sample set for the target domain.

[0009] Step 2: Extract the width features of the source and target domain image data respectively through a width learning network, and concatenate them to construct a joint width feature matrix;

[0010] Step 3: Construct an inter-domain difference metric matrix based on the maximum mean difference method to describe the deviation of the mean distribution between the source domain and the target domain. Introduce a domain alignment transformation matrix to transform the joint width feature matrix to obtain the domain alignment term.

[0011] Step 4: Construct the direction matrix and non-negative drag matrix to create the label enhancement matrix;

[0012] Step 5: Calculate the source domain recognition reconstruction term through the label enhancement matrix, establish the structural correlation between the recognition projection matrix and the domain alignment transformation matrix through the structural consistency matrix, calculate the structural consistency term, and construct a relaxed domain adaptive width learning model that includes the domain alignment term, the source domain recognition reconstruction term, and the structural consistency term.

[0013] Step 6: Use an alternating optimization strategy to jointly optimize the relaxed domain adaptive width learning model, obtain the domain alignment transformation matrix and the recognition projection matrix, and use the recognition transformation matrix to predict the image samples in the test sample set of the target domain to determine the category of the target domain samples.

[0014] Preferably, the preprocessing includes unifying the image size and normalizing the pixel value range; the feature extraction uses a VGG16 deep network to extract high-level semantic features of the image.

[0015] The VGG16 deep network is used to extract the source domain feature matrix from the preprocessed source domain image data, which is then used as the training sample set for the source domain. The VGG16 deep network is used to extract the preprocessed target domain image data to obtain the original feature matrix of the target domain, which serves as the test sample set for the target domain. .in, and These represent the number of samples in the source domain and the target domain, respectively. For source domain feature dimensions, The feature dimension of the target domain;

[0016] Construct a source domain label matrix based on the true category labels of the source domain image data. ;in, The total number of categories is represented by the actual category labels, which are encoded in binary form.

[0017] Preferably, the method for constructing the joint width feature matrix is ​​as follows:

[0018] The training sample set of the source domain The input width learning network is processed by concatenating randomly mapped nodes with augmentation nodes to obtain its width feature output, thus yielding the source domain width feature matrix. ;in, The number of samples in the source domain. The width feature dimension output by the width learning network;

[0019] Test sample set of the target domain The target domain samples are processed by a learning network of the same width to obtain the target domain width feature matrix. ;in, The number of samples in the target domain;

[0020] The source domain width feature matrix and the target domain width feature matrix are concatenated vertically to obtain the joint feature matrix. for:

[0021] ;

[0022] in, Represents the source domain width feature matrix. The width feature dimension; This represents the feature matrix representing the width of the target domain.

[0023] Preferably, the domain alignment term is ;in, Indicates size is The domain alignment transformation matrix, The feature dimension after domain alignment; Represents the trace operation of a matrix. This is the inter-domain difference measure matrix obtained based on the maximum mean difference method.

[0024] Preferably, the tag enhancement matrix ;in, The source domain label matrix is ​​in binary format. Represents the element-wise product of matrices. It is a non-negative drag matrix. This is the direction matrix for dragging the labels. This represents the total number of categories.

[0025] Preferably, the inter-domain difference metric matrix for:

[0026] ;

[0027] in, It is a column vector of all 1s. This is the transpose of the matrix;

[0028] The direction matrix The line, number Column elements: ;in, Represents the source domain label matrix The Middle line, number The numerical values ​​of the column elements;

[0029] The drag matrix The values ​​are obtained using random initialization, and each element is a non-negative value.

[0030] Preferably, the relaxed domain adaptive width learning model is

[0031] ,

[0032] in, This represents the recognition projection matrix that needs to be learned; Represents the structural consistency matrix. Representation domain alignment transformation matrix, The feature dimension after domain alignment; Represents the trace operation of a matrix. This is the inter-domain difference measurement matrix. For adjusting source domain identification reconstruction terms Weight in the overall objectives; For adjusting structural consistency terms The weights; Let F be the norm of the matrix. Reconstruction terms identified for the source domain For the joint characteristic matrix, This represents the source domain width feature matrix.

[0033] Preferably, the weights are determined using a grid search method based on the joint feature matrix. and The range of values; an alternating optimization strategy is used to optimize the recognition projection matrix respectively. Domain alignment transformation matrix Structural consistency matrix and non-negative drag matrix Update the solution:

[0034] Solving for the projection matrix Fixed-domain alignment transformation matrix Structural consistency matrix , Then, regarding the identification of the projection matrix The subproblem degenerates into a regularized least-squares regression, which yields the identification projection matrix:

[0035] ;

[0036] Solve the domain alignment transformation matrix : In a fixed recognition projection matrix Structural consistency matrix Then, the domain alignment transformation matrix The optimization subproblem consists of two parts: one is based on the inter-domain difference metric matrix. The first is the domain alignment term; the second is the structural consistency constraint term; the optimization subproblem is related to the domain alignment transformation matrix. Taking the partial derivatives and setting them to zero, we obtain the Sylvester-type matrix equations: ;

[0037] The Sylvester-type matrix equation is decomposed and transformed to obtain the domain alignment transformation matrix. for:

[0038] ;

[0039] in, The dimension is The identity matrix; Representation matrix The Column vector, and ;

[0040] Solving the drag matrix Fixed recognition projection matrix Introducing a residual term representing the deviation between the model prediction and the true label. ;in, The source domain's label matrix; to make the dragged labels closer to the prediction direction, drag the matrix. The subproblem is transformed into: ;

[0041] drag matrix The subproblem is solved by taking partial derivatives and combining them with nonnegativity constraints for projection. The drag matrix is:

[0042] ;

[0043] Solving the structural consistency matrix Fixed-domain alignment transformation matrix With recognition projection matrix Structural consistency moments It is a least-squares fitting problem, and the structural consistency moments are: ;

[0044] in, The dimension is The identity matrix; It is the regularization coefficient of the structural consistency term.

[0045] Preferably, for any image sample in the target domain The width feature vector is obtained through a width learning network. ,in The width feature dimension is used; the obtained recognition projection matrix is ​​then used. For width feature vector Perform a linear transformation to obtain the class response vector, and the class response vector's th... Each component represents the category. The predicted response value; the target domain sample is obtained by selecting the index with the largest predicted response value. The final predicted category.

[0046] Preferably, weight and The range of values ​​for are as follows: , ;

[0047] The method for decomposing and transforming Sylvester-type matrix equations is as follows: [The rest of the text appears to be incomplete and requires further context.] Perform eigenvalue decomposition:

[0048] ;

[0049] in, For matrix The corresponding eigenvector matrix, For matrix The eigenvalue diagonal matrix, For matrix The One eigenvalue;

[0050] Category response vector ,in, Total number of categories;

[0051] Final prediction category ;in, Indicates category The predicted response value, Indicates that the return value is a category Predicted response value The category number that yields the maximum value.

[0052] This invention proposes a novel cross-domain image recognition method by combining domain alignment and label enhancement mechanisms. First, image data from both the source and target domains are preprocessed and width features are extracted to construct a joint feature representation of the two domains. Then, a domain alignment mechanism is designed based on the maximum mean difference, introducing a domain alignment transformation matrix to reduce the distribution difference between the two domains in the image feature space. Next, a label enhancement mechanism is employed, constructing a direction matrix and a non-negative dragging matrix to moderately relax the source domain labels, thereby expanding the inter-class label spacing and improving the discriminative ability of image recognition. Finally, a recognition reconstruction term and a structural consistency term are designed to ensure the discriminative ability of the source domain and maintain the structural correlation between domain alignment and recognition transformation. The optimal discriminative transformation matrix is ​​obtained by jointly optimizing the model using an alternating minimization strategy and applied to the target domain image recognition task. Experimental results show that this invention effectively alleviates the distribution offset problem in cross-domain image recognition, significantly improving recognition accuracy and the model's generalization ability. The beneficial effects of this invention are:

[0053] (1) This invention effectively reduces the difference in image feature distribution between the source and target domains by designing a domain alignment mechanism for cross-domain image recognition, thereby significantly improving the transfer performance of cross-domain image recognition. In the specific implementation, a width learning network is first used to extract feature representations of image samples from the two domains. Then, a domain difference metric is constructed based on the maximum mean difference, and a domain alignment transformation matrix is ​​introduced to map the image features of the source and target domains to the same subspace, making the two domains more structurally consistent. This mapping process can effectively alleviate the cross-domain image distribution shift caused by factors such as imaging conditions and device differences, and improve the model's recognition generalization ability on the target domain image.

[0054] (2) This invention proposes a label enhancement mechanism for cross-domain image recognition, which reasonably relaxes the original image category labels in the source domain to enhance the separability between image categories and improve the model's discrimination ability in cross-domain scenarios. This mechanism constructs a direction matrix and a non-negative drag matrix to generate controlled offsets in the image feature space while maintaining the consistency of image semantic labels, thereby expanding the inter-class interval and enhancing the ability to distinguish difficult-to-identify image samples. In addition, through structural consistency constraints, the label enhancement and domain alignment processes work together, which is conducive to forming a clearer and more stable recognition decision boundary in cross-domain image data.

[0055] (3) This invention constructs an analytically friendly optimization solution framework, making the training process of the cross-domain image recognition model more efficient and stable, and suitable for large-scale cross-domain image recognition scenarios. In the joint objective, each variable is constructed to have a closed-form solution or can be directly solved by matrix equations, thereby avoiding complex gradient iteration processes, significantly improving the efficiency of image feature alignment and recognizer training, and reducing optimization difficulty. This optimization structure has good numerical stability and interpretability, is suitable for deployment in engineering environments, and meets the application requirements that are sensitive to computing resources and response speed. Attached Figure Description

[0056] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0057] Figure 1 This is a flowchart of the present invention.

[0058] Figure 2 These are example images from the cross-domain image recognition dataset (Office+Caltech10) used in this invention; where (a) is an Amazon domain image sample; (b) is a Caltech10 domain image sample; (c) is a DSLR domain image sample; and (d) is a Webcam domain image sample.

[0059] Figure 3 This is a flowchart illustrating the application of the present invention in an example of cross-domain image recognition. Detailed Implementation

[0060] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0061] To address the issue of inconsistent feature distributions caused by differences in imaging conditions in cross-domain image recognition, this invention proposes a cross-domain image recognition method based on relaxed domain adaptive width learning. This method extracts image feature representations from the source and target domains through a width learning network and combines domain difference measurement and domain alignment mechanisms to ensure that samples from different domains achieve a consistent distribution structure in a unified feature space. Simultaneously, label relaxation and structural consistency constraints are introduced to enable the label enhancement process and the domain alignment process to work synergistically, thereby improving the model's discriminative ability and cross-domain generalization performance. This invention provides an efficient, stable, and easily solvable solution for image recognition tasks in cross-domain scenarios. The specific implementation process is as follows: Figure 1 As shown.

[0062] Step 1: Preprocess and extract features from the source and target domain image data to obtain the training sample set for the source domain and the test sample set for the target domain.

[0063] Preprocessing includes unifying image size and normalizing pixel value ranges to reduce interference from different imaging conditions. Subsequently, a VGG16 deep network is used to extract high-level semantic features from the image to reduce low-level visual noise interference, enhance cross-domain robustness, and support heterogeneous domain data alignment and model generalization. Finally, the source domain feature matrix is ​​obtained as the training sample set. and the original feature matrix of the target domain as the test sample set .in, and These represent the number of samples in the source domain and the target domain, respectively. For source domain feature dimensions, This represents the feature dimension of the target domain. Simultaneously, a source domain label matrix is ​​constructed based on the true class labels of the source domain samples. .in, The total number of categories is represented by labels encoded in binary form; target domain samples do not contain labels and are only used for model prediction and evaluation.

[0064] Figure 2 The images shown are example images from the cross-domain image recognition dataset used in this invention. These image samples come from commonly used Office series cross-domain datasets, including image samples from four domains: Amazon (A), Caltech10 (C), DSLR (D), and Webcam (W). For example... Figure 2 As shown in (a), the Amazon domain images are mainly product images scraped from e-commerce platforms, containing rich and standardized shooting angles, but exhibiting significant stylistic differences; such as Figure 2 As shown in (b), the Caltech 10 domain images were taken in Caltech's natural environment, with complex backgrounds and large lighting variations, exhibiting obvious characteristics of a real scene; such as Figure 2As shown in (c), the DSLR domain is acquired by a high-resolution SLR camera, resulting in high image resolution and clear details, but with varied viewing angles; such as Figure 2 As shown in (d), the Webcam domain originates from low-resolution camera devices, exhibiting high noise levels and unstable lighting, resulting in significant low-quality imaging characteristics. These four domains differ significantly in imaging devices, lighting conditions, shooting angles, and background textures, constituting a typical cross-domain distribution shift scenario. This inter-domain difference leads to inconsistent feature representations of samples of the same category in different domains, making it difficult to directly transfer traditional recognition models.

[0065] Step 2: Extract the width features of the source and target domain image data respectively through a width learning network, and concatenate them to construct a joint width feature matrix for subsequent modeling.

[0066] First, the training sample set of the source domain The input width learning network is processed by concatenating randomly mapped nodes with augmentation nodes to obtain its width feature output, resulting in the source domain width feature matrix: .in, The number of samples in the source domain. The width feature dimension of the output of the width learning network; the source domain width feature matrix. Each row corresponds to the width feature vector of an image sample in the source domain. Similarly, the test sample set in the target domain... The target domain samples are processed by a learning network of the same width to obtain the target domain width feature matrix: .in, The number of samples in the target domain, the width of the target domain feature matrix Each row corresponds to the width feature vector of a target domain sample.

[0067] The two width feature matrices from the source and target domains are concatenated vertically to obtain the joint feature matrix. for:

[0068] ;

[0069] in, Represents the source domain width feature matrix. The number of samples in the source domain. To utilize the width feature dimension obtained from the width learning network; Represents the feature matrix of the target domain width. This represents the number of samples in the target domain. Since it's a vertical concatenation, the number of columns in the matrix remains unchanged, i.e., the width feature dimension. It remains unchanged. The joint width feature matrix... number of rows This is the sum of the number of samples in the source domain and the target domain.

[0070] By using a width-learning network to further extract high-level semantic features from the VGG16 deep network, the feature representation capability can be enhanced while preserving the original high-level semantic information. This process can also uncover potential structural relationships between features and improve the discriminative power of the features. A joint feature matrix containing information from both the source and target domains is constructed based on the features extracted again by the width-learning network, providing a unified feature representation foundation for subsequent inter-domain distribution alignment.

[0071] Step 3: Construct an inter-domain difference measure based on the maximum mean difference method. Introduce a domain alignment transformation matrix to transform the joint width feature matrix to obtain a domain alignment term, so as to minimize the feature distribution difference between the source domain and the target domain.

[0072] Domain alignment is achieved by minimizing the Maximum Mean Discrepancy (MMD), which measures the deviation of the mean distributions of the source and target domains in the width feature space. This aims to make the width feature distributions of the source and target domains as close as possible, thereby improving the model's generalization ability in the target domain. First, the inter-domain discrepancy measurement matrix... Defined as:

[0073] ,

[0074] in, It is a column vector consisting entirely of 1s; For size The inter-domain difference metric matrix is ​​used to describe the deviation of the mean distribution between the source domain and the target domain.

[0075] In the constructed inter-domain difference matrix Building upon this foundation, a domain alignment transformation matrix is ​​introduced, and the source and target domain distributions are aligned by minimizing the inter-domain difference (MMD). The MMD measures the difference in the mean distributions of samples from two domains in the feature space, and its general form can be expressed as the square norm of the difference between the mean vectors of the source and target domains after mapping. To facilitate matrix-based solutions, this difference is represented as a matrix based on the inter-domain difference metric. The quadratic form of the expression is obtained by introducing the field alignment transformation matrix. Afterwards, the MMD item can be represented as:

[0076] ,

[0077] in, Indicates size is The domain alignment transformation matrix, For the width joint feature dimension, The feature dimension after domain alignment; Representation matrix The trace operation, i.e., the sum of the elements on the diagonal of the matrix ( ); For the domain alignment transformation matrix The quadratic form of the difference between the mean vectors of the source and target domains is used. The smaller the value, the closer the means of the two domains are after mapping. Thus, distribution alignment is achieved to minimize the distribution difference between the two domains.

[0078] Step 4: Construct the direction matrix and the non-negative drag matrix to establish the label enhancement matrix, thereby relaxing the binary label constraints of the source domain and improving discriminability, thus enhancing the separability between categories.

[0079] First, define the direction matrix for tag dragging. for:

[0080]

[0081] in, Represents the source domain label matrix The Middle line, number The numerical values ​​of the column elements. To relax the fixed constraints of traditional binary labels, a non-negative drag-and-drop matrix is ​​introduced. The drag matrix It can be obtained by random initialization, and all elements are non-negative values, i.e. ,in This represents the total number of categories. This is achieved by introducing a matrix. This allows the label representation to be extended from a fixed binary form to an adjustable form. Therefore, the enhanced label enhancement matrix can be represented as:

[0082] ,

[0083] in, The source domain label matrix is ​​the original binary data. For the enhanced label enhancement matrix, This represents the element-wise product of matrices. In the technical implementation, it is achieved by dragging and dropping the direction matrix. Determine the adjustment direction of the tag elements, where the source domain tag matrix... Elements with a value of 1 are adjusted in the positive direction, and elements with a value of 0 are adjusted in the negative direction; this is combined with a non-negative drag matrix. By controlling the adjustment range of each element, the original binary label is expanded from a fixed form to an adjustable form, thereby increasing the inter-class interval and alleviating the discrimination limitations caused by the fixed binary label.

[0084] Step 5: Calculate the source domain recognition reconstruction term using the label enhancement matrix, establish the structural correlation between the recognition projection matrix and the domain alignment transformation matrix using the structural consistency matrix, calculate the structural consistency term, and construct a relaxed domain adaptive width learning model that includes the domain alignment term, the source domain recognition reconstruction term, and the structural consistency term.

[0085] To facilitate subsequent analysis, some key variables will be explained first: This represents the recognition projection matrix that needs to be learned; The structural consistency matrix is ​​used to establish a linear structural mapping between the domain-aligned feature space and the discriminant space, making the recognition matrix... Structurally aligned with domain transformation To maintain consistency, the final relaxed-domain adaptive width learning model for cross-domain image recognition is designed as follows:

[0086] ,

[0087] in, Used to adjust source domain identification reconstruction items The weights in the overall objective control the strength of the model's fitting to the source domain enhanced labels, thereby affecting the discriminative ability; Used to adjust structural consistency items Weights, constraint recognition projection matrix Domain-aligned transformation matrix The degree of consistency between them, in terms of the synergistic relationship in the stable domain alignment and discrimination process. Matrix The F-norm is defined as This is used to ensure the differentiability, analytical nature, and global structural constraints of the optimization objective. For the first term of the model... Indicates the field alignment item; Reconstruction terms for source domain identification are used to ensure that the model has sufficient discriminative ability and to avoid domain alignment causing the inter-class distance to decrease; This is a structural consistency term used to establish the recognition projection matrix. Domain-aligned transformation matrix The structural relationships between them maintain the stability of the identification structure while achieving domain alignment.

[0088] Step 6: Use an alternating optimization strategy to achieve joint training of the joint optimization objective of the relaxed domain adaptive width learning model, obtain the domain alignment transformation matrix and the recognition projection matrix, and use the recognition transformation matrix to predict the samples in the test sample set of the target domain to determine the category of the target domain samples.

[0089] This invention discloses the parameter value ranges of the model during specific implementation. The model mainly includes two key regularization hyperparameters, whose value ranges are as follows: , For different cross-domain recognition tasks, the corresponding source domain samples and target domain samples are preprocessed, feature extracted, and mapped using a width learning network to obtain a joint feature matrix. Based on the joint feature matrix, a grid search method is used to determine the optimal parameter configuration.

[0090] This invention employs an alternating optimization strategy to identify the projection matrix. Domain alignment transformation matrix Structural consistency matrix and nonnegative drag matrix Update the solution.

[0091] Solving for the projection matrix :fixed , , Then, regarding the identification of the projection matrix The subproblem degenerates into a regularized least squares regression, let the label enhancement matrix... The final solution is:

[0092] .

[0093] Solve the domain alignment transformation matrix : in fixed , Then, the domain alignment transformation matrix The optimization subproblem consists of two parts: one is based on the inter-domain difference metric matrix. The first is the MMD term, used to quantify the distribution difference between the source and target domains; the second is the structural consistency constraint term, used to maintain the recognition projection matrix. Domain-aligned transformation matrix The structural consistency of the subproblem. Regarding the domain alignment transformation matrix... Taking the partial derivatives and setting them to zero, we obtain the Sylvester-type matrix equations:

[0094] .

[0095] To directly solve the domain alignment transformation matrix The analytical solution, firstly for the matrix Perform eigenvalue decomposition:

[0096] ;

[0097] in, For matrix The corresponding eigenvector matrix, For matrix The eigenvalue diagonal matrix, For matrix The 1 eigenvalue, Substituting the original Sylvester equation into the above decomposition and transforming it, we can obtain the domain alignment transformation matrix. The direct matrix analytic expression:

[0098] ;

[0099] in, The dimension is The identity matrix. Representation matrix The Column vector, and .

[0100] Solving the drag matrix Fixed recognition projection matrix First, let's introduce a residual term:

[0101] ,

[0102] in, The label matrix of the source domain. This represents the deviation between the model's prediction and the true label. The goal of the drag matrix is ​​to make the dragged label closer to the predicted direction. The subproblem is transformed into:

[0103] ,

[0104] Considering drag matrix Let represent the drag distance. Utilizing the element-wise decoupling property of this objective function, we solve it by taking its partial derivative and combining it with non-negativity constraints through projection. Therefore, the final update form is:

[0105] .

[0106] Solving the structural consistency matrix :fixed With recognition projection matrix At that time, structural consistency moments The update is actually a simple least-squares fitting problem:

[0107] ,

[0108] in, The dimension is The identity matrix, The feature dimension after domain alignment; It is the regularization coefficient of the structural consistency term, preferably taken as... ,join in The matrix then becomes strictly positive definite, thus ensuring the stable existence of the closed-form solution.

[0109] Using the obtained recognition projection matrix Predict the category of a target domain sample. After model training, for any target domain sample... First, its width feature vector is obtained through a width learning network, denoted as: ,in The width feature dimension is then used. The recognition projection matrix obtained during training is then utilized. For width feature vector Perform a linear transformation to obtain the category response vector:

[0110] ,

[0111] in, For the total number of categories, the th category response vector Each component represents the model's relationship to the category. The predicted response value, in mathematical form, is:

[0112] ,

[0113] In the formula, The feature vector representing the sample width of the target domain The Each feature component Represents the recognition projection matrix The Middle line, number The weighted element of the column, used to measure the weight of the first column. The width feature for the first The contribution level of each category is considered. The final predicted category is obtained by selecting the index with the largest response value, expressed as:

[0114] ,

[0115] in, Indicates category The predicted response value, Indicates that the return value is a category Predicted response value Category number that gets the maximum value .therefore, For target domain samples The final predicted category.

[0116] In a specific embodiment, the maximum number of training rounds is set to 30, and the recognition accuracy based on the target domain test sample set is used as the main metric. By comparing the performance of the present invention with that of existing cross-domain image recognition and domain adaptation algorithms under the same experimental conditions, the recognition performance of the present invention on multiple cross-domain image recognition benchmark datasets, such as Office+Caltech10 and ImageNet+VOC2007, is evaluated.

[0117] Figure 3 The application process of this invention in a cross-domain image recognition example is demonstrated, and the specific implementation steps are as follows:

[0118] First, the source and target domain image datasets are preprocessed and feature extracted. Then, a joint width feature matrix containing information from both domains is constructed using a width learning network. Subsequently, a domain difference metric matrix is ​​constructed based on the maximum mean difference to align the distribution, and a direction matrix and a non-negative drag matrix are constructed using a label dragging technique to generate a label enhancement matrix for the source domain. The joint features, the domain difference metric matrix, and the enhanced labels are then introduced into a relaxed domain adaptive width learning framework to construct a joint optimization model containing a dual transformation matrix. After obtaining the optimal recognition projection matrix by solving the objective function through an alternating optimization strategy, the target domain test sample is input, and prediction is performed based on the optimal solution to output the recognition result.

[0119] This invention uses recognition accuracy as an evaluation metric for model performance, and its calculation formula is as follows:

[0120] ,

[0121] in, This represents the number of samples correctly identified by the model in the target domain test sample set. This indicates the total number of samples in the target domain test sample set. The value range of this metric is... The higher the value, the better the model's recognition performance in cross-domain tasks.

[0122] Furthermore, in order to comprehensively measure the overall generalization performance of the model on all cross-domain transfer tasks under a specific dataset, this invention uses average recognition accuracy as a comprehensive evaluation metric, and its calculation formula is as follows:

[0123] ,

[0124] in, This represents the total number of cross-domain migration tasks built under this dataset (e.g., in the Office+Caltech10 dataset). In the ImageNet+VOC2007 dataset, ); The model represents the first time. The recognition accuracy on a specific cross-domain task, here Number the task. This metric reflects the average performance of the model under different combinations of source and target domains. The higher the value, the stronger the robustness and stability of the model in varied cross-domain scenarios.

[0125] The specific composition and data partitioning of the dataset are as follows:

[0126] The Office+Caltech10 dataset contains four domains with different distribution characteristics: Amazon (A) from an online marketplace, Caltech10 (C) from a general image library, high-resolution images taken by a DSLR camera (D) and webcam images taken by a low-resolution webcam (W). In the experiment, a full permutation strategy was used to construct the transfer tasks. That is, each time, one of the four domains was selected as the source domain (training set) and the other as the target domain (test set). A total of 12 sets of image cross-domain recognition tasks were constructed, including: tasks with Amazon as the source domain: Amazon→Caltech10 (A→C), Amazon→DSLR (A→D), Amazon→Webcam (A→W); tasks with Caltech10 as the source domain: Caltech10→Amazon (C→A), Caltech10→DSLR (C→D), Caltech10→Webcam (C→W); tasks with DSLR as the source domain: DSLR→Amazon (D→A), DSLR→Caltech10 (D→C), DSLR→Webcam (D→W); and tasks with Webcam as the source domain: Webcam→Amazon (W→A), Webcam→Caltech10 (W→C), Webcam→DSLR (W→D).

[0127] For the COIL20 object recognition dataset: To evaluate the model's adaptability to changes in viewpoint, the dataset was divided according to the rotation angle of the object's image. The shooting angle was... and The image is divided into a subset CO1 (corresponding to the first and third quadrants), and the shooting angle is... and The images are divided into a subset CO2 (corresponding to the second and fourth quadrants). Based on this, two cross-view transfer tasks are constructed: CO1→CO2 and CO2→CO1.

[0128] For the ImageNet+VOC2007 dataset: a subset was constructed using five categories (birds, cats, dogs, people, and motorcycles) common to both datasets. Experiments used VOC2007 (V) and ImageNet (I) alternately as the source and target domains, respectively, to construct two large-scale cross-domain image recognition tasks: V→I and I→V, to evaluate the model's generalization ability under massive datasets.

[0129] Tables 1 and 2 present the performance comparison results of the RDABLS method of the present invention on the Office+Caltech10 image cross-domain recognition task dataset, compared with various advanced recognition methods. The methods compared in Table 1 include Transfer Component Analysis (TCA), Joint Distribution Adaptation (JDA), Cross-domain Extreme Learning Machines (CDELM), Cross-domain Convolutional Deep Belief Networks (CD-CDBN), and Transfer Extreme Learning Machine with Output Weight Alignment (TELM-OWA). Table 2 lists the BLS methods compared, including BLS, Semi-supervised Broad Learning System (SS-BLS), Subspace Distribution Adaptation based Broad Learning System (BLS-SDA), and Domain Adaptive Broad Learning System (DABLS).

[0130] Table 1 Comparison results of non-BLS methods on the Office+Caltech10 dataset

[0131]

[0132] Table 2 Comparison results of the BLS method on the Office+Caltech10 dataset

[0133]

[0134] Tables 3 and 4 show the performance comparison results of the RDABLS method of the present invention with various advanced recognition methods on the COIL20 image cross-domain recognition task dataset. The methods compared in Table 3 include TCA, JDA, CDELM, and CD-CDBN; Table 4 lists the BLS methods compared, including BLS, Automated Broad-Transfer Learning (AUBTL), SS-BLS, BLS-SDA, and DABLS.

[0135] Table 3 Comparison results of non-BLS methods on the COIL20 dataset

[0136]

[0137] Table 4 Comparison results of the BLS method on the COIL20 dataset

[0138]

[0139] Table 5 shows the performance comparison results of the RDABLS method of the present invention with a variety of advanced recognition methods on the ImageNet+VOC2007 image cross-domain recognition task dataset. The methods compared in Table 5 include BLS, SS-BLS, TCA, Optimal Discriminative Extreme Learning Machine (ODELM), and DABLS.

[0140] Table 5. Comparison of methods on the ImageNet+VOC2007 dataset

[0141]

[0142] As shown in Tables 1-5, the RDABLS method of this invention exhibits excellent recognition performance on all three cross-domain image recognition datasets, with its accuracy significantly outperforming other comparative methods. Specifically, in the Office+Caltech10 dataset (Tables 1 and 2), RDABLS achieves an average recognition accuracy of 49.78%, achieving the best results in most cross-domain tasks, and its recognition performance is significantly better than TCA, JDA, and the basic BLS model. In the COIL20 dataset (Tables 3 and 4), RDABLS achieves a recognition accuracy close to 90% (89.86%), performing best among all comparative methods. In the ImageNet+VOC2007 dataset (Table 5), RDABLS continues to lead, with an average accuracy of 75.27%, significantly higher than DABLS and SS-BLS. Overall, the results demonstrate that this invention possesses excellent generalization ability and stability when processing cross-domain image recognition data.

[0143] In summary, this invention proposes an image cross-domain recognition method based on relaxed domain adaptive width learning. By introducing a dual transformation matrix strategy, the domain distribution alignment process is decoupled from the discrimination modeling process, effectively avoiding the problem of mutual interference between alignment and discrimination targets in traditional single transformation scenarios. Simultaneously, a label enhancement matrix is ​​constructed using a label drag-and-drop technique, relaxing the strict binary label constraints and significantly expanding the inter-class margin. Experimental results demonstrate that this method effectively improves image recognition accuracy and model generalization ability in cross-domain scenarios while maintaining computational efficiency, providing an efficient and accurate solution for image cross-domain recognition tasks in complex environments.

[0144] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. 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. An image cross-domain recognition method based on relaxed domain adaptive width learning, characterized in that, The steps include: Step 1: Preprocess and extract features from the source and target domain image data to obtain the training sample set for the source domain and the test sample set for the target domain. Step 2: Extract the width features of the source and target domain image data respectively through a width learning network, and concatenate them to construct a joint width feature matrix; Step 3: Construct an inter-domain difference metric matrix based on the maximum mean difference method to describe the deviation of the mean distribution between the source domain and the target domain. Introduce a domain alignment transformation matrix to transform the joint width feature matrix to obtain the domain alignment term. Step 4: Construct the direction matrix and non-negative drag matrix to create the label enhancement matrix; Step 5: Calculate the source domain recognition reconstruction term through the label enhancement matrix, establish the structural correlation between the recognition projection matrix and the domain alignment transformation matrix through the structural consistency matrix, calculate the structural consistency term, and construct a relaxed domain adaptive width learning model that includes the domain alignment term, the source domain recognition reconstruction term, and the structural consistency term. Step 6: Use an alternating optimization strategy to jointly optimize the relaxed domain adaptive width learning model, obtain the domain alignment transformation matrix and the recognition projection matrix, and use the recognition transformation matrix to predict the image samples in the test sample set of the target domain to determine the category of the target domain samples.

2. The image cross-domain recognition method based on relaxed domain adaptive width learning according to claim 1, characterized in that, The preprocessing includes unifying the image size and normalizing the pixel value range; The feature extraction is performed by using a VGG16 deep network to extract high-level semantic features of the image. The VGG16 deep network is used to extract the source domain feature matrix from the preprocessed source domain image data, which is then used as the training sample set for the source domain. ; The VGG16 deep network is used to extract the preprocessed target domain image data to obtain the original feature matrix of the target domain, which serves as the test sample set for the target domain. .in, and These represent the number of samples in the source domain and the target domain, respectively. For source domain feature dimensions, The feature dimension of the target domain; Construct a source domain label matrix based on the true category labels of the source domain image data. ;in, The total number of categories is represented by the actual category labels, which are encoded in binary form.

3. The image cross-domain recognition method based on relaxed domain adaptive width learning according to claim 1 or 2, characterized in that, The method for constructing the joint width feature matrix is ​​as follows: The training sample set of the source domain The input width learning network is processed by concatenating randomly mapped nodes with augmentation nodes to obtain its width feature output, thus yielding the source domain width feature matrix. ;in, The number of samples in the source domain. The width feature dimension output by the width learning network; Test sample set of the target domain The target domain samples are processed by a learning network of the same width to obtain the target domain width feature matrix. ;in, The number of samples in the target domain; The source domain width feature matrix and the target domain width feature matrix are concatenated vertically to obtain the joint feature matrix. for: ； in, Represents the source domain width feature matrix. The width feature dimension; This represents the feature matrix representing the width of the target domain.

4. The image cross-domain recognition method based on relaxed domain adaptive width learning according to claim 3, characterized in that, The domain alignment item is ;in, Indicates size is The domain alignment transformation matrix, The feature dimension after domain alignment; Represents the trace operation of a matrix. This is the inter-domain difference measure matrix obtained based on the maximum mean difference method.

5. The image cross-domain recognition method based on relaxed domain adaptive width learning according to claim 4, characterized in that, The label enhancement matrix ;in, The source domain label matrix is ​​in binary format. Represents the element-wise product of matrices. It is a non-negative drag matrix. This is the direction matrix for dragging the labels. This represents the total number of categories.

6. The image cross-domain recognition method based on relaxed domain adaptive width learning according to claim 5, characterized in that, The inter-domain difference measurement matrix for: ； in, It is a column vector of all 1s. This is the transpose of the matrix; The direction matrix The line, number Column elements: ;in, Represents the source domain label matrix The Middle line, number The numerical values ​​of the column elements; The drag matrix The values ​​are obtained using random initialization, and each element is a non-negative value.

7. The image cross-domain recognition method based on relaxed domain adaptive width learning according to claim 5 or 6, characterized in that, The relaxed region adaptive width learning model is ， in, This represents the recognition projection matrix that needs to be learned; Represents the structural consistency matrix. Representation domain alignment transformation matrix, The feature dimension after domain alignment; Represents the trace operation of a matrix. This is the inter-domain difference measurement matrix. For adjusting source domain identification reconstruction terms Weight in the overall objectives; For adjusting structural consistency terms The weights; Let F be the norm of the matrix. Reconstruction terms identified for the source domain For the joint characteristic matrix, This represents the source domain width feature matrix.

8. The image cross-domain recognition method based on relaxed domain adaptive width learning according to claim 7, characterized in that, Weights are determined using a grid search method based on the joint feature matrix. and The range of values; an alternating optimization strategy is used to optimize the recognition projection matrix respectively. Domain alignment transformation matrix Structural consistency matrix and non-negative drag matrix Update the solution: Solving for the projection matrix Fixed-domain alignment transformation matrix Structural consistency matrix , Then, regarding the identification of the projection matrix The subproblem degenerates into a regularized least-squares regression, which yields the identification projection matrix: ； Solve the domain alignment transformation matrix : In a fixed recognition projection matrix Structural consistency matrix Then, the domain alignment transformation matrix The optimization subproblem consists of two parts: one is based on the inter-domain difference metric matrix. The first is the domain alignment term; the second is the structural consistency constraint term. The optimization subproblem with respect to the domain alignment transformation matrix Taking the partial derivatives and setting them to zero, we obtain the Sylvester-type matrix equations: ; The Sylvester-type matrix equation is decomposed and transformed to obtain the domain alignment transformation matrix. for: ； in, The dimension is The identity matrix; Representation matrix The Column vector, and ; Solving the drag matrix Fixed recognition projection matrix Introducing a residual term representing the deviation between the model prediction and the true label. ;in, The source domain's label matrix; to make the dragged labels closer to the prediction direction, drag the matrix. The subproblem is transformed into: ; drag matrix The subproblem is solved by taking partial derivatives and combining them with nonnegativity constraints for projection. The drag matrix is: ； Solving the structural consistency matrix Fixed-domain alignment transformation matrix With recognition projection matrix Structural consistency moments It is a least-squares fitting problem, and the structural consistency moments are: ; in, The dimension is The identity matrix; It is the regularization coefficient of the structural consistency term.

9. The image cross-domain recognition method based on relaxed domain adaptive width learning according to claim 8, characterized in that, For any image sample in the target domain The width feature vector is obtained through a width learning network. ,in The width feature dimension; Using the obtained recognition projection matrix For width feature vector Perform a linear transformation to obtain the class response vector, and the class response vector's th... Each component represents the category. The predicted response value; the target domain sample is obtained by selecting the index with the largest predicted response value. The final predicted category.

10. The image cross-domain recognition method based on relaxed domain adaptive width learning according to claim 9, characterized in that, Weight and The range of values ​​for are as follows: , ; The method for decomposing and transforming Sylvester-type matrix equations is as follows: [The rest of the text appears to be incomplete and requires further context.] Perform eigenvalue decomposition: ； in, For matrix The corresponding eigenvector matrix, For matrix The eigenvalue diagonal matrix, For matrix The One eigenvalue; Category response vector ,in, Total number of categories; Final prediction category ;in, Indicate category The predicted response value, Indicates that the return value is a category Predicted response value The category number that yields the maximum value.

Images