Textile multi-modal feature selection method and system based on hessian manifold regularization

Abstract

The application discloses a textile multi-modal feature selection method and system based on a Hessian manifold regularization, and relates to the technical field of computer vision, which comprises the following steps: constructing multi-modal data by a jacquard chart image and a process single parameter, and obtaining each modality pseudo label matrix through subspace projection; introducing L2, 1 norm sparse constraint and Hessian manifold regularization to maintain local curvature structure and improve sparsity; generating cross-modal shared pseudo labels by using adaptive weighted fusion, and optimizing in cooperation to enhance consistency; further tensorizing the pseudo labels, applying tensor kernel norm low rank constraint, and taking into account modality commonality and specificity; finally, all components are optimized in a unified framework, and a high discriminability and structure robust feature subset is output. The application realizes efficient feature selection without manual annotation by fusing subspace learning, Hessian manifold regularization and tensor kernel norm analysis.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, specifically to a method and system for selecting multimodal features of textiles based on Hessian manifold regularization. Background Technology

[0002] With the intelligent development of the textile industry, textile quality inspection and classification are gradually shifting from traditional manual inspection to automation and intelligence. In actual production, textile data typically exists in multiple modalities, such as Jacquard images, process parameters, and texture features. These multimodal data describe the characteristics of textiles from different perspectives and contain rich complementary information. How to effectively integrate and utilize this multimodal data has become a key issue in the field of intelligent classification and quality control of textiles.

[0003] However, multimodal textile data are typically high-dimensional and highly redundant, containing a large number of noisy features irrelevant to the classification task. These redundant features not only increase computational complexity but may also degrade the performance of classification models. Therefore, feature selection, as an effective data preprocessing technique, is of great significance in multimodal textile data analysis. Feature selection aims to choose the most discriminative subset of features from the original high-dimensional feature space to improve the efficiency and accuracy of subsequent learning tasks.

[0004] Currently, feature selection methods are mainly divided into three categories: supervised, semi-supervised, and unsupervised. Supervised feature selection methods rely on sufficient label information to guide feature selection. However, in actual textile production scenarios, obtaining a large number of accurate labels often requires manual annotation by professionals, which is costly and time-consuming. Therefore, unsupervised feature selection methods have attracted widespread attention due to their advantage of not relying on label information.

[0005] Most existing unsupervised feature selection methods are designed based on single-modal data, making it difficult to effectively handle the correlation and complementarity between multimodal data. While some multimodal feature selection methods consider the relationships between different modalities, they suffer from the following shortcomings: First, they fail to fully exploit the manifold structure information of the data, ignoring the geometric characteristics of the data distribution; second, while maintaining the consistency of multimodal data, they struggle to effectively preserve the specific information of each modality; third, traditional graph Laplacian regularization methods can only capture first-order local information of the data, failing to accurately characterize the curvature properties of complex manifolds. Summary of the Invention

[0006] To address the aforementioned issues, this invention proposes a method and system for selecting multimodal features of textiles based on Hessian manifold regularization. By coupling Hessian manifold regularization with sparse low-rank multimodal collaborative learning, this invention achieves accurate selection of key features that preserve the structure of multi-source heterogeneous features of textiles, maintain cross-modal consistency, and are physically interpretable.

[0007] On the one hand, a multimodal feature selection method for textiles based on Hessian manifold regularization includes:

[0008] S1. Generate multimodal textile data based on Jacquard images and process parameters of textiles. Construct feature matrices for each modality based on the multimodal textile data. Learn the low-dimensional representation of the feature matrices for each modality through subspace projection to obtain the pseudo-label matrix for each modality.

[0009] S2, apply L2,1 norm sparse regularization to the pseudo-label matrix of each mode and introduce Hessian manifold regularization term to capture the local curvature information of the data and obtain pseudo-label matrix that maintains the manifold structure continuity and sparseness for each mode.

[0010] S3 integrates the continuous and sparse pseudo-label matrices of each modality that maintain the manifold structure through an adaptive weighting mechanism. Based on the continuous and sparse pseudo-label matrices of each modality that maintain the manifold structure, a cross-modal shared pseudo-label matrix is ​​generated. The cross-modal shared pseudo-label matrix and the preset shared matrix are learned collaboratively to obtain a pseudo-label matrix with the representation of information of each modality.

[0011] S4 introduces a tensor kernel norm to impose a low-rank constraint on the pseudo-label matrix with modal information representation, and after maintaining cross-modal consistency while preserving the specific information of each modality, constructs a textile multimodal feature selection objective function based on Hessian manifold regularization to output the final feature selection result.

[0012] Furthermore, by learning the low-dimensional representation of the feature matrix of each modality through subspace projection, the pseudo-label matrix of each modality is obtained, and the calculation formula is as follows:

[0013] ;

[0014] in, The feature matrix representing each mode; Represents the subspace projection matrix; This represents the pseudo-label matrix generated for each modality; Represents the Frobenius norm; This represents the total number of all modes; This represents a modal index.

[0015] Furthermore, the formula for calculating the pseudo-label matrix that maintains the continuity and sparseness of the manifold structure for each mode is as follows:

[0016] ;

[0017] in, This represents the Hessian popularization regularization weighting coefficient; Represents the Hessian popular regularization matrix; Indicates the trade-off parameters; This represents a consensus pseudo-label matrix that maintains a continuous and sparse manifold structure across all modes. This represents the sum of the diagonal elements of the matrix.

[0018] Furthermore, a cross-modal shared pseudo-label matrix is ​​generated based on the pseudo-label matrix that maintains the continuity and sparseness of the manifold structure for each modality. The calculation formula is as follows;

[0019] ;

[0020] ;

[0021] in, st represents the constraint condition; Represents the identity matrix; This is a pseudo-label matrix shared across modalities.

[0022] Furthermore, a low-rank constraint is imposed on the pseudo-label matrix containing modal information by introducing the tensor nuclear norm. The calculation formula is as follows:

[0023] ;

[0024] ;

[0025] in, ; Represents a mapping function; These represent pseudo-label matrices that represent each modal information. express Tensors stacked along three dimensions.

[0026] Furthermore, a multimodal feature selection objective function for textiles based on Hessian manifold regularization is constructed to output the final feature selection result. The calculation formula of the objective function is as follows:

[0027] ;

[0028] .

[0029] On the other hand, a textile multimodal feature selection system based on Hessian manifold regularization includes:

[0030] The pseudo-label matrix construction module is used to generate multimodal textile data based on Jacquard images and process parameters of textiles, construct feature matrices for each modality based on the multimodal textile data, learn the low-dimensional representation of the feature matrices of each modality through subspace projection, and obtain the pseudo-label matrix of each modality.

[0031] The Hessian manifold regularization module is used to apply L2,1 norm sparsity regularization to the pseudo-label matrix of each mode and introduces the Hessian manifold regularization term to capture the local curvature information of the data, so as to obtain a pseudo-label matrix that maintains the continuity of the manifold structure and is sparse for each mode.

[0032] The cross-modal shared pseudo-label matrix construction module is used to integrate the continuous and sparse pseudo-label matrices of each modality that maintain the manifold structure through an adaptive weighting mechanism. Based on the continuous and sparse pseudo-label matrices of each modality that maintain the manifold structure, a cross-modal shared pseudo-label matrix is ​​generated. The cross-modal shared pseudo-label matrix and the preset shared matrix are learned collaboratively to obtain a pseudo-label matrix with information representation of each modality.

[0033] The feature selection module is used to introduce tensor kernel norms to impose low-rank constraints on pseudo-label matrices that represent each modality. After maintaining cross-modal consistency and preserving the specific information of each modality, it constructs a textile multimodal feature selection objective function based on Hessian manifold regularization to output the final feature selection result.

[0034] The present invention adopts the above technical solution and has the following beneficial effects:

[0035] (1) This invention performs subspace projection on multimodal textile data such as Jacquard images and process parameters, and extracts pseudo-label matrices for each modality, effectively capturing deep correlation information between multimodal data, thereby significantly improving the accuracy and efficiency of feature extraction;

[0036] (2) This invention employs a feature selection model with L2,1 norm sparsity regularization and Hessian manifold regularization. While ensuring the sparsity of the selected features, it effectively captures the local curvature information of the data and maintains the continuity of the manifold structure. It is suitable for processing textile data with complex geometric structures.

[0037] (3) This invention integrates the pseudo-label matrices of each modality through an adaptive weighting mechanism to generate a cross-modal shared pseudo-label matrix, and learns in collaboration with the shared matrix. While maintaining cross-modal consistency, it also maintains the specific information of each modality, thereby achieving consistent representation of multimodal information.

[0038] (4) This invention applies a low-rank constraint to the pseudo-label matrix by introducing a tensor kernel norm, which effectively mines the deep correlation between multiple modalities, reduces data dimensionality and computational complexity, and improves the accuracy of intelligent classification and process matching of textiles.

[0039] (5) This invention effectively integrates subspace learning, Hessian manifold regularization and tensor nuclear norm analysis techniques, achieving efficient feature selection without manual annotation, significantly reducing data dimensionality and computational complexity, and providing technical support for intelligent textile manufacturing. Attached Figure Description

[0040] Figure 1 This is a flowchart of the multimodal feature selection method for textiles based on Hessen manifold regularization, as described in an embodiment of the present invention.

[0041] Figure 2 This is a diagram of a textile multimodal feature selection system based on Hessen manifold regularization, according to an embodiment of the present invention. Detailed Implementation

[0042] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto.

[0043] like Figure 1 As shown, the present invention provides a method for selecting multimodal features of textiles based on Hessian manifold regularization, comprising:

[0044] S1 generates multimodal textile data based on Jacquard images and process parameters of textiles. Based on the multimodal textile data, a feature matrix of each modality is constructed. The low-dimensional representation of the feature matrix of each modality is learned through subspace projection to obtain the pseudo-label matrix of each modality.

[0045] Specifically, the low-dimensional representation of the feature matrix of each modality is learned through subspace projection, and the pseudo-label matrix of each modality is obtained. The calculation formula is as follows:

[0046] ;

[0047] in, The feature matrix representing each mode; Represents the subspace projection matrix; This represents the pseudo-label matrix generated for each modality; Represents the Frobenius norm; This represents the total number of all modes; This represents a modal index.

[0048] Specifically, in this embodiment, a subspace projection matrix is ​​generated for the data of each view. This subspace projection matrix is ​​then used to project the data of each view, thereby generating a pseudo-label matrix. In multimodal data analysis of textiles, data from different modalities have different feature spaces and data distributions. To achieve effective feature selection, it is first necessary to project the data of each modality into a unified low-dimensional subspace. Specifically, let the original data matrix of the v-th view be... ,in Let represent the feature dimension of the v-th view, and n represent the number of samples. This is achieved by learning the subspace projection matrix. The original high-dimensional data is projected onto a c-dimensional low-dimensional space to generate a pseudo-label matrix. , where c represents the number of clusters or the feature dimension after dimensionality reduction. By minimizing the reconstruction error between the projected data and the pseudo-label matrix, a pseudo-label matrix that can effectively characterize the essential structure of each modality's data can be obtained, laying the foundation for subsequent cross-modal information fusion.

[0049] S2 applies L2,1 norm sparse regularization to the pseudo-label matrix of each mode and introduces Hessian manifold regularization term to capture the local curvature information of the data, thereby obtaining a pseudo-label matrix that maintains the continuity and sparseness of the manifold structure for each mode.

[0050] Specifically, the formula for calculating the pseudo-label matrix that maintains the continuity and sparseness of the manifold structure for each mode is as follows:

[0051] ;

[0052] in, This represents the Hessian popularization regularization weighting coefficient; Represents the Hessian popular regularization matrix; Indicates the trade-off parameters; This represents a consensus pseudo-label matrix that maintains a continuous and sparse manifold structure across all modes. This represents the sum of the diagonal elements of the matrix.

[0053] Specifically, in actual textile data, there are often a large number of redundant and noisy features. These features not only increase the computational burden but may also affect classification performance. To achieve effective feature selection, this invention adjusts the projection matrix... Imposing L2,1 norm constraints induces row sparsity in the projection matrix, thereby enabling automatic feature selection. Furthermore, traditional graph Laplacian regularization can only capture first-order local information of the data manifold, making it difficult to accurately characterize the curvature properties of complex manifold structures. To more accurately preserve the manifold structure of the data, this invention introduces a Hessian manifold regularization term. Hessian manifold regularization captures the local curvature of the manifold through second-order derivative information, enabling a more accurate description of the data's geometric structure.

[0054] S3 integrates the pseudo-label matrices of each modality that maintain the continuity and sparseness of the manifold structure through an adaptive weighting mechanism. Based on the pseudo-label matrices of each modality that maintain the continuity and sparseness of the manifold structure, a cross-modal shared pseudo-label matrix is ​​generated. The cross-modal shared pseudo-label matrix and the preset shared matrix are learned collaboratively to obtain a pseudo-label matrix with the representation of information of each modality.

[0055] Specifically, a cross-modal shared pseudo-label matrix is ​​generated based on a pseudo-label matrix that maintains the continuity and sparseness of the manifold structure for each modality. The calculation formula is as follows;

[0056] ;

[0057] ;

[0058] in, st represents the constraint condition; Represents the identity matrix; This is a pseudo-label matrix shared across modalities.

[0059] Specifically, in this embodiment, an adaptive learning mechanism is used to integrate the pseudo-label matrices of each modality to generate a cross-view shared pseudo-label matrix. The pseudo-label matrices of each modality and the shared matrix learn collaboratively to achieve a consistent representation of multimodal information, and orthogonal and non-negativity constraints are applied to further enhance the interpretability of the embedding.

[0060] By minimizing the pseudo-label matrix of each modality With shared pseudo-label matrix By addressing the differences between modalities and achieving consistent representation of multimodal information, an adaptive learning mechanism is used to automatically adjust the contribution weight of each modality in shared representation learning based on its data quality and importance, thereby achieving more robust multimodal information fusion.

[0061] S4 introduces a tensor kernel norm to impose a low-rank constraint on the pseudo-label matrix with modal information representation, and after maintaining cross-modal consistency while preserving the specific information of each modality, constructs a textile multimodal feature selection objective function based on Hessian manifold regularization to output the final feature selection result.

[0062] Specifically, a low-rank constraint is imposed on the pseudo-label matrix containing modal information by introducing the tensor nuclear norm. The calculation formula is as follows:

[0063] ;

[0064] ;

[0065] in, ; Represents a mapping function; These represent pseudo-label matrices that represent each modal information. express Tensors stacked along three dimensions.

[0066] Specifically, a multimodal feature selection objective function for textiles based on Hessian manifold regularization is constructed to output the final feature selection result. The calculation formula of the objective function is as follows:

[0067] ;

[0068] .

[0069] Specifically, in this embodiment, a tensor kernel norm is introduced to impose a low-rank constraint on the pseudo-label matrix, maintaining cross-modal consistency while preserving the specific information of each modality. Although a shared pseudo-label matrix can capture consistency information between multiple modalities, data from different modalities often possess their own unique specific information. To maintain consistency while preserving specificity, a tensor kernel norm is introduced to impose a low-rank constraint on the pseudo-label matrix. As a convex relaxation of the tensor rank, the tensor kernel norm can effectively mine the low-rank structure between multimodal data, allowing each modality to retain its specific information while maintaining cross-modal consistency. Through the tensor low-rank constraint, this invention can effectively balance the consistency and specificity of multimodal data, avoiding excessive information fusion leading to the loss of specific information, and also avoiding excessive separation leading to the ineffective utilization of consistency information. Finally, the tensor kernel norm, subspace learning, and Hessian regularization are embedded in the same framework to output the final feature selection result.

[0070] Specifically, the unsupervised feature selection method for multimodal textile data based on Hessian manifold regularization of this invention achieves efficient feature extraction from multimodal textile data through the application of subspace projection learning technology, constructing a high-performance and computationally efficient model. It effectively integrates complementary information from multiple modalities, such as Jacquard images and process parameters, while balancing cross-modal consistency and modal specificity through an adaptive weighting mechanism. It combines L2,1 norm and Hessian manifold regularization to enhance feature sparsity, capturing local curvature information of the data while maintaining manifold structure continuity. The introduction of tensor kernel norm to apply low-rank constraints effectively mines deep correlations between multimodalities, prevents overfitting, and improves the model's stability and generalization ability.

[0071] As shown in Table 1, this invention selected seven publicly available multi-view benchmark datasets for experimental verification, including: handwritten digit dataset, face image dataset, Yale face dataset, outdoor scene recognition dataset, web page text dataset, image segmentation dataset containing 7 types of objects, and object recognition dataset. To comprehensively evaluate the feature selection effect of the proposed method, this invention uses clustering accuracy and normalized mutual information (NMI) as performance evaluation metrics. Specifically, the ACC metric measures the degree of matching between the clustering results and the true labels, with a value range of [0,1], where a larger value indicates a better clustering effect; the NMI metric evaluates the clustering quality by calculating the mutual information between the cluster partitions and the true partitions, also with a value range of [0,1]. When NMI is 0, it indicates that the two partitions are completely independent; when NMI is 1, it indicates that the two partitions are completely consistent.

[0072] Table 1. Comparison of measurements of normalized mutual information;

[0073]

[0074] As shown in Table 1, the method of the present invention ranks highly in the evaluation based on the NMI index on seven datasets. It has excellent performance and stability in the multi-view feature selection task in the textile industry and can effectively extract key information and perform feature selection. The underlined data represents the suboptimal index, and the bolded part represents the optimal index.

[0075] Table 2 Comparative analysis of accuracy measurements;

[0076]

[0077] Furthermore, as shown in Table 2, the method corresponding to this invention performed excellently in evaluations based on specific metrics across seven datasets, achieving the highest average value. Therefore, it can be proven that the textile multimodal feature selection method based on Hessian manifold regularization proposed in this application, through the application of subspace projection learning technology, achieves efficient feature extraction from multimodal textile data, constructing a high-performance and computationally efficient model. It effectively integrates complementary information from multiple modalities, such as Jacquard images and process single parameters, while balancing cross-modal consistency and modal specificity through an adaptive weighting mechanism. Combining L2,1 norm and Hessian manifold regularization enhances feature sparsity, capturing local curvature information of the data while maintaining manifold structure continuity. Introducing tensor kernel norm to apply low-rank constraints effectively mines deep correlations between multimodalities, prevents overfitting, and improves model stability and generalization ability, demonstrating the promising prospects of innovatively applying Hessian manifold regularization to unsupervised multi-view feature selection tasks. In this paper, underlined data represents suboptimal metrics, and bolded data represents optimal metrics.

[0078] like Figure 2 As shown, this embodiment also discloses a textile multimodal feature selection system based on Hessian manifold regularization, including:

[0079] The pseudo-label matrix construction module 21 is used to generate multimodal textile data based on the Jacquard image of textiles and process single parameters, construct feature matrices of each modality based on the multimodal textile data, learn the low-dimensional representation of the feature matrices of each modality through subspace projection, and obtain the pseudo-label matrix of each modality.

[0080] Hessian manifold regularization module 22 is used to apply L2,1 norm sparsity regularization to the pseudo-label matrix of each mode and introduce Hessian manifold regularization term to capture the local curvature information of the data and obtain pseudo-label matrix that maintains the continuity and sparsity of the manifold structure for each mode.

[0081] The cross-modal shared pseudo-label matrix construction module 23 is used to integrate the continuous and sparse pseudo-label matrices of each modality that maintain the manifold structure through an adaptive weighting mechanism, generate the cross-modal shared pseudo-label matrix based on the continuous and sparse pseudo-label matrices of each modality, and learn the cross-modal shared pseudo-label matrix and the preset shared matrix together to obtain the pseudo-label matrix with the information representation of each modality.

[0082] Feature selection module 24 introduces tensor kernel norm to impose low-rank constraints on pseudo-label matrices with modal information representations, and after maintaining cross-modal consistency while preserving the specific information of each modality, constructs a textile multimodal feature selection objective function based on Hessian manifold regularization to output the final feature selection result.

[0083] The specific implementation of the textile multimodal feature selection system based on Hessian manifold regularization is the same as the textile multimodal feature selection method based on Hessian manifold regularization, and will not be described again in this embodiment.

[0084] Although the invention has been specifically shown and described in conjunction with preferred embodiments, those skilled in the art will understand that various changes in form and detail may be made to the invention without departing from the spirit and scope of the invention as defined in the appended claims, all of which shall be within the scope of protection of the invention.

Claims

1. A method for selecting multimodal features of textiles based on Hessian manifold regularization, characterized in that, Includes the following steps: S1. Multimodal textile data is generated based on Jacquard images and process parameters of textiles. Feature matrices for each modality are constructed based on the multimodal textile data. Low-dimensional representations of the feature matrices for each modality are learned through subspace projection to obtain the pseudo-label matrix for each modality. The calculation formula is as follows: ； in, The feature matrix representing each mode; Represents the subspace projection matrix; This represents the pseudo-label matrix generated for each modality; Represents the Frobenius norm; This represents the total number of all modes; Indicates modal index; S2. Apply L2,1 norm sparse regularization to the pseudo-label matrix of each mode and introduce a Hessian manifold regularization term to capture the local curvature information of the data, thereby obtaining a pseudo-label matrix that maintains the continuity of the manifold structure and is sparse for each mode. The calculation formula is as follows: ； in, This represents the Hessian popularization weighting coefficient; Represents the Hessian popular regularization matrix; Indicates the trade-off parameters; This represents a consensus pseudo-label matrix that maintains a continuous and sparse manifold structure across all modes. This represents the sum of the elements on the diagonal of the matrix; S3 integrates the pseudo-label matrices of each mode that maintain the continuity and sparseness of the manifold structure through an adaptive weighting mechanism, and generates a cross-modal shared pseudo-label matrix based on the pseudo-label matrices of each mode that maintain the continuity and sparseness of the manifold structure. The calculation formula is as follows. ； ； in, st represents the constraint condition; Represents the identity matrix; The pseudo-label matrix is ​​shared across modalities; the pseudo-label matrix shared across modalities and the preset shared matrix are learned collaboratively to obtain a pseudo-label matrix with representations of information from each modality. S4, introduce the tensor nuclear norm to impose a low-rank constraint on the pseudo-label matrix that represents each modality, and the calculation formula is as follows: ； ； in, ; Represents a mapping function; These represent pseudo-label matrices that represent each modal information. express Tensors stacked along three dimensions; After maintaining cross-modal consistency and preserving the specific information of each modality, a multimodal feature selection objective function for textiles based on Hessian manifold regularization is constructed to output the final feature selection result. The calculation formula of the objective function is as follows: ； 。 2. A multimodal feature selection system for textiles based on Hessian manifold regularization, characterized in that, include: The pseudo-label matrix construction module is used to generate multimodal textile data based on Jacquard images and process parameters of textiles. It constructs feature matrices for each modality based on the multimodal textile data, learns the low-dimensional representation of the feature matrices for each modality through subspace projection, and obtains the pseudo-label matrix for each modality. The calculation formula is as follows: ； in, The feature matrix representing each mode; Represents the subspace projection matrix; This represents the pseudo-label matrix generated for each modality; Represents the Frobenius norm; This represents the total number of all modes; Indicates modal index; The Hessian manifold regularization module applies L2,1 norm sparsity regularization to the pseudo-label matrix of each mode and introduces a Hessian manifold regularization term to capture the local curvature information of the data, obtaining a pseudo-label matrix that maintains the continuity and sparseness of the manifold structure for each mode. The calculation formula is as follows: ； in, This represents the Hessian popularization weighting coefficient; Represents the Hessian popular regularization matrix; Indicates the trade-off parameters; This represents a consensus pseudo-label matrix that maintains a continuous and sparse manifold structure across all modes. This represents the sum of the elements on the diagonal of the matrix; The cross-modal shared pseudo-label matrix construction module is used to integrate the continuous and sparse pseudo-label matrices of each modality through an adaptive weighting mechanism. Based on the continuous and sparse pseudo-label matrices of each modality, a cross-modal shared pseudo-label matrix is ​​generated. The calculation formula is as follows. ； ； in, st represents the constraint condition; Represents the identity matrix; The pseudo-label matrix is ​​shared across modalities; the pseudo-label matrix shared across modalities and the preset shared matrix are learned collaboratively to obtain a pseudo-label matrix with representations of information from each modality. The feature selection module introduces the tensor kernel norm to impose a low-rank constraint on the pseudo-label matrix containing modal information representations. The calculation formula is as follows: ； ； in, ; Represents a mapping function; These represent pseudo-label matrices that represent each modal information. express Tensors stacked along three dimensions; After maintaining cross-modal consistency and preserving the specific information of each modality, a multimodal feature selection objective function for textiles based on Hessian manifold regularization is constructed to output the final feature selection result. The calculation formula of the objective function is as follows: ； 。

Images