An image texture classification method and system

Abstract

The present invention discloses an image texture classification method and system, comprising the following steps: extracting image texture features from an original image to form a feature vector of the original image; using a Gaussian kernel function to map the feature vector of the original image to a high-dimensional feature space to obtain a feature expression of the high-dimensional space; mapping the feature vector of the original image to a low-dimensional space by a random mapping method to obtain a feature expression of the low-dimensional space; fusing the feature expression of the high-dimensional space with the feature expression of the low-dimensional space to form a comprehensive feature representation; based on the comprehensive feature representation, using a classifier to classify the image texture. The present invention comprehensively applies high-dimensional space expression and low-dimensional space expression, fully utilizes the advantages of features of different dimensions, and improves the accuracy and efficiency of texture classification. At the same time, the present invention also has strong versatility and scalability, and can adapt to texture classification needs in different industries and scenarios.

Description

Technical Field

[0001] The present invention belongs to the fields of image processing, artificial intelligence, and pattern recognition, and particularly relates to an image texture classification method and system. Background Art

[0002] With the rapid development of computer vision and artificial intelligence technologies, image processing technologies are increasingly widely applied in various industries. Among them, image texture classification, as a key branch in the field of image processing, plays a crucial role in material recognition, product quality control, and automated production. Texture, as a basic attribute of an image, can intuitively reflect the structural information and visual characteristics of the object surface. Therefore, accurate and efficient classification of image textures has always been a research hotspot in the field of image processing.

[0003] Traditional image texture classification methods mainly rely on manually designed feature extractors, such as gray-level co-occurrence matrices, Gabor filter features, local binary patterns, etc. Although these traditional image texture classification methods have achieved certain results, they have problems such as incomplete feature extraction, limited classification performance, and weak generalization ability. With the rise of machine learning and deep learning technologies, learning-based image texture classification methods have gradually received the favor of researchers. These methods significantly improve the performance and generalization ability of texture classification by automatically learning the deep features of images. However, such methods often require a large amount of labeled data and face challenges such as the curse of dimensionality brought by high-dimensional data.

[0004] Currently, most research on image texture classification methods focuses on the low-dimensional representation or high-dimensional representation of texture features, and there are few solutions that comprehensively consider the advantages of both. Low-dimensional features are simple to calculate and easy to understand, but often lose some key information; while high-dimensional features, although rich in information, have high computational complexity and are difficult to intuitively interpret. Therefore, how to effectively combine the high-dimensional and low-dimensional spatial expressions of image textures to improve the accuracy and efficiency of texture classification has become an urgent problem to be solved.

[0005] Especially in the textile and wood processing industries, the accuracy of texture classification is directly related to the appearance quality and market competitiveness of products. Due to the diversity and complexity of material textures in these industries, traditional image texture classification methods often fail to meet the actual application requirements. Therefore, developing a new type of image texture classification method to improve the accuracy and efficiency of material texture classification in the textile and wood processing industries has important research significance and application value.

[0006] In summary, the existing image texture classification methods have many deficiencies, especially in dealing with complex textures and meeting the actual application requirements. Therefore, the present invention aims to propose an image texture classification method and system that comprehensively applies a variety of advanced technologies. Summary of the Invention

[0007] Aiming at the problems of low accuracy of image texture classification and low efficiency in processing high-dimensional data in the prior art, the present invention proposes an image texture classification method and system, which comprehensively applies high-dimensional space expression and low-dimensional space expression, makes full use of the advantages of different-dimensional features, and improves the accuracy and efficiency of texture classification. At the same time, the present invention also has strong versatility and scalability, and can adapt to the texture classification requirements in different industries and scenarios.

[0008] To achieve the above object, the present invention provides the following solutions:

[0009] An image texture classification method, comprising the following steps:

[0010] Extract image texture features from the original image to form a feature vector of the original image;

[0011] Use a Gaussian kernel function to map the feature vector of the original image into a high-dimensional feature space to obtain a feature expression in the high-dimensional space;

[0012] Map the feature vector of the original image into a low-dimensional space by a random mapping method to obtain a feature expression in the low-dimensional space;

[0013] Fuse the feature expression in the high-dimensional space and the feature expression in the low-dimensional space to form a comprehensive feature representation;

[0014] Based on the comprehensive feature representation, use a classifier to classify the image texture.

[0015] Preferably, the method for extracting image texture features from the original image includes: gray-level co-occurrence matrix, local binary pattern, Gabor filter, wavelet transform, histogram of oriented gradients, and deep learning method.

[0016] Preferably, the method for using a Gaussian kernel function to map the feature vector of the original image into a high-dimensional feature space to obtain a feature expression in the high-dimensional space includes:

[0017] Use a Gaussian kernel function to calculate the kernel value K between each pair of feature vectors x i and x j in the feature vector set of the original image; ij ;

[0018] Combine all the kernel values K ij to form a kernel matrix K, where each element K of the kernel matrix K ij represents the inner product of the feature vectors x i and x j in the high-dimensional space;

[0019] Using the kernel matrix K, the original feature vectors are implicitly mapped into a high-dimensional feature space to obtain the feature representation in the high-dimensional space;

[0020] Among them, the expression of the Gaussian kernel function is:

[0021]

[0022] In the formula, K represents the Gaussian kernel function; x i and x j are two feature vectors; σ is the bandwidth parameter of the kernel function; exp() represents the exponential function, ||x i -x j || 2 represents the Euclidean distance between x i and x j .

[0023] Preferably, the method for mapping the feature vectors of the original image into a low-dimensional space by the random mapping method to obtain the feature representation in the low-dimensional space includes:

[0024] Based on the number of original feature vectors and the complexity of the expected classification task, a method based on the proportion of data variance explanation is used to determine the dimension of the low-dimensional space;

[0025] Select the random projection method, and according to the dimension of the low-dimensional space, use the random.normal function in the NumPy library to generate a random projection matrix;

[0026] Multiply the feature vectors of the original image by the random projection matrix to obtain the mapped low-dimensional feature vectors, that is, the feature representation in the low-dimensional space.

[0027] Preferably, the method for fusing the feature representation in the high-dimensional space and the feature representation in the low-dimensional space to form a comprehensive feature representation includes:

[0028] The feature representation in the high-dimensional space and the feature representation in the low-dimensional space are connected end-to-end by the splicing method to form a comprehensive feature vector, that is, the comprehensive feature representation;

[0029] Based on the comprehensive feature representation, the method for classifying image textures using a classifier includes: classification methods based on deep learning, classification methods based on multi-kernel learning, classification methods based on manifold learning, classification methods based on metric learning, and classification methods based on quantum computing.

[0030] The present invention also provides an image texture classification system, including: a preprocessing module, a high-dimensional feature extraction module, a low-dimensional feature extraction module, and a feature fusion and classification module;

[0031] The preprocessing module is used to extract image texture features from the original image and form a feature vector of the original image;

[0032] The high-dimensional feature extraction module is used to map the feature vector of the original image into a high-dimensional feature space by using a Gaussian kernel function to obtain a feature representation in the high-dimensional space;

[0033] The low-dimensional feature extraction module is used to map the feature vector of the original image into a low-dimensional space by a random mapping method to obtain a feature representation in the low-dimensional space;

[0034] The feature fusion and classification module is used to fuse the feature representation in the high-dimensional space and the feature representation in the low-dimensional space to form a comprehensive feature representation, and based on the comprehensive feature representation, a classifier is used to classify the image texture.

[0035] Preferably, the process of extracting image texture features from the original image includes: gray-level co-occurrence matrix, local binary pattern, Gabor filter, wavelet transform, histogram of oriented gradients, and deep learning method.

[0036] Preferably, the high-dimensional feature extraction module includes: a kernel value calculation unit, a combination unit, and an implicit mapping unit;

[0037] The kernel value calculation unit is used to calculate the kernel value K between each pair of feature vectors x i and x j in the feature vector set of the original image by using a Gaussian kernel function; ij ;

[0038] The combination unit is used to combine all the kernel values K ij to form a kernel matrix K, where each element K of the kernel matrix K ij represents the inner product of the feature vectors x i and x j in the high-dimensional space;

[0039] The implicit mapping unit is used to implicitly map the original feature vector into the high-dimensional feature space by using the kernel matrix K to obtain a feature representation in the high-dimensional space;

[0040] wherein, the expression of the Gaussian kernel function is:

[0041]

[0042] In the formula, K represents the Gaussian kernel function; x i and x j are two feature vectors; σ is the bandwidth parameter of the kernel function; exp() represents the exponential function, ||x i -x j || 2Denote x i and x j 's Euclidean distance.

[0043] Preferably, the low-dimensional feature extraction module includes: a dimension calculation unit, a projection unit, and a product unit;

[0044] The dimension calculation unit is used to determine the dimension of the low-dimensional space based on the number of original feature vectors and the expected complexity of the classification task, using a method based on the proportion of data variance explained.

[0045] The projection unit is used to select the method of random projection. According to the dimension of the low-dimensional space, the random.normal function is called using the NumPy library to generate a random projection matrix.

[0046] The product unit is used to multiply the feature vector of the original image by the random projection matrix to obtain the mapped low-dimensional feature vector, that is, the feature representation in the low-dimensional space.

[0047] Preferably, in the feature fusion and classification module, the process of fusing the feature representation in the high-dimensional space and the feature representation in the low-dimensional space to form a comprehensive feature representation includes:

[0048] The feature representation in the high-dimensional space and the feature representation in the low-dimensional space are connected end-to-end by the splicing method to form a comprehensive feature vector, that is, a comprehensive feature representation.

[0049] Based on the comprehensive feature representation, the process of classifying image textures using a classifier includes: a deep learning-based classifier, a multi-kernel learning-based classifier, a manifold learning-based classifier, a metric learning-based classifier, and a quantum computing-based classifier.

[0050] Compared with the prior art, the beneficial effects of the present invention are:

[0051] First, by combining the mapping of high-dimensional and low-dimensional feature spaces and various feature extraction and fusion techniques, the present invention can capture the subtle differences in image textures more comprehensively, thereby significantly improving the accuracy of texture classification. Second, the multi-kernel learning and manifold learning methods adopted by the present invention enable the classification model to not only adapt to the complexity of texture features but also exhibit strong robustness to image noise and deformation. Third, the method framework of the present invention is flexible, easy to integrate with existing image processing systems, and also convenient for customization and expansion according to different application requirements. The technical solution of the present invention is applicable to various industries, especially in the field of material classification such as textiles and wood processing, and is of great significance for improving product quality and the level of industrial automation. In summary, the present invention not only achieves multiple innovation points technically but also demonstrates high practical value and good market prospects in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] To more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0053] Figure 1 It is a schematic flowchart of an image texture classification method according to an embodiment of the present invention;

[0054] Figure 2 It is a schematic structural diagram of an image texture classification system according to an embodiment of the present invention;

[0055] Figure 3 It is a schematic flowchart of an image texture classification method based on quantum computing according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0056] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of them. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0057] To make the above objects, features, and advantages of the present invention more obvious and understandable, the present invention will be further described in detail below with reference to the drawings and specific embodiments.

[0058] Embodiment 1

[0059] As Figure 1 shown, the present invention provides an image texture classification method, including the following steps:

[0060] Extract the image texture features from the original image to form the feature vector of the original image;

[0061] Map the feature vector of the original image into a high-dimensional feature space using a Gaussian kernel function to obtain the feature representation in the high-dimensional space;

[0062] Map the feature vector of the original image into a low-dimensional space through a random mapping method to obtain the feature representation in the low-dimensional space;

[0063] Fuse the feature representation in the high-dimensional space and the feature representation in the low-dimensional space to form a comprehensive feature representation;

[0064] Based on the comprehensive feature representation, use a classifier to classify the image texture.

[0065] In this embodiment, the present invention uses one or more texture feature extraction methods to process the image to obtain one or more texture features, and the foregoing one or more extracted texture features constitute the feature vector of the original image. This step ensures that the key information representing the texture characteristics can be extracted from the original image.

[0066] The methods for extracting image texture features from the original image include: gray-level co-occurrence matrix, local binary pattern, Gabor filter, wavelet transform, histogram of oriented gradients, and deep learning methods.

[0067] In the specific implementation of the present invention, first preprocess the input image, including size normalization, noise removal, etc., to ensure the accuracy of subsequent feature extraction. Then, the following multiple texture feature extraction methods are comprehensively used, specifically including:

[0068] Gray-level co-occurrence matrix: Calculate the gray-level relationship between pixel points in the image and their neighboring pixel points to generate a gray-level co-occurrence matrix. Select pixel pairs with a neighboring pixel spacing of d, where d = 1, 2, 3, and calculate their co-occurrence frequencies to generate a gray-level co-occurrence matrix for capturing the spatial dependence of the texture.

[0069] Local binary pattern: Generate a local binary pattern by comparing the gray value of the central pixel with those of its surrounding neighboring pixels. Within a 3x3 window, compare the gray value of the central pixel with those of the surrounding 8 pixels to generate an 8-bit binary number. This process is repeated for all valid windows in the image.

[0070] Gabor filter: Use Gabor filters at different scales and orientations to filter the image and extract multi-scale and multi-directional texture features. Apply Gabor filters in four directions (0°, 45°, 90°, 135°) and two scales to extract multi-scale and multi-directional texture features.

[0071] Wavelet transform: Perform wavelet transform on the image to obtain texture information at different resolutions. Apply Haar wavelet transform and extract low-frequency and high-frequency components at at least two decomposition levels to obtain texture information at different resolutions.

[0072] Histogram of Oriented Gradients (HOG): Calculate the histogram of gradient directions in local regions of the image to capture shape and texture edge information. Segment the image into several small regions and calculate the gradient histogram within each region to capture shape and texture edge information.

[0073] Deep learning method: Use a Convolutional Neural Network (CNN) to automatically learn and extract high-level texture features of the image. Design a convolutional neural network with three convolutional layers and two pooling layers and train it with a large amount of labeled data to automatically extract deep features.

[0074] The multiple texture features extracted are combined into a feature vector through feature concatenation or weighted fusion techniques to ensure that the information between different features is maximally retained and utilized.

[0075] Through the above methods, not only rich texture features are extracted, but also through the effective combination of features, the accuracy and efficiency of texture classification are significantly improved. The feature extraction strategy of the present invention fully considers the complementarity of different texture features and provides strong feature support for subsequent classification steps.

[0076] In this embodiment, the present invention uses a Gaussian kernel function to map the feature vector of the original image into a higher-dimensional feature space to obtain the feature expression in the high-dimensional space, making the originally linearly inseparable data linearly separable in the new feature space. By calculating the kernel matrix, the present invention realizes the expression of the feature vector in the high-dimensional space and selects appropriate kernel function parameters through the cross-validation method to ensure that the mapped feature space can fully reflect the structure and relationship of the original features.

[0077] After obtaining the feature vector of the original image, the present invention uses a Gaussian kernel function to map it into a high-dimensional feature space to handle the complex non-linear relationship between features. The specific implementation steps of the mapping are as follows:

[0078] Select the Gaussian kernel function: Select the standard form of the Gaussian kernel function

[0079]

[0080] where K represents the Gaussian kernel function; x i and x j are two feature vectors; σ is the bandwidth parameter of the kernel function; exp() represents the exponential function, ||x i -xj || 2 Represents x i and x j 's Euclidean distance.

[0081] Calculate the kernel matrix: For each pair of feature vectors x in the set of feature vectors of the original image i and x j , use the above Gaussian kernel function to calculate the kernel value K(x i , x j ). All kernel values K ij are combined to form the kernel matrix K, where each element K ij of the matrix represents the inner product of the feature vectors x i and x j in the high-dimensional space.

[0082] Parameter selection: Use the k-fold cross-validation method to determine the optimal bandwidth parameter σ. It is recommended to set k = 10 for actual calculation. Calculate the cross-validation error for different σ values and select the σ value that minimizes the error.

[0083] Feature mapping: Utilize the kernel matrix K to implicitly map the original feature vectors to a high-dimensional feature space. This enables the use of linear learning algorithms in the high-dimensional space to handle the non-linear relationships in the original feature space without directly calculating the coordinates in the high-dimensional space.

[0084] Through the above steps, the present invention can effectively map the image texture features into a high-dimensional space, enhancing the model's ability to handle complex texture classification problems. In addition, the influence of different σ values on the classification accuracy is verified through experiments, ensuring that the mapped high-dimensional feature space can better reflect the structure and relationships of the original features. Reasonable selection of kernel function parameters helps to improve the classification accuracy.

[0085] In this embodiment, in order to balance the computational efficiency and the retention of feature information, the present invention further maps the feature vectors of the original image into a low-dimensional space to obtain the feature representation in the low-dimensional space. Through random mapping methods such as random projection or hashing techniques, the present invention realizes the dimensionality reduction processing of the feature vectors of the original image. During the mapping process, the present invention adjusts the mapping parameters and evaluates the information retention situation to ensure that while reducing the dimension, the key information useful for the classification task is retained to the greatest extent.

[0086] To improve the computational efficiency and facilitate subsequent processing, the present invention maps the feature vectors of the original image into a low-dimensional space. The specific implementation steps of the low-dimensional mapping are as follows:

[0087] Determine the dimension of the low-dimensional space: Based on the number of original feature vectors and the expected complexity of the classification task, the dimension of the low-dimensional space is determined using a method based on the proportion of data variance explained. The minimum dimension that can explain 90% of the variance of the original data is selected as the target dimension.

[0088] Select random mapping method: Select random projection as the random mapping method. The advantage of random projection is that it can approximately maintain the Euclidean distance between vectors in the original high-dimensional space while reducing the dimension.

[0089] Generate a random projection matrix: Based on the determined low-dimensional space dimension, use the NumPy library to call the random.normal function to generate a random projection matrix. The elements of the matrix are randomly drawn from the standard normal distribution N(0,1) to ensure the randomness and effectiveness of the mapping.

[0090] Apply random mapping: Multiply the feature vector of the original image by the random projection matrix to obtain the mapped low-dimensional feature vector. This step realizes the conversion of the feature vector from high dimension to low dimension by introducing randomness.

[0091] Evaluate information preservation: The amount of information preserved by the low-dimensional feature vector is evaluated by calculating the reconstruction error and the retained variance ratio. The reconstruction error is calculated by projecting the low-dimensional feature vector back into the high-dimensional space and comparing it to the original feature vector. The variance ratio is calculated based on the ratio of the variance of the low-dimensional feature vector to the variance of the original feature vector.

[0092] Through the above steps, the present invention effectively reduces the dimension of the feature vector while retaining as much key information as possible that is useful for the classification task. This process not only reduces the computational complexity of subsequent processing, but also improves the practicality and flexibility of the image texture classification method.

[0093] In this embodiment, the present invention fuses the feature expression of high-dimensional space and the feature expression of low-dimensional space to form a comprehensive feature representation. This fusion feature is intended to combine the rich information of high-dimensional features and the computational efficiency advantages of low-dimensional features to provide more comprehensive and efficient information for image texture classification. Based on feature fusion, the present invention uses a classifier to classify image textures. The classifier can use a machine learning algorithm, such as a support vector machine, a random forest or a deep learning model, etc., and is selected according to specific application scenarios and requirements.

[0094] Specifically, in terms of feature fusion: the feature expression of the high-dimensional space and the feature expression of the low-dimensional space are connected end-to-end through a splicing method to form a comprehensive feature vector, i.e., a comprehensive feature representation. In addition, according to the importance of the features, a weighted fusion strategy is also adopted to assign weights to different features to highlight the features that are more important to the classification task.

[0095] In terms of training and classification: Using a labeled training dataset, a supervised learning method is employed to train the classifier. During the training process, the cross-entropy loss function is used to optimize the model parameters, and metrics such as accuracy, recall, and F1-score are used to evaluate the performance of the classifier on an independent test dataset.

[0096] Classifier selection: For different application scenarios and requirements, an appropriate classifier is selected. The selected classifiers include: deep learning-based classification methods, multi-kernel learning-based classification methods, manifold learning-based classification methods, metric learning-based classification methods, and quantum computing-based classification methods. Implementation examples of these several classification methods are described separately below.

[0097] a) Implementation example of deep learning-based classification method

[0098] In the present invention, we designed and implemented a deep learning-based classification model, which is particularly suitable for processing and classifying image textures. The architecture of this model includes two input branches, which are specifically designed to process the mapped high-dimensional features and low-dimensional features respectively. The following are the implementation details:

[0099] Model architecture: Construct a deep neural network model with two parallel input branches. Each branch consists of several convolutional layers and pooling layers to effectively extract and process their respective feature sets (i.e., high-dimensional and low-dimensional features). These features are then merged in the network through a specific fusion layer, which can be a concatenation layer or a weighted sum layer, in order to integrate feature information from different sources.

[0100] Fusion strategy: At the fusion layer, the high-dimensional and low-dimensional features are connected end-to-end or weighted and summed according to their importance in the classification task. The purpose of this step is to form a comprehensive feature representation that can fully reflect the key information of the image texture.

[0101] Output branch: The fused features are processed through the output branch, which includes several fully connected layers and an output layer. The output layer uses the softmax activation function to facilitate the classification of multi-class image textures.

[0102] Training process: The model is trained using a labeled training dataset. During the training process, the cross-entropy loss function is used to optimize the model parameters, and metrics such as accuracy, recall, and F1-score are used to evaluate the performance of the model on an independent test dataset.

[0103] Through the above design and implementation process, the deep learning model of the present invention not only improves the efficiency and accuracy of image texture classification, but also enhances the applicability and robustness of the model in various practical application scenarios through a flexible feature fusion strategy and a robust classification strategy.

[0104] b) Example of the implementation of the classification method based on multi-kernel learning

[0105] The present invention adopts a technology based on multi-kernel learning (MKL), which is a machine learning method that combines multiple different kernel functions. Its goal is to improve the performance of image texture classification by automatically selecting and combining various kernel functions. In the implementation of the present invention, first, appropriate kernel functions are selected according to different feature subsets, such as Gaussian kernel, polynomial kernel, and linear kernel, etc. Each kernel function can capture specific information in the feature space. Then, a learning process is used to optimize the weights of each kernel function, and this process can adopt a heuristic-based method or an optimization algorithm-based method, such as gradient descent or sequential minimal optimization algorithm (SMO).

[0106] In the present invention, a set of kernel functions is defined, and an initial weight is assigned to each kernel function. Then, a training data set with labels is used to train the multi-kernel learning model, and the weights of each kernel function are adjusted by minimizing the classification error. This optimization process aims to find the best combination of kernel functions to effectively integrate the information of high-dimensional and low-dimensional feature spaces, so as to achieve accurate image texture classification.

[0107] The well-trained multi-kernel learning model will be used for new image data to predict the texture category of the image by integrating the information of different kernel functions. This method not only improves the classification accuracy but also enhances the generalization ability of the model, enabling it to adapt to various different texture classification tasks.

[0108] c) Example of the implementation of the classification method based on manifold learning

[0109] The present invention analyzes and processes the fused features by using manifold learning methods to achieve effective classification of image textures. Manifold learning aims to reveal the low-dimensional manifold structure in the high-dimensional data space, thereby discovering the intrinsic structure of the data.

[0110] The specific implementation method is as follows:

[0111] (1) Select a manifold learning algorithm: According to the local continuity and global distribution characteristics of the image texture features, select one of the local linear embedding (LLE), isometric mapping (Isomap), and Laplacian Eigenmaps algorithms. These algorithms can explore the intrinsic geometric structure of the data from different angles and effectively map the high-dimensional data to the low-dimensional manifold space.

[0112] (2) Construct an adjacency graph: For the selected manifold learning algorithm, construct an adjacency graph to characterize the proximity relationship between feature vectors. The construction of the adjacency graph takes into account the Euclidean distance or cosine similarity between features, ensuring that the construction of the graph reflects the true proximity relationship of the data.

[0113] (3) Calculate the low-dimensional embedding: Apply the manifold learning algorithm to map the high-dimensional fused features to the low-dimensional manifold space while preserving the intrinsic geometric structure and relationships in the original feature space as much as possible.

[0114] (4) Train the classification model: In the low-dimensional manifold space, use the mapped features to train the classification model. Here, classifiers that can be used include support vector machine (SVM), random forest, k-nearest neighbor (k-NN), etc., to determine the classifier most suitable for processing low-dimensional data.

[0115] (5) Model optimization and validation: Optimize the parameters of the classification model through cross-validation and other model evaluation techniques, and verify the classification performance of the model. According to the experimental results, the parameters of the manifold learning algorithm may be adjusted or the classifier may be replaced to achieve the optimal classification effect.

[0116] Through the above steps, the classification method based on manifold learning can effectively utilize the intrinsic geometric structure of image texture features, thereby improving the accuracy and robustness of image texture classification. This method is particularly suitable for processing image data with complex textures and shape variations, providing an effective technical means for applications such as material classification, quality control, and product design in the textile and wood processing industries.

[0117] d) Example of implementing the classification method based on metric learning

[0118] Metric learning is a machine learning method aimed at learning the distance or similarity metric between data points. It optimizes the distance function so that the distance between data points of the same class is smaller, while the distance between data points of different classes is larger. Metric learning can be used in combination with deep learning or implemented as an independent method, especially suitable for tasks that require fine-grained distance discrimination. In the present invention, metric learning is described as an independent classification method.

[0119] The specific implementation method is as follows:

[0120] (1) Define the distance metric: Select or define a distance metric function suitable for image texture features. The distance metric functions include Mahalanobis distance, Euclidean distance, and Cosine similarity.

[0121] (2) Construct similarity and dissimilarity constraints: Based on the training dataset, a series of similarity and dissimilarity constraints are constructed. The similarity constraints require that the texture features of images of the same class are closer under the distance metric, while the dissimilarity constraints require that the texture features of images of different classes are farther apart.

[0122] (3) Learn the distance metric: Using metric learning algorithms, including Nearest Neighbor Component Analysis (NCA), Large Margin Nearest Neighbor (LMNN), or Triplet loss, learn a distance metric that can best satisfy the above constraints.

[0123] (4) Feature space transformation: Through the learned distance metric, transform the fused feature space so that in the transformed space, the distance between feature points of the same class is reduced, and the distance between feature points of different classes is increased.

[0124] (5) Classifier training: In the transformed feature space, use any standard classifier, such as k-Nearest Neighbor (k-NN), Support Vector Machine (SVM), etc., to train the classifier.

[0125] (6) Model evaluation and optimization: Evaluate the performance of the classification model through methods such as cross-validation, and adjust the parameters of the metric learning algorithm and the settings of the classifier according to the evaluation results to further improve the classification accuracy. This step is carried out iteratively with the aim of refining the distance metric and the classifier configuration until the optimal classification performance is achieved.

[0126] By implementing the above steps, the classification method based on metric learning can effectively improve the accuracy of image texture classification, especially suitable for cases where the differences between classes are subtle. The distance metric obtained by metric learning not only helps with the classification task but can also be applied to related tasks such as clustering and retrieval, further enhancing the application value of the present invention in the textile and wood processing industries. The successful implementation of this method provides a new technical approach for dealing with complex image texture classification problems, demonstrating the powerful potential of metric learning in the field of image processing.

[0127] e) Example of the implementation of the classification method based on quantum computing

[0128] The present invention proposes an image texture classification method based on quantum computing, which utilizes the principles of quantum mechanics to encode and operate on data through qubits (quantum bits) to process large-scale and complex data and improve the speed and efficiency of classification.

[0129] The specific implementation method is as follows:

[0130] (1) Quantum data encoding: We use quantum encoding techniques such as amplitude encoding to encode the image texture feature vectors onto qubits, effectively utilizing the high-dimensional characteristics of the quantum system.

[0131] (2) Quantum feature mapping: By designing a quantum circuit to operate on the encoded quantum state, the quantum feature mapping of the eigenvector is realized. This mapping can be linear or non-linear to enhance the ability of the quantum classifier to process complex data.

[0132] (3) Quantum classifier construction: We construct quantum classifiers such as quantum support vector machines (QSVMs), quantum neural networks (QNNs), or quantum decision trees. Quantum classifiers utilize the superposition and entanglement properties of quantum states to achieve effective data classification.

[0133] (4) Quantum state measurement: After the quantum circuit execution is completed, the classification result is obtained by measuring the quantum state. The probability distribution obtained from the measurement is used to determine the texture category of the image.

[0134] (5) Quantum optimization and training: We use quantum optimization algorithms to optimize and train the parameters of the quantum classifier to enhance its classification accuracy.

[0135] (6) Model evaluation and model optimization: The performance of the quantum classifier is evaluated by methods such as cross-validation, and the parameters of the quantum classifier are adjusted according to the evaluation results to achieve the best classification effect.

[0136] The classification method based on quantum computing can theoretically process larger-scale data sets and achieve performance beyond traditional computing methods under specific conditions. This novel classification method brings unprecedented possibilities to the field of image texture classification, especially in application scenarios with huge data volume and high complexity.

[0137] Through the above specific implementation manners, the present invention realizes efficient and accurate classification of image textures. By comprehensively applying high-dimensional space expression and low-dimensional space expression, the present invention makes full use of the advantages of features in different dimensions, improving the accuracy and efficiency of texture classification. At the same time, the present invention also has strong generality and scalability, and can adapt to the texture classification requirements in different industries and scenarios.

[0138] Embodiment 2

[0139] As Figure 2 shown, the present invention also provides an image texture classification system, including: a preprocessing module, a high-dimensional feature extraction module, a low-dimensional feature extraction module, and a feature fusion and classification module;

[0140] The preprocessing module is used to extract image texture features from the original image to form the feature vector of the original image;

[0141] The high-dimensional feature extraction module is used to map the feature vector of the original image into a high-dimensional feature space by using a Gaussian kernel function to obtain the feature expression in the high-dimensional space;

[0142] The low-dimensional feature extraction module is used to map the feature vectors of the original image into a low-dimensional space through a random mapping method to obtain the feature representation in the low-dimensional space;

[0143] The feature fusion and classification module is used to fuse the feature representation in the high-dimensional space and the feature representation in the low-dimensional space to form a comprehensive feature representation. Based on the comprehensive feature representation, a classifier is used to classify the image texture.

[0144] In this embodiment, the process of extracting image texture features from the original image includes: gray-level co-occurrence matrix, local binary pattern, Gabor filter, wavelet transform, histogram of oriented gradients, and deep learning method.

[0145] In this embodiment, the high-dimensional feature extraction module includes: a kernel value calculation unit, a combination unit, and an implicit mapping unit;

[0146] The kernel value calculation unit is used to calculate the kernel value K between each pair of feature vectors x in the feature vector set of the original image using the Gaussian kernel function i and x j ; ij ;

[0147] The combination unit is used to combine all the kernel values K ij to form a kernel matrix K, where each element K of the kernel matrix K ij represents the inner product of the feature vectors x i and x j in the high-dimensional space;

[0148] The implicit mapping unit is used to use the kernel matrix K to implicitly map the original feature vectors into a high-dimensional feature space to obtain the feature representation in the high-dimensional space;

[0149] Among them, the expression of the Gaussian kernel function is:

[0150]

[0151] In the formula, K represents the Gaussian kernel function; x i and x j are two feature vectors; σ is the bandwidth parameter of the kernel function; exp() represents the exponential function, ||x i -x j || 2 represents the Euclidean distance between x i and x j ;

[0152] In this embodiment, the low-dimensional feature extraction module includes: a dimension calculation unit, a projection unit, and a product unit;

[0153] The dimension calculation unit is used to determine the dimension of the low-dimensional space by using a method based on the proportion of data variance interpretation, based on the number of original feature vectors and the expected complexity of the classification task;

[0154] The projection unit is used to select the method of random projection. According to the dimension of the low-dimensional space, the random.normal function is called using the NumPy library to generate a random projection matrix;

[0155] The product unit is used to multiply the feature vector of the original image by the random projection matrix to obtain the mapped low-dimensional feature vector, which is the feature representation in the low-dimensional space.

[0156] In this embodiment, in the feature fusion and classification module, the process of fusing the feature representation in the high-dimensional space and the feature representation in the low-dimensional space to form a comprehensive feature representation includes:

[0157] The feature representation in the high-dimensional space and the feature representation in the low-dimensional space are connected end-to-end by the splicing method to form a comprehensive feature vector, which is the comprehensive feature representation;

[0158] Based on the comprehensive feature representation, the process of classifying image textures using a classifier includes: a classifier based on deep learning, a classifier based on multi-kernel learning, a classifier based on manifold learning, a classifier based on metric learning, and a classifier based on quantum computing.

[0159] Embodiment III

[0160] As Figure 3 shown, the present invention also provides an image texture classification method based on quantum computing. The steps include: quantum data encoding, quantum feature mapping, construction of a sub-classifier, quantum state measurement, quantum optimization and training, model evaluation, and model optimization;

[0161] Quantum data encoding: The image texture feature vector is encoded onto qubits using amplitude encoding technology. First, the original image is preprocessed and feature extracted to obtain the image texture feature vector. Each element of these image texture feature vectors will correspond to an amplitude value, and then these amplitude values are encoded onto the quantum state to form quantum data. Amplitude encoding uses fewer qubits to represent more classical information, thus effectively utilizing the high-dimensional characteristics of the quantum system.

[0162] By extracting image texture features from the original image, a feature vector of the original image is constructed; the Gaussian kernel function is used to map the feature vector of the original image into a high-dimensional feature space to obtain the feature representation in the high-dimensional space; the feature vector of the original image is mapped into a low-dimensional space by the random mapping method to obtain the feature representation in the low-dimensional space. The aforementioned image texture feature vector can adopt one or a combination of multiple ones among the feature vector of the original image, the feature representation in the high-dimensional space, and the feature representation in the low-dimensional space.

[0163] Quantum feature mapping: The operation on the encoded quantum state is completed by a specifically designed quantum circuit. The quantum circuit consists of a series of quantum logic gates, such as Hadamard gates, controlled-NOT gates, etc. The encoded quantum data is input into the quantum circuit, and through the superposition and entanglement of qubits, a non-linear mapping is performed on the feature vector, thereby realizing the expansion of the quantum feature space. Quantum feature mapping can effectively map data into a high-dimensional feature space to solve the problem of complex data structures that are difficult to handle by classical computing.

[0164] Construction of the quantum classifier: Adopt quantum machine learning algorithms such as quantum support vector machine (QSVM), quantum neural network (QNN), or quantum decision tree. In this method, QSVM is selected to calculate the kernel function using a quantum circuit and classify data using the superposition and entanglement states of qubits. QSVM is a method for linear classification in a high-dimensional quantum feature space and can effectively handle non-linearly separable data sets.

[0165] Quantum state measurement: After the quantum classifier processes the data, the quantum state needs to be measured to obtain the classification result. By measuring the qubits, the result obtained is a probability distribution, which can reflect the likelihood of different classes. According to the probability distribution obtained by measurement, the texture class to which the image belongs can be determined. Quantum measurement can provide statistical data on the information of the quantum state for making decisions on the classification result.

[0166] Quantum optimization and training: Use the quantum optimization algorithm, specifically the quantum approximate optimization algorithm (QAOA) to optimize and train the parameters of the quantum classifier. Through the quantum computing advantage, the global optimal solution or a high-quality approximate solution can be found faster, improving the classification accuracy of the model. During the training process, the quantum algorithm continuously adjusts the parameters of the quantum circuit to make the prediction result as close as possible to the true label.

[0167] Model evaluation and model optimization: The performance evaluation of the model is completed through the cross-validation method, that is, the data set is divided into a training set and a validation set, and the classification accuracy and generalization ability of the quantum classifier are estimated through repeated training and validation processes. According to the evaluation results, the parameters and model structure of the quantum classifier can be further adjusted to achieve the best classification effect. The evaluation metrics include accuracy, recall, precision, and F1 score, etc.

[0168] Through the above steps, the method of the present invention effectively utilizes the unique advantages of quantum computing to achieve image texture classification, especially suitable for complex pattern recognition tasks, and solves the image processing problems that are difficult to handle by classical computers.

[0169] The present invention discloses an efficient and accurate image texture classification method and system, aiming to solve the problems of low accuracy of image texture classification and low efficiency in processing high-dimensional data in the prior art. The present invention realizes the effective classification of image texture by comprehensively applying high-dimensional space expression and low-dimensional space expression. First, the image is processed by using one or more texture feature extraction methods to obtain one or more texture features, which constitute the feature vector of the original image. Subsequently, the Gaussian kernel function is used to map the feature vector of the original image into a higher-dimensional feature space to obtain the feature expression in the high-dimensional space. At the same time, in order to balance the computational efficiency and the retention of feature information, the feature vector of the original image is further mapped into a low-dimensional space to obtain the feature expression in the low-dimensional space. Finally, a comprehensive feature representation is formed through the feature fusion technology, and a classifier is used for the classification of image texture. The present invention combines the rich information of high-dimensional features and the computational efficiency advantage of low-dimensional features, provides more comprehensive and efficient information for image texture classification, and significantly improves the performance and generalization ability of texture classification. At the same time, the present invention has strong versatility and scalability, can adapt to the texture classification needs in different industries and scenarios, especially suitable for material classification fields such as textiles and wood processing, and is of great significance for improving product quality and industrial automation level. It should be noted that the specific implementation manner of the present invention can be flexibly adjusted and optimized according to the actual application scenario and requirements. For example, in the texture feature extraction stage, an appropriate feature extraction method can be selected according to the specific characteristics of the image; in the high-dimensional and low-dimensional feature space mapping stage, the mapping parameters can be adjusted according to the data distribution and the requirements of the classification task; in the feature fusion and classification stage, an appropriate fusion strategy and classification algorithm can be selected according to the performance and computational efficiency requirements of the classifier. Through these adjustments and optimizations, the present invention can better adapt to various complex texture classification tasks and improve the accuracy and efficiency of classification.

[0170] The embodiments described above are only descriptions of the preferred embodiments of the present invention, and do not limit the scope of the present invention. Without departing from the design spirit of the present invention, various deformations and improvements made by those of ordinary skill in the art to the technical solutions of the present invention shall fall within the protection scope determined by the claims of the present invention.

Claims

1. An image texture classification method, characterized in that, It includes the following steps: Extract the image texture features from the original image to form the feature vector of the original image; Use the Gaussian kernel function to map the feature vector of the original image into a high-dimensional feature space to obtain the feature representation in the high-dimensional space; Map the feature vector of the original image into a low-dimensional space by the random mapping method to obtain the feature representation in the low-dimensional space; Fuse the feature representation in the high-dimensional space and the feature representation in the low-dimensional space to form a comprehensive feature representation; Based on the comprehensive feature representation, use a classifier to classify the image texture; The methods for extracting the image texture features from the original image include: gray-level co-occurrence matrix, local binary pattern, Gabor filter, wavelet transform, histogram of oriented gradients, and deep learning method; First, preprocess the input image, including size normalization and noise removal; then, comprehensively use a variety of texture feature extraction methods, specifically including: Gray-level co-occurrence matrix: Calculate the gray-level relationship between the pixel points in the image and their neighboring pixel points to generate the gray-level co-occurrence matrix; Select the pixel pairs with the adjacent pixel spacing of d, d = 1, 2, 3, and calculate the co-occurrence frequency of the pixel pairs to generate the gray-level co-occurrence matrix for capturing the spatial dependence of the texture; Local binary pattern: Generate the local binary pattern by comparing the gray value of the central pixel with the gray values of its surrounding neighboring pixels; In a 3x3 window, compare the gray value of the central pixel with the gray values of the surrounding 8 pixels to generate an 8-bit binary number; This process is repeated for all valid windows in the image; Gabor filter: Filter the image using Gabor filters with different scales and directions to extract the texture features of multiple scales and multiple directions; Apply Gabor filters with four directions 0°, 45°, 90°, 135° and two scales to extract the texture features of multiple scales and multiple directions; Wavelet transform: Perform wavelet transform on the image to obtain the texture information at different resolutions; Apply Haar wavelet transform and extract the low-frequency and high-frequency components of at least two decomposition levels to obtain the texture information at different resolutions; Histogram of oriented gradients HOG: Calculate the gradient direction histogram of the local area of the image for capturing the shape and texture edge information; Divide the image into several small regions and calculate the gradient histogram in each region to capture the shape and texture edge information; Deep learning method: Use a convolutional neural network CNN to automatically learn and extract the high-level texture features of the image; Design a convolutional neural network containing three convolutional layers and two pooling layers and train it with a large amount of labeled data to automatically extract the deep features; The methods for fusing the feature representation in the high-dimensional space and the feature representation in the low-dimensional space to form a comprehensive feature representation include: Connect the feature representation in the high-dimensional space and the feature representation in the low-dimensional space end-to-end by the splicing method to form a comprehensive feature vector, that is, the comprehensive feature representation; The method for classifying image textures using a classifier based on the comprehensive feature representation includes: a classification method based on deep learning, a classification method based on multi-kernel learning, a classification method based on manifold learning, a classification method based on metric learning, and a classification method based on quantum computing; a) Classification method based on deep learning: A classification model based on deep learning was designed and implemented. The architecture of this model includes two input branches, which are specifically used to process the mapped high-dimensional features and low-dimensional features respectively; Model architecture: Construct a deep neural network model with two parallel input branches; each branch consists of several convolutional layers and pooling layers to extract and process their respective feature sets, namely high-dimensional and low-dimensional features; subsequently, they are merged in the network through a specific fusion layer, and the fusion layer is either a concatenation layer or a weighted sum layer to integrate feature information from different sources; Fusion strategy: In the fusion layer, the high-dimensional and low-dimensional features are end-to-end connected or weighted and summed according to their importance in the classification task to form a comprehensive feature representation that can comprehensively reflect the key information of the image texture; Output branch: The fused features are processed through the output branch, which includes several fully connected layers and an output layer; the output layer uses the softmax activation function to facilitate the classification of multi-class image textures; Training process: The model is trained using a labeled training dataset; during the training process, the cross-entropy loss function is used to optimize the model parameters, and accuracy, recall, and F1-score metrics are used to evaluate the performance of the model on an independent test dataset.

2. The image texture classification method according to claim 1, wherein, The method for mapping the feature vector of the original image to a high-dimensional feature space using a Gaussian kernel function to obtain the feature expression in the high-dimensional space includes: Calculate the kernel value K between each pair of feature vectors x in the set of feature vectors of the original image using the Gaussian kernel function i and x j ; ij ; Combine all the kernel values K ij to form a kernel matrix K, where each element K of the kernel matrix K ij represents the inner product of the feature vectors x i and x j in the high-dimensional space; Using the kernel matrix K to implicitly map the original feature vector to a high-dimensional feature space to obtain the feature expression in the high-dimensional space; where the expression of the Gaussian kernel function is: Where K represents the Gaussian kernel function; x i and x j are two feature vectors; σ is the bandwidth parameter of the kernel function; exp() represents the exponential function, ||x i -x j || 2 denotes the Euclidean distance between x i and x j .

3. The image texture classification method according to claim 1, characterized in that The method for mapping the feature vector of the original image to a low-dimensional space using a random mapping method to obtain the feature expression in the low-dimensional space includes: Based on the number of original feature vectors and the complexity of the expected classification task, a method based on the proportion of data variance explained is used to determine the dimension of the low-dimensional space; Select the method of random projection. According to the dimension of the low-dimensional space, use the NumPy library to call the random.normal function to generate a random projection matrix; Multiply the feature vector of the original image by the random projection matrix to obtain the mapped low-dimensional feature vector, which is the feature expression in the low-dimensional space.

4. An image texture classification system for implementing an image texture classification method according to any one of claims 1-3, characterized in that, Includes: A preprocessing module, a high-dimensional feature extraction module, a low-dimensional feature extraction module, and a feature fusion and classification module; The preprocessing module is used to extract image texture features from the original image to form the feature vector of the original image; The high-dimensional feature extraction module is used to map the feature vector of the original image to a high-dimensional feature space using a Gaussian kernel function to obtain the feature expression in the high-dimensional space; The low-dimensional feature extraction module is used to map the feature vector of the original image into a low-dimensional space by a random mapping method to obtain the feature representation in the low-dimensional space; The feature fusion and classification module is used to fuse the feature representation in the high-dimensional space and the feature representation in the low-dimensional space to form a comprehensive feature representation, and based on the comprehensive feature representation, a classifier is used to classify the image texture.

5. The image texture classification system according to claim 4, wherein In the preprocessing module, the process of extracting image texture features from the original image includes: gray-level co-occurrence matrix, local binary pattern, Gabor filter, wavelet transform, histogram of oriented gradients, and deep learning method.

6. The image texture classification system according to claim 4, characterized in that The high-dimensional feature extraction module includes: a kernel value calculation unit, a combination unit, and an implicit mapping unit; The kernel value calculation unit is used to calculate the kernel value K ij between each pair of feature vectors x i and x j in the feature vector set of the original image by using the Gaussian kernel function; i and x j between the kernel value K ij ; The combined unit is used to combine all kernel values K ij to form a kernel matrix K, where each element K of the kernel matrix K ij represents the inner product of the feature vectors x i and x j in the high-dimensional space; The implicit mapping unit is used to implicitly map the original feature vector to a high-dimensional feature space by using the kernel matrix K to obtain the feature representation in the high-dimensional space; Among them, the expression of the Gaussian kernel function is: Where K represents the Gaussian kernel function; x i and x j are two feature vectors; σ is the bandwidth parameter of the kernel function; exp() represents the exponential function, ||x i -x j || 2 represents the Euclidean distance between x i and x j .

7. The image texture classification system according to claim 4, characterized in that, The low-dimensional feature extraction module includes: a dimension calculation unit, a projection unit, and a product unit; The dimension calculation unit is used to determine the dimension of the low-dimensional space by using a method based on the proportion of data variance explanation, based on the number of original feature vectors and the complexity of the expected classification task; The projection unit is used to select a random projection method, and according to the dimension of the low-dimensional space, call the random.normal function of the NumPy library to generate a random projection matrix; The product unit is used to multiply the feature vector of the original image by the random projection matrix to obtain the mapped low-dimensional feature vector, that is, the feature representation in the low-dimensional space.

8. The image texture classification system according to claim 4, characterized in that, In the feature fusion and classification module, the process of fusing the feature representation in the high-dimensional space and the feature representation in the low-dimensional space to form a comprehensive feature representation includes: Connect the feature representation in the high-dimensional space and the feature representation in the low-dimensional space end-to-end by a splicing method to form a comprehensive feature vector, that is, a comprehensive feature representation; Based on the comprehensive feature representation, the process of classifying the image texture by using a classifier includes: a deep learning-based classifier, a multi-kernel learning-based classifier, a manifold learning-based classifier, a metric learning-based classifier, and a quantum computing-based classifier.

Images