Artistic pattern recognition system based on artificial intelligence

Abstract

The invention relates to the technical field of artistic pattern recognition, and provides an artistic pattern recognition system based on artificial intelligence, which comprises a data acquisition module, a feature extraction module, a feature selection module, a texture analysis module and a result evaluation module.The artistic pattern recognition system overcomes the defects in the prior art and is reasonable in design and high in practicability. A digital image of the surface of an artwork is obtained through high-resolution digital imaging equipment, denoising, white balance correction and image enhancement processing are carried out to ensure image quality, then texture feature vectors are extracted through a rotation invariant local binary pattern method, texture features can be effectively extracted, rotation invariance is achieved, and classification robustness is improved; the screened features are input into the pre-trained deep learning model, the artistic pattern category probability is output, the deep learning model is used for learning texture features, and the generalization ability and classification performance of the model are improved.

Description

Technical Field

[0001] This invention relates to the field of art pattern recognition technology, and more specifically to an art pattern recognition system based on artificial intelligence. Background Technology

[0002] In the field of art recognition, traditional texture analysis methods mainly rely on statistical features, frequency domain analysis, and model basis methods. For example, the Gray-Level Co-occurrence Matrix (GLCM) extracts texture features by calculating the co-occurrence frequency of pixel pairs at different gray levels, while wavelet transform provides multi-scale texture information, enabling a more comprehensive description of texture details. Furthermore, methods based on Local Binary Patterns (LBP) are also widely used for texture feature extraction, generating binary codes by comparing the gray values ​​of the center pixel with those of surrounding pixels, thus forming a texture feature vector. However, these methods often face problems such as incomplete feature extraction, low computational efficiency, and sensitivity to noise when dealing with complex textures and high-resolution images.

[0003] Existing problems:

[0004] Incomplete feature extraction: Although traditional methods such as GLCM and LBP can extract texture features, they are difficult to capture high-order texture information, which limits the classification effect.

[0005] Low computational efficiency: Many texture analysis methods require a lot of computational resources, especially when processing high-resolution images, resulting in high computational costs.

[0006] Sensitive to noise: Traditional methods are easily interfered with when processing images containing noise, affecting the accuracy of classification.

[0007] Feature selection is challenging: After feature extraction, selecting the most representative features for classification is a complex problem. Existing methods often rely on manual selection, lacking automated and intelligent feature selection mechanisms.

[0008] To address the aforementioned problems, we propose an artificial intelligence-based art pattern recognition system. Summary of the Invention

[0009] To address the shortcomings of existing technologies, this invention provides an art pattern recognition system based on artificial intelligence. This system overcomes the deficiencies of existing technologies, is rationally designed, and solves the problems of image adaptability, rotation robustness, and small sample generalization in art pattern recognition.

[0010] To achieve the above objectives, the present invention provides the following technical solution:

[0011] An artificial intelligence-based art pattern recognition system includes:

[0012] Data acquisition module: used to acquire high-resolution digital images of the target area on the surface of the artwork, and to perform noise reduction, white balance and enhancement processing;

[0013] Feature extraction module: Converts the processed image into a grayscale image and extracts texture feature vectors using the rotation-invariant local binary mode method;

[0014] Feature selection module: Standardizes and performs significance analysis on feature vectors to select a subset of key features;

[0015] Texture analysis module: Inputs the filtered features into the pre-trained classification model and outputs the probability of the artistic pattern category;

[0016] Results evaluation module: Visualizes the classification results.

[0017] Preferably, the execution steps of the data acquisition module are as follows:

[0018] Select a specific target area with textured features on the surface of an artwork, including but not limited to paintings, fabrics, sculptures, and murals; acquire a digital image of the selected area using a high-resolution digital imaging device; and preprocess the acquired digital image.

[0019] Preferably, the execution steps of the feature extraction module are as follows:

[0020] The preprocessed digital image is acquired and converted into a grayscale image. During the conversion process, the input digital image is verified to be in standard RGB format. The verification first checks the number of channels in the digital image. If the number of channels is not equal to three (non-RGB) or the color space label is missing, the color space conversion is triggered. The verification is performed by acquiring the number of channels and detecting the color space label.

[0021] For digital images in non-standard RGB format, they are uniformly converted to the sRGB standard color space; if CMYK or Lab color space is encountered, they are first converted to XYZ color space and then mapped to sRGB.

[0022] Grayscale conversion extracts the values ​​of the red, green, and blue channels of each pixel from the sRGB color space, calculates them according to the human eye's sensitivity to different colors, and finally outputs a single-channel grayscale image.

[0023] Preferably, the execution steps of the local binary mode are as follows:

[0024] It receives a single-channel grayscale image as input data, establishes a coordinate system with the current pixel as the center, calculates n uniformly distributed sampling points on the circumference based on the radius R, and uses bilinear interpolation to obtain accurate grayscale values ​​for non-integer coordinate points.

[0025] Each sample point grayscale is compared with the center point to generate binary bits. The n comparison results are combined to form an initial binary code. The smallest decimal value is found by cyclic shifting.

[0026] Traverse all valid pixels in the entire image, map the local binary pattern of each point to histogram intervals, count the frequency of occurrence in each interval to generate feature vectors, and then output the normalized LBP feature histogram.

[0027] Preferably, the execution steps of the feature selection module are as follows:

[0028] Input feature vectors, eliminate dimensional differences through z-score standardization, making the feature mean 0 and variance 1, and output standardized feature vectors;

[0029] z-score eliminates dimensions and calculates the arithmetic mean of all samples separately for each feature. ,in, The number of samples is the total number of valid pixels in the image that are used in the calculation. The original data values ​​are the feature codes obtained by LBP calculation for each pixel; then, the dispersion of the data relative to the mean is calculated. Perform a standardization transformation on each data point to convert the original values... Convert to standardized value , ,in, This indicates the elimination of data position offset; This indicates the elimination of dimensional differences in the data;

[0030] Input a standardized feature vector, use statistical tests to calculate the p-value and t-value of each feature, and output a list of feature significance indicators;

[0031] Statistical test process: Input two sets of feature vectors, calculate the mean and standard deviation of the two sets of feature vectors respectively, calculate the combined standard error, divide the difference between the two sets of means by the combined standard error to obtain the t-value, calculate the p-value according to the t-value distribution table, and output the p-value and t-value.

[0032] Preferably, a significance threshold α is preset, and the p-value is compared with α: if p≤α, the difference is significant; if p>α, the difference is not significant; the sign of the t-value indicates the direction of the difference; features that satisfy p≤α are retained, and a preliminary feature subset is output.

[0033] Input the initial feature subset: if the number of features is less than 10, decrease α to increase the number of features; if the number of features is greater than 100, increase α to improve the screening criteria; output the optimized feature subset and the adjusted threshold.

[0034] Input multiple sets of pre-selected thresholds, each threshold corresponding to a feature subset of multiple thresholds. Calculate the F1 value of each subset through cross-validation, and output the final feature subset and threshold. The process of calculating the F1 value of each subset through cross-validation involves inputting multiple thresholds and their corresponding feature subsets. For each feature subset corresponding to a threshold, perform five-fold cross-validation. Each time, train the model with four sets of data, test with the remaining set, and calculate the F1 value. Finally, take the average of the five F1 values ​​and select the feature subset with the highest average F1 value and its corresponding threshold.

[0035] Preferably, the texture analysis module performs the following steps:

[0036] Receive an optimized feature subset from the feature selection module, with dimensions n×m, where n is the number of samples and m is the number of features;

[0037] The m-dimensional features are reorganized into a (h×w×c) pseudo-image tensor and scaled to a standard size using bilinear interpolation. The image is then input into a pre-trained deep learning model, and high-order texture semantic features are extracted using the convolutional layers of the deep learning model.

[0038] The high-order features extracted by the convolutional layer are integrated by the fully connected layer at the end. The activation function gives the high-order features non-linear discriminative power and outputs the probability distribution vector of the art pattern category (p1,p2,...,pk), where k is the total number of predefined categories.

[0039] Preferably, the result evaluation module performs the following steps:

[0040] Receive category probability distribution data output by the texture analysis module;

[0041] Select the category with the highest probability as the predicted label and record the confidence level;

[0042] A heatmap of texture feature saliency is generated based on the category probability distribution, which maps the decision weights of different texture regions;

[0043] Output a confusion matrix and classification report, and label the incorrectly identified samples in each category;

[0044] The visualization results are integrated and output to the interactive interface.

[0045] This invention provides an art pattern recognition system based on artificial intelligence. It has the following beneficial effects:

[0046] Digital images of the artwork surface are acquired using high-resolution digital imaging equipment, and then denoising, white balance correction, and image enhancement are performed to ensure image quality. Subsequently, the rotation-invariant local binary mode (LBP) method is used to extract texture feature vectors, which can effectively extract texture features and also has rotation invariance, thus improving the robustness of classification.

[0047] By inputting the filtered features into a pre-trained deep learning model, the model outputs the probability of the art pattern category. The deep learning model learns the texture features, thereby improving the model's generalization ability and classification performance.

[0048] Through the preprocessing steps of the data acquisition module, including noise reduction, white balance correction, and image enhancement, noise interference is effectively removed, image quality is improved, and thus the accuracy of classification is enhanced.

[0049] The feature selection module performs standardization and significance analysis on the extracted feature vectors, filters out key feature subsets, eliminates dimensional differences through z-score standardization, and evaluates the significance of features using statistical tests. Finally, features with significant differences are retained, thus automating feature selection. Attached Figure Description

[0050] Figure 1 This is a schematic diagram of the process of the present invention;

[0051] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0052] See attached document Figure 1 An artificial intelligence-based art pattern recognition system includes the following steps:

[0053] Data acquisition module: Used to acquire high-resolution digital images of the target area on the surface of the artwork, and perform noise reduction, white balance, and enhancement processing; the execution steps of the data acquisition module are as follows:

[0054] Select a specific target area with textured features on the surface of an artwork, including but not limited to paintings, fabrics, sculptures, and murals; acquire a digital image of the selected area using a high-resolution digital imaging device; and preprocess the acquired digital image.

[0055] Feature extraction module: Converts the processed image to a grayscale image and extracts texture feature vectors using the rotation-invariant local binary mode method; the execution steps of the feature extraction module are as follows:

[0056] The preprocessed digital image is acquired and converted into a grayscale image. During the conversion process, the input digital image is verified to be in standard RGB format. The verification first checks the number of channels in the digital image. If the number of channels is not equal to three (non-RGB) or the color space label is missing, the color space conversion is triggered. The verification is performed by acquiring the number of channels and detecting the color space label.

[0057] For digital images in non-standard RGB format, they are uniformly converted to the sRGB standard color space; if CMYK or Lab color space is encountered, they are first converted to XYZ color space and then mapped to sRGB.

[0058] Grayscale conversion extracts the values ​​of the red, green, and blue channels of each pixel from the sRGB color space, and calculates them according to the human eye's sensitivity to different colors (the green component has the largest weight, and the blue component has the smallest weight), and finally outputs a single-channel grayscale image.

[0059] The execution steps of the local binary pattern are as follows:

[0060] It receives a single-channel grayscale image as input data, establishes a coordinate system with the current pixel as the center, calculates n uniformly distributed sampling points on the circumference based on the radius R, and uses bilinear interpolation to obtain accurate grayscale values ​​for non-integer coordinate points.

[0061] Each sample point's grayscale value is compared with the center point to generate binary bits. The n comparison results are combined to form an initial binary code. The smallest decimal value is found through cyclic shifting to achieve rotation invariance.

[0062] Traverse all valid pixels in the entire image, map the local binary pattern of each point to histogram intervals, count the frequency of occurrence in each interval to generate feature vectors, and then output the normalized LBP feature histogram.

[0063] Feature selection module: Performs standardization and significance analysis on feature vectors to filter out a subset of key features; the execution steps of the feature selection module are as follows:

[0064] Input feature vectors, eliminate dimensional differences through z-score standardization, making the feature mean 0 and variance 1, and output standardized feature vectors;

[0065] z-score eliminates dimensions and calculates the arithmetic mean of all samples separately for each feature. ,in, The number of samples is the total number of valid pixels in the image that are used in the calculation. The original data values ​​are the feature codes obtained by LBP calculation for each pixel; then, the dispersion of the data relative to the mean is calculated. Perform a standardization transformation on each data point to convert the original values... Convert to standardized value , ,in, This indicates the elimination of data position offset (centralization). This means eliminating the dimensional differences in the data (dimensionless transformation).

[0066] Input a standardized feature vector, use statistical tests to calculate the p-value and t-value of each feature, and output a list of feature significance indicators;

[0067] Statistical test process: Input two sets of feature vectors (standardized feature vectors), calculate the mean and standard deviation of the two sets of feature vectors respectively, calculate the combined standard error, divide the difference between the two sets of means by the combined standard error to obtain the t-value, calculate the p-value according to the t-value distribution table, and output the p-value and t-value;

[0068] A significance threshold α is preset, and the p-value is compared with α: if p≤α, the difference is significant; if p>α, the difference is not significant; the sign of the t-value indicates the direction of the difference; features that satisfy p≤α are retained, and a preliminary feature subset is output.

[0069] Input the initial feature subset: if the number of features is less than 10, decrease α (e.g., 0.05 → 0.1) to increase the number of features; if the number of features is greater than 100, increase α (e.g., 0.05 → 0.01) to improve the screening criteria; output the optimized feature subset and the adjusted threshold.

[0070] Input multiple sets of pre-selected thresholds, each threshold corresponding to a feature subset of multiple thresholds. Calculate the F1 value of each subset through cross-validation, and output the final feature subset and threshold. The process of calculating the F1 value of each subset through cross-validation involves inputting multiple thresholds and their corresponding feature subsets. For each feature subset corresponding to a threshold, perform five-fold cross-validation. Each time, train the model with four sets of data, test with the remaining set, and calculate the F1 value. Finally, take the average of the five F1 values ​​and select the feature subset with the highest average F1 value and its corresponding threshold.

[0071] Texture analysis module: Inputs the filtered features into the pre-trained classification model and outputs the probability of the artistic pattern category; the texture analysis module performs the following steps:

[0072] Receive an optimized feature subset from the feature selection module, with dimensions n×m, where n is the number of samples and m is the number of features;

[0073] The m-dimensional features are reorganized into a (h×w×c) pseudo-image tensor and scaled to a standard size (224×224×3) by bilinear interpolation. The image is then input into a pre-trained deep learning model, and high-order texture semantic features are extracted using the convolutional layers of the deep learning model.

[0074] The high-order features (such as object contours and texture combinations) extracted by the convolutional layers are integrated through a fully connected layer at the end. The activation function then imparts non-linear discriminative power to these high-order features and outputs an artistic pattern category probability distribution vector (p1, p2, ..., pk), where k is the total number of predefined categories. The result evaluation module executes the following steps:

[0075] Receive category probability distribution data output by the texture analysis module;

[0076] Select the category with the highest probability as the predicted label and record the confidence level;

[0077] A heatmap of texture feature saliency is generated based on the category probability distribution, which maps the decision weights of different texture regions;

[0078] Output a confusion matrix and classification report, and label the incorrectly identified samples in each category;

[0079] The visualization results are integrated and output to the interactive interface.

[0080] Example: Data acquisition module: Select the scroll pattern area on the surface of the silk fragment, use a 24-megapixel macro lens to capture digital images, automatically correct the yellowing caused by aging, and eliminate fabric fiber noise.

[0081] Feature extraction module: The detected image is in RGB color space, first converted to XYZ color space and then mapped to sRGB standard space;

[0082] The grayscale conversion uses a visual weighting formula: Grayscale value = Red component × 0.299 + Green component × 0.587 + Blue component × 0.114;

[0083] The rotation-invariant LBP algorithm is adopted: with a radius of 3 pixels and 24 sampling points, bilinear interpolation is used to complete the edges of the incomplete pattern, generating a 256-dimensional normalized feature histogram;

[0084] Feature selection module: Standardizes the initial features, calculates the mean and standard deviation of each feature dimension, and filters through statistical tests: preset threshold α=0.05; retains 79 key features with p-values ​​<0.05.

[0085] Dynamic adjustment and optimization: Due to the excessive number of features (79>50), the threshold was increased to 0.01, and cross-validation determined that the optimal subset was 37-dimensional features;

[0086] Texture analysis module: Reconstructs 37-dimensional features into a 6×6×1 pseudo-image; scales it to a standard size of 224×224×3 using bilinear interpolation, and inputs it into a pre-trained deep learning model: Convolutional layers extract texture topology features; Fully connected layers output class probabilities: scrolling grass texture: 92.7%, cloud and thunder texture: 5.1%, others: 2.2%;

[0087] Results Evaluation Module: Generates a heatmap: Highlights the main pattern area, marks two break points that cause a decrease in confidence, and outputs a classification report: Precision 91.3%, Recall 89.7%;

[0088] Interactive interface display: The original icon marks the identification area on the left, the heat map is overlaid with a 3D texture model on the right, and the pattern evolution diagram is displayed at the bottom.

[0089] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An art pattern recognition system based on artificial intelligence, characterized in that, include: Data acquisition module: used to acquire high-resolution digital images of the target area on the surface of the artwork, and to perform noise reduction, white balance and enhancement processing; Feature extraction module: Converts the processed image into a grayscale image and extracts texture feature vectors using the rotation-invariant local binary mode method; Feature selection module: Standardizes and performs significance analysis on feature vectors to select a subset of key features; Texture analysis module: Inputs the filtered features into the pre-trained classification model and outputs the probability of the artistic pattern category; Results evaluation module: Visualizes the classification results.

2. The art pattern recognition system based on artificial intelligence according to claim 1, characterized in that, The execution steps of the data acquisition module are as follows: Select a specific target area with textured features on the surface of an artwork, including but not limited to paintings, fabrics, sculptures, and murals; acquire a digital image of the selected area using a high-resolution digital imaging device; and preprocess the acquired digital image.

3. The art pattern recognition system based on artificial intelligence according to claim 1, characterized in that, The execution steps of the feature extraction module are as follows: The preprocessed digital image is acquired and converted into a grayscale image. During the conversion process, the input digital image is verified to be in standard RGB format. The verification first checks the number of channels in the digital image. If the number of channels is not equal to three (non-RGB) or the color space label is missing, the color space conversion is triggered. The verification is performed by acquiring the number of channels and detecting the color space label. For digital images in non-standard RGB format, they are uniformly converted to the sRGB standard color space; if CMYK or Lab color space is encountered, they are first converted to XYZ color space and then mapped to sRGB. Grayscale conversion extracts the values ​​of the red, green, and blue channels of each pixel from the sRGB color space, calculates them according to the human eye's sensitivity to different colors, and finally outputs a single-channel grayscale image.

4. The art pattern recognition system based on artificial intelligence according to claim 3, characterized in that, The execution steps of the local binary pattern are as follows: It receives a single-channel grayscale image as input data, establishes a coordinate system with the current pixel as the center, calculates n uniformly distributed sampling points on the circumference based on the radius R, and uses bilinear interpolation to obtain accurate grayscale values ​​for non-integer coordinate points. Each sample point grayscale is compared with the center point to generate binary bits. The n comparison results are combined to form an initial binary code. The smallest decimal value is found by cyclic shifting. Traverse all valid pixels in the entire image, map the local binary pattern of each point to histogram intervals, and generate feature vectors by counting the frequency of occurrence in each interval. Then output the normalized LBP feature histogram.

5. The art pattern recognition system based on artificial intelligence according to claim 1, characterized in that, The execution steps of the feature selection module are as follows: Input feature vectors, eliminate dimensional differences through z-score standardization, making the feature mean 0 and variance 1, and output standardized feature vectors; z-score eliminates dimensions and calculates the arithmetic mean of all samples separately for each feature. ,in, The number of samples is the total number of valid pixels in the image that are used in the calculation. The original data values ​​are the feature codes obtained by LBP calculation for each pixel; then, the dispersion of the data relative to the mean is calculated. Perform a standardization transformation on each data point to convert the original value Convert to standardized value , ,in, This indicates the elimination of data position offset; This indicates the elimination of dimensional differences in the data; Input a standardized feature vector, use statistical tests to calculate the p-value and t-value of each feature, and output a list of feature significance indicators; Statistical test process: Input two sets of feature vectors, calculate the mean and standard deviation of the two sets of feature vectors respectively, calculate the combined standard error, divide the difference between the two sets of means by the combined standard error to obtain the t-value, calculate the p-value according to the t-value distribution table, and output the p-value and t-value.

6. The art pattern recognition system based on artificial intelligence according to claim 5, characterized in that, A significance threshold α is preset, and the p-value is compared with α: if p≤α, the difference is significant; if p>α, the difference is not significant; the sign of the t-value indicates the direction of the difference; features that satisfy p≤α are retained, and a preliminary feature subset is output. Input the initial feature subset: if the number of features is less than 10, decrease α to increase the number of features; if the number of features is greater than 100, increase α to improve the screening criteria; output the optimized feature subset and the adjusted threshold. Input multiple sets of pre-selected thresholds, each threshold corresponding to a feature subset of multiple thresholds. Calculate the F1 value of each subset through cross-validation and output the final feature subset and threshold. The process of cross-validation to calculate the F1 value of each subset involves inputting multiple thresholds and their corresponding feature subsets. For each feature subset corresponding to a threshold, five-fold cross-validation is performed. Each time, four sets of data are used to train the model, and the remaining set is used for testing and calculating the F1 value. Finally, the average of the five F1 values ​​is taken, and the feature subset with the highest average F1 value and its corresponding threshold are selected.

7. The art pattern recognition system based on artificial intelligence according to claim 1, characterized in that, The texture analysis module performs the following steps: Receive an optimized feature subset from the feature selection module, with dimensions n×m, where n is the number of samples and m is the number of features; The m-dimensional features are reorganized into a (h×w×c) pseudo-image tensor and scaled to a standard size using bilinear interpolation. The image is then input into a pre-trained deep learning model, and high-order texture semantic features are extracted using the convolutional layers of the deep learning model. The high-order features extracted by the convolutional layer are integrated by the fully connected layer at the end. The activation function gives the high-order features non-linear discriminative power and outputs the probability distribution vector of the art pattern category (p1,p2,...,pk), where k is the total number of predefined categories.

8. The art pattern recognition system based on artificial intelligence according to claim 1, characterized in that, The result evaluation module performs the following steps: Receive category probability distribution data output by the texture analysis module; Select the category with the highest probability as the predicted label and record the confidence level; A heatmap of texture feature saliency is generated based on the category probability distribution, which maps the decision weights of different texture regions; Output a confusion matrix and classification report, and label the incorrectly identified samples in each category; The visualization results are integrated and output to the interactive interface.

Images