Wood texture aesthetic feature extraction and personalized matching system

Abstract

The invention discloses a wood texture aesthetic feature extraction and personalized matching system, which belongs to the technical field of wood texture analysis and calculation aesthetics and comprises an image acquisition module, an aesthetic feature extraction module, a preference learning module, a similarity calculation module, a personalized recommendation module and a texture continuity optimization module. The method comprises the following steps: quantitatively extracting multi-dimensional aesthetic characteristic parameters of wood through multi-scale characteristic analysis and visual perception, constructing a personalized preference model based on a user selection history, calculating weighted similarity between to-be-matched wood and user preference, generating a personalized recommendation list, and providing an optimal arrangement scheme for splicing multiple pieces of wood. The technical problems that in the prior art, wood texture aesthetic feature quantification is not comprehensive, a personalized recommendation mechanism is lacked, and the multi-piece splicing effect is poor are solved, and the method is suitable for the fields of high-end furniture manufacturing, interior design, wood product personalized customization and the like.

Description

Technical Field

[0001] This invention relates to the field of wood texture analysis and computational aesthetics, and in particular to a system for extracting and personalized matching aesthetic features of wood texture. Background Technology

[0002] As an important building and decorative material, the aesthetic characteristics of wood's surface texture directly influence the market value of products and consumer preferences. In fields such as furniture manufacturing, interior design, and custom wood products, how to scientifically evaluate the aesthetic quality of wood texture and achieve personalized recommendations has always been a key issue of concern in the industry.

[0003] CN113610187A discloses a method for extracting and classifying wood textures based on image technology. This method extracts binary images of wood textures by preprocessing, grayscale processing, filtering, differential processing, and binarization of the wood image. The wood textures are then classified according to feature parameters such as the aspect ratio of the texture's bounding rectangle, primarily categorizing them into straight textures and mountain textures. While this method can achieve automatic classification of wood textures, it has the following shortcomings:

[0004] First, this method only focuses on the geometric features of the texture, such as aspect ratio, and lacks in-depth quantitative analysis of the aesthetic features of wood texture. The aesthetic value of wood texture depends not only on its geometric shape, but also on multi-dimensional aesthetic features such as contrast, complexity, regularity, and directionality. Simple geometric features cannot fully reflect the aesthetic quality of wood.

[0005] Secondly, this method uses a fixed classification standard to simply divide wood grain into two categories: straight grain and mountain grain, lacking consideration for the personalized aesthetic preferences of different consumers. Different users have significantly different aesthetic preferences for wood grain; some users prefer fine and uniform grain, while others prefer rough and bold grain. The fixed classification standard cannot meet the needs of personalized selection.

[0006] Third, this method does not address the issue of optimizing the texture continuity when splicing multiple pieces of wood. In practical applications, especially in the production of large-area wood flooring or furniture panels, the splicing effect of multiple pieces of wood directly affects the overall visual aesthetics. How to select and arrange the wood to ensure a natural and harmonious transition of textures after splicing is a technical problem that this method fails to solve.

[0007] Fourth, this method relies on binarization and simple statistical feature calculations, lacking an aesthetic quantification mechanism based on human visual perception models. Human perception of texture exhibits non-linear characteristics, and simple statistical methods are insufficient to accurately reflect human visual perception and aesthetic judgment.

[0008] Therefore, there is an urgent need for a wood texture aesthetic evaluation and matching system that can deeply extract the aesthetic features of wood texture, support personalized preference learning, optimize the splicing effect of multiple pieces, and perform scientific quantification based on a visual perception model. Summary of the Invention

[0009] The purpose of this invention is to provide a system for extracting and matching aesthetic features of wood texture, in order to solve the technical problems of incomplete quantification of aesthetic features of wood texture, lack of personalized recommendation mechanism, and poor splicing effect of multiple pieces in the prior art.

[0010] To achieve the above objectives, this invention provides a system for extracting and personalized matching aesthetic features of wood grain, comprising:

[0011] The image acquisition module is used to acquire raw texture image data of the wood surface;

[0012] An aesthetic feature extraction module is used to extract multi-dimensional aesthetic feature parameters from the original texture image data. The multi-dimensional aesthetic feature parameters include contrast features, complexity features, regularity features, and directional features.

[0013] The preference learning module is used to acquire users' historical data on wood texture selection and to build a personalized preference model based on the historical data.

[0014] The similarity calculation module is used to calculate the aesthetic similarity between the wood to be matched and the user's preferences;

[0015] A personalized recommendation module is used to generate a personalized recommendation list based on the aesthetic similarity.

[0016] The texture continuity optimization module is used to generate the best splicing arrangement when a request to splice multiple pieces of wood is received.

[0017] In one possible implementation, the aesthetic feature extraction module includes a multi-scale feature analysis unit, a visual perception quantization unit, and an aesthetic parameter calculation unit. The multi-scale feature analysis unit decomposes the original texture image data within a preset range of multiple spatial scales to obtain texture feature components at different scale levels. The visual perception quantization unit calculates the visual saliency weights of the texture feature components based on a human visual perception model. If the visual saliency weights meet a first target weight range, the texture feature components are weighted to generate a visual genome feature vector. The aesthetic parameter calculation unit calculates multi-dimensional aesthetic feature parameters based on the visual genome feature vector.

[0018] In one possible implementation, the preference learning module acquires the user's historical data on wood texture selection. If the number of samples of historical selection data meets the second target number range, a personalized preference model is constructed based on the historical selection data. The personalized preference model includes the user's preference weights for different aesthetic feature dimensions.

[0019] In one possible implementation, the similarity calculation module includes a feature space mapping unit and a preference matching unit. The feature space mapping unit maps the texture feature vector of the wood to be matched to a preset aesthetic feature space. If the dimension of the mapped feature vector conforms to the third target dimension range, the feature vector is normalized. The preference matching unit calculates the weighted similarity between the normalized feature vector and the user's preference features based on the preference weights in the personalized preference model. If the weighted similarity conforms to the fourth target similarity range, the wood to be matched is marked as a candidate recommended wood.

[0020] In one possible implementation, the texture continuity optimization module acquires the texture boundary features of the wood to be spliced, calculates the texture transition continuity index between adjacent woods, and if the texture transition continuity index meets the fifth target continuity range, then the optimal splicing arrangement scheme is generated. The optimal splicing arrangement scheme includes the splicing order and rotation angle of the wood.

[0021] Compared with the prior art, the beneficial effects of the present invention include:

[0022] First, by constructing a multi-dimensional aesthetic feature parameter system, including contrast features, complexity features, regularity features, and directional features, this invention achieves a comprehensive quantitative evaluation of the aesthetic features of wood texture, making up for the shortcomings of existing technologies that only focus on geometric features, and can more accurately reflect the aesthetic quality of wood texture.

[0023] Second, this invention introduces a quantification mechanism based on a visual perception model. By using visual saliency weights and visual genome feature vectors, the extraction of aesthetic features is made more in line with the visual perception characteristics of the human eye, thereby improving the scientificity and accuracy of aesthetic evaluation.

[0024] Third, this invention designs a personalized preference learning mechanism that can build a personalized preference model based on users' historical selection data, realizing the transformation from general aesthetic standards to personalized aesthetic recommendations and meeting the personalized aesthetic needs of different users.

[0025] Fourth, this invention innovatively designs a texture continuity optimization module, which can provide the best arrangement scheme for splicing multiple pieces of wood. By calculating the texture transition continuity index, it optimizes the splicing effect, solves the technical problem of texture incoordination when splicing wood, and improves the overall aesthetic quality of the final product.

[0026] Fifth, the present invention employs condition judgment and range checking mechanisms, such as the first target weight range and the second target quantity range, which enable the system to adaptively process according to different input conditions, thereby improving the robustness and applicability of the system. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the overall structure of the wood texture aesthetic feature extraction and personalized matching system described in this invention.

[0028] Figure 2 This is a schematic diagram of the internal structure of the aesthetic feature extraction module described in this invention.

[0029] Figure 3 This is a schematic diagram of the internal structure of the similarity calculation module described in this invention.

[0030] Figure 4 This is a schematic diagram illustrating the principle of multi-scale feature decomposition as described in this invention.

[0031] Figure 5 This is a schematic diagram of the texture continuity optimization process described in this invention. Detailed Implementation

[0032] Please refer to the attached document. Figures 1-5 The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings. It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other.

[0033] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings. It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other.

[0034] like Figure 1 As shown, the present invention provides a wood texture aesthetic feature extraction and personalized matching system, including an image acquisition module 1, an aesthetic feature extraction module 2, a preference learning module 3, a similarity calculation module 4, a personalized recommendation module 5, and a texture continuity optimization module 6.

[0035] Image acquisition module 1 is used to acquire raw texture image data of the wood surface. In one specific embodiment, image acquisition module 1 uses an industrial-grade high-resolution camera with a resolution of no less than 3840×2160 pixels, a shooting distance of 30 cm to 80 cm, and uses natural light or a standard lighting source with a color temperature of 5000K to 6500K to ensure the clarity and color accuracy of the wood texture image. The wood surface should be kept flat during shooting, and the camera lens should be perpendicular to the wood surface to avoid perspective distortion. The acquired raw texture image data is stored in RGB three-channel format, with each channel having a bit depth of 8 bits.

[0036] like Figure 2 As shown, the aesthetic feature extraction module 2 is used to extract multi-dimensional aesthetic feature parameters from the original texture image data, including contrast features, complexity features, regularity features, and directional features. The aesthetic feature extraction module 2 includes a multi-scale feature analysis unit 21, a visual perception quantization unit 22, and an aesthetic parameter calculation unit 23.

[0037] The multi-scale feature analysis unit 21 decomposes the original texture image data within a preset range of multiple spatial scales to obtain texture feature components at different scale levels. In one specific embodiment, the multi-scale feature analysis uses wavelet transform to decompose the original texture image into three scale levels: detail texture layer, mid-level texture layer, and overall texture layer.

[0038] The mathematical expression for wavelet decomposition is:

[0039] ,

[0040] in: These are the wavelet transform coefficients; For scale parameters, ; These are displacement parameters; The input texture signal; These are wavelet basis functions; It is the conjugate of the wavelet basis functions; For time or spatial location variables.

[0041] For two-dimensional image textures, the wavelet transform is extended as follows:

[0042] ,

[0043] in: For the original texture image at position Pixel value at; These are the displacement parameters in the horizontal direction; The displacement parameter is in the vertical direction; These are two-dimensional wavelet basis functions.

[0044] In the specific implementation, Daubechies wavelet or Haar wavelet is used as the wavelet basis function, and the number of decomposition layers is set to 3. The first layer corresponds to the detail texture layer, and the scale parameter is... The second layer corresponds to the middle texture layer, with scale parameters... The third layer corresponds to the overall texture layer, with scale parameters... Each layer of decomposition yields high-frequency subbands in three directions: horizontal, vertical, and diagonal, as well as a low-frequency subband.

[0045] The visual perception quantization unit 22 calculates the visual saliency weights of texture feature components based on the human eye's visual perception model. In one specific embodiment, visual perception quantization employs a perception model based on Weber's law and Fechner's law.

[0046] Weber's Law is expressed as follows:

[0047] ,

[0048] in: The smallest perceptible difference in brightness; Set the background brightness; Let be the Weber constant, for the human visual system, to .

[0049] Fechner's law is expressed as follows:

[0050] ,

[0051] in: For perceived intensity; Stimulus intensity (e.g., brightness); The stimulus threshold; It is a constant.

[0052] Based on the above law, visual saliency weight The calculation formula is:

[0053] ,

[0054] in: Image location The visual saliency weight at each location ranges from 0 to 1. For position Local contrast at that location; This represents the average contrast of the entire image. The standard deviation of the contrast ratio; This is the steepness coefficient, typically ranging from 2.0 to 5.0; It is an exponential function.

[0055] If visual salience weight Meets the first target weight range ,in , Then, the texture feature components are weighted. The formula for weighting is:

[0056] ,

[0057] in: These are the weighted feature components; These are the original feature components.

[0058] Visual genome feature vector Generates by integrating weighted feature components from multiple scale levels:

[0059] ,

[0060] in: The feature vector of the detail texture layer; The feature vector of the middle texture layer; The feature vector of the overall texture layer; superscript This represents the vector transpose operation; The dimension is The typical value ranges from 128 to 512.

[0061] The aesthetic parameter calculation unit 23 calculates multidimensional aesthetic feature parameters based on the visual genome feature vector. In one specific embodiment, the multidimensional aesthetic feature parameters include contrast features. Complexity features Regularity characteristics and directional features .

[0062] Contrast characteristics The calculation formula is:

[0063] ,

[0064] in: The local grayscale standard deviation; This represents the global grayscale standard deviation.

[0065] Local grayscale standard deviation The calculation formula is:

[0066] ,

[0067] in: This refers to the number of local windows; For the first Each local window, typically 16×16 or 32×32 pixels in size; For window The number of pixels within; For position The pixel grayscale value at that location; For window The average gray value within, .

[0068] Global grayscale standard deviation The calculation formula is:

[0069] ,

[0070] in: The total number of pixels in the image. ; Image height (pixels); Image width (pixels); The average gray value of the entire image. .

[0071] If the contrast characteristic value Does not meet the preset contrast threshold range ,in , Then, adaptive enhancement processing is applied to the contrast feature values. The enhancement processing formula is:

[0072] ,

[0073] in: These are the enhanced contrast feature values; This represents the minimum value of the contrast feature in the dataset; This represents the maximum value of the contrast feature in the dataset.

[0074] Complexity features Calculated using the entropy method:

[0075] ,

[0076] in: For an 8-bit image, the grayscale level is [number of gray levels]. ; grayscale The probability of occurrence ; grayscale The number of pixels; It is a logarithm with base 2.

[0077] Regularity features Calculated using the autocorrelation function:

[0078] ,

[0079] in: For delay The autocorrelation coefficient at time; This is the maximum delay value, typically 10% of the image width; The normalization constant is .

[0080] Autocorrelation coefficient The calculation formula is:

[0081] ,

[0082] in: This represents the number of pixels of delay in the horizontal direction.

[0083] Directional features Calculated using the Histogram of Oriented Gradients (HOG) statistic:

[0084] ,

[0085] in: For the first The cumulative gradient magnitude across all directional intervals; This represents the number of direction intervals, typically 8 or 12. This is for retrieving the maximum value.

[0086] gradient magnitude and gradient direction The calculation formula is:

[0087] ,

[0088] ,

[0089] in: The gradient component is in the horizontal direction. ; This represents the gradient component in the vertical direction. ; It is the arctangent function.

[0090] Preference learning module 3 is used to acquire users' historical data on wood texture selection. If the number of samples in the historical selection data meets the second target range, a personalized preference model is constructed based on the historical selection data. In one specific embodiment, the historical selection data includes aesthetic feature parameters of the wood samples selected by the user and the user's rating or select / reject label. Second target range for That is, when the number of historical samples of a user's choices is between 10 and 500, it is considered that the sample size is sufficient to build a personalized preference model.

[0091] like Figure 5 As shown, personalized preference models employ either weighted linear regression or neural network models. For the weighted linear regression model, preference scores... The prediction formula is:

[0092] ,

[0093] in: The preference weights for each aesthetic feature dimension; This is a bias term.

[0094] Preference weights are solved using the least squares method:

[0095] ,

[0096] in: This is the weight vector; This is the optimal weight vector; To select the number of historical data samples; For the first User ratings or tags for a sample; For the first Feature vectors of each sample; This is the regularization coefficient, typically ranging from 0.01 to 0.1; The square of the L2 norm. .

[0097] When the number of historical data samples selected is less than the lower limit of the second target number range. At that time, a cold start strategy is adopted. The cold start strategy is based on a general aesthetic preference model for initial recommendations, and the weight of the general aesthetic preference model is... , , , , These weights are derived from statistical analysis of the average preferences of a large number of users.

[0098] The personalized preference model employs an attention mechanism to dynamically weight different aesthetic feature dimensions. Attention weights. The calculation formula is:

[0099] ,

[0100] in: For the first Attention scores for each feature dimension; It is an exponential function; These correspond to contrast, complexity, regularity, and directionality, respectively.

[0101] Attention Score The calculation formula is:

[0102] ,

[0103] in: This is the attention vector; This is the attention weight matrix; For the first Feature vectors with 1 feature dimension; It is the bias vector; It is the hyperbolic tangent activation function.

[0104] Preference weights are adaptively updated based on users' recent selections. The update strategy uses a sliding window mechanism, with an update cycle of every 5 to 50 user selections. The weight update formula is:

[0105] ,

[0106] in: For the first The weight vector at the next update; For the first The weight vector at the next update; The weight vector is obtained by retraining based on the most recently selected data; This is the momentum coefficient, typically ranging from 0.7 to 0.9.

[0107] like Figure 3 As shown, the similarity calculation module 4 includes a feature space mapping unit 41 and a preference matching unit 42. The feature space mapping unit 41 maps the texture feature vector of the wood to be matched to a preset aesthetic feature space. In one specific embodiment, the mapping uses a fully connected neural network or principal component analysis (PCA) dimensionality reduction method.

[0108] For fully connected neural network mapping, the mapping formula is:

[0109] ,

[0110] in: The mapped feature vector has a dimension of ; The mapping weight matrix has dimensions of . ; The visual genomic feature vector of the wood to be matched has a dimension of ; Let be the bias vector, with dimension . ; The activation function can be either ReLU or tanh.

[0111] The third objective dimension range is: If the dimension of the mapped feature vector is... If the feature vectors conform to the range of the third target dimension, then normalization is performed. The normalization process uses the L2 norm normalization method.

[0112] ,

[0113] in: These are the normalized feature vectors; The L2 norm of the eigenvectors ; For feature vectors The Each component.

[0114] The preference matching unit 42 calculates the weighted similarity between the normalized feature vector and the user's preference features based on the preference weights in the personalized preference model. In one specific embodiment, the weighted similarity is measured using the cosine similarity method.

[0115] ,

[0116] in: For weighted similarity, the value ranges from -1 to 1, but in practical applications it is usually from 0 to 1; The user preference feature vector is obtained by weighting and averaging the feature vectors of timber selected by users in history. for The transpose of .

[0117] Considering that the L2 norm of the normalized vector is 1, the weighted similarity formula simplifies to:

[0118] ,

[0119] With the introduction of preference weights, the formula for calculating weighted similarity is further optimized as follows:

[0120] ,

[0121] in: Weighted similarity; For the first Attention weights (preference weights) for each feature dimension; For the wood to be matched in the first Normalized feature values ​​on each feature dimension; For user preferences in the first Feature values ​​in each feature dimension; For the first Similarity across multiple dimensions can be achieved using... Or other similarity measurement methods.

[0122] The fourth target similarity range is If weighted similarity It meets the fourth target similarity range, that is If the wood to be matched is selected, then it is marked as a candidate recommended wood.

[0123] The personalized recommendation module 5 sorts the candidate timber species based on weighted similarity, generating a personalized recommendation list. In one specific embodiment, the sorting is in descending order, with the timber species having the highest similarity listed first. The sorting formula is:

[0124] ,

[0125] in: For the first Recommended ranking of candidate timbers; The total number of candidate timbers recommended; This is a function that returns the ranking of similarity scores in descending order.

[0126] The personalized recommendation module 5 further ranks the candidate timber recommendations based on weighted similarity and available timber inventory information. (Comprehensive ranking score) The calculation formula is:

[0127] ,

[0128] in: This is the similarity weighting coefficient, typically ranging from 0.7 to 0.9; This is an inventory availability indicator, with a value ranging from 0 to 1. ; This refers to the available inventory quantity of timber. This is the inventory threshold, typically ranging from 10 to 100.

[0129] The personalized recommendation list contains at least five recommendations, each including a wood identifier, a similarity score, and a description of aesthetic features. The recommendation list is represented as follows:

[0130] ,

[0131] in: For the first A label for recommended timber; For the first The overall ranking score of the recommended timber; For the first A recommended description of the aesthetic characteristics of the wood, including textual descriptions of its contrast, complexity, regularity, and directional features; The total number of recommended items. .

[0132] The texture continuity optimization module 6 is used to generate the optimal splicing arrangement when a request to splice multiple pieces of wood is received. In one specific embodiment, the texture continuity optimization module 6 obtains the texture boundary features of the wood to be spliced ​​and calculates the texture transition continuity index between adjacent pieces of wood.

[0133] The texture boundary feature extraction method involves extracting the left and right boundary regions of the wood to be spliced, with a typical boundary region width of 5 to 20 pixels. (Boundary texture features) and These represent the feature vectors of the left and right boundaries of the wood, respectively. The feature vectors include the average gray value, gray variance, dominant texture direction, and color histogram statistics of the boundary region.

[0134] Texture direction difference The calculation formula is:

[0135] ,

[0136] in: For the first The dominant grain direction at the right boundary of the wood slab; For the first The dominant grain direction at the left boundary of the wood slab; This is for taking the absolute value.

[0137] Dominant texture direction Determined by the peak value of the gradient direction histogram:

[0138] ,

[0139] in: The operation that returns the argument that makes the function take the maximum value.

[0140] Color difference Using Euclidean distance metric:

[0141] ,

[0142] in: For the first Average RGB color components of the right boundary of the wood piece; For the first The average RGB color component of the left boundary of the wood piece.

[0143] Texture transition continuity index The calculation formula is:

[0144] ,

[0145] in: These are the weighting coefficients for texture direction, color, and grayscale. Typical value , , ; For the RGB color space, to achieve the maximum color difference, ; and The first Film and the first The average gray value of the boundary area of ​​the wood piece.

[0146] Preset color difference threshold Typical values ​​range from 30 to 60. The texture direction difference corresponding to the fifth target continuity range is 0 to 15 degrees, i.e. And color differences At that time, the texture transition continuity is determined to meet the requirements.

[0147] The optimal arrangement scheme is determined using an optimization algorithm. The objective function is:

[0148] ,

[0149] in: The arrangement order of the wood is a set. A permutation; This represents the total quantity of timber to be spliced. Let be the rotation angle vector of each piece of wood. ; For the first Film and the first Wood slices at rotation angle and The texture transition continuity index below.

[0150] Optimization algorithms can include genetic algorithms, simulated annealing algorithms, or greedy algorithms. For small-scale problems ( For large-scale problems, dynamic programming or branch and bound methods can be used to find the exact optimal solution. Heuristic algorithms are used to find the approximate optimal solution.

[0151] The optimal arrangement scheme is represented as follows:

[0152] ,

[0153] Where: each element Indicates the first concatenation sequence The position should be placed with the first Slice the wood and rotate it. angle.

[0154] In one specific embodiment, a greedy algorithm is used to generate the splicing and permutation scheme. The steps of the greedy algorithm are as follows:

[0155] Step 1: Randomly select a piece of wood as the starting piece, denoted as . Set the initial arrangement Initial rotation angle .

[0156] Step 2: For the first Second choice ( ), in the remaining unselected timber collection In the middle, traversing each piece of wood and all possible rotation angles Calculate and select the first Wood chips Texture transition continuity index .

[0157] Step 3: Select to use The largest combination of timber and rotation angle is denoted as and Update the arrangement .

[0158] Step 4: from Remove from the pile, and repeat steps 2 to 3 until all the wood has been selected.

[0159] Step 5: Output the optimal splicing arrangement scheme .

[0160] In another embodiment, to further improve the splicing effect, the system can also consider factors such as the size, texture density, and grade of the wood to construct a multi-objective optimization model:

[0161] ,

[0162] in: Let be the weight coefficients of each objective function, and ; For the first Quality grade rating of wood chips; For arrangement The size uniformity index measures the degree of matching of the dimensions of each piece of wood.

[0163] Dimensional uniformity index The calculation formula is:

[0164] ,

[0165] in: For the first The length of the piece of wood; This is the average length of all timbers. .

[0166] This invention presents a wood texture aesthetic feature extraction and personalized matching system. Through multi-scale feature analysis, visual perception quantification, and aesthetic parameter calculation, it achieves comprehensive quantification of wood texture aesthetic features; through preference learning and similarity calculation, it enables personalized wood recommendations; and through texture continuity optimization, it achieves the optimal arrangement for splicing multiple pieces of wood. The system can be widely applied in high-end furniture manufacturing, interior design, and personalized wood product customization, significantly enhancing the aesthetic value and user satisfaction of wood products.

[0167] It should be noted that the specific values ​​and parameter ranges in the embodiments of the present invention are merely examples, and can be adjusted according to specific needs and scenarios in actual applications. The scope of protection of the present invention is not limited to the above embodiments. Any modifications and improvements made by those skilled in the art without departing from the concept of the present invention should fall within the scope of protection defined by the claims of the present invention.

Claims

1. A system for extracting and personalized matching the aesthetic features of wood grain, characterized in that, include: The image acquisition module is used to acquire raw texture image data of the wood surface; An aesthetic feature extraction module is used to extract multi-dimensional aesthetic feature parameters from the original texture image data. These multi-dimensional aesthetic feature parameters include contrast features, complexity features, regularity features, and directional features. The aesthetic feature extraction module includes: The multi-scale feature analysis unit is used to decompose the original texture image data within a preset range of multiple spatial scales to obtain texture feature components at different scale levels. The visual perception quantization unit is used to calculate the visual saliency weight of the texture feature component based on the human eye visual perception model. If the visual saliency weight meets the first target weight range, the texture feature component is weighted to generate a visual genome feature vector. The aesthetic parameter calculation unit is used to calculate the multidimensional aesthetic feature parameters based on the visual genome feature vector. If the contrast feature value in the multidimensional aesthetic feature parameters does not meet the preset contrast threshold range, the contrast feature value is subjected to adaptive enhancement processing. The preference learning module is used to acquire users' historical data on wood texture selection. If the number of samples of the historical selection data meets the second target number range, a personalized preference model is constructed based on the historical selection data. The personalized preference model includes the user's preference weights for different aesthetic feature dimensions. A similarity calculation module is used to obtain the texture feature vector of the wood to be matched and the personalized preference model, and to calculate the aesthetic similarity between the wood to be matched and the user's preference. The similarity calculation module includes: The feature space mapping unit is used to map the texture feature vector of the wood to be matched to a preset aesthetic feature space. If the dimension of the mapped feature vector conforms to the range of the third target dimension, the feature vector is normalized. The preference matching unit is used to calculate the weighted similarity between the normalized feature vector and the user preference features based on the preference weights in the personalized preference model. If the weighted similarity meets the fourth target similarity range, the wood to be matched is marked as a candidate recommended wood. The personalized recommendation module is used to sort the candidate recommended timber according to the weighted similarity and generate a personalized recommendation list; The texture continuity optimization module is used to obtain the texture boundary features of the wood to be spliced ​​when a request to splice multiple pieces of wood is received, calculate the texture transition continuity index between adjacent wood pieces, and generate the best splicing arrangement scheme if the texture transition continuity index meets the fifth target continuity range. The best splicing arrangement scheme includes the splicing order and rotation angle of the wood pieces.

2. The wood texture aesthetic feature extraction and personalized matching system according to claim 1, characterized in that, The multi-scale feature analysis unit is further used to: decompose the original texture image data into at least three different spatial scale levels, the spatial scale levels including a detail texture layer, a mid-level texture layer and a global texture layer, each scale level corresponding to a different feature extraction filter group.

3. The wood texture aesthetic feature extraction and personalized matching system according to claim 1, characterized in that, The visual perception quantization unit is further used to: establish a visual perception model based on Weber's law and Fechner's law, the visual perception model being used to quantify the nonlinear perception characteristics of the human eye for texture contrast and complexity, the first target weight range being 0.15 to 0.

85.

4. The wood texture aesthetic feature extraction and personalized matching system according to claim 1, characterized in that, The aesthetic parameter calculation unit is further used to: when calculating the contrast feature, use the ratio of local grayscale standard deviation to global grayscale standard deviation, the preset contrast threshold range is 0.05 to 0.95, and when the contrast feature value is lower than 0.05, perform contrast enhancement processing.

5. The wood texture aesthetic feature extraction and personalized matching system according to claim 1, characterized in that, The preference learning module is further configured to: adopt a cold start strategy when the number of samples of selected historical data is less than the lower limit of the second target number range, the cold start strategy including initial recommendation based on a general aesthetic preference model, the second target number range being 10 to 500.

6. The wood texture aesthetic feature extraction and personalized matching system according to claim 1, characterized in that, The personalized preference model uses an attention mechanism to dynamically weight different aesthetic feature dimensions. The preference weights are adaptively updated based on the user's most recent selection behavior, with an update cycle of every 5 to 50 user selection operations.

7. The wood texture aesthetic feature extraction and personalized matching system according to claim 1, characterized in that, The feature space mapping unit is further used to: map the texture feature vector of the wood to be matched to an aesthetic feature space of 128 to 512 dimensions, wherein the third target dimension ranges from 128 to 512, and the normalization process adopts the L2 norm normalization method.

8. The wood texture aesthetic feature extraction and personalized matching system according to claim 1, characterized in that, When the preference matching unit calculates the weighted similarity, it uses the cosine similarity measurement method. The fourth target similarity ranges from 0.60 to 1.

00. When the weighted similarity is greater than or equal to 0.60, it is marked as a candidate recommended wood.

9. The wood texture aesthetic feature extraction and personalized matching system according to claim 1, characterized in that, The personalized recommendation module is further used to: comprehensively sort the candidate recommended timber based on the weighted similarity and the available inventory information of timber, and generate a personalized recommendation list containing at least 5 recommendation items, each recommendation item including timber identifier, similarity score and aesthetic feature description.

10. The wood texture aesthetic feature extraction and personalized matching system according to claim 1, characterized in that, The texture continuity optimization module is further used to: extract the texture features of the left and right boundaries of the wood to be spliced, calculate the texture direction difference and color difference of adjacent wood boundaries, and if the texture direction difference is less than 15 degrees and the color difference is less than a preset color difference threshold, then the texture transition continuity is determined to meet the requirements. The texture direction difference corresponding to the fifth target continuity range is 0 to 15 degrees.

Images