Cloth pattern similarity comparison method and device

Abstract

The invention discloses a cloth pattern similarity comparison method and device, and the method comprises the steps: firstly, employing a Laplacian pyramid texture removal algorithm to remove texture details from a cloth pattern, extracting a pattern with a maintained structure, carrying out the processing of the pattern, and combining the color region distribution characteristics of a structure image, re-coloring the structure image by adopting a color segmentation and reconstruction technology based on multi-region fuzzy competition; secondly, judging whether content similarity is true or not by adopting a feature extraction technology based on a convolutional neural network, and calculating image perception loss features; and finally, calculating a similarity comparison score by adopting a segmented accumulation method. According to the method, the influence of textures and colors on the similarity score of the pattern structure can be effectively reduced, and meanwhile, conditions for judging whether similarity comparison is established or not are introduced, so that similarity comparison score calculation is more effectively carried out on different types of cloth patterns. According to the method, the problem that the existing cloth pattern similarity comparison score depends on subjective evaluation is solved, and the similarity comparison score between the patterns can be objectively calculated.

Description

Technical Field

[0001] The invention relates to a cloth pattern similarity comparison method and device. Background Art

[0002] Fabric patterns are an important feature of the appearance design of patterned textiles and have important economic value. In the field of copyright protection of fabric patterns, there has been a lack of methods and systems for effectively calculating objective similarity scores. Affected by factors such as the color, texture, and combination of elements of the textile appearance, it is difficult for both manual and existing computer methods to accurately calculate the similarity score between two patterns. As the textile and apparel industry has an increasingly strong demand for intellectual property protection, the similarity evaluation of fabric patterns should try to avoid the influence of subjective factors, so an objective similarity comparison method and system are urgently needed to solve the above problems. Summary of the invention

[0003] The present invention aims to overcome the above-mentioned deficiencies in the prior art and provide a method and device capable of performing cloth pattern similarity comparison and scoring calculation.

[0004] A first aspect of the present invention relates to a method for comparing similarity of fabric patterns, comprising the following steps:

[0005] Step 1: Process the texture and color of the fabric pattern. To reduce the impact of texture detail fluctuations on image calculation, the Laplacian pyramid texture removal algorithm is used to remove texture details from the fabric pattern to obtain a pattern with intact structure. The color segmentation and reconstruction technology based on multi-region fuzzy competition is used to recolor the structural pattern using the same palette to obtain a colored structural pattern with smaller regional color differences.

[0006] Step 2, pattern content similarity judgment. The feature extraction technology based on convolutional neural network is used to judge whether the content similarity of the re-colored structural pattern is established. First, different pattern types are trained based on the large data set of fabric patterns to obtain a structural pattern classification convolutional neural network model. Then, the structural pattern is feature extracted and the distance between pattern features is calculated, and the distance value ranges between 0 and 1. A certain threshold is set to judge whether the content similarity between the structural patterns is established, and the threshold range is between 0.5 and 0.7.

[0007] Step 3: Calculate the similarity comparison score. If the structural pattern content similarity is not established, the characteristic distance value of the structural pattern calculated in step 2 is used as the similarity comparison score. If the structural pattern content similarity is established, the image perception loss feature is calculated, and based on the different emphasis of the pattern structure calculation method, a segmented accumulation method is used to calculate the final similarity comparison score.

[0008] Wherein, step 1 comprises the following steps:

[0009] Step 1.1, remove the Laplacian pyramid texture.

[0010] The cloth pattern is modeled using an image denoising algorithm based on joint bilateral filtering:

[0011]

[0012]

[0013] Where I is the input fabric pattern, is the guide cloth pattern, p and q are pixel coordinate values, is a d×d neighborhood centered on p, g σ (x) = exp(-x 2 / 2σ 2 ) is a Gaussian kernel function with standard deviation σ, s and σ r The parameters control the pattern support space and edge sensitivity respectively. A joint bilateral upsampling algorithm is used to upsample the low-resolution image in an edge-aware mode under a high-resolution guidance image. This method is denoted as JBF↑.

[0014] In order to obtain better texture denoising and structure edge preservation effects, Gaussian and Laplace pyramid hierarchical processing calculations are introduced:

[0015]

[0016]

[0017] In the formula, R k represents a smooth texture pattern with an intermediate level of k (0<k≤N, N is the coarsest level), G k-1 is a guiding image of a Gaussian pyramid at level k-1. The output pattern obtained by using the JBF↑ upsampling algorithm retains G k-1 The pattern structure in the guiding image, L K-1 Represents an image at level k-1 of the Laplacian pyramid. Superimposed pattern structure and guiding pattern The pattern structure without texture noise is approximated, and such joint bilateral filtering processing is more conducive to fabric pattern texture denoising and structure edge preservation.

[0018] Step 1.2, perform multi-region fuzzy competitive segmentation and recoloring on the fabric structure pattern;

[0019] The cloth pattern obtained in step 1.1 is used as the input image of the multi-region fuzzy competitive pattern segmentation model. Assume is a bounded open region, I(x):Ω→R 3 is the target image, which is divided into N sub-regions, denoted as Ω1,...,Ω N ; I(x) in the region Ω i The image in b(x)c i Indicates that c i represents the intermediate color value of the area, b(x):Ω→R is the color offset field of the cloth pattern, and I(x)≈c i ; The multi-region fuzzy competition pattern segmentation model is as follows:

[0020]

[0021]

[0022] The above formula satisfies the following conditions:

[0023]

[0024] In the formula, the function is a Gaussian window function, σ is a parameter that defines the window size, λ and α are empirical control parameters, and b(y) is the value of b(x) in the region Ω. i An approximate local surrogate function of ; region Ω i The membership function is u i (x) is represented by the alternating direction minimization method. The above equations are solved by optimizing the variables to be solved, and the Chambolle projection algorithm and augmented Lagrange multipliers are used for fast solution.

[0025] For the pattern obtained by the multi-region fuzzy competitive segmentation model, the central color of the original fabric pattern color offset field is replaced by a given color palette. The calculation method is as follows:

[0026]

[0027] In the formula, nc i represents the color value of a given palette, and F(x) represents the cloth pattern after recoloring.

[0028] Wherein, step 2 comprises the following steps:

[0029] Step 2.1, training of deep learning model for fabric image classification. Based on the large fabric pattern dataset, different pattern types are trained to obtain a fabric pattern classification convolutional neural network. The ResNet152 deep residual network is used, batch normalization is used to accelerate the training, and the trained model is used as the fabric pattern classification model. Since the features are obtained by training and calculation on a large-scale dataset, the feature extraction calculation of different pattern categories has a certain degree of universality.

[0030] Step 2.2, calculate the feature distance between patterns and determine whether the pattern content similarity is established. The residual network model trained in step 2.1 is used to extract features from the structural pattern and calculate the feature distance between the two patterns. The feature distance calculation method adopts the Euclidean distance calculation method. The distance value obtained by normalized data processing is between 0 and 1, and the distance value is represented by d1. A certain threshold is set to determine whether the content similarity between the patterns is established. The threshold range is between 0.5 and 0.7, and the threshold is represented by θ1. If the feature distance is greater than or equal to the threshold, the pattern content similarity is established. If the feature distance is less than the threshold, the pattern content similarity is not established.

[0031] Wherein, step 3 comprises the following steps:

[0032] Step 3.1: When the pattern content similarity is not established, the similarity comparison score is calculated. If the calculation value in step 2.2 determines that the fabric pattern content similarity is not established, the fabric pattern classification feature distance value is used as the similarity comparison score.

[0033] Step 3.2, when the pattern content similarity is established, calculate the image perception loss feature distance. The image perception loss feature calculation is based on the deep learning neural network training model, and the calculation method is as follows:

[0034]

[0035] Where d(x,x0) represents the perceptual loss feature between pattern x and pattern x0, F is the AlexNet neural network model trained with a large-scale dataset of fabric patterns, represents the normalized features extracted from the lth layer of the neural network model (a total of L layers), It is an activation parameter located in the channel of the lth layer, and ⊙ represents the multiplication operation on the channel.

[0036] Step 3.3, when the pattern content similarity is established, the similarity comparison score is calculated. When the pattern content similarity is established, based on the different emphasis of the pattern comparison score calculation method, a segmented accumulation method is used to calculate the final similarity comparison score. The calculation method is:

[0037] d s =αd1+βd (7)

[0038] Where, d s represents the similarity comparison score of the two patterns, d1 and d are calculated by step 2.2 and step 3.2 respectively, α and β are a set of empirical parameters, and satisfy α+β=1.

[0039] A second aspect of the present invention relates to a fabric pattern similarity comparison device, comprising a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, they are used to implement a fabric pattern similarity comparison method of the present invention.

[0040] A third aspect of the present invention relates to a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements a method for comparing similarity of fabric patterns of the present invention.

[0041] The present invention firstly uses the Laplace pyramid texture removal algorithm to remove texture details from the fabric pattern, extracts the pattern that maintains the structure for processing, combines the color area distribution characteristics of the structural image, and uses the color segmentation and reconstruction technology based on multi-region fuzzy competition to recolor the structural image; then, uses the feature extraction technology based on the convolutional neural network to determine whether the content similarity is established, and calculates the image perception loss characteristics; finally, uses a segmented accumulation method to calculate the similarity comparison score. The present invention can effectively reduce the influence of texture and color on the pattern structure similarity score, and at the same time introduces the conditions for judging whether the similarity comparison is established, so as to more effectively calculate the similarity comparison scores of different types of fabric patterns. The present invention solves the problem that the existing fabric pattern similarity comparison score relies on subjective evaluation, and can objectively calculate the similarity comparison scores between patterns.

[0042] The working principle of the present invention is: using a Laplace pyramid texture removal algorithm to obtain a structural pattern with texture removed, performing multi-region fuzzy competitive segmentation and reconstruction on the structural pattern to obtain a structural pattern with similar color region features, using a convolutional neural network feature extraction technology to determine whether the similarity of the structural patterns is established, combining image perception loss features, and using a segmented accumulation method to calculate the similarity comparison score.

[0043] The advantages of the present invention are: by removing and recoloring the texture and color of the fabric pattern respectively, a fabric structure pattern that is less affected by texture details and color differences can be obtained. The present invention can effectively reduce the influence of texture and color on the calculation of pattern structure similarity, and at the same time introduce the judgment of whether the pattern structure similarity is established, and combine the difference in image perception characteristics, and more effectively calculate the similarity comparison score according to the fabric pattern type. The present invention solves the problem that the existing fabric pattern similarity comparison scoring relies on subjective evaluation, and can objectively calculate the similarity comparison score between fabric patterns. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.

[0045] Figure 1 The figure is a flow chart of a method for comparing similarity of fabric patterns according to the present invention. DETAILED DESCRIPTION

[0046] Example 1

[0048] This embodiment relates to a method for comparing similarity of fabric patterns, comprising the following steps:

[0049] Step 1: Process the texture and color of the fabric pattern. To reduce the impact of texture detail fluctuations on image calculation, the Laplacian pyramid texture removal algorithm is used to remove texture details from the fabric pattern to obtain a pattern with intact structure. The color segmentation and reconstruction technology based on multi-region fuzzy competition is used to recolor the structural pattern using the same palette to obtain a colored structural pattern with smaller regional color differences.

[0050] Step 2, pattern content similarity judgment. The feature extraction technology based on convolutional neural network is used to judge whether the content similarity of the re-colored structural pattern is established. First, different pattern types are trained based on the large data set of fabric patterns to obtain a structural pattern classification convolutional neural network model. Then, the structural pattern is feature extracted and the distance between pattern features is calculated, and the distance value ranges between 0 and 1. A certain threshold is set to judge whether the content similarity between the structural patterns is established, and the threshold range is between 0.5 and 0.7.

[0051] Step 3: Calculate the similarity comparison score. If the structural pattern content similarity is not established, the characteristic distance value of the structural pattern calculated in step 2 is used as the similarity comparison score. If the structural pattern content similarity is established, the image perception loss feature is calculated, and based on the different emphasis of the pattern structure calculation method, a segmented accumulation method is used to calculate the final similarity comparison score.

[0052] Wherein, step 1 comprises the following steps:

[0053] Step 1.1, remove the Laplacian pyramid texture.

[0054] The cloth pattern is modeled using an image denoising algorithm based on joint bilateral filtering:

[0055]

[0056]

[0057] Where I is the input fabric pattern, is the guide cloth pattern, p and q are pixel coordinate values, is a d×d neighborhood centered on p, g σ (x) = exp(-x 2 / 2σ 2) is a Gaussian kernel function with standard deviation σ, s and σ r The parameters control the pattern support space and edge sensitivity respectively. A joint bilateral upsampling algorithm is used to upsample the low-resolution image in an edge-aware mode under a high-resolution guidance image. This method is denoted as JBF↑.

[0058] In order to obtain better texture denoising and structure edge preservation effects, Gaussian and Laplace pyramid hierarchical processing calculations are introduced:

[0059]

[0060]

[0061] In the formula, R k represents a smooth texture pattern with an intermediate level of k (0<k≤N, N is the coarsest level), G k-1 is a guiding image of a Gaussian pyramid at level k-1. The output pattern obtained by using the JBF↑ upsampling algorithm retains G k-1 The pattern structure in the guiding image, L K-1 Represents an image at level k-1 of the Laplacian pyramid. Superimposed pattern structure and guiding pattern The pattern structure without texture noise is approximated, and such joint bilateral filtering processing is more conducive to fabric pattern texture denoising and structure edge preservation.

[0062] Step 1.2, perform multi-region fuzzy competitive segmentation and recoloring on the fabric structure pattern;

[0063] The cloth pattern obtained in step 1.1 is used as the input image of the multi-region fuzzy competitive pattern segmentation model. Assume is a bounded open region, I(x):Ω→R 3 is the target image, which is divided into N sub-regions, denoted as Ω1,...,Ω N ; I(x) in the region Ω i The image in b(x)c i Indicates that c i represents the intermediate color value of the area, b(x):Ω→R is the color offset field of the cloth pattern, and I(x)≈c i ; The multi-region fuzzy competition pattern segmentation model is as follows:

[0064]

[0065] The above formula satisfies the following conditions:

[0066]

[0067] In the formula, the function is a Gaussian window function, σ is a parameter that defines the window size, λ and α are empirical control parameters, and b(y) is the value of b(x) in the region Ω. i An approximate local surrogate function of ; region Ω i The membership function is u i (x) is represented by the alternating direction minimization method. The above equations are solved by optimizing the variables to be solved, and the Chambolle projection algorithm and augmented Lagrange multipliers are used for fast solution.

[0068] For the pattern obtained by the multi-region fuzzy competitive segmentation model, the central color of the original fabric pattern color offset field is replaced by a given color palette. The calculation method is as follows:

[0069]

[0070] In the formula, nc i represents the color value of a given palette, and F(x) represents the cloth pattern after recoloring.

[0071] Wherein, step 2 comprises the following steps:

[0072] Step 2.1, training of deep learning model for fabric image classification. Based on the large fabric pattern dataset, different pattern types are trained to obtain a fabric pattern classification convolutional neural network. The ResNet152 deep residual network is used, batch normalization is used to accelerate the training, and the trained model is used as the fabric pattern classification model. Since the features are obtained by training and calculation on a large-scale dataset, the feature extraction calculation of different pattern categories has a certain degree of universality.

[0073] Step 2.2, calculate the feature distance between patterns and determine whether the pattern content similarity is established. The residual network model trained in step 2.1 is used to extract features from the structural pattern and calculate the feature distance between the two patterns. The feature distance calculation method adopts the Euclidean distance calculation method. The distance value obtained by normalized data processing is between 0 and 1, and the distance value is represented by d1. A certain threshold is set to determine whether the content similarity between the patterns is established. The threshold range is between 0.5 and 0.7, and the threshold is represented by θ1. If the feature distance is greater than or equal to the threshold, the pattern content similarity is established. If the feature distance is less than the threshold, the pattern content similarity is not established.

[0074] Wherein, step 3 comprises the following steps:

[0075] Step 3.1: When the pattern content similarity is not established, the similarity comparison score is calculated. If the calculation value in step 2.2 determines that the fabric pattern content similarity is not established, the fabric pattern classification feature distance value is used as the similarity comparison score.

[0076] Step 3.2, when the pattern content similarity is established, calculate the image perception loss feature distance. The image perception loss feature calculation is based on the deep learning neural network training model, and the calculation method is as follows:

[0077]

[0078] Where d(x,x0) represents the perceptual loss feature between pattern x and pattern x0, F is the AlexNet neural network model trained with a large-scale dataset of fabric patterns, represents the normalized features extracted from the lth layer of the neural network model (a total of L layers), It is an activation parameter located in the channel of the lth layer, and ⊙ represents the multiplication operation on the channel.

[0079] Step 3.3, when the pattern content similarity is established, the similarity comparison score is calculated. When the pattern content similarity is established, based on the different emphasis of the pattern comparison score calculation method, a segmented accumulation method is used to calculate the final similarity comparison score. The calculation method is:

[0080] d s =αd1+βd (14)

[0081] Where, d s represents the similarity comparison score of the two patterns, d1 and d are calculated by step 2.2 and step 3.2 respectively, α and β are a set of empirical parameters, and satisfy α+β=1.

[0082] Example 2

[0083] The present embodiment provides a fabric pattern similarity comparison device, comprising a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, the fabric pattern similarity comparison method described in Embodiment 1 is implemented.

[0084] Example 3

[0085] This embodiment relates to a computer-readable storage medium on which a program is stored. When the program is executed by a processor, the method for comparing similarity of fabric patterns described in Embodiment 1 is implemented.

[0086] The contents described in the embodiments of this specification are merely an enumeration of the implementation forms of the inventive concept. The protection scope of the present invention should not be regarded as limited to the specific forms described in the embodiments. The protection scope of the present invention also extends to equivalent technical means that can be conceived by those skilled in the art based on the inventive concept.

Claims

1. A method for comparing similarity of fabric patterns, comprising the following steps: Step 1: Process the texture and color of the cloth pattern; in order to reduce the influence of texture detail fluctuation on image calculation, the Laplacian pyramid texture removal algorithm is used to remove the texture details from the cloth pattern to obtain a pattern with intact structure; The color segmentation and reconstruction technology based on multi-region fuzzy competition is adopted to recolor the structural patterns respectively using the same color palette to obtain colored structural patterns with smaller regional color differences. Step 2, pattern content similarity judgment: use the feature extraction technology based on convolutional neural network to judge whether the content similarity of the re-colored structural pattern is established; first, based on the large data set of fabric patterns, different pattern types are trained to obtain a structural pattern classification convolutional neural network model; then, feature extraction is performed on the structural pattern and the distance between pattern features is calculated, and the distance value range is between 0 and 1; a threshold is set to judge whether the content similarity between the structural patterns is established, and the threshold range is between 0.5 and 0.7; Step 3, similarity comparison score calculation; if the structural pattern content similarity is not established, the characteristic distance value of the structural pattern calculated in step 2 is used as the similarity comparison score; If the structural pattern content similarity holds, the image perception loss features are calculated. Based on the different emphases of the pattern structure calculation method, a segmented accumulation method is used to calculate the final similarity comparison score.

2. A method for comparing fabric patterns according to claim 1, characterized in that: Step 1 includes the following steps: Step 1.1, remove the Laplacian pyramid texture; The cloth pattern is modeled using an image denoising algorithm based on joint bilateral filtering: Where I is the input fabric pattern, is the guide cloth pattern, p and q are pixel coordinate values, is a d×d neighborhood centered on p, g σ (x) = exp(-x 2 / 2σ 2 ) is a Gaussian kernel function with standard deviation σ, s and σ r The parameters control the pattern support space and edge sensitivity respectively; the joint bilateral upsampling algorithm is used to upsample the low-resolution image in edge-aware mode under a high-resolution guide image, and this method is denoted as JBF↑; In order to obtain better texture denoising and structure edge preservation effects, Gaussian and Laplace pyramid hierarchical processing calculations are introduced: In the formula, R k represents a smooth texture pattern with an intermediate level of k (0<k≤N, N is the coarsest level), G k-1 is a guiding image of a Gaussian pyramid at level k-1. The output pattern obtained by using the JBF↑ upsampling algorithm retains G k-1 The pattern structure in the guiding image, L K-1 Represents an image at level k-1 of the Laplacian pyramid; Superimposed pattern structure and guiding pattern The pattern structure without texture noise is approximated. Such joint bilateral filtering is more conducive to fabric pattern texture denoising and structure edge preservation. Step 1.2, perform multi-region fuzzy competitive segmentation and recoloring on the fabric structure pattern; The cloth pattern obtained in step 1.1 is used as the input image of the multi-region fuzzy competitive pattern segmentation model; assuming is a bounded open region, I(x):Ω→R 3 is the target image, which is divided into N sub-regions, denoted as Ω1,...,Ω N ; I(x) in the region Ω i The image in b(x)c i Indicates that c i represents the intermediate color value of the area, b(x):Ω→R is the color offset field of the cloth pattern, and I(x)≈c i ; The multi-region fuzzy competition pattern segmentation model is as follows: The above formula satisfies the following conditions: In the formula, the function is a Gaussian window function, σ is a parameter that defines the window size, λ and α are empirical control parameters, and b(y) is the value of b(x) in the region Ω. i An approximate local surrogate function of ; region Ω i The membership function is u i (x) represents; the above equation is solved by using the alternating direction minimization method to optimize the variables to be solved, and the Chambolle projection algorithm and augmented Lagrange multiplier are used for fast solution; For the pattern obtained by the multi-region fuzzy competitive segmentation model, the central color of the original fabric pattern color offset field is replaced by a given color palette. The calculation method is as follows: In the formula, nc i represents the color value of a given palette, and F(x) represents the cloth pattern after recoloring.

3. A method for comparing similarity of fabric patterns according to claim 1, characterized in that: Step 2 includes the following steps: Step 2.1, training of deep learning model for fabric image classification; training different pattern types based on a large fabric pattern dataset to obtain a fabric pattern classification convolutional neural network; using ResNet152 deep residual network, batch normalization is used to accelerate training, and the trained model is used as the fabric pattern classification model. Since the features are obtained by training and calculation on a large-scale dataset, the feature extraction calculation of different pattern categories has a certain versatility; Step 2.2, calculate the feature distance between patterns and determine whether the pattern content similarity holds; use the residual network model trained in step 2.1 to extract features from the structural pattern and calculate the feature distance between the two patterns; the feature distance calculation method uses the Euclidean distance calculation method, and the distance value obtained by normalized data processing is between 0 and 1, and the distance value is represented by d1; set a certain threshold to determine whether the content similarity between the patterns holds, the threshold range is between 0.5 and 0.7, and the threshold is represented by θ1; if the feature distance is greater than or equal to the threshold, the pattern content similarity holds; if the feature distance is less than the threshold, the pattern content similarity does not hold.

4. A method for comparing similarity of fabric patterns according to claim 1, characterized in that: Step 3 includes the following steps: Step 3.1, when the pattern content similarity is not established, the similarity comparison score is calculated; if the calculation value in step 2.2 determines that the fabric pattern content similarity is not established, the fabric pattern classification feature distance value is used as the similarity comparison score; Step 3.2, when the pattern content similarity is established, the image perception loss feature distance is calculated; the image perception loss feature calculation is based on the deep learning neural network training model, and the calculation method is as follows: Where d(x,x0) represents the perceptual loss feature between pattern x and pattern x0, F is the AlexNet neural network model trained with a large-scale dataset of fabric patterns, represents the normalized features extracted from the lth layer of the neural network model (a total of L layers), is an activation parameter located in the channel of the lth layer, ⊙ represents the multiplication operation on the channel; Step 3.3, when the pattern content similarity is established, the similarity comparison score is calculated; when the pattern content similarity is established, based on the different emphasis of the pattern comparison score calculation method, a segmented accumulation method is used to calculate the final similarity comparison score, and the calculation method is: d s =αd1+βd (7) Where, d s represents the similarity comparison score of the two patterns, d1 and d are calculated by step 2.2 and step 3.2 respectively, α and β are a set of empirical parameters, and satisfy α+β=1.

5. A cloth pattern similarity comparison device, characterized in that: The invention comprises a memory and one or more processors, wherein the memory stores executable codes, and when the one or more processors execute the executable codes, the method for comparing similarity of fabric patterns according to any one of claims 1 to 4 is implemented.

6. A computer-readable storage medium, characterized in that: A program is stored thereon, and when the program is executed by a processor, a cloth pattern similarity comparison method according to any one of claims 1 to 4 is implemented.

Images