Objective quality evaluation method for embroidery-style-imitated printing design image

Abstract

The invention discloses a quality objective evaluation method for a printing design image with an imitated embroidery style, and belongs to the technical field of digital printing. The method comprises the steps that printing design patterns processed through a degradation algorithm and corresponding grade labels are collected to serve as a data set, the grade labels are divided into five grades and correspond to different degradation degrees of images respectively, and the grade labels and image quality are in negative correlation; and training a constructed embroidery-imitating style printing design image quality objective evaluation model through the data set, further constraining the model through a loss function, optimizing parameters of the model by using a gradient descent method, and finally inputting a simulation image generated by printing image generation software into the evaluation model to obtain a degradation level of the simulation image, thereby realizing the evaluation of the quality of the printing design image. And evaluation on the quality of the analog image is further obtained.

Description

Technical Field

[0001] The present invention relates to an objective evaluation method for the quality of a printed design image in an embroidered style, belonging to the technical field of digital printing. Background Art

[0002] As a fabric art form with a long history, the traditional production process of the embroidered style highly relies on manual operation. With the rapid development of digital printing technology, it has become the mainstream way of modern printed fabric design, which can not only quickly achieve the embroidered effect, but also significantly save resources and improve production efficiency. In this context, the Chinese patent with the publication number CN119227549A proposes a method for generating editable conditional printed images based on deep learning, bringing new technological breakthroughs to the field of printed design.

[0003] However, the current evaluation process of printed image design still faces many challenges. The traditional evaluation method mainly relies on experts to judge the quality of the design draft by comparing samples and manual evaluation. There are two significant problems with this evaluation mode: First, the evaluation process is complex, time-consuming, and laborious; Second, different experts may obtain different evaluation results, making it difficult to form a unified and stable evaluation standard. Therefore, it is particularly important to establish a scientific and objective quality objective evaluation method, which can not only unify the evaluation standard, significantly improve the evaluation efficiency, but also provide reliable technical support for the generation and optimization of embroidered style images.

[0004] Currently, no patent in the field of printed images has achieved the function of objective quality evaluation. It is worth noting that in other related fields, some researchers have proposed objective evaluation methods for tactile quality, and these research results may provide valuable references for the objective quality evaluation of printed images.

[0005] The Chinese patent CN106023208A discloses an objective evaluation method for image quality. This method formulates calculation rules for degradation, requires the input of the original image and the distorted image, and calculates the similarity of the image content to measure the degree of degradation. However, the evaluation index of this scheme is mainly based on the similarity of contours, which cannot comprehensively reflect the image quality and has certain limitations.

[0006] A Chinese patent with the publication number CN106408035A discloses a method for objectively evaluating the realism of force - tactile reproduction based on human tactile perception characteristics. By collecting force - tactile data in real and virtual environments to generate "force - tactile images", using perceptual filtering to map the data into the perceptual space, and performing similarity analysis to complete the evaluation. However, this method involves complex experimental designs and data recording processes, and it is difficult to achieve flexible and efficient acquisition when the embroidery set data is too large. A Chinese patent with the publication number CN110764619A discloses a method for quantitatively evaluating the obvious realism of the tactile reproduction contour based on feature similarity. By recording the perceptual experimental data of real and virtual objects and constructing a matrix, using principal component analysis to obtain the principal eigenvectors, and calculating the similarity to quantitatively evaluate the realism of tactile reproduction, which is particularly suitable for 3D sine bump rendering of electrostatic tactile feedback devices. However, the principal component analysis method is a linear dimensionality reduction method and may not be able to capture the complex features in embroidery images.

[0007] In summary, the above methods still do not objectively evaluate the quality of printed design images in the imitation embroidery style maturely. Summary of the Invention

[0008] To solve the problems existing in the above - mentioned prior art, the present invention discloses an objective evaluation method for the quality of printed design images in the imitation embroidery style. By adopting a quantitative degradation method, several degradation levels are preset as labels for generating a deep - learning training data set. After training with a deep - learning model, this method can be used to evaluate the quality of printed design images. In addition, the present invention does not involve the input of original images and distorted images, but directly uses the label data generated by quantitative degradation for model training and image quality evaluation.

[0009] The technical solutions adopted by the present invention include:

[0010] Step 1: Collect printed design images as original images, process the original images through an image degradation algorithm to obtain degraded image samples and corresponding grade labels, and construct a data set using the degraded images and corresponding grade labels;

[0011] Step 2: Construct an objective evaluation model for the quality of printed design images in the imitation embroidery style, train the objective evaluation model using the data set obtained in Step 1, introduce an embroidery degradation level error adjustment loss function to constrain the model, and optimize the parameters of the model using the gradient descent method;

[0012] Step 3: Input the printed embroidery virtual images generated by a printed image generation software into the evaluation model obtained in Step 2, and output the evaluation grade for the printed embroidery virtual images.

[0013] The specific construction of the data set in Step 1 includes:

[0014] Step S1: Collect the printed design image in the style of imitating embroidery as the original image;

[0015] Step S2: Process the original image using an image degradation algorithm to obtain a degraded image sample, and mark the degradation level as a label. The higher the degradation level, the lower the image quality. Specifically, it includes:

[0016] Color degradation: Adjust the image saturation, which is divided into 5 levels according to the following formula:

[0017]

[0018] where, I color is the color-degraded image, I is the original image, and I gray is the grayscale image. The color degradation level parameter l is represented by discrete grading, and its value range is an integer from 0 to 4, corresponding to the 1-5 level degradation degrees respectively. Among them, l = 0 represents the non-degraded state, that is, level 1; l = 4 represents the maximum degree of degradation, that is, level 5;

[0019] Texture degradation: Add noise at 5 levels to damage the texture of the image. The expression is:

[0020] I texture = I·G h

[0021] where, I texture is the texture-degraded image, G h is the Gaussian filter kernel with size h, and h = 1 + 2w; where, w is the texture degradation level parameter, and its value range is an integer from 0 to 4, corresponding to the 1-5 level degradation degrees respectively. w = 0 represents the non-degraded state, that is, level 1; w = 4 represents the maximum degree of degradation, that is, level 5;

[0022] Light and shadow degradation: Adjust the brightness and contrast of the image. The expression is:

[0023] I light = γ·(I - μ) + μ

[0024] where, I light is the light and shadow-degraded image, γ represents the light and shadow degradation level parameter, and the specific values are 1.0, 0.9, 0.8, 0.7, 0.6, corresponding to the 1-5 level degradation levels respectively. 1.0 corresponds to level 1, representing the non-degraded state; 0.6 corresponds to level 5, representing the maximum degree of degradation, and μ represents the average brightness of the image.

[0025] The objective evaluation model for the quality of the printed design image in the style of imitating embroidery in Step 2 includes an input module, a color quality evaluation module, a texture quality evaluation module, a light and shadow quality evaluation module, and an output module;

[0026] Among them, the input module includes the reading of the printed design image in the style of embroidered pattern and the virtual image of printed embroidery;

[0027] The color quality evaluation module includes the embroidered local color level enhancement algorithm, convolutional layer, and fully connected layer;

[0028] The texture quality evaluation module includes the embroidered texture extraction algorithm, convolutional layer, and fully connected layer;

[0029] The light and shadow quality evaluation module includes the embroidered light and shadow characteristic enhancement algorithm, convolutional layer, and fully connected layer;

[0030] The color quality evaluation module, texture quality evaluation module, and light and shadow quality evaluation module respectively output the degradation levels of color, texture, and light and shadow. Finally, the average value of the three degradation levels is taken as the output of the output module.

[0031] Among them, the calculation process of the color quality evaluation module is as follows:

[0032] Step A1: For the color degradation image I color Considering the characteristics of rich local details and diverse colors in embroidered colors, local entropy is introduced; by calculating the gradient difference of color channels, in the regions with rich information, the key information of local colors is extracted, and combined with the local color mean, the two act together to strengthen the color hierarchy and three-dimensional sense, constructing an embroidered local color level enhancement algorithm to enhance the printed design image in the style of embroidered pattern. The calculation formula is:

[0033]

[0034] Among them, C(x,y) represents the color filtering result, W a ×H b is the size of the regional color window, W a represents the width of the regional color window, h b represents the height of the regional color window, and W a ,H b is set to one-tenth of the color degradation image I color , where a ∈ [1,10], b ∈ [1,10]; (x,y) represents the coordinates of the local image, x ∈ [0,W a , y ∈ [0,H b , M local (x,y) represents the color mean of the local image, H local (x,y) represents the entropy of the local image, and respectively represent the gradients of the c 1 channel and c 2 channel of the regional color window, c 1 ,c 2 belong to the RGB three channels;

[0035] Step A2: Perform convolution operation and fully connected operation on the color filtering result to obtain the color quality degradation level, and its expression is:

[0036] X cokor = FC(CNN(C(x, y)))

[0037] where X color represents the degradation level output by the color quality evaluation module.

[0038] The calculation process of the texture quality evaluation module is as follows:

[0039] Step B1: Perform Fourier transform on the texture degraded image I texture to obtain the frequency domain representation F(u, v);

[0040] Step B2: Regarding the unique frequency distribution characteristics of the embroidery texture in the image, we can enhance the texture features by analyzing the high-frequency energy distribution in the frequency domain. Since the stitch texture in the embroidery image usually shows high-frequency components, the distribution of high-frequency information in the frequency domain can be adjusted to make it tilt towards the higher frequency direction as a whole, so as to better adapt to the frequency characteristics of the embroidery texture.

[0041] Specifically, the stitch details (such as lines, edges, and complex patterns) in the embroidery image are mainly concentrated in the high-frequency region, while the background or smooth regions correspond to low-frequency components. To highlight these high-frequency texture features, we can perform non-linear enhancement on the high-frequency energy in the frequency domain while suppressing the low-frequency components. In this way, an embroidery texture extraction algorithm is constructed, and the frequency distribution will shift towards the higher frequency direction, so as to more accurately reflect the unique frequency characteristics of the embroidery texture. The calculation formula is:

[0042]

[0043] where τ is an adaptive parameter, representing the ratio of high-frequency energy to total energy in the frequency domain, m is a regulation coefficient, set to 1, D(u, v) represents the distance from the frequency domain point to the center point, D 0 represents the average distance from the frequency domain point to the center point, T(u, v) represents the texture filtering result, D max represents the maximum possible distance in the frequency domain. β is a texture index coefficient, set to 1.5, and k is a slope control parameter, set to 1;

[0044] Step B3: Perform convolution operation and fully connected operation on the texture filtering result to obtain the texture quality degradation level;

[0045] X texture = FC(CNN(T(u, v)))

[0046] Among them, X texture represents the degradation level output by the texture quality evaluation module.

[0047] The calculation process of the light and shadow quality evaluation module is as follows:

[0048] Step C1: Statistically calculate the local gradient magnitude, local gradient direction, and local contrast of the light and shadow degraded image within the local light and shadow window:

[0049]

[0050] Among them, O w represents the mean value of the local gradient magnitude, W c ×H d represents the size of the local light and shadow window, W c represents the width of the local light and shadow window, H d represents the height of the local light and shadow window, and W c ,H d is one-tenth of the width and height of the light and shadow degraded image I light c ∈ [1, 10], d ∈ [1, 10], I light (i, j) represents the value of the light and shadow degraded image I light at the point (i, j) within the local light and shadow window W c ×H d , represents the gradient of the image at the position (i, j), which is obtained by calculating the difference between adjacent pixel values, θ w represents the average direction of the local gradient, D w represents the local contrast, and represent the maximum and minimum gray values within the window, and ε is a very small constant to prevent division by zero;

[0051] Step C2: According to the characteristics of the light and shadow area in the embroidery image, that is, the gradient changes gently and the direction is relatively consistent, define the local consistency factor:

[0052]

[0053] Among them, C w represents the local consistency factor;

[0054] Step C3: By multiplying the consistency factor by the local contrast, it avoids the enhancement of areas with gentle gradients such as the background, and filters the input image through the embroidery light and shadow feature enhancement algorithm. The calculation formula is as follows:

[0055] I enh (i, j) = I light (i, j)[1 + λC w ·D w ​

[0056] Among them, I enh (i, j) represents the light and shadow filtering result, and λ represents the weight of the consistency factor;

[0057] Step C4: Perform a convolution operation and a fully connected operation on the light and shadow filtering result to obtain the light and shadow quality degradation level;

[0058] X light = FC(CNN(I enh (i, j)))

[0059] Among them, X light represents the degradation level output by the light and shadow quality evaluation module.

[0060] The expression of the embroidery degradation level error adjustment loss function adopted in Step 2 is:

[0061]

[0062] Among them, y i represents the true degradation level of the i-th sample, represents the predicted degradation level of the i-th sample, y j represents the true degradation level of the j-th sample, represents the predicted degradation level of the j-th sample. Through the loss function, the error between the predicted value and the true value of each module's score is reduced. At the same time, the inconsistency between the module scores is punished to avoid errors caused by a very large scoring error.

[0063] In Step 3, the printed embroidery virtual image is input into the evaluation model obtained in Step 2, and the evaluation model outputs the evaluation level of the printed embroidery virtual image.

[0064] The beneficial effects of the present invention are:

[0065] The present invention proposes an objective evaluation method for the quality of printed design images in an embroidered style. The original image is processed through an image degradation algorithm, and the degradation degree is divided into 5 levels. The printed design images after degradation processing and the corresponding labels are collected to construct a printed design image dataset. The objective evaluation model for the quality of the printed design images in the embroidered style is trained through the printed design image dataset, and the embroidery degradation level error adjustment loss function and the gradient descent method are respectively used to constrain and optimize the model. Finally, the quality of the printed design images is evaluated; it solves the problem that the quality of the printed design images in the embroidered style cannot be objectively evaluated in the prior art, effectively reduces the demand for manual designers or experts, significantly reduces the input of human resources, and at the same time, avoids the fluctuation of the evaluation results caused by the replacement of designers, thereby improving the stability and reliability of the evaluation. Description of the Drawings

[0066] To more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required for description in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0067] Figure 1 It is a schematic structural diagram of an objective evaluation model for the quality of an embroidered-style printed design image constructed in the objective evaluation method for the quality of an embroidered-style printed design image proposed in the first embodiment of the present invention;

[0068] Figure 2 It is a virtual schematic diagram of a printed embroidery to be evaluated proposed in the second embodiment of the present invention. Detailed implementation manners

[0069] To make the objectives, technical solutions, and advantages of the present invention clearer, the following will further describe the embodiments of the present invention in detail with reference to the drawings.

[0070] Embodiment 1

[0071] This embodiment provides an objective evaluation method for the quality of an embroidered-style printed design image. This method evaluates the quality of the printed design image by constructing an objective evaluation model for the quality of the embroidered-style printed design image. The schematic structural diagram of the model is as Figure 1 shown, including an input module, a color quality evaluation module, a texture quality evaluation module, a light and shadow quality evaluation module, and an output module;

[0072] The specific method includes:

[0073] Step 1: Collect printed design images as original images, process the original images by using an image degradation algorithm to obtain degraded image samples and corresponding evaluation labels, and construct a data set by using the degraded images and the corresponding evaluation labels;

[0074] Step 2: Construct an objective evaluation model for the quality of an embroidered-style printed design image, train the objective evaluation model by using the data set obtained in Step 1, introduce an embroidery degradation level error adjustment loss function to constrain the model, and optimize the parameters of the model by using the gradient descent method;

[0075] Step 3: Input the virtual printed embroidery image generated by the printed image generation software into the evaluation model obtained in Step 2, and output the evaluation level of the virtual printed embroidery image.

[0076] The construction of the data set in Step 1 specifically includes:

[0077] Step S1: Collect the printed design image in the style of imitating embroidery as the original image;

[0078] Step S2: Process the original image using an image degradation algorithm to obtain a degraded image sample, and mark the degradation level as a label. The higher the degradation level, the lower the image quality. Specifically, it includes:

[0079] Color degradation: Adjust the image saturation, which is divided into 5 levels according to the following formula:

[0080]

[0081] where, I color is the color-degraded image, I is the original color image, I gray is the grayscale image. The color degradation level parameter l is represented by discrete grading, and its value range is an integer from 0 to 4, corresponding to the 1-5 level degradation degrees respectively. Among them, l = 0 represents the non-degraded state (level 1), and l = 4 represents the maximum degradation (level 5);

[0082] Texture degradation: Add noise at 5 levels to damage the texture of the image, and the expression is:

[0083] I texture = I·G h

[0084] where, I texture is the texture-degraded image, G h is the Gaussian filter kernel with size h, and h = 1 + 2w; where, w is the texture degradation level parameter, and its value range is an integer from 0 to 4, corresponding to the 1-5 level degradation degrees respectively. w = 0 represents the non-degraded state, that is, level 1; w = 4 represents the maximum degradation, that is, level 5;

[0085] Light and shadow degradation: Adjust the brightness and contrast of the image, and the expression is:

[0086] I light = γ·(I - μ) + μ

[0087] where, I light is the light and shadow-degraded image, γ represents the light and shadow degradation level parameter, and the specific values are 1.0, 0.9, 0.8, 0.7, 0.6, corresponding to the 1-5 level degradation levels respectively. 1.0 corresponds to level 1, representing the non-degraded state; 0.6 corresponds to level 5, representing the maximum degradation, and μ represents the average brightness of the image.

[0088] The objective evaluation model for the quality of the printed design image in the style of imitating embroidery in Step 2 includes an input module, a color quality evaluation module, a texture quality evaluation module, a light and shadow quality evaluation module, and an output module;

[0089] Among them, the input module includes the reading of the printed design image in the style of embroidered patterns;

[0090] The color quality evaluation module includes an embroidered local color level enhancement algorithm, a convolutional layer, and a fully connected layer;

[0091] The texture quality evaluation module includes an embroidered texture extraction algorithm, a convolutional layer, and a fully connected layer;

[0092] The light and shadow quality evaluation module includes an embroidered light and shadow characteristic enhancement algorithm, a convolutional layer, and a fully connected layer;

[0093] The color quality evaluation module, the texture quality evaluation module, and the light and shadow quality evaluation module respectively output the degradation levels of color, texture, and light and shadow. Finally, the average value of the three degradation levels is taken as the output of the output module.

[0094] Among them, the calculation process of the color quality evaluation module is as follows:

[0095] Step A1: Considering the characteristics of rich local details and diverse colors in the embroidered colors of the color-degraded image, local entropy is introduced to make the algorithm automatically focus on the information-rich areas; by calculating the gradient difference of the color channels, in the information-rich areas, the key information of the local colors is extracted, and combined with the local color mean, the comprehensive effect is used to strengthen the color hierarchy and three-dimensional sense, and an embroidered local color level enhancement algorithm is constructed to enhance the printed design image in the style of embroidered patterns. The calculation formula is:

[0096]

[0097] Among them, C(x, y) represents the color filtering result, W a ×H b is the size of the regional color window, W a represents the width of the regional color window, H b represents the height of the regional color window, and W a , H b is set to one-tenth of the color-degraded image I color . Among them, a ∈ [1, 10], b ∈ [1, 10]; (x, y) represents the coordinates of the local image, x ∈ [0, W a , y ∈ [0, H b , M local (x, y) represents the color mean of the local image, H local (x, y) represents the entropy of the local image, and respectively represent the gradients of the regional color window c 1 channel and the regional color window c 2 channel, c 1 , c 2 belong to the RGB three channels;

[0098] Step A2: Perform a convolution operation and a fully connected operation on the color filtering result to obtain the color quality degradation level, and its expression is:

[0099] X color = fC(CNN(C(x, y)))

[0100] Among them, X color represents the degradation level output by the color quality evaluation module.

[0101] The calculation process of the texture quality evaluation module is as follows:

[0102] Step B1: Perform a Fourier transform on the texture degraded image I texture to obtain the frequency domain representation F(u, v);

[0103] Step B2: Regarding the unique frequency distribution characteristics of the embroidery texture in the image, we can enhance the texture features by analyzing the high-frequency energy distribution in the frequency domain. Since the stitch textures in the embroidery image usually show high-frequency components, the distribution of high-frequency information in the frequency domain can be adjusted to make it tilt towards a higher frequency direction as a whole, so as to better adapt to the frequency characteristics of the embroidery texture.

[0104] Specifically, the stitch details (such as lines, edges, and complex patterns) in the embroidery image are mainly concentrated in the high-frequency region, while the background or smooth regions correspond to low-frequency components. To highlight these high-frequency texture features, we can perform non-linear enhancement on the high-frequency energy in the frequency domain while suppressing the low-frequency components. In this way, an embroidery texture extraction algorithm is constructed, and the frequency distribution will shift towards a higher frequency direction, so as to more accurately reflect the unique frequency characteristics of the embroidery texture. The calculation formula is:

[0105]

[0106] Among them, τ is an adaptive parameter, representing the ratio of the high-frequency energy to the total energy in the frequency domain, m is a regulation coefficient, set to 1, D(u, v) represents the distance from the frequency domain point to the center point, D 0 represents the average distance from the frequency domain point to the center point, T(u, v) represents the texture filtering result, D max represents the maximum possible distance in the frequency domain. β is a texture index coefficient, set to 1.5;

[0107] Step B3: Perform a convolution operation and a fully connected operation on the texture filtering result to obtain the texture quality degradation level;

[0108] X texture = FC(CNN(T(u, v)))

[0109] Among them, X textureRepresents the degradation level output by the texture quality evaluation module.

[0110] The calculation process of the light and shadow quality evaluation module is as follows:

[0111] Step C1: Statistically calculate the local gradient magnitude, local gradient direction, and local contrast of the light and shadow degraded image I light in the local light and shadow window:

[0112]

[0113] Among them, O w represents the mean value of the local gradient magnitude, W c ×H d represents the size of the local light and shadow window, W c represents the width of the local light and shadow window, H c represents the height of the local light and shadow window, and W c , H d is one-tenth of the width and height of the light and shadow degraded image I light where c ∈ [1, 10], d ∈ [1, 10], and I light (i, j) represents the value of the light and shadow degraded image I light at the point (i, j) in the local light and shadow window W c ×H d . represents the gradient of the light and shadow degraded image I light at the position (i, j), which is obtained by calculating the difference between adjacent pixel values, θ w represents the average direction of the local gradient, and D w represents the local contrast, and represent the maximum and minimum gray values in the window, and ε is a very small constant to prevent division by zero;

[0114] Step C2: Considering the characteristics of the light and shadow areas in the embroidery image, namely, the gradient changes smoothly and the directions are relatively consistent, define the local consistency factor:

[0115]

[0116] Among them, C w represents the local consistency factor;

[0117] Step C3: By multiplying the consistency factor by the local contrast, it avoids enhancing the areas with smooth gradients such as the background, and filters the input image through the embroidery light and shadow feature enhancement algorithm. The calculation formula is as follows:

[0118] I enh (i, j) = I light (i, j)[1 + λC w·D w

[0119] Among them, I enh (i, j) represents the light and shadow filtering result, and λ represents the weight of the consistency factor;

[0120] Step C4: Perform a convolution operation and a fully connected operation on the light and shadow filtering result to obtain the light and shadow quality degradation level;

[0121] X light = FC(CNN(I enh (i, j)))

[0122] Among them, X light represents the degradation level output by the light and shadow quality evaluation module.

[0123] The calculation process of the embroidery degradation level error adjustment loss function adopted in Step 2 is as follows:

[0124]

[0125] Among them, y i represents the true degradation level of the i-th sample, represents the predicted degradation level of the i-th sample, y j represents the true degradation level of the j-th sample, represents the predicted degradation level of the j-th sample.

[0126] In Step 3, input the printed embroidery virtual image generated by the printed image generation software into the objectively evaluated model of the printed design image in the imitation embroidery style trained in Step 2 to obtain the quality degradation level, and further obtain the evaluation of the image quality.

[0127] Example 2

[0128] This example provides an objectively evaluated method for the quality of a printed design image in the imitation embroidery style. This method realizes the evaluation of the quality of the printed design image based on the objectively evaluated model of the printed design image in the imitation embroidery style described in Example 1, and specifically includes:

[0129] Input the printed embroidery virtual image to be evaluated into the objectively evaluated model of the printed design image in the imitation embroidery style for evaluation. The printed embroidery virtual image to be evaluated is as Figure 2 shown.

[0130] After being processed by the objectively evaluated model of the printed design image in the imitation embroidery style; among them, the output result of the color quality evaluation module is 3.6, the output result of the texture quality evaluation module is 2.8, the output result of the light and shadow quality evaluation model is 4.2, and the output result of the final model is 3.5. ​

[0131] The evaluation of the quality of the printed pattern design image is realized through the model of the present invention, effectively reducing the need for manual designers or experts, significantly reducing the input of human resources. At the same time, it avoids the fluctuations in the evaluation results caused by the replacement of designers, thereby improving the stability and reliability of the evaluation.

[0132] Some steps in the embodiments of the present invention can be implemented by software, and the corresponding software program can be stored in a readable storage medium, such as an optical disc or a hard disk, etc.

[0133] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. An objective evaluation method for the quality of an embroidery-style printed design image, characterized in that: The method comprises: Step 1: Collect imitation embroidery style print design images as original images, use image degradation algorithm to process the original images to obtain degraded image samples and corresponding grade labels, and use the degraded image samples and their corresponding grade labels to construct an imitation embroidery style print design image dataset; Step 2: Construct an objective evaluation model for the image quality of embroidery-style print design. The objective evaluation model is trained using the embroidery-style print design image dataset obtained in step 1. The embroidery degradation level error adjustment loss function is introduced to constrain the model and optimize the model parameters. Step 3: Input the printed embroidery virtual image generated by the printed image generation software into the imitation embroidery style printed design image quality objective evaluation model trained in step 2 to obtain the quality degradation level, and then obtain the evaluation level of the image quality; The objective evaluation model for the image quality of the imitation embroidery style print design in step 2 comprises an input module, a color quality evaluation module, a texture quality evaluation module, a light and shadow quality evaluation module and an output module; The input module includes reading of imitation embroidery style printing design images and printing embroidery virtual images; The color quality evaluation module includes an embroidery local color level enhancement algorithm, a convolutional layer, and a fully connected layer; The texture quality evaluation module includes an embroidery texture extraction algorithm, a convolutional layer, and a fully connected layer; The light and shadow quality evaluation module includes an embroidery light and shadow characteristic enhancement algorithm, a convolutional layer, and a fully connected layer; The color quality evaluation module, texture quality evaluation module and light and shadow quality evaluation module output degradation levels of color, texture and light and shadow respectively, and the average of the three degradation levels is taken as the output of the output module.

2. The method according to claim 1, characterized in that The construction of the imitation embroidery style printing design image dataset includes: Step S1: collecting a printed design image of an imitation embroidery style as an original image; Step S2: Degrade the original image to obtain a degraded image sample, and mark the degradation level as a label. The degradation level is negatively correlated with the image quality, specifically including: Color degradation; adjust the image saturation, divided into 5 levels according to the following formula: Among them I color is the color degraded image, I is the original image, and I gray For grayscale images, the color degradation level parameter l is expressed in discrete levels, and its value range is an integer from 0 to 4, corresponding to the degradation levels of 1-5, where l = 0 represents the non-degraded state, i.e., level 1; l = 4 represents the maximum degree of degradation, i.e., level 5; Texture degradation: Add 5 levels of noise to destroy the texture of the image. The formula is as follows: I texture =I·G h Among them, I texture is the texture degraded image, G h is a Gaussian filter kernel of size h, h=1+2w; wherein w is a texture degradation level parameter, and its value range is an integer from 0 to 4, corresponding to the degradation levels 1-5, w=0 indicates the non-degraded state, i.e., level 1; w=4 indicates the maximum degradation, i.e., level 5; Light and shadow degradation; adjust the brightness and contrast of the image, the formula is as follows: I light =γ·(I-μ)+μ Among them, I light is a light and shadow degraded image, γ represents the light and shadow degradation level parameter, and its specific values ​​are 1.0, 0.9, 0.8, 0.7, and 0.6, corresponding to degradation levels 1 to 5, respectively. 1.0 corresponds to level 1, indicating an undegraded state; 0.6 corresponds to level 5, indicating the maximum degree of degradation, and μ represents the average brightness of the image.

3. The method according to claim 2, characterized in that The calculation process of the color quality evaluation module in the objective evaluation model of the imitation embroidery style print design image quality is as follows: Step A1: For the color degraded image I color The local entropy is introduced into the local embroidery color. The local color information is extracted by calculating the gradient difference of the color channel. Combined with the local color mean, an embroidery local color level enhancement algorithm is constructed. The expression is: Among them, C(x,y) represents the color filtering result, W a ×H b is the regional color window size, W a Represents the width of the regional color window, H b Represents the height of the regional color window, and W a ,H b Set to color degraded image I color One tenth of the width and height, where a∈[1,10], b∈[1,10]; (x,y) represents the coordinates of the local image, x∈[0,W a ],y∈[0,H b ],M local (x, y) represents the color mean of the local image, H local (x,y) represents the entropy of the local image, Represents the gradient of the c1 channel of the regional color window, Represents the gradient of the c2 channel of the regional color window, c1 and c2 belong to the RGB three channels; Step A2: performing convolution operation and full connection operation on the color filtering result to obtain the color quality degradation level, which is expressed as: X color =FC(CNN(C(x,y))) Among them, X color Indicates the degradation level of the color quality assessment module output.

4. The method according to claim 3, characterized in that: The calculation process of the texture quality evaluation module in the objective evaluation model of the imitation embroidery style print design image quality is as follows: Step B1: Texture degraded image I texture Perform Fourier transform to obtain frequency domain representation F(u,v); Step B2: Construct an embroidery texture extraction algorithm, which is expressed as: Among them, τ is an adaptive parameter, which indicates the ratio of high frequency energy to total energy in the frequency domain, m is an adjustment coefficient, which is set to 1, d(u,v) indicates the distance from the frequency domain point to the center point, d0 indicates the mean distance from the frequency domain point to the center point, T(u,v) indicates the texture filtering result, and D max represents the maximum possible distance in the frequency domain, β is the texture index coefficient, which is set to 1.5, and k is the slope control parameter, which is set to 1; Step 3: performing convolution operation and full connection operation on the texture filtering result to obtain the texture quality degradation level; X texture =FC(CNN(T(u,v))) Among them, X texture Indicates the degradation level output by the texture quality assessment module.

5. The method according to claim 3, characterized in that: The calculation process of the light and shadow quality evaluation module in the objective evaluation model of the embroidery-style print design image quality is as follows: Step C1: Count the light and shadow degradation image I in the local light and shadow window light The mean of the local gradient amplitude, the average direction of the local gradient, and the local contrast: Among them, O w represents the mean of the local gradient amplitude, W c ×H d Indicates the size of the local light and shadow window, W c Represents the width of the local light and shadow window, H d Represents the height of the local light and shadow window, and W c ,H d is the light-shadow degraded image I light One tenth of the width and height, c∈[1,10], d∈[1,10], I light (i,j) represents the light-shadow degraded image I light In the local light and shadow window W c ×H d The value of the interior point (i,j), Represents the light-shadow degraded image I light The gradient at position (i, j) is obtained by calculating the difference between adjacent pixel values, θ w represents the average direction of the local gradient, D w represents the local contrast, and Respectively represent the maximum and minimum grayscale values ​​in the window, ε is a constant to prevent zero division; Step C2: Define the local consistency factor C for the light and shadow areas in the embroidery image, where the gradient changes smoothly and the direction is relatively consistent w , whose expression is: Step C3: By multiplying the consistency factor with the local contrast, the area with a gentle gradient is prevented from being enhanced. The input image is filtered by the embroidery light and shadow characteristic enhancement algorithm. The calculation formula is as follows: I enh (i,j)=I light (i,j)[1+λC w ·D w ] Among them, I enh (i, j) represents the light and shadow filtering result, and λ represents the weight of the consistency factor; Step C4: performing convolution operation and full connection operation on the light and shadow filtering result to obtain the light and shadow quality degradation level, which is expressed as: X light =FC(CNN(I enh (i,j))) Among them, X light Indicates the degradation level of the output of the lighting quality assessment module.

6. The method according to claim 1, characterized in that The expression of the embroidery degradation level error adjustment loss function is: Among them, y i represents the true degradation level of the i-th sample, represents the predicted degradation level of the i-th sample, y j represents the true degradation level of the jth sample, represents the predicted degradation level of the jth sample.

7. The method according to claim 1, characterized in that In the step 2, when the objective evaluation model for the image quality of the imitation embroidery style print design is trained to the point where the embroidery degradation level error adjustment loss function no longer decreases with the training process, the training is terminated; and the model parameters are optimized using the gradient descent method.

8. The method according to claim 1, characterized in that: The image degradation algorithm in step 1 is one of Gaussian blur, motion blur or image downsampling.

Images