Method for Evaluating Realism of Imitation Embroidery Design Images Based on Degraded Image Adversarial Learning

Abstract

The present invention discloses a method for evaluating the realism of an embroidered design image based on adversarial learning of degraded images, belonging to the field of digital printing technology. The method includes: constructing an adversarial learning evaluation model for the embroidered design image, which includes a multi-scale degradation algorithm module, an embroidered design image generation sub-model, and an embroidered design image evaluation sub-model; realizing the evaluation of the design realism of different regions of the embroidered design image through image degradation, image reconstruction, and a region-by-region evaluation process, in combination with adversarial technology.

Description

Technical Field

[0001] The present invention relates to a method for evaluating the realism of embroidered design images based on adversarial learning of degraded images, belonging to the field of digital printing technology. Background Art

[0002] Embroidery, as a fabric art form with a long history, its traditional production process highly depends on manual operation. With the rapid development of digital printing technology, with its efficient production capacity and resource-saving advantages, it has become the mainstream way of modern printed fabric design. In this context, the Chinese patent with the publication number CN119227549A proposes a method for generating editable conditional printing images based on deep learning, bringing new technological breakthroughs to the field of digital printing.

[0003] However, there are still technical problems to be solved in the evaluation of current embroidered image designs. First of all, the existing image realism evaluation methods are limited to global evaluation and lack the ability to analyze specific regions locally, while embroidered designs precisely require local realism evaluation to help designers accurately locate problem areas. Secondly, data is difficult to collect; the data of designers' planar draft drawings as input for the generation model has problems such as limited collection scale and difficulty in format standardization, and it is also difficult to collect a large number of software-generated embroidered simulation images for the evaluation model.

[0004] Currently, no patent in the field of printed images has realized the function of objective quality evaluation. It should be noted that in other related fields, some researchers have proposed methods for objective evaluation of tactile quality, and these research results may provide valuable reference for the objective evaluation of the quality of printed images.

[0005] The Chinese patent with the publication number CN106023208A discloses an objective evaluation method for image quality. This method requires inputting both a standard image and a distorted image at the same time, and through a preset distortion quantization algorithm, it focuses on analyzing the consistency of their edge contour features to evaluate the image degradation level. However, this evaluation method emphasizes the authenticity of the graphic structure in the image. On the one hand, it cannot achieve regional realism evaluation, and on the other hand, it is not suitable for evaluating the color and texture of embroidered images. The Chinese patent with the publication number CN118015383A discloses a training method for an image evaluation model, an image evaluation method and related devices. This method trains an image evaluation model, and after inputting the image to be evaluated, it obtains a label and generates an evaluation result, which belongs to a general-purpose image evaluation model. However, for the evaluation of embroidered images, it is difficult to collect image data and cannot achieve a relatively fine evaluation of local color and texture regions.

[0006] A Chinese patent with the publication number CN106408035A discloses a method for evaluating the realism of force - tactile reproduction based on human tactile perception characteristics. By comparing the collected real force - sense signals and the feedback data of the virtual environment, they are converted into computable "force - tactile images". After being processed by a perception filter, similarity calculation is performed in the perception - dimension space to evaluate the simulation effect. However, this solution relies on a large amount of experimental data collection, and it is difficult to collect sample data of embroidery design images, so it is difficult to apply. 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 constructing a feature matrix of real and virtual tactile experimental data, principal - component extraction technology is used to obtain core feature components, and finally, the quantization evaluation of the tactile rendering effect is realized through feature - similarity calculation, which is applicable to the evaluation of the three - dimensional convex rendering effect of electrostatic tactile devices. However, as a linear feature - extraction method, the traditional principal - component analysis method has limitations in expressing features when dealing with complex data such as embroidery images that contain non - linear features.

[0007] In summary, the above methods still do not maturely evaluate the realism of embroidered - design - image simulations. Summary of the Invention

[0008] To solve the above problems existing in the current prior art, the present invention provides a method for evaluating the realism of embroidered - design - image simulations based on degraded - image adversarial learning. Aiming at the above two problems, on the one hand, a region - division - based evaluation mechanism is developed, which can perform local realism evaluation on design images, provide more accurate problem - location for designers, and introduce a region - division strategy that conforms to real - world visual perception characteristics. By comprehensively considering visual elements such as color distribution, spatial relationship, and pattern boundaries, a region - division effect that is more in line with human visual perception is achieved; on the other hand, a multi - scale super - pixel degradation technology combined with an adversarial learning model is adopted, so that the method only needs to input real embroidery images to complete the training of the evaluation model, which can effectively reduce the dependence of the method on raw data. The technical solutions adopted by this method are as follows:

[0009] Step 1: Collect real embroidery images to construct a real - embroidery - image data set;

[0010] Step 2: Construct an embroidered - design - image adversarial - learning evaluation model. The embroidered - design - image adversarial - learning evaluation model includes a multi - scale super - pixel degradation algorithm module, an embroidered - design - image generation sub - model, and an embroidered - design - image evaluation sub - model;

[0011] Step 3: Use the data set obtained in Step 1 to train the embroidered - design - image adversarial - learning evaluation model;

[0012] Step 4: Construct the loss function of the embroidery design image generation sub-model and the loss function of the embroidery design image evaluation sub-model, and optimize the embroidery design image generation sub-model and the embroidery design image evaluation sub-model respectively during the training process;

[0013] Step 5: Input the embroidery-like design image to be evaluated into the optimized embroidery design image evaluation sub-model to obtain the evaluation result.

[0014] The real embroidery image dataset constructed in Step 1 is denoted as , where represents the serial number of the real embroidery image, represents the number of real embroidery images;

[0015] In the multi-scale superpixel degradation algorithm module in Step 2, by using the edge intensity feature to simulate the human eye's attention to the object edge, and using the multi-scale regional consistency feature to achieve the degradation effect with pattern size adaptability; and using the multi-scale superpixel segmentation algorithm to realize the reverse design process from the real embroidery image simulation to generate the superpixel degradation image; in addition, using the superpixel region division to obtain the region coordinate set approximately dividing the real embroidery pattern region;

[0016] The specific steps are as follows:

[0017] Step S1: Execute the large-scale superpixel degradation algorithm , and obtain the large-scale superpixel image A; where is the large-scale parameter, representing the number of regions into which the real embroidery image is divided in the large-scale superpixel degradation algorithm;

[0018] Step S2: Based on the large-scale superpixel image A, execute the small-scale superpixel degradation algorithm , and obtain the superpixel degradation image ; where is the small-scale parameter, representing the number of regions into which the real embroidery image is divided in the small-scale superpixel degradation algorithm;

[0019] Step S3: According to the superpixel degradation image , obtain the pixel coordinate set of the region where each superpixel is located, where , represents the superpixel serial number, represents the number of superpixel regions.

[0020] Among them, the processing processes of the large-scale superpixel degradation algorithm and the small-scale superpixel degradation algorithm include:

[0021] Step 1: Image feature extraction;

[0022] a. Take each pixel point in the real embroidery image as the spatial feature of this pixel point, denoted as ; ;

[0023] b. Convert the real embroidery image to the Lab color space, and take the channel values at the position of each pixel point in the Lab color space as the color feature of the pixel point , denoted as ;

[0024] c. Process the input real embroidery image through a pre-trained UNet model to obtain an edge intensity feature map , and take the value at the position of the pixel point in it as the edge intensity feature of the pixel point , denoted as ; among them, the UNet model is pre-trained through a dataset composed of natural object images and their corresponding edge images;

[0025] Step 2: Initialization of superpixel centers; Divide the real embroidery image uniformly into square grid regions, which are the initial superpixel regions, where represents the scale parameter; Let the central pixel point of each grid region be the initial superpixel center point , , represents the serial number of the initial superpixel region;

[0026] Step 3: Calculation of superpixel center features; Calculate the mean values of the spatial features and color features of the pixels in the initial superpixel region as the spatial feature and color feature of the superpixel center point;

[0027] Step 4: Calculation of the distance between pixels and superpixel centers; For each pixel point in the real embroidery image , in a window centered on this pixel point with a length and width of , calculate the comprehensive feature distance between the pixel point and each clustering center point in this window, and its expression is:

[0028]

[0029] wherein, is the Euclidean distance of the color features between the two, is the Euclidean distance of the spatial features between the two, is the superpixel region consistency factor, is the pixel point 's edge intensity feature, are the spatial feature distance weight parameter, the superpixel region consistency factor weight parameter, and the edge intensity feature weight parameter respectively;

[0030] wherein, when performing the large-scale superpixel degradation algorithm, the value of the superpixel region consistency factor is 0, and when performing the small-scale superpixel degradation algorithm, the value of the superpixel region consistency factor is as follows:

[0031]

[0032] Step 5: Update the superpixel center; Let each pixel point be grouped into the same superpixel with the superpixel center point having the smallest comprehensive feature distance to it, obtaining the current superpixel image , if the current iteration number has not reached the set value , then return to Step 3;

[0033] Step 6: Output the result; Output the superpixel degradation image of the th iteration.

[0034] In Step 2, the expression of the embroidery design image generation sub-model is:

[0035]

[0036] It includes a downsampling stage and an upsampling stage, wherein, represents the operation of the embroidery design image generation sub-model, represents the parameters of the embroidery design image generation sub-model;

[0037] In the downsampling stage: Input the superpixel draft , and extract the deep region features through three layers of gradually downsampling convolutional pooling modules. Each layer of convolutional pooling module sequentially includes a convolutional layer, a batch normalization layer, an activation function layer, and a pooling layer;

[0038] In the upsampling stage: Gradually upsample and reconstruct the deep region features extracted in the downsampling stage, and use three upsampling modules to restore the spatial resolution and generate the simulated design image , where each layer of the module sequentially includes an upsampling layer, a convolutional layer, a batch normalization layer, and an activation function layer.

[0039] In step two, the expression of the embroidery design image evaluation sub-model is:

[0040]

[0041] The embroidery design image evaluation sub-model includes a texture feature extraction module and an embroidery design image evaluation module, where represents the operation of the embroidery design image evaluation sub-model, represents the parameters of the embroidery design image evaluation sub-model;

[0042] The processing process of the texture feature extraction module includes:

[0043] Step a1: By adjusting the direction and scale parameters of the Gabor filter, a filter bank containing Gabor filters is constructed, and the real embroidery image and the simulated design image are respectively filtered to obtain the real embroidery image texture filtered image and the simulated design image texture filtered image , where represents the filter serial number;

[0044] Step a2: The real embroidery image texture feature and the simulated design image texture feature are respectively calculated using the real embroidery image texture filtered image and the simulated design image texture filtered image ;

[0045] The texture feature calculation process is as follows. According to the superpixel degraded image and the pixel coordinate set corresponding to each superpixel region in it, where , represents the superpixel region serial number, represents the number of superpixel regions; the mean and variance of the pixel values in the corresponding regions of the real embroidery image texture filtered image and the simulated design image texture filtered image are extracted to obtain the real embroidery image texture feature with a length of and the simulated design image texture feature ;

[0046] The processing process of the embroidery design image evaluation module includes:

[0047] The texture features of the real embroidery fabric image and the texture features of the simulated design image are respectively input into a fully-connected neural network with a sigmoid function as the output layer, and the realness score estimates of the real embroidery fabric image and the simulated design image in each superpixel region are obtained. The closer the value is to 1, the higher the realness of the region. and The closer to 1, the higher the realness of the region.

[0048] In step two, for the embroidered design images with different resolutions, following the principle of ensuring that each superpixel contains no less than 400 pixel points, the large-scale parameter and the small-scale parameter of the multi-scale superpixel degradation algorithm module are flexibly set according to this principle in the application.

[0049] The training process of step three is as follows:

[0050] The embroidery design image generation sub-model and the embroidery design image evaluation sub-model realize the training of the embroidered design image adversarial learning evaluation model through adversarial learning;

[0051] The multi-scale superpixel degradation algorithm module processes the real embroidery image to obtain the superpixel degraded image and the superpixel region division therein. The embroidery design image generation sub-model reconstructs the superpixel degraded image to obtain the simulated design image, and the embroidery design image evaluation sub-model discriminates the authenticity of the real embroidery image and the simulated design image region by region;

[0052] For the divided superpixel regions, the loss function of the embroidery design image generation sub-model and the loss function of the embroidery design image evaluation sub-model are alternately trained and optimized.

[0053] In step four:

[0054] The loss function of the embroidery design image generation sub-model is constructed using the adversarial loss function of the simulated design image and the real embroidery image and the reconstruction loss function ; The expressions of the adversarial loss function , the reconstruction loss function and the loss function of the embroidery design image generation sub-model are respectively:

[0055]

[0056]

[0057]

[0058] Among them, represents the weight coefficient of the adversarial loss function ; represents the operation of the embroidery design image generation sub-model; represents the operation of the embroidery design image evaluation sub-model;

[0059] The loss function of the embroidery design image generation sub-model updates the parameters of the embroidery design image generation sub-model .

[0060] The loss function of the embroidery design image evaluation sub-model is calculated through the estimated value of the realism score of each superpixel region and and , and its expression is:

[0061]

[0062] Among them, represents the number of superpixel regions;

[0063] The loss function of the embroidery design image evaluation sub-model updates the parameters of the embroidery design image evaluation sub-model .

[0064] When the loss function of the embroidery design image generation sub-model and the loss function of the embroidery design image evaluation sub-model no longer decrease with the training process, the training of the anti-embroidery design image adversarial learning evaluation model is completed.

[0065] In step five: The anti-embroidery design image to be evaluated is input into the optimized embroidery design image evaluation sub-model to obtain the evaluation result.

[0066] The beneficial effects of the present invention are:

[0067] The present invention proposes a method for evaluating the realism of embroidered design images based on degraded image adversarial learning. On the one hand, by developing an evaluation mechanism based on region division, it is possible to conduct a local realism assessment of the design image, providing more accurate problem positioning for designers. And a region division strategy that conforms to the characteristics of real visual perception is introduced. By comprehensively considering visual elements such as color distribution, spatial relationship, and pattern boundary, a region division effect that is more in line with human visual perception is achieved. On the other hand, by using multi-scale superpixel degradation technology combined with an adversarial learning model, the method can complete the training of the evaluation model only by inputting real embroidered fabric images, solving the problem that it is difficult to obtain paired data of embroidered design drawings and real embroidered fabric images. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings 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.

[0069] Figure 1 is a flowchart of a method for evaluating the realism of embroidered design images based on degraded image adversarial learning provided in Embodiment 1 of the present invention;

[0070] Figure 2 is a schematic diagram of an embroidered design image to be evaluated input in a method for evaluating the realism of embroidered design images based on degraded image adversarial learning provided in Embodiment 2 of the present invention;

[0071] Figure 3 is a schematic diagram of a superpixel degraded image in a method for evaluating the realism of embroidered design images based on degraded image adversarial learning provided in Embodiment 2 of the present invention;

[0072] Figure 4 is a schematic diagram of the scoring results region by region in a method for evaluating the realism of embroidered design images based on degraded image adversarial learning provided in Embodiment 2 of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0073] 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.

[0074] Embodiment 1

[0075] This embodiment provides a method for evaluating the realism of embroidered design images based on degraded image adversarial learning, which is implemented by constructing an adversarial learning evaluation model for embroidered design images. The working process of this model is as Figure 1As shown, the model specifically includes a multi-scale superpixel degradation algorithm module, an embroidery design image generation sub-model, and an embroidery design image evaluation sub-model;

[0076] First, the input real embroidery image is processed by the multi-scale superpixel degradation algorithm module to obtain a superpixel degraded image and the superpixel region division therein; then, a simulated design image is reconstructed through the embroidery design image generation sub-model ; then, the texture features of the real embroidery fabric image and the simulated design image are respectively extracted, and the authenticity of the simulated design image is evaluated region by region through the embroidery design image evaluation sub-model; then, the generation sub-model and the evaluation sub-model are trained in an adversarial learning manner, that is, for each divided image region, the generation sub-model constructs a generation loss function with the goal that the generated embroidery texture generation image is difficult to be distinguished by the evaluation sub-model, and the evaluation sub-model constructs an evaluation loss function with the goal of accurately distinguishing the real embroidery fabric image from the embroidery texture generation image, and then performs alternating training optimization; finally, the optimized embroidery design image evaluation sub-model is obtained.

[0077] The method specifically includes:

[0078] Step 1: Collect real embroidery images to construct a real embroidery image dataset;

[0079] Step 2: Construct an anti-embroidery design image adversarial learning evaluation model, which includes a multi-scale superpixel degradation algorithm module, an embroidery design image generation sub-model, and an embroidery design image evaluation sub-model;

[0080] Step 3: Use the dataset obtained in Step 1 to train the anti-embroidery design image adversarial learning evaluation model;

[0081] Step 4: Construct a loss function for the embroidery design image generation sub-model and a loss function for the embroidery design image evaluation sub-model, and optimize the embroidery design image generation sub-model and the embroidery design image evaluation sub-model respectively during the training process;

[0082] Step 5: Input the anti-embroidery design image to be evaluated into the optimized embroidery design image evaluation sub-model to obtain an evaluation result.

[0083] The real embroidery image dataset constructed in Step 1 is denoted as where represents that there are a total of 200 images in the real embroidery image dataset. The real embroidery image dataset consists of embroidery fabric images taken by a color camera, and the images have a unified size of 512×512;

[0084] In the multi-scale superpixel degradation algorithm module in Step 2, by utilizing edge intensity features to simulate the human eye's attention to object edges and using multi-scale regional consistency features to achieve a degradation effect with pattern size adaptability; and using the multi-scale superpixel segmentation algorithm to implement the reverse design process from the real embroidery image simulation to generate a superpixel degradation image; in addition, using superpixel region division to obtain a set of region coordinates approximately dividing the real embroidery pattern area;

[0085] The specific steps are as follows:

[0086] Step S1: Execute the large-scale superpixel degradation algorithm , to obtain a large-scale superpixel image A; where, is the large-scale parameter, representing the number of regions into which the real embroidery image is divided in the large-scale superpixel degradation algorithm;

[0087] Step S2: Based on the large-scale superpixel image A, execute the small-scale superpixel degradation algorithm , to obtain the superpixel degradation image ; where, is the small-scale parameter, representing the number of regions into which the real embroidery image is divided in the small-scale superpixel degradation algorithm;

[0088] Step S3: According to the superpixel degradation image , obtain the set of pixel coordinates of the region where each superpixel is located, where , k represents the superpixel serial number, represents the number of superpixel regions;

[0089] Among them, the processing processes of the large-scale superpixel degradation algorithm and the small-scale superpixel degradation algorithm include:

[0090] Step 1: Image feature extraction;

[0091] a. Take the coordinates of each pixel point in the real embroidery image as the spatial feature of this pixel point, expressed as ;

[0092] b. Convert the real embroidery image to the Lab space, and take the channel values at the position of the pixel point in the Lab space as the color feature of the pixel point , expressed as ;

[0093] c. Process the input real embroidery image through a pre-trained UNet model to obtain an edge intensity feature map . Take the value at the position of the pixel as the edge intensity feature of the pixel , denoted as ; among them, the UNet model is pre-trained through a dataset composed of natural object images and their corresponding edge images;

[0094] Step 2: Initialization of superpixel centers; Divide the real embroidery image uniformly into square grid regions, which are the initial superpixel regions, where represents the scale parameter; Let the central pixel of each grid region be the initial superpixel center point , , represents the serial number of the initial superpixel region;

[0095] Step 3: Calculation of superpixel center features; Calculate the mean of the spatial features and color features of the pixels within the initial superpixel region as the spatial feature and color feature of the superpixel center point;

[0096] Step 4: Calculation of the distance between a pixel and a superpixel center; For each pixel point in the real embroidery image , within the window with the pixel point as the center and the length and width of , calculate the comprehensive feature distance between the pixel point and each clustering center point within the window, and its expression is: where,

[0097]

[0098] Among them, is the Euclidean distance of the color features between the two, is the Euclidean distance of the spatial features between the two, is the superpixel region consistency factor, is the edge intensity feature of the pixel point , are the spatial feature distance weight parameter, the superpixel region consistency factor weight parameter, and the edge intensity feature weight parameter respectively;

[0099] Among them, when the large-scale superpixel degradation algorithm is executed, the superpixel region consistency factor takes the value of 0. When the small-scale superpixel degradation algorithm is executed, the superpixel region consistency factor takes the following values:

[0100]

[0101] Step 5: Update the superpixel center; Let each pixel point be grouped into the same superpixel with the superpixel center point having the smallest comprehensive feature distance to it, obtaining the current superpixel image . If the current iteration number has not reached the set value , then return to Step 3;

[0102] Step 6: Result output; Output the superpixel degradation image of the th iteration .

[0103] In Step 2, the expression of the embroidery design image generation sub-model is:

[0104]

[0105] It includes a downsampling stage and an upsampling stage. Among them, represents the operation of the embroidery design image generation sub-model, represents the parameters of the embroidery design image generation sub-model;

[0106] In the downsampling stage: Input the superpixel draft , and extract the deep region features through three layers of gradually downsampling convolutional pooling modules. Each layer of convolutional pooling module sequentially includes a convolutional layer, a batch normalization layer, an activation function layer, and a pooling layer;

[0107] In the upsampling stage: Gradually upsample and reconstruct the deep region features extracted in the downsampling stage, and use three upsampling modules to restore the spatial resolution and generate the simulated design image , where each layer of the module sequentially includes an upsampling layer, a convolutional layer, a batch normalization layer, and an activation function layer.

[0108] In Step 2, the expression of the embroidery design image evaluation sub-model is:

[0109]

[0110] This embroidery design image evaluation sub-model includes a texture feature extraction module and an embroidery design image evaluation module. Among them, represents the operation of the embroidery design image evaluation sub-model, represents the parameters of the embroidery design image evaluation sub-model;

[0111] The processing process of the texture feature extraction module includes:

[0112] Step a1: By adjusting the orientation and scale parameters of the Gabor filter, a filter bank consisting of Gabor filters is formed, and the real embroidery image and the simulated design image are respectively filtered to obtain the real embroidery image texture filtered image and the simulated design image texture filtered image , where represents the filter serial number;

[0113] Step a2: The real embroidery image texture features and the simulated design image texture features are respectively calculated using the real embroidery image texture filtered image and the simulated design image texture filtered image ;

[0114] The texture feature calculation process is as follows: According to the set of pixel coordinates corresponding to each superpixel region in the superpixel degraded image , where , , represents the superpixel region serial number, represents the number of superpixel regions; The mean and variance of the pixel values in the corresponding regions of the real embroidery image texture filtered image and of the simulated design image texture filtered image are extracted to obtain the real embroidery image texture features and the simulated design image texture features with a length of ;

[0115] The processing process of the embroidery design image evaluation module includes:

[0116] The real embroidery fabric image texture features and the simulated design image texture features are respectively input into a fully connected neural network with a sigmoid function as the output layer to obtain the realness score estimation values of the real embroidery fabric image and the simulated design image in each superpixel region and . The closer to 1, the higher the realness of the region.

[0117] In step two, for the simulated embroidery design images with different resolutions, following the principle that each superpixel contains no less than 400 pixel points, the large-scale parameter in the multi-scale superpixel degradation algorithm module and small-scale parameters In applications, it is flexibly set based on this principle.

[0118] The training process of Step 3 is as follows:

[0119] The embroidery design image generation sub-model and the embroidery design image evaluation sub-model implement the training of the anti-embroidery design image adversarial learning evaluation model through adversarial learning;

[0120] The multi-scale superpixel degradation algorithm module processes the real embroidery image to obtain the superpixel degradation image and the superpixel region division therein. The embroidery design image generation sub-model reconstructs the superpixel degradation image to obtain a simulated design image, and the embroidery design image evaluation sub-model discriminates the authenticity of each region between the real embroidery image and the simulated design image;

[0121] For the divided superpixel regions, alternating training optimization is performed by constructing the loss function of the embroidery design image generation sub-model and the loss function of the embroidery design image evaluation sub-model.

[0122] In Step 4:

[0123] The loss function of the embroidery design image generation sub-model Using the simulated design image and the real embroidery image of the adversarial loss function and the reconstruction loss function are constructed; the adversarial loss function , the reconstruction loss function and the loss function of the embroidery design image generation sub-model are expressed as follows:

[0124]

[0125]

[0126]

[0127] Among them, represents the weight coefficient of the adversarial loss function ; represents the operation of the embroidery design image generation sub-model; represents the operation of the embroidery design image evaluation sub-model;

[0128] Through the loss function of the embroidery design image generation sub-model update the parameters of the embroidery design image generation sub-model .

[0129] Loss function of the embroidery design image evaluation sub-model Through each superpixel region Estimated value of the sense of reality score And Calculate, and its expression is:

[0130]

[0131] Among them, Represents the number of superpixel regions;

[0132] Update the parameters of the embroidery design image evaluation sub-model through the loss function of the embroidery design image evaluation sub-model .

[0133] When the loss function of the embroidery design image generation sub-model And the loss function of the embroidery design image evaluation sub-model No longer decreases with the training process, the training of the anti-embroidery design image adversarial learning evaluation model is completed.

[0134] In step five: Input the anti-embroidery design image to be evaluated into the optimized embroidery design image evaluation sub-model to obtain the evaluation result.

[0135] Embodiment 2

[0136] This embodiment provides a method for evaluating the sense of reality of an anti-embroidery design image by multi-scale superpixel adversarial learning, which is implemented based on the anti-embroidery design image sense of reality evaluation model described in Embodiment 1;

[0137] The method specifically includes:

[0138] Input the embroidery design image to be evaluated into the anti-embroidery design image sense of reality evaluation model, and first obtain the superpixel degraded image through the processing of the multi-scale superpixel degradation algorithm module, where the embroidery design image and the superpixel degraded image are respectively as Figure 2 And Figure 3 Shown;

[0139] Utilize the superpixel region coordinate set in the superpixel degraded image , , Represents the superpixel serial number, Indicates the number of superpixel regions, and the embroidery design image evaluation sub-model performs per-region sense of reality evaluation on the embroidery design image. The regional score result is as Figure 4 Shown, the score of each superpixel Is , and the average sense of reality score calculation formula is:

[0140] ​

[0141] From the above average realism score calculation formula, the average realism score is obtained as 0.7120.

[0142] 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.

[0143] 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 principles of the present invention shall be included within the protection scope of the present invention.

Claims

1. A method for evaluating the realism of embroidered design images based on adversarial learning of degraded images, characterized in that The method includes: Step 1: Collect real embroidery images to construct a real embroidery image dataset, denoted as , where represents the serial number of the real embroidery fabric image, represents the number of real embroidery fabric images; Step 2: Construct an anti - learning evaluation model for embroidered design images. The anti - learning evaluation model for embroidered design images includes a multi - scale super - pixel degradation algorithm module, an embroidered design image generation sub - model, and an embroidered design image evaluation sub - model; Step 3: Use the data set obtained in Step 1 to train the anti - learning evaluation model for embroidered design images; Step 4: Construct the loss functions of the embroidered design image generation sub - model and the embroidered design image evaluation sub - model, and optimize the embroidered design image generation sub - model and the embroidered design image evaluation sub - model respectively during the training process; Step 5: Input the embroidered design image to be evaluated into the optimized embroidered design image evaluation sub - model to obtain an evaluation result; The processing process of the multi - scale super - pixel degradation algorithm module for the input real embroidered image includes: Step S1: Execute the large-scale superpixel degradation algorithm , and obtain the large-scale superpixel image A; among them, is the large-scale parameter, indicating the number of regions into which the real embroidery image is divided in the large-scale superpixel degradation algorithm; Step S2: Based on the large-scale superpixel image A, execute the small-scale superpixel degradation algorithm , to obtain a superpixel degradation image ; where is a small-scale parameter, representing the number of regions into which the real embroidery image is divided in the small-scale superpixel degradation algorithm; Step S3: Based on the superpixel degraded image , obtain the pixel coordinate set of the region where each superpixel is located, where , k represents the superpixel serial number, represents the number of superpixel regions; The large-scale superpixel degradation algorithm and the small-scale superpixel degradation algorithm The processing process includes: Step 1: Image feature extraction; a. Obtain the real embroidery image For each pixel point in the coordinate of which is the spatial feature of this pixel point, expressed as ; b. Convert the real embroidery image to the Lab color space and obtain the channel values at the pixel position as the color feature of the pixel , denoted as ; c. Process the input real embroidery image through a pre-trained UNet model to obtain an edge intensity feature map , and take the value at the position of the pixel as the edge intensity feature of the pixel , denoted as ; wherein, the pre-trained UNet model is pre-trained through a data set composed of natural object images and their corresponding edge images; Step 2: Initialization of superpixel centers; Divide the real embroidery image evenly into square grid regions, which are the initial superpixel regions, where represents the scale parameter; Let the central pixel point of each grid region be the initial superpixel center point , , represents the serial number of the initial superpixel region; Step 3: Superpixel center feature calculation; Calculate the initial superpixel region The mean of the spatial features and color features of the pixels within it is used as the spatial feature and color feature ; Step 4: Calculate the distance between pixels and superpixel centers; for the real embroidery image For each pixel point in it, in the window centered on this pixel point with length and width being , calculate the comprehensive feature distance between the pixel point and each cluster center point within this window , and its expression is: Among them, is the Euclidean distance of color features between the two, is the Euclidean distance of spatial features between the two, is the superpixel region consistency factor, is the pixel point 's edge intensity feature, , and are the spatial feature distance weight parameter, the superpixel region consistency factor weight parameter, and the edge intensity feature weight parameter, respectively; Among them, when the large-scale superpixel degradation algorithm is executed, the value of the superpixel region consistency factor is 0. When the small-scale superpixel degradation algorithm is executed, the value of the superpixel region consistency factor is as follows: Step 5: Update the superpixel center; Let each pixel point be grouped into the same superpixel with the superpixel center point that has the smallest comprehensive feature distance from it, obtaining the current superpixel image . If the current iteration number has not reached the set value , then return to Step 3; Step 6: Result output; output the superpixel degradation image of the th iteration .

2. The method according to claim 1, wherein In Step 2, the embroidered design image generation sub - model includes a down - sampling stage and an up - sampling stage, and its expression is: Among them, represents the operation of the embroidery design image generation sub-model, represents the parameters of the embroidery design image generation sub-model; In the downsampling stage: the input superpixel artwork passes through a convolutional pooling module with three levels of progressive downsampling to extract deep region features. Each convolutional pooling module in each layer sequentially includes a convolutional layer, a batch normalization layer, an activation function layer, and a pooling layer; In the upsampling stage: gradually upsample and reconstruct the deep regional features extracted in the downsampling stage, and use a three-layer upsampling module to restore the spatial resolution and generate a simulated design image. Each layer of the module sequentially includes an upsampling layer, a convolutional layer, a batch normalization layer, and an activation function layer.

3. The method according to claim 1, wherein In Step 2, the embroidered design image evaluation sub - model includes a texture feature extraction module and an embroidered design image evaluation module, and its expression is: Among them, represents the operation of the embroidery design image evaluation sub-model, represents the parameters of the embroidery design image evaluation sub-model; The processing process of the texture feature extraction module includes: Step a1: By adjusting the orientation and scale parameters of the Gabor filter, a filter bank consisting of Gabor filters is formed, and the real embroidery image and the simulated design image are respectively filtered to obtain the texture filtered image of the real embroidery image and the texture filtered image of the simulated design image , where represents the serial number of the parameter group; Step a2: Use the texture filtered image of the real embroidery image and the texture filtered image of the simulated design image to calculate the texture features of the real embroidery image and the texture features of the simulated design image ; The texture feature calculation process is as follows: Based on the superpixel degraded image for the pixel coordinate set corresponding to each superpixel region where , represents the superpixel region number, represents the number of superpixel regions; extract the pixel value means and variances of the corresponding regions of the texture filtered image of the real embroidery image and the texture filtered image of the simulated design image to obtain the real embroidery image texture features and the simulated design image texture features with a length of ; The processing process of the embroidered design image evaluation module includes: The texture features of the real embroidery fabric image and the texture features of the simulated design image are respectively input into a fully connected neural network with a sigmoid function as the output layer, and the real embroidery fabric image and the simulated design image are obtained. The estimated values of the realism scores in each superpixel region and are obtained. The closer the value is to 1, the higher the realism degree of the region is.

4. The method according to claim 1, characterized in that The training process of Step 3 is: The embroidered design image generation sub - model and the embroidered design image evaluation sub - model realize the training of the anti - learning evaluation model for embroidered design images through an adversarial learning method; The multi - scale super - pixel degradation algorithm module processes the real embroidered image to obtain a super - pixel degraded image and the super - pixel region division therein. The embroidered design image generation sub - model reconstructs the super - pixel degraded image to obtain a simulated design image. The embroidered design image evaluation sub - model discriminates the authenticity of each region between the real embroidered image and the simulated design image; For the divided super - pixel regions, the loss functions of the embroidered design image generation sub - model and the embroidered design image evaluation sub - model are constructed for alternating training and optimization.

5. The method according to claim 1, wherein The loss function of the embroidery design image generation sub-model Using the simulated design image and the real embroidery image of the adversarial loss function and the reconstruction loss function are constructed; the adversarial loss function , the reconstruction loss function and the loss function of the embroidery design image generation sub-model are respectively expressed as: Among them, represents the weight coefficient of the adversarial loss function ; represents the operation of the embroidery design image generation sub-model; represents the operation of the embroidery design image evaluation sub-model; Loss function of the embroidery design image generation sub-model Update the parameters of the embroidery design image generation sub-model .

6. The method according to claim 1, wherein The loss function of the embroidery design image evaluation sub-model Through each superpixel region The estimated value of the realism score And Calculate, and its expression is: Among them, represents the number of superpixel regions; Loss function of the embroidery design image evaluation sub-model Update the parameters of the embroidery design image evaluation sub-model 。 7. The method according to claim 1, characterized in that The loss function of the embroidery design image generation sub-model and the loss function of the embroidery design image evaluation sub-model When it no longer decreases with the training process, the training of the embroidery design image adversarial learning evaluation model is completed.

Images