Garment model image generation method and system

Abstract

The invention relates to the field of image processing, and relates to a clothing model image generation method and system, and the method comprises the steps: obtaining a clothing commodity graph, and extracting different types of features of corresponding clothing according to the clothing commodity graph; obtaining fusion features; generating a to-be-selected model set, and further obtaining candidate models; using the 3D human body posture model to adjust the body shape parameters of each candidate model, and adjusting the posture parameters of each candidate model; respectively calculating the similarity of the multiple types of features between each candidate model and the clothing commodity graph, and carrying out weighted summation on the similarity of each type of features so as to obtain the comprehensive similarity corresponding to each candidate model; and taking the model of which the comprehensive similarity is greater than a similarity threshold as an optimal model. By adopting the method of the invention, the generation efficiency of the clothing model image can be improved, and the difference between the clothing of the clothing model image and the clothing on the original clothing commodity image can be reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing. More particularly, the present application relates to a method and system for generating a clothing model image. BACKGROUND

[0002] Clothing product images are images specifically designed to showcase and sell clothing products. These images are used to highlight key elements of the clothing product. Through high-quality images, customers can clearly see the style, color, material, and details of the clothing, making a purchase decision.

[0003] When a merchant lists clothing products on e-commerce platforms such as Taobao and Pinduoduo, they usually need to upload clothing product images. The traditional method of making clothing product images is to have a real model try on the clothes and take pictures to obtain the clothing product images, but this method is costly.

[0004] In addition, a model template in a clothing model gallery can be used in combination with image synthesis technology to obtain a clothing model image. This method requires first obtaining a person wearing an image; then recalling a model template with a similar pose to the human body in the person wearing an image from the model gallery; then determining the deformation feature parameters of the original clothing image in the person wearing an image relative to the body image, transforming the original clothing image according to the deformation feature parameters to obtain a clothing transformation image; finally, synthesizing the clothing transformation image into the body image of the model template to obtain a model wearing an image. Since this method requires taking a person wearing an image before synthesizing a clothing model image, it results in low efficiency in generating a clothing model image. In addition, this method requires transforming the original clothing image using deformation feature parameters, which changes the features of the original clothing in terms of shape and size, resulting in a large difference between the clothing in the generated clothing model image and the original clothing on the clothing product image. SUMMARY

[0005] To solve the technical problems of low efficiency in generating a clothing model and large differences between the clothing in the generated clothing model image and the original clothing on the clothing product image in the prior art, the present application provides solutions in the following aspects.

[0006] In a first aspect, the present application provides a method for generating a clothing model image, comprising: obtaining a clothing product image and extracting different types of features of the corresponding clothing according to the clothing product image, the types of features extracted including color features, material features, size features, and style features; performing weighted fusion on the different types of features extracted to obtain fused features; input the initial model set and the fusion features into the generative adversarial network model to generate a candidate model set, and then calculate a loss function value corresponding to each candidate model, and take the candidate model with a loss function value less than a loss threshold as a candidate model; adjust the body shape parameters of each candidate model to match the garment looseness corresponding to the garment image by using the 3D human pose model, and adjust the pose parameters of each candidate model to simulate the dynamic effect of the garment; respectively calculate the similarity of each type of feature between each candidate model and the garment image, and obtain the comprehensive similarity of each candidate model by weighted summing the similarity of each type of feature; and take the model with a comprehensive similarity greater than a similarity threshold as the optimal model.

[0007] Preferably, the calculation expression of the comprehensive similarity is: ; In the formula, Similarity represents the comprehensive similarity, N represents the total number of feature types, represents the weight of the kth type of feature, represents the feature vector of the kth type of feature of the garment image, represents the feature vector of the kth type of feature of the candidate model, represents the norm of the vector.

[0008] Preferably, the multiple types of features between the candidate model and the garment image include color features and size features.

[0009] Preferably, if the garment type corresponding to the garment image is loose, the value of the similarity threshold is 0.8, otherwise, the value of the similarity threshold is 0.85.

[0010] Preferably, the method for extracting the color features of the garment includes extracting the color distribution, the main color, the auxiliary color, the proportion of cold color, the color semantic features of the garment image, and the color compatibility between the garment image and the model background, and the color semantic features include "cold color system", "warm color system" and "clashing color design".

[0011] Preferably, the method for extracting the material features of the garment includes: obtaining the texture features of the garment image by using a gray level co-occurrence matrix, and obtaining the material type of the garment image by using a VGG-16 model; and taking the texture features and the material type as the texture features of the garment image.

[0012] Preferably, the method for extracting the size features of the garment includes: inputting the garment image into a YOLOV5 model to obtain a series of predicted bounding boxes; for each predicted bounding box, calculating the intersection over union between the predicted bounding box and the real garment boundary. From all predicted bounding boxes, the bounding box with the highest intersection-union ratio (IU) is selected as the high-confidence bounding box. The pixel-level dimensions of the high-confidence bounding box and the clothing product image are then obtained. The dimensions of the high-confidence bounding box are then standardized to obtain standard clothing dimensions. The calculation expression is as follows: ; In the formula, Indicates standard clothing size. This represents the pixel-level size of the high-confidence bounding box. This indicates the pixel-level dimensions of the clothing product image. Indicates a reference dimension.

[0013] Preferably, extracting the style features of the clothing includes: inputting a clothing image into a preset Mask R-CNN model to obtain the attributes of each detected clothing component, wherein the clothing component includes one or more of collar, sleeve, button and zipper, wherein the attributes of the collar include collar shape; the attributes of the sleeve include sleeve length type; and the attributes of the buttons and zippers include whether they are closed or not.

[0014] Preferably, the extracted features also include: matching information for clothing items and clothing style tags.

[0015] Preferably, the extracted features also include: applicable scenario information for the clothing, which includes applicable season and applicable occasion.

[0016] The beneficial effects of this invention are as follows: when generating clothing models, it is not necessary to take images of the person beforehand; only images of the clothing itself are needed, thereby improving the efficiency of clothing model image generation. Since the size and shape of the clothing in the original clothing image are not altered, the difference between the clothing in the clothing model image and the clothing in the original clothing product image is reduced. Furthermore, the influence of the color, material, size, and style characteristics of the clothing in the clothing product image on the generated clothing model image is fully considered, resulting in the generated clothing model displaying clothing features that are closer to those in the clothing product image. When generating clothing model images, instead of relying on a single initial model, a set of initial models (i.e., multiple initial models) is set up to match the most suitable model in terms of gender, height, and body type based on the clothing product image, thereby further improving the generation effect of clothing model images. Attached Figure Description

[0017] The above and other objects, features, and advantages of exemplary embodiments of the present invention will become readily apparent upon reading the following detailed description with reference to the accompanying drawings. In the drawings, several embodiments of the invention are illustrated by way of example and not limitation, and like or corresponding reference numerals denote like or corresponding parts, wherein: Figure 1 is a flow chart of a clothes model image generation method according to an embodiment of the present application; Figure 2 is a structural schematic diagram of a clothes model image generation system according to an embodiment of the present application. DETAILED DESCRIPTION

[0018] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some but not all of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the protection scope of the present application.

[0019] The specific embodiments of the present application will be described in detail below with reference to the drawings.

[0020] Embodiment of clothes model image generation method: As shown in Figure 1 , the present application provides a clothes model image generation method, comprising: S101, obtaining a clothes commodity image.

[0021] The obtained clothes commodity image can be a single clothes commodity image or multiple different clothes commodity images. The clothes commodity image can be obtained through various e-commerce platforms or can be obtained by camera shooting.

[0022] S102, extracting the clothes features of the clothes commodity image, specifically: extracting different types of features of the corresponding clothes according to the clothes commodity image, the types of the extracted features including: color features, material features, size features and style features.

[0023] The color features and material features can ensure that the clothes on the generated model maintain a certain color and material as much as possible with the clothes on the clothes commodity image.

[0024] The size features can determine the body shape matching of the generated model in the subsequent steps. For example: a long skirt needs a tall model.

[0025] The style features can constrain the posture of the generated model in the subsequent steps, for example: if the collar shape is V-neck, the neck of the model needs to be displayed; the closure method of the zipper or button affects the dynamic effect of the generated model.

[0026] S103, weighting and fusing the different types of features extracted to obtain fused features.

[0027] In this embodiment, the attention mechanism can be used to weight and fuse the extracted different types of features. The attention mechanism can dynamically adjust the weight of the features, so that the model pays more attention to the features related to the task, and generates more discriminative comprehensive feature representation.

[0028] S104, generating a candidate model, specifically: setting a generative adversarial network, inputting the initial model set and the fused features into the generative adversarial network model to generate a set of candidate models, and then calculating the loss function value corresponding to each candidate model, and taking the candidate model with a loss function value less than a loss threshold as a candidate model.

[0029] The models in the initial model set include models of various ages, various heights, various postures, and various degrees of fatness.

[0030] The calculation expression of the loss function is:

[0031] In the formula, L represents the total loss function of the generative adversarial network, L represents the loss term of the discriminator for real data, L represents the loss term of the generator, L represents the expectation of random noise z and conditional information F, L represents the logarithm of the complement of the predicted value of the discriminator for the generated data G(z, F), L1 is a regularization term, which is used to constrain the difference between the generated data and the real data in the extracted various types of features. Lambda is a hyperparameter that balances the importance of the generative adversarial loss and the L1 regularization loss. L represents the L1 distance between the generated data and the real data.

[0032] The generative adversarial network can be used to generate realistic images by using the adversarial loss. By introducing the L1 loss in the loss function, the pixel-level similarity between the generated image and the real garment can be constrained, and the blur can be reduced.

[0033] S105, adjusting the body shape parameters and posture parameters of each candidate model, specifically: adjusting the body shape parameters of each candidate model using a 3D human posture model to match the garment looseness corresponding to the garment item image, and adjusting the posture parameters of each candidate model to simulate the dynamic effect of the garment.

[0034] The expression corresponding to the 3D human posture model is:

[0035] In the formula, θ represents the posture parameter, β represents the body shape parameter, M(θ, β) represents the 3D human mesh model, T(β) represents the template network, J(β) represents the joint position, and W represents the linear skinning weight.

[0036] The dynamic effect of the garment includes a skirt swing.

[0037] S106, acquire the comprehensive similarity corresponding to each candidate model, specifically, calculate the similarity of each type of feature between each candidate model and the garment commodity image, and acquire the comprehensive similarity corresponding to each candidate model by weighting and summing the similarity of each type of feature.

[0038] In this embodiment, the calculation expression of the comprehensive similarity is:

[0039] In the formula, Similarity represents the comprehensive similarity, N represents the total number of feature types, represents the weight of the kth type of feature, represents the feature vector of the kth type of feature of the garment commodity image, represents the feature vector of the kth type of feature of the candidate model, represents the norm of the vector.

[0040] S7, the model with the comprehensive similarity greater than the similarity threshold value is taken as the optimal model.

[0041] In this embodiment, if the garment type corresponding to the garment commodity image is loose, the value of the similarity threshold value is 0.8, otherwise, the value of the similarity threshold value is 0.85.

[0042] Since the tolerance of the loose type to size matching is higher, the threshold value is appropriately reduced to enhance the diversity of the optimal model screened out.

[0043] The garment model image generation method of this embodiment does not need to pre-shoot a person's highlight image when generating a garment model, and only needs to shoot an image of the garment itself, thereby improving the generation efficiency of the garment model image; in addition, the color feature, material feature, size feature and style feature of the garment of the garment commodity image are fully considered to affect the image effect of the generated garment model, so that the garment features displayed by the generated garment model are closer to the real features of the garment; when generating the garment model image, not according to one initial model, but setting an initial model set (i.e. multiple initial models), so as to match the most suitable model in terms of gender, height and body shape according to the garment commodity image, thereby further improving the generation effect of the garment model image.

[0044] In one embodiment, the total number of feature types is 2, which are color feature and size feature.

[0045] Since color is the first visual perception element, size mismatch will cause the generated result to be unavailable, therefore, considering the color feature similarity and size feature similarity between the candidate model and the clothing commodity image can effectively match the model with the best image effect.

[0046] In one embodiment, further comprising: pre-processing the clothing commodity image after obtaining it, the pre-processing including: resolution standardization and denoising.

[0047] By performing resolution standardization on the clothing commodity image, all input clothing commodity images can have the same clarity and detail level. This ensures that the generation model receives consistent quality input, which helps the model learn clothing features more stably. In addition, consistent resolution helps the generation model more accurately capture details and features of the clothing, such as texture, color, and shape. This in turn improves the model's ability to generate high-quality clothing model images.

[0048] By denoising the clothing commodity image, random noise or unwanted details in the image can be eliminated, such as sensor noise, poor lighting conditions, etc.; thereby significantly improving the image quality, making the clothing details clearer, and helping the clothing model generation model more accurately extract clothing features such as edges, contours, and textures.

[0049] In one embodiment, extracting the color features of the clothing includes: extracting the color distribution, dominant color, secondary color, cold color proportion, color semantic features of the clothing commodity image, and the color compatibility of the clothing commodity image and the model background, wherein the color semantic features include "cold color system", "warm color system", and "clashing color design".

[0050] The dominant color determines the overall visual coordination of the generated model, the secondary color is used for the detail matching of the model, and the color system affects the seasonal adaptability (cold color adapts to summer, warm color adapts to winter). By extracting the dominant color, secondary color, and cold color proportion of the clothing commodity image, the clothes on the subsequently generated model is closer to the clothing on the clothing commodity image.

[0051] In the present embodiment, the method for extracting the color distribution of the clothing commodity image includes: obtaining the color histogram of the clothing commodity image and normalizing the color histogram; when obtaining the histogram, the calculation expression of the number of pixel points with pixel value i is:

[0052] In the formula, denotes the Kronecker function, which returns 1 when I(x, y) = i, otherwise returns 0, denotes the number of pixels with pixel value i in the image; I(x, y) denotes the pixel value of the pixel point at pixel position (x, y); i denotes the pixel value.

[0053] The color histogram of the clothing commodity image can be used to count the pixel distribution of each color channel (R / G / B), and the normalization of the color histogram can eliminate the influence of different image resolutions on the comparison of the histogram.

[0054] The dominant color and the auxiliary color of the clothing commodity image can be obtained according to the color histogram. Since the K-means clustering can adaptively extract the dominant color, the k-means clustering algorithm can be used in combination with the color histogram to obtain the dominant color and the auxiliary color of the clothing commodity image.

[0055] The proportion of cold colors can be calculated by using the color histogram in combination with the proportion of cold colors formula, and the calculation formula is:

[0056] In the formula, C represents the proportion of cold colors, and N represents the total number of pixel points of the clothing commodity image. represents the number of cold color pixel points, represents the total number of pixel points of the clothing commodity image.

[0057] In this embodiment, the ResNet-50 classification model can be used to extract the color semantic features of the clothing commodity image.

[0058] In this embodiment, the color compatibility of the clothing commodity image and the model background includes: The color distribution of the clothing commodity image and the color distribution of the model background are both regarded as probability distributions, and the KL divergence is used to quantify the difference between the color distribution of the clothing commodity image and the color distribution of the model background, which is used as the color compatibility of the clothing commodity image and the model background. The calculation expression of the KL divergence is:

[0059] In the formula, D represents the KL divergence, and P represents the probability of the i th color level in the color distribution of the clothing commodity image. represents the KL divergence, represents the probability of the i th color level in the color distribution of the clothing commodity image, represents the probability of the i th color level in the color distribution of the model background, and lg represents the logarithmic function with base 10.

[0060] In one embodiment, the method for extracting the material features of the clothing includes: The texture features of the clothing commodity image are obtained by using the gray level co-occurrence matrix, and the material type of the clothing commodity image is obtained by using the VGG-16 model; and the texture features and the material type are used as the texture features of the clothing commodity image.

[0061] When the gray level co-occurrence matrix is used to obtain the texture features of the clothing commodity image, the calculation expression of the contrast is:

[0062] The calculation expression of the correlation is:

[0063] In the above two formulas, Contrast represents the contrast, and Index represents the index of the gray value, Correlation represents the correlation, and and respectively represent the average of the gray values of all pixel pairs that satisfy I(x,y)=i and I(x+d,y+d)=j. Where I(x,y) represents the pixel gray value of the image at position (x,y), and d is the distance of the pixel pair.

[0064] In one embodiment, extracting the size features of the garment includes: S201, inputting the garment commodity image into the YOLOV5 model to obtain a series of predicted bounding boxes; For each predicted bounding box, calculate its intersection over union with the true garment boundary, and the calculation formula is:

[0065] In the formula, IoU represents the intersection over union of the predicted bounding box and the true garment boundary, represents the predicted bounding box, represents the true garment boundary, and Area represents the area.

[0066] S202, select the bounding box with the largest intersection over union from all predicted bounding boxes as the high-confidence bounding box, and obtain the pixel-level size of the high-confidence bounding box and the pixel-level size of the garment commodity image, and then standardize the size of the high-confidence bounding box to obtain the standard garment size, and the calculation expression is:

[0067] In the formula, represents the standard garment size, represents the pixel-level size of the high-confidence bounding box, represents the pixel-level size of the garment commodity image, represents the reference size.

[0068] The size features of the garment include the length of the garment and the looseness threshold, and the calculation expression of the length of the garment L is:

[0069] In the formula, and respectively represent the maximum and minimum values of the y-axis coordinates of the high-confidence bounding box.

[0070] The calculation expression of the looseness threshold value R is:

[0071] In the formula, and respectively represent the clothing area and the body area of the model.

[0072] By setting the looseness threshold value, the deformation of the clothing on the subsequently generated candidate model can be avoided.

[0073] In one embodiment, the style features of the clothing are extracted by inputting the clothing product image into a preset Mask R-CNN model to obtain the attributes of each detected clothing component, the clothing component including one or more of a collar, a sleeve, a button, and a zipper, wherein the attributes of the collar include a collar shape; the attributes of the sleeve include a sleeve length type; and the attributes of the button and the zipper include closed or not closed.

[0074] The training process of the Mask R-CNN model includes: S301, collecting clothing product images containing different collar types, different sleeve lengths, and different button closure states; Wherein, the collar type includes round collar, V-neck, square collar, etc.; the sleeve length includes, for example, short sleeve, long sleeve, seven-eighth sleeve, etc.; and the button closure state includes, for example, fully open, half open, closed, etc.

[0075] S302, performing component-level representation on each clothing product image, drawing an accurate segmentation mask for each clothing component, and labeling its attributes, thereby obtaining a data set; and dividing the data set into a training set and a validation set; S303, training the Mask R-CNN model using the training set, and optimizing the parameters of the model during the training process to improve the segmentation accuracy and attribute classification accuracy.

[0076] S304, evaluating the performance of the model using the validation set, including segmentation accuracy, attribute classification accuracy, and other indicators. And according to the evaluation results, adjust the model parameters or training strategy to improve the performance of the model.

[0077] In one embodiment, the type of extracted features also includes set matching information and clothing style labels corresponding to the clothing product image.

[0078] The set matching information acquisition method includes: Using a target detection model to identify different components in the clothing product image; Calculating the intersection over union (LOU) of different component regions, and the calculation formula is:

[0079] In the formula, A and B respectively represent the areas of the two identified components, A B represents the area of the overlapping part of the two components, A B represents the total area of the two components.

[0080] The clothing item image with an intersection-over-union greater than 0.6 is determined as a suit.

[0081] The title information of the clothing item image can be parsed using an NLP (natural language processing) model to obtain a clothing style label.

[0082] The NLP can supplement the semantic information missing from the visual features.

[0083] The style label can be used to guide the generation of the background of the model (e.g., a formal dress requires a simple scene).

[0084] A suit usually includes multiple clothing items that match each other, and these items have certain coordination in terms of style, color, and material. Identifying whether the clothing in the clothing item image is a suit helps the generative adversarial network to maintain the style coordination and integrity between the clothing items when generating model images, so that the combination of clothing in the generated model images is more natural, harmonious, and in line with fashion rules and aesthetic standards.

[0085] In one embodiment, the types of extracted features further include: application scene information of the clothing, the application scene information including an applicable season and an applicable occasion.

[0086] The method for obtaining the application scene information is: inputting the clothing item image into a preset ResNet-18 model to obtain the probability distribution of the season classification and the occasion classification. The class with the highest probability is selected as the final result, i.e., the season (spring / summer / autumn / winter) and the occasion (sports / business) to which the clothing is applicable.

[0087] In this embodiment, the training process of the ResNet-18 model is as follows: S401, constructing a data set: collecting clothing item images containing different seasons (summer / winter) and occasions (sports / business) and labeling them. Ensure that the data set covers a variety of clothing styles, colors, and materials.

[0088] S401, data division: divide the data set into a training set and a validation set, with a ratio of 8:2 or 7:3, to ensure reasonable data distribution.

[0089] S403, pre-processing the data in the training set and the validation set, including image enhancement and standardization processing.

[0090] Image augmentation: Apply image augmentation operations such as random cropping, flipping, adjusting brightness / contrast, etc. on the training set to increase data diversity and improve model generalization.

[0091] Standardization: Uniformly adjust the image size to the default size of the ResNet-18 model input (e.g. 224x224), and perform normalization processing (e.g. scale the pixel value to the range of [0, 1]).

[0092] S404, model training, including: 1) Load pre-trained model: Use the pre-trained ResNet-18 model on a large dataset (such as ImageNet) as a starting point, and use transfer learning technology to accelerate model convergence.

[0093] 2) Adjust model structure: Modify the output layer of the ResNet-18 model, replace the last fully connected layer with two independent classification heads, respectively for seasonal classification (summer / winter) and occasion classification (sports / business). The number of neurons in each classification head is set according to the number of categories (seasonal classification is 2, occasion classification is 2).

[0094] 3) Set hyperparameters: Optimizer: Choose Adam or AdamW optimizer, initial learning rate can be set to 1e-4 or smaller, adjust according to training conditions.

[0095] Loss function: Use cross-entropy loss function, calculate the loss of seasonal classification and occasion classification respectively, and perform weighted sum (such as each weight is 0.5).

[0096] Batch size: Set according to the size of the memory, generally 32 or 64.

[0097] Training period: Adjust according to the performance of the validation set, can be set to 10-20 cycles.

[0098] 4) Model training: Input the preprocessed training set into the model for training, evaluate the model performance on the validation set after each cycle. Adjust hyperparameters such as learning rate decay strategy, data augmentation method, etc. according to the validation set results.

[0099] By obtaining the applicable season, the environment light of the generated model can be constrained (such as warm color in winter), and by obtaining the applicable occasion, the posture of the generated model can be matched (such as dynamic posture for sports wear).

[0100] In one embodiment, the types of extracted features also include: logo position of the clothing, patterns on the clothing, and functional labels of the clothing, the logo position includes two positions of left chest and sleeve, the logo pattern includes two patterns of stripes and prints; the functional label of the clothing includes waterproof.

[0101] By obtaining the logo position, the clothing logo can be clearly visible in the generated model image.

[0102] Since complex patterns require high-resolution rendering, by obtaining the complex pattern, it can be ensured that the complex pattern is clearly displayed in the generated model image.

[0103] Since functional labels affect material rendering parameters, by identifying functional labels, the rendering parameters of the generated model image can be adjusted. For example: the waterproof function needs to apply a matte effect to the model image.

[0104] In this embodiment, the method for identifying the logo position is: S501, using Tesseract model to perform OCR identification on the clothing commodity image to extract text information in the clothing commodity image; S501, extracting the coordinate information of the text from the recognition result of the Tesseract model, so as to locate the logo position.

[0105] In this embodiment, the SIFT model can be used to obtain the pattern on the clothing commodity image, including: 1) Install OpenCV library.

[0106] 2) Read the clothing commodity image, use the imread function of OpenCV to read the clothing commodity image.

[0107] 3) Create a SIFT object and detect feature points: use the SIFT_create function of OpenCV to create a SIFT object, and then use the detectAndCompute method to detect key points and calculate descriptors: 4) Draw key points: you can use the drawKeypoints function of OpenCV to draw key points on the original image 5) Use descriptors for pattern matching: use the SIFT algorithm to extract key point descriptors of two images, and then use feature matching algorithms (such as FLANN or BFMatcher) for matching.

[0108] 6) After obtaining the pattern on the clothing commodity image, use the entropy formula to quantify the complexity of the pattern.

[0109] By quantifying the complexity of the pattern by the entropy formula, it can be ensured that the pattern details are not lost when generating the model.

[0110] The entropy formula is:

[0111] In the formula, H represents the entropy value, P (i) denotes the probability of the random variable i, and log denotes a logarithm function.

[0112] Embodiments of a garment model image generation system are provided. The present application also provides a garment model image generation system, as shown in the accompanying drawings, comprising a processor and a memory, the memory storing computer program instructions which, when executed by the processor, implement the garment model image generation method in the above garment model image generation method embodiments. Figure 2

[0113] The system also includes a communication bus and a communication interface and other components well known to those skilled in the art, the settings and functions of which are known in the art, and thus will not be described here.

[0114] In the description of the present application, the meaning of "a plurality of" is at least two, for example two, three or more, etc., unless otherwise explicitly specified.

[0115] Although the present application has shown and described the preferred embodiments of the present application, it will be apparent to those skilled in the art that many modifications, changes and substitutions can be made thereto without departing from the spirit and scope of the present application. It is to be understood that various alternatives to the embodiments of the present application described herein can be employed in practicing the present application.​

Claims

1. A method of generating a garment mannequin image, characterized by, The method comprises the following steps: obtaining a clothing commodity image and extracting different types of features of the corresponding clothing from the clothing commodity image, the types of the extracted features including color features, material features, size features, and style features; performing weighted fusion on the different types of features extracted to obtain fused features; inputting the initial model set and the fused features into a generative adversarial network model to generate a set of candidate models, and then calculating the loss function value corresponding to each candidate model, and taking the candidate model with a loss function value less than a loss threshold as a candidate model; adjusting the body shape parameters of each candidate model to match the clothing looseness corresponding to the clothing commodity image using a 3D human pose model, and adjusting the pose parameters of each candidate model to simulate the dynamic effect of the clothing; calculating the similarity of each type of feature between each candidate model and the clothing commodity image, and performing weighted summation on the similarity of each type of feature to obtain the comprehensive similarity corresponding to each candidate model; and taking the model with a comprehensive similarity greater than a similarity threshold as the optimal model.

2. The garment model image generation method according to claim 1, wherein The calculation expression of the comprehensive similarity is: ； In the formula, Similarity represents the comprehensive similarity, N represents the total number of feature types, represents the weight of the kth type of feature, represents the feature vector of the kth type of feature of the garment commodity image, represents the feature vector of the kth type of feature of the candidate model, represents the norm of the vector.

3. The garment model image generation method according to claim 1, wherein The multiple types of features between the candidate model and the clothing commodity image include color features and size features.

4. The garment model image generation method according to claim 1, wherein The calculation expression of the similarity threshold is: if the clothing type corresponding to the clothing commodity image is loose, the value of the similarity threshold is 0.8, otherwise, the value of the similarity threshold is 0.

85.

5. The garment model image generation method according to claim 1, wherein The extraction of the color features of the clothing includes: extracting the color distribution, main color, auxiliary color, cold color proportion, color semantic features of the clothing commodity image, and the color compatibility between the clothing commodity image and the model background, the color semantic features including "cold color system", "warm color system", and "clashing color design".

6. The garment model image generation method according to claim 1, wherein The method for extracting the material features of the clothing includes: obtaining the texture features of the clothing commodity image using a gray level co-occurrence matrix, and obtaining the material type of the clothing commodity image using a VGG-16 model; and taking the texture features and the material type as the texture features of the clothing commodity image.

7. The garment model image generation method according to claim 1, wherein The extraction of the size features of the clothing includes: inputting the clothing commodity image into a YOLOV5 model to obtain a series of predicted bounding boxes; for each predicted bounding box, calculating its intersection over union with the real clothing boundary; selecting the bounding box with the largest intersection over union from all the predicted bounding boxes as a high-confidence bounding box, obtaining the pixel-level size of the high-confidence bounding box and the pixel-level size of the clothing commodity image, and then standardizing the size of the high-confidence bounding box to obtain a standard clothing size, the calculation expression being: wherein, represents a standard garment size, represents a pixel level size of a high confidence bounding box, represents a pixel level size of a garment product image, represents a reference size.

8. The garment model image generation method according to claim 1, characterized by, The extraction of the style features of the clothing includes: inputting the clothing commodity image into a preset Mask R-CNN model to obtain the attributes of each detected clothing component, the clothing component including one or more of a collar, a sleeve, a button, and a zipper, wherein the attributes of the collar include collar shape; the attributes of the sleeve include sleeve length type; and the attributes of the button and the zipper include closed or not closed.

9. The garment model image generation method according to any one of claims 1 to 8, characterized by, The types of the extracted features also include: complete matching information and clothing style labels corresponding to the clothing commodity image.

10. A garment model image generation system comprising a memory and a processor, the memory storing computer program instructions, characterized in that, The program instructions, when executed by the processor, implement the clothing model image generation method of any one of claims 1-9.

Images