A clothing design and customization method and system based on artificial intelligence

Abstract

The invention discloses a clothing design and customization method based on artificial intelligence, which relates to the field of clothing design and customization, and comprises: performing big data retrieval on clothing design drawings and finished product drawings based on input keywords, establishing a clothing design sample library, and marking sample library image data; selecting preprocessed sample library image data samples for feature extraction, and establishing a combined feature vector of the sample data; selecting a generative adversarial network, using training data to train the model, and deploying the model to a system application; identifying the generated clothing customization design drawings through an open source clothing recognition model, and screening qualified design drawings according to the keyword hit rate of the recognition result; selecting the qualified design drawings, and performing specific customized design according to the input size, thereby realizing automatic recognition of clothing keyword input based on artificial intelligence to generate clothing designs.

Description

Technical Field

[0001] The present invention relates to the field of clothing design and customization, and in particular to a clothing design and customization method and system based on artificial intelligence. Background Art

[0002] According to the wearer's individual situation, the clothes are made in one piece, tailored to the wearer's body. Most clothing stores operate in this way. In foreign countries, the business model of custom-made clothes is also called personalized clothing design. Usually, the fabrics, patterns and styles of the clothes are selected according to the wearer's individual body shape, skin color, occupation, temperament, hobbies, etc.

[0003] There is a large gap in the current market for customized clothing, which requires designers to have knowledge of clothing design, art, and mathematics, so there is a high learning threshold. In addition, manual customized clothing design generally has the characteristics of long production cycle, high cost, and high correlation with the designer's personal aesthetic, making it difficult to produce in large quantities.

[0004] In order to solve the above problems, this solution provides a method for automatically designing and generating customized clothing drawings based on artificial intelligence by inputting clothing design keywords, thereby solving the problems of high cost and low output of manual design. Summary of the invention

[0005] In order to solve the above technical problems, a clothing design and customization method based on artificial intelligence is provided. This technical solution solves the problems raised in the above background technology.

[0006] In order to achieve the above purpose, the technical solution adopted by the present invention is:

[0007] A clothing design and customization method based on artificial intelligence, comprising:

[0008] Based on the input keywords, the clothing design drawings and finished product drawings are retrieved from big data, a clothing design sample library is established, and the sample library image data is marked;

[0009] Select the preprocessed sample library image data samples for feature extraction and establish a combined feature vector of the sample data;

[0010] Select a generative adversarial network, train the model using training data, and deploy the model to the system application;

[0011] The generated custom clothing design drawings are identified through the open source clothing recognition model, and qualified design drawings are screened according to the keyword hit rate of the recognition results;

[0012] Select a qualified design drawing and make a specific custom design based on the input size.

[0013] Preferably, the step of performing big data retrieval on clothing design drawings and finished product drawings based on input keywords, establishing a clothing design sample library, and marking the sample library image data specifically includes:

[0014] Get input to retrieve clothing keywords, retrieve clothing design drawings and finished product drawings related to the keywords through big data, and establish the association between clothing design drawings and keywords;

[0015] Select at least one clothing design drawing as a clothing design sample and establish a clothing design sample library;

[0016] Invalid and defective samples in the clothing design sample library are screened out, and all samples in the clothing design sample library are batch labeled based on the association between clothing design drawings and keywords.

[0017] Preferably, the selecting of the preprocessed sample library image data samples for feature extraction and establishing a combined feature vector of the sample data specifically includes:

[0018] Perform image normalization and size standardization on the sample library image data;

[0019] A pre-trained model based on a convolutional neural network is selected to extract features from the pre-processed sample library image data, and the extracted low-dimensional features of the image are represented as feature vectors D;

[0020] The sample library image data is analyzed, the style supplement feature vector S of each sample is calculated, the feature vector D is directly concatenated with the style supplement feature vector S, and a combined feature vector C is established, where C = [D, S].

[0021] Preferably, the calculating of the style supplement feature vector of each sample specifically includes:

[0022] The components in the style supplementary feature vector include the texture dispersion evaluation value and the overall coordination evaluation value, which are obtained as follows:

[0023] Convert all sample images into grayscale images, apply two-dimensional Fourier transform to the grayscale images, calculate the intensity of each frequency component in the frequency domain, and construct the amplitude spectrum;

[0024] Logarithmically transform the amplitude spectrum and create a frequency histogram;

[0025] The histogram is normalized to obtain the frequency distribution P(i) corresponding to each frequency. The entropy value of the histogram is calculated based on the probability distribution. The entropy value calculation expression of the histogram is:

[0026]

[0027] Where N is the total number of frequency partitions, and H is the calculated entropy value;

[0028] Calculate the variance of the frequency histogram probability distribution and label it as σ 2 , perform weighted summation on the entropy value of the histogram and the variance of the probability distribution, and use the result of the weighted summation as the texture dispersion evaluation value;

[0029] Take the lower left vertex of the image as the origin, establish a plane rectangular coordinate system based on the width and height of the image, and calculate the coordinate position of the golden section point of the image Where (P x , P y ) is the coordinate of the golden section point, W, H are the width and height of the image respectively, and φ is the golden section ratio;

[0030] Calculate the average Euclidean distance of all pixels in the image relative to the golden section point as the overall coordination evaluation value of the image;

[0031] The texture dispersion evaluation value and the overall coordination evaluation value are combined to establish a style supplementary feature vector.

[0032] Preferably, the selecting of a generative adversarial network, training the model using training data, and deploying the model to a system application specifically includes:

[0033] Obtain a sample library data set corresponding to the combined feature vector, and divide the sample library data set into a training set, a validation set, and a test set according to a set allocation ratio;

[0034] Select a generative adversarial network model, train the model based on the training set data, calculate the adversarial loss between the generated image and the real image, update the generator and discriminator parameters, draw and record the loss curves of the generator and discriminator, and regularly generate image samples to monitor the training progress;

[0035] The model is evaluated based on the validation set data. The difference between the feature maps of the convolutional layer of the generated image and the real image is calculated through the pre-trained network model to evaluate the visual similarity. The optimal hyperparameter configuration combination is selected through cross-validation.

[0036] The trained model is tested based on the test set data. The generalization ability of the model is evaluated by generating feature similarity between new images and the test set data. The standard value of the model generalization ability is set. If the generalization ability of the model does not meet the standard, the training data is adjusted to retrain and verify the model. If the generalization ability of the model meets the standard, the model is deployed to the system application.

[0037] Preferably, identifying the generated custom clothing design drawings by using an open source clothing recognition model, and screening qualified design drawings according to the keyword hit rate of the recognition result specifically includes:

[0038] Select a pre-trained open source clothing recognition model, deploy it to recognize the generated clothing custom design pictures, obtain the original scores of all recognition result keywords, and convert the original scores of keywords into probability distribution through the Softmax function;

[0039] According to the hit rate of the identified keywords compared with the input keywords, the matching of the generated design drawings is evaluated;

[0040] The probability of recognizing keywords is accumulated, and the product of the accumulated result and the hit rate of the keywords in the recognition result is used as the matching evaluation value of the generated design drawing. The matching evaluation value calculation expression is: Where M is the matching evaluation value of the generated design diagram, m and L are the number of keyword hits and the total number of input keywords, respectively, and pr j is the probability distribution of the raw score of the jth identified keyword;

[0041] A matching evaluation value threshold of the generated design drawing is set. When the matching evaluation value is greater than or equal to the matching evaluation value threshold, the generated design drawing is marked as a qualified clothing customization design drawing. Conversely, when the matching evaluation value is less than the matching evaluation value threshold, the generated design drawing is marked as an unqualified clothing customization design drawing.

[0042] Preferably, the selecting of a qualified design drawing and performing a specific customized design according to the input size specifically includes:

[0043] Obtain design drawings that have passed the matching evaluation, select a set number of design drawings with the largest matching evaluation values, and perform specific custom designs on the clothing based on the design drawings and the input required dimensions.

[0044] Furthermore, a clothing design and customization system based on artificial intelligence is proposed, which is used to implement the above-mentioned clothing design and customization method based on artificial intelligence, including:

[0045] A sample retrieval module, which retrieves clothing category keywords according to input, retrieves clothing design drawings and finished product drawings related to the keywords through big data, establishes the association between clothing design drawings and keywords and establishes a clothing design sample library, and batch marks all samples in the clothing design sample library based on the association between clothing design drawings and keywords;

[0046] A feature extraction module, wherein the feature extraction module performs image normalization and size standardization on the sample library data, selects a pre-trained model based on a convolutional neural network, extracts features from the pre-processed sample library data, analyzes the sample library data, calculates a style supplement feature vector for each sample, directly concatenates the feature vector with the style supplement feature vector, and establishes a combined feature vector, wherein the style supplement feature vector includes a texture dispersion evaluation value and an overall coordination evaluation value;

[0047] A model training and deployment module, which divides the sample library data set according to a set allocation ratio, selects a generative adversarial network model, trains the model based on the training set data, selects the optimal hyperparameter configuration combination through cross-validation, and deploys the model to the system application when the model generalization ability meets the standard;

[0048] A design drawing screening module is generated, wherein the generated design drawing screening module identifies the generated custom clothing design drawing, obtains the probability distribution of all recognition result keywords, calculates the matching evaluation of the generated design drawing according to the hit rate of the recognized keywords relative to the input keywords, and screens the design drawings that meet the matching evaluation.

[0049] Compared with the prior art, the present invention has the following beneficial effects:

[0050] Based on the input keywords, big data retrieval of clothing design drawings and finished product drawings is performed, a clothing design sample library is established, and the sample library image data is labeled; the pre-processed sample library image data samples are selected for feature extraction, and a combined feature vector of the sample data is established; a generative adversarial network is selected, the model is trained using training data, and the model is deployed to the system application; the generated clothing custom design drawings are identified through an open source clothing recognition model, and qualified design drawings are screened according to the keyword hit rate of the recognition results; qualified design drawings are selected, and specific customized designs are performed according to the input size.

[0051] According to the above process steps, the automatic generation and design of customized clothing drawings based on existing artificial intelligence technology is realized, which solves the problems of high threshold, high cost and low output of manual design. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 A flowchart of a clothing design and customization method based on artificial intelligence of the present invention;

[0053] Figure 2 A flowchart of the present invention for performing big data retrieval of clothing design drawings and finished product drawings based on input keywords, establishing a clothing design sample library, and marking the sample library image data;

[0054] Figure 3The present invention selects the sample library image data samples after preprocessing to perform feature extraction and establishes a combined feature vector flow chart of the sample data;

[0055] Figure 4 A flow chart of calculating the style supplementary feature vector of each sample of the present invention;

[0056] Figure 5 A flow chart of selecting a generative adversarial network for the present invention, training the model using training data, and deploying the model into a system application;

[0057] Figure 6 The invention is to identify the generated clothing customization design drawings through the open source clothing recognition model, and screen the qualified design drawings according to the keyword hit rate of the recognition result;

[0058] Figure 7 For the present invention, a qualified design drawing is selected and a specific customized design flow chart is made according to the input size;

[0059] Figure 8 The figure is a schematic diagram of the structure of a clothing design and customization system based on artificial intelligence of the present invention. DETAILED DESCRIPTION

[0060] The following description is used to disclose the present invention so that those skilled in the art can implement the present invention. The preferred embodiments described below are only examples, and those skilled in the art may think of other obvious variations.

[0061] Reference Figure 1 As shown, a clothing design and customization method based on artificial intelligence includes:

[0062] Based on the input keywords, the clothing design drawings and finished product drawings are retrieved from big data, a clothing design sample library is established, and the sample library image data is marked;

[0063] Select the preprocessed sample library image data samples for feature extraction and establish a combined feature vector of the sample data;

[0064] Select a generative adversarial network, train the model using training data, and deploy the model to the system application;

[0065] The generated custom clothing design drawings are identified through the open source clothing recognition model, and qualified design drawings are screened according to the keyword hit rate of the recognition results;

[0066] Select a qualified design drawing and make a specific custom design based on the input size.

[0067] Reference Figure 2As shown in the figure, based on the input keywords, the clothing design drawings and finished product drawings are retrieved from big data, a clothing design sample library is established, and the image data of the sample library is marked, which specifically includes:

[0068] Get input to retrieve clothing keywords, retrieve clothing design drawings and finished product drawings related to the keywords through big data, and establish the association between clothing design drawings and keywords;

[0069] Select at least one clothing design drawing as a clothing design sample and establish a clothing design sample library;

[0070] Invalid and defective samples in the clothing design sample library are screened out, and all samples in the clothing design sample library are batch labeled based on the association between clothing design drawings and keywords.

[0071] The relationship between clothing design drawings and keywords is a many-to-many relationship. During batch labeling, each search keyword needs to be labeled in a non-overlapping manner to ensure the accuracy of the clothing design sample library data in model training.

[0072] Reference Figure 3 As shown, the preprocessed sample library image data samples are selected for feature extraction, and the combined feature vector of the sample data is established, which specifically includes:

[0073] The sample library image data is normalized and resized. Image normalization and resize standardization belong to the data preprocessing process. Image normalization is to convert the image pixel value into an interval range of 0 to 1 or -1 to 1. Resize standardization is to adjust the image size to a fixed size, the default is 256*256 pixels.

[0074] Select a pre-trained model based on a convolutional neural network, extract features from the pre-processed sample library image data, and represent the extracted low-dimensional features of the image as a feature vector D;

[0075] Among them, the pre-trained models of convolutional neural networks can be selected from models that have been trained in large-scale general data sets, such as ResNet, VGG, Inception, etc. These models can be directly used for image feature extraction without additional training. In this solution, they are mainly used to extract the feature vectors of the convolutional layer for subsequent use.

[0076] The sample library image data is analyzed, and the style supplement feature vector S of each sample is calculated. The feature vector D is directly concatenated with the style supplement feature vector S to establish a combined feature vector C, C = [D, S]. By adding new features to the low-dimensional feature vector, the performance of the model is improved.

[0077] Reference Figure 4 As shown, calculating the style supplementary feature vector of each sample specifically includes:

[0078] The components in the style supplementary feature vector include the texture dispersion evaluation value and the overall coordination evaluation value, which are obtained as follows:

[0079] Convert all sample images into grayscale images, apply two-dimensional Fourier transform to the grayscale images, calculate the intensity of each frequency component in the frequency domain, and construct the amplitude spectrum;

[0080] Logarithmically transform the amplitude spectrum and create a frequency histogram;

[0081] The histogram is normalized to obtain the frequency distribution P(i) corresponding to each frequency. The entropy value of the histogram is calculated based on the probability distribution. The entropy value calculation expression of the histogram is:

[0082]

[0083] Where N is the total number of frequency partitions, and H is the calculated entropy value;

[0084] Through Fourier transform, the spatial domain information of the image can be converted into frequency domain information, the distribution of different frequency components in the image can be analyzed, and the frequency histogram can be calculated to evaluate the statistical characteristics of its variance and entropy. The dispersion of the image texture can be expressed, thereby improving the detail complexity of the image and the performance of texture features in the model.

[0085] The frequency distribution P(i) is calculated as: Where H(i) is the frequency of the lowest i frequency ranges, and N is the total frequency of the frequency range.

[0086] Calculate the variance of the frequency histogram probability distribution and label it as σ 2 , perform weighted summation on the entropy value of the histogram and the variance of the probability distribution, and use the result of the weighted summation as the texture dispersion evaluation value;

[0087] Checking the harmony of image composition through the golden section rule is a common method used in art and design to create visual harmony. Placing the main elements on the golden section point can better satisfy the visual balance.

[0088] Take the lower left vertex of the image as the origin, establish a plane rectangular coordinate system based on the width and height of the image, and calculate the coordinate position of the golden section point of the image Where (P x , P y ) is the coordinate of the golden section point, W, H are the width and height of the image respectively, and φ is the golden section ratio;

[0089] Calculate the average Euclidean distance of all pixels in the image relative to the golden section point as the overall coordination evaluation value of the image;

[0090] Among them, the calculation method of the overall coordination evaluation value of the image is: Where COO is the overall coordination evaluation value, k and Z are the kth pixel and the total number of pixels, respectively.

[0091] The texture dispersion evaluation value and the overall coordination evaluation value are combined to establish a style supplementary feature vector.

[0092] Reference Figure 5 As shown, a generative adversarial network is selected, and the model is trained using training data. The deployment of the model into the system application specifically includes:

[0093] Obtain a sample library data set corresponding to the combined feature vector, and divide the sample library data set into a training set, a validation set, and a test set according to a set allocation ratio;

[0094] Select a generative adversarial network model, train the model based on the training set data, calculate the adversarial loss between the generated image and the real image, update the generator and discriminator parameters, draw and record the loss curves of the generator and discriminator, and regularly generate image samples to monitor the training progress;

[0095] Among them, StyleGAN / StyleGAN2 is selected by default for the generative adversarial network because of its unique capabilities of personalized generation and high-resolution image generation. It is more suitable for scenarios that require fine control of generated content and is widely used in image generation and artistic creation.

[0096] The model is evaluated based on the validation set data. The difference between the feature maps of the convolutional layer of the generated image and the real image is calculated through the pre-trained network model to evaluate the visual similarity. The optimal hyperparameter configuration combination is selected through cross-validation.

[0097] Analyze the training loss and validation loss recorded during the training process. If the training loss is lower than the validation loss, make supplementary adjustments to the training set data in time to prevent overfitting.

[0098] The trained model is tested based on the test set data. The generalization ability of the model is evaluated by generating feature similarity between new images and the test set data. The standard value of the model generalization ability is set. If the generalization ability of the model does not meet the standard, the training data is adjusted to retrain and verify the model. If the generalization ability of the model meets the standard, the model is deployed to the system application.

[0099] Among them, whether the generalization ability of the model meets the standard is comprehensively evaluated by generating the feature similarity between the new image and the test set data and the accuracy, precision and recall rate of the training set, which will not be elaborated here.

[0100] Reference Figure 6As shown in the figure, the generated clothing customization design drawings are identified through the open source clothing recognition model. According to the keyword hit rate of the recognition result, the qualified design drawings are screened, including:

[0101] Select a pre-trained open source clothing recognition model, deploy it, and recognize the generated clothing customization design pictures. Obtain the original scores of all recognition result keywords, and convert the original scores of keywords into probability distribution through the Softmax function. The specific expression is: Where exp(i) is the original score of the ith keyword, U is the number of keywords identified, and pr j is the probability distribution of the original score of the jth identified keyword;.

[0102] According to the hit rate of the identified keywords compared with the input keywords, the matching of the generated design drawings is evaluated;

[0103] The probability of recognizing keywords is accumulated, and the product of the accumulated result and the hit rate of the keywords in the recognition result is used as the matching evaluation value of the generated design drawing. The matching evaluation value calculation expression is: Where M is the matching evaluation value of the generated design diagram, and m and L are the number of keyword hits and the total number of input keywords, respectively.

[0104] A matching evaluation value threshold of the generated design drawing is set. When the matching evaluation value is greater than or equal to the matching evaluation value threshold, the generated design drawing is marked as a qualified clothing customization design drawing. Conversely, when the matching evaluation value is less than the matching evaluation value threshold, the generated design drawing is marked as an unqualified clothing customization design drawing.

[0105] The matching evaluation value threshold formula is set by professionals in this field after comprehensively considering the model generation capability, recognition accuracy, equipment computing performance bottleneck and specific clothing design requirements, and will not be repeated here.

[0106] Reference Figure 7 As shown, select a qualified design drawing and make a specific customized design based on the input size, including:

[0107] Obtain design drawings that have passed the matching evaluation, select a set number of design drawings with the largest matching evaluation values, and perform specific custom designs on the clothing based on the design drawings and the input required dimensions.

[0108] For further information, see Figure 8 As shown, based on the same inventive concept as the above-mentioned clothing design and customization method based on artificial intelligence, this solution also proposes a clothing design and customization system based on artificial intelligence, including:

[0109] A sample retrieval module, which retrieves clothing category keywords according to input, retrieves clothing design drawings and finished product drawings related to the keywords through big data, establishes the association between clothing design drawings and keywords and establishes a clothing design sample library, and batch marks all samples in the clothing design sample library based on the association between clothing design drawings and keywords;

[0110] A feature extraction module, wherein the feature extraction module performs image normalization and size standardization on the sample library data, selects a pre-trained model based on a convolutional neural network, extracts features from the pre-processed sample library data, analyzes the sample library data, calculates a style supplement feature vector for each sample, directly concatenates the feature vector with the style supplement feature vector, and establishes a combined feature vector, wherein the style supplement feature vector includes a texture dispersion evaluation value and an overall coordination evaluation value;

[0111] A model training and deployment module, which divides the sample library data set according to a set allocation ratio, selects a generative adversarial network model, trains the model based on the training set data, selects the optimal hyperparameter configuration combination through cross-validation, and deploys the model to the system application when the model generalization ability meets the standard;

[0112] A design drawing screening module is generated, wherein the generated design drawing screening module identifies the generated custom clothing design drawing, obtains the probability distribution of all recognition result keywords, calculates the matching evaluation of the generated design drawing according to the hit rate of the recognized keywords relative to the input keywords, and screens the design drawings that meet the matching evaluation.

[0113] Furthermore, the present solution also proposes a storage medium for a clothing design and customization method based on artificial intelligence, on which a computer-readable program is stored. When the computer-readable program is called, the above-mentioned clothing design and customization method based on artificial intelligence is executed.

[0114] It is understandable that the storage medium may be a magnetic medium, such as a floppy disk, a hard disk, or a magnetic tape; an optical medium, such as a DVD; or a semiconductor medium, such as a solid state drive (SSD).

[0115] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions only describe the principles of the present invention. The present invention may be subject to various changes and improvements without departing from the spirit and scope of the present invention. These changes and improvements fall within the scope of the present invention. The scope of protection claimed by the present invention is defined by the attached claims and their equivalents.

Claims

1. A clothing design and customization method based on artificial intelligence, characterized in that: include: Based on the input keywords, the clothing design drawings and finished product drawings are retrieved from big data, a clothing design sample library is established, and the sample library image data is marked; Select the preprocessed sample library image data samples for feature extraction and establish a combined feature vector of the sample data; Select a generative adversarial network, train the model using training data, and deploy the model to the system application; The generated custom clothing design drawings are identified through the open source clothing recognition model, and qualified design drawings are screened according to the keyword hit rate of the recognition results; Select qualified design drawings and make specific customized designs according to the input size; The selecting of the preprocessed sample library image data samples for feature extraction and establishing the combined feature vector of the sample data specifically includes: Perform image normalization and size standardization on the sample library image data; Select a pre-trained model based on a convolutional neural network, extract features from the pre-processed sample library image data, and represent the extracted low-dimensional features of the image as a feature vector D; Analyze the sample library image data, calculate the style supplement feature vector S of each sample, directly concatenate the feature vector D with the style supplement feature vector S, and establish a combined feature vector C, C = [D, S]; The calculating of the style supplement feature vector of each sample specifically includes: The components in the style supplementary feature vector include the texture dispersion evaluation value and the overall coordination evaluation value, which are obtained as follows: Convert all sample images into grayscale images, apply two-dimensional Fourier transform to the grayscale images, calculate the intensity of each frequency component in the frequency domain, and construct the amplitude spectrum; Logarithmically transform the amplitude spectrum and create a frequency histogram; The histogram is normalized to obtain the frequency distribution P(i) corresponding to each frequency. The entropy value of the histogram is calculated based on the probability distribution. The entropy value calculation expression of the histogram is: Where N is the total number of frequency partitions, and H is the calculated entropy value; Calculate the variance of the frequency histogram probability distribution and label it as σ 2 , perform weighted summation on the entropy value of the histogram and the variance of the probability distribution, and use the result of the weighted summation as the texture dispersion evaluation value; Take the lower left vertex of the image as the origin, establish a plane rectangular coordinate system based on the width and height of the image, and calculate the coordinate position of the golden section point of the image Where (P x , P y ) is the coordinate of the golden section point, W, H are the width and height of the image respectively, and φ is the golden section ratio; Calculate the average Euclidean distance of all pixels in the image relative to the golden section point as the overall coordination evaluation value of the image; The texture dispersion evaluation value and the overall coordination evaluation value are combined to establish a style supplementary feature vector.

2. The clothing design and customization method based on artificial intelligence according to claim 1 is characterized in that: The method of performing big data retrieval on clothing design drawings and finished product drawings based on input keywords, establishing a clothing design sample library, and marking the sample library image data specifically includes: Get input to retrieve clothing keywords, retrieve clothing design drawings and finished product drawings related to the keywords through big data, and establish the association between clothing design drawings and keywords; Select at least one clothing design drawing as a clothing design sample and establish a clothing design sample library; Invalid and defective samples in the clothing design sample library are screened out, and all samples in the clothing design sample library are batch labeled based on the association between clothing design drawings and keywords.

3. The clothing design and customization method based on artificial intelligence according to claim 2 is characterized in that: The selecting of a generative adversarial network, training the model using training data, and deploying the model to a system application specifically includes: Obtain a sample library data set corresponding to the combined feature vector, and divide the sample library data set into a training set, a validation set, and a test set according to a set allocation ratio; Select a generative adversarial network model, train the model based on the training set data, calculate the adversarial loss between the generated image and the real image, update the generator and discriminator parameters, draw and record the loss curves of the generator and discriminator, and regularly generate image samples to monitor the training progress; The model is evaluated based on the validation set data. The difference between the feature maps of the convolutional layer of the generated image and the real image is calculated through the pre-trained network model to evaluate the visual similarity. The optimal hyperparameter configuration combination is selected through cross-validation. The trained model is tested based on the test set data. The generalization ability of the model is evaluated by generating feature similarity between new images and the test set data. The standard value of the model generalization ability is set. If the generalization ability of the model does not meet the standard, the training data is adjusted to retrain and verify the model. If the generalization ability of the model meets the standard, the model is deployed to the system application.

4. The clothing design and customization method based on artificial intelligence according to claim 3 is characterized in that: The method of identifying the generated custom clothing design drawings by using the open source clothing recognition model and screening qualified design drawings according to the keyword hit rate of the recognition results specifically includes: Select a pre-trained open source clothing recognition model, deploy it to recognize the generated clothing custom design pictures, obtain the original scores of all recognition result keywords, and convert the original scores of keywords into probability distribution through the Softmax function; According to the hit rate of the identified keywords compared with the input keywords, the matching of the generated design drawings is evaluated; The probability of recognizing keywords is accumulated, and the product of the accumulated result and the hit rate of the keywords in the recognition result is used as the matching evaluation value of the generated design drawing. The matching evaluation value calculation expression is: Where M is the matching evaluation value of the generated design diagram, m and L are the number of keyword hits and the total number of input keywords, respectively, and pr j is the probability distribution of the raw score of the jth identified keyword; A matching evaluation value threshold of the generated design drawing is set. When the matching evaluation value is greater than or equal to the matching evaluation value threshold, the generated design drawing is marked as a qualified clothing customization design drawing. Conversely, when the matching evaluation value is less than the matching evaluation value threshold, the generated design drawing is marked as an unqualified clothing customization design drawing.

5. The clothing design and customization method based on artificial intelligence according to claim 4 is characterized in that: The selecting of qualified design drawings and carrying out customized design according to the input size specifically includes: Obtain design drawings that have passed the matching evaluation, select a set number of design drawings with the largest matching evaluation values, and perform specific custom designs on the clothing based on the design drawings and the input required dimensions.

6. A clothing design and customization system based on artificial intelligence, used to implement a clothing design and customization method based on artificial intelligence as described in any one of claims 1 to 5, characterized in that: include: A sample retrieval module, which retrieves clothing category keywords according to input, retrieves clothing design drawings and finished product drawings related to the keywords through big data, establishes the association between clothing design drawings and keywords and establishes a clothing design sample library, and batch marks all samples in the clothing design sample library based on the association between clothing design drawings and keywords; A feature extraction module, wherein the feature extraction module performs image normalization and size standardization on the sample library data, selects a pre-trained model based on a convolutional neural network, extracts features from the pre-processed sample library data, analyzes the sample library data, calculates a style supplement feature vector for each sample, directly concatenates the feature vector with the style supplement feature vector, and establishes a combined feature vector, wherein the style supplement feature vector includes a texture dispersion evaluation value and an overall coordination evaluation value; A model training and deployment module, which divides the sample library data set according to a set allocation ratio, selects a generative adversarial network model, trains the model based on the training set data, selects the optimal hyperparameter configuration combination through cross-validation, and deploys the model to the system application when the model generalization ability meets the standard; A design drawing screening module is generated, wherein the generated design drawing screening module identifies the generated custom clothing design drawing, obtains the probability distribution of all recognition result keywords, calculates the matching evaluation of the generated design drawing according to the hit rate of the recognized keywords relative to the input keywords, and screens the design drawings that meet the matching evaluation.

Images