Garment customization system and method based on Internet of Things

Abstract

The invention relates to a garment customization system and method based on the Internet of Things, and belongs to the field of garment customization. The garment customization system based on the Internet of Things comprises a terminal, a size data generation module and a garment customization module. The terminal is used for collecting images and size data uploaded by a user; the size data generation module is used for carrying out feature extraction on the uploaded image and then reconstructing a human body image to generate size data of each part; the clothing customization module is used for matching the image output by the terminal to generate a virtual fitting picture, and the virtual fitting picture is used as a customization effect picture to be sent to the terminal after being subjected to aesthetic evaluation, secondary creation and aesthetic reevaluation; and the terminal combines the customization effect picture with the size data of each part of the human body to generate a garment customization order. According to the invention, based on the virtual fitting technology and the multi-dimensional aesthetic evaluation technology, the highly personalized customization service is provided for the user, and the accuracy of the online customization effect picture and the satisfaction degree of the user are effectively improved.

Description

Technical Field

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

[0002] With the rapid development of the e-commerce industry, people's shopping methods have quietly changed, and more and more consumers prefer to order various daily necessities online. However, while e-commerce is developing rapidly, consumer services have not kept up in time, which has led to serious returns and exchanges of goods. According to public data, among the returned goods, the return rate of clothing is as high as 7.89%. According to the survey, the main reasons for consumers to return and exchange clothes after purchasing clothes online are: inappropriate size, dissatisfaction with the style and quality problems. After receiving the clothes, many consumers found that the purchased clothes were not very fit, or the actual dressing effect was too different from what they expected, so they returned and exchanged them. The main reason for this phenomenon is that consumers estimate their approximate size without specific real data. In addition, the dressing effect of clothes varies from person to person. The body shape and temperament of the models who try on online are quite different from those of consumers. Therefore, the actual try-on effect of consumers is very different from the promotional pictures on the product details page. The new business model that combines clothing customization with online e-commerce simulates the actual effect of customized clothing on users through digital platforms, allowing users to choose different clothing colors and patterns according to their preferences; users can preview the effect of their designed clothing on the target body online, increasing user participation and satisfaction in the clothing customization process. Clothing companies can also reduce labor costs and customer return rates by using virtual fitting technology, while reducing sales costs by allowing users to determine shopping needs in advance. Remote body size measurement can provide consumers with more convenient, more accurate, and more personalized clothing customization services, and clothing companies can use this technology to reduce production costs and inventory risks and improve production quality. Virtual fitting and remote body measurement are crucial to increasing users' confidence and satisfaction in purchasing customized clothing and improving brand influence.

[0003] At present, with the booming e-commerce industry, the development prospects of combining clothing customization with Internet e-commerce are very broad. Domestic and foreign companies specializing in online clothing customization have solved the problem of personalized clothing design to a certain extent, but they are still lacking in user fitting experience and size recommendation. Summary of the invention

[0004] Based on this, it is necessary to provide a clothing customization system and method based on the Internet of Things to address the problems of difficulty in obtaining size data, slow customization speed and poor fitting experience in existing online clothing customization technologies.

[0005] The present invention is implemented by the following technical solution: A clothing customization system based on the Internet of Things, characterized in that it includes: The terminal is used to register users, log in to accounts and change personal information; users upload multiple target clothing images, multi-view images of target bodies and necessary size data; The size data generation module includes a feature extraction network 1, a fusion layer 1, a fully connected layer 1, an output layer 1 and an output layer 2; the feature extraction network 1 is used to pre-process the multi-view images and necessary size data of the terminal output user to extract multi-view features; the fusion layer 1 is used to add the features of the multi-view to obtain a fusion feature vector; the fully connected layer 1 is used to map the high-dimensional fusion feature vector to the low-dimensional output space to obtain the morphological parameters of SMPL, and the output layer 1 is used to drive the SMPL standard model transformation through the morphological parameters of SMPL to form target human images of different body shapes and postures; the output layer 2 uses the section method to measure the target human image and output the size data of each part; The clothing customization module includes a design unit, a secondary creation unit, and an evaluation module; the design unit is used to match the target clothing image output by the terminal with the target human body image to generate a number of virtual fitting pictures; the aesthetic evaluation unit is used to perform aesthetic evaluation on the several virtual fitting pictures and send the a virtual fitting pictures with the highest scores to the secondary creation unit; the secondary creation unit is used for the user to perform personalized customization on a virtual fitting picture to generate a customized preview picture; the aesthetic evaluation unit simultaneously performs aesthetic evaluation on the virtual fitting picture and the customized preview picture and sends the a picture with the highest score to the terminal as a customized effect picture; The management module is used to manage orders placed by customers, import or delete algorithm models into the system, and manage basic information of anthropometric models.

[0006] In one of the embodiments, the design unit includes a collection module, a geometric matching module and a virtual fitting module; the collection module performs feature extraction on the input clothing image and the target human body image, the geometric matching module is used to distort and deform the clothing according to the shape and posture characteristics of the target human body; the virtual fitting module is used to fuse the distorted clothing with the target human body characteristics to obtain a virtual fitting image.

[0007] In one embodiment, the aesthetic evaluation unit includes feature extraction network 3, fusion layer 2, fully connected layer 2, and output layer 3; feature extraction network 3 selects ResNet-50 feature extraction network to extract aesthetic features of virtual fitting pictures, and ResNet-50 feature extraction network includes 5 stages, which are used to extract aesthetic features of different dimensions; fusion layer 2 reduces the dimension of high-dimensional features through global average pooling, and then fuses multiple features after dimensionality reduction to generate fusion features; fully connected layer 2 is used to map fusion features to different output layers 3; output layer 3 is used to output aesthetic evaluation results through quality regression model. This module can quickly and effectively evaluate the aesthetic quality of virtual fitting pictures, and the aesthetic evaluation method is subjective, ensuring that it is easily recognized by the public.

[0008] In one embodiment, the feature extraction network 1 uses U-Net as the feature extraction network for semantic segmentation, and uses ResNet as the encoder to extract the deep features of the human body contour map. This method can effectively learn high-level semantic information, avoid the problem of gradient disappearance through residual connection, and enhance the target human image reconstruction performance. In one embodiment, the section method used in the output layer 2 is as follows: a section is intersected with the human body model, and an optimal curve is selected from a plurality of closed curves obtained by the intersection for measurement.

[0009] In one embodiment, the geometric matching module includes an input layer 1, a fusion layer 3, a regression layer and an output layer 4; the fusion layer 3 fuses the input human body representation feature r and the target clothing feature f to generate a fusion feature, the regression layer is used to generate a TPS transformation parameter θ through the regression network, and the output layer 4 performs a TPS transformation on the input θ and the target clothing feature f to obtain a distorted clothing f w ; The virtual fitting module includes input layer 2, U-shaped network layer and output layer 5. The U-shaped network layer is based on f w With r as input, the U-shaped encoder-decoder network is used to generate the rendered image Tr and the predicted fusion mask M. The output layer 5 fuses Ir and M through the mask to obtain the final fitting result. , try-on results The calculation formula is as follows:

[0010] In the formula represents the Hadamard product.

[0011] In one embodiment, the penalty term calculation formula of the loss function of the geometric matching module is as follows:

[0012] In the formula, express , express ; Select VGG19 network as the pre-trained network to distort the clothing image f w The image f parsed from the target image p The perceptual loss between is calculated as follows:

[0013] In the formula represent Input the i-th layer feature map obtained by the pre-trained network, Represents the weight of the layer, and N represents the number of selected feature layers; The final loss function loss calculation formula of the geometric matching module is as follows

[0014] In the formula , is the hyperparameter of the model.

[0015] By constructing a distorted image of clothing f 𝑤 The real clothing image f parsed from the target image p The perceptual loss between them not only ensures the overall clothing alignment, but also takes into account the local distortion of the clothing, thereby achieving the effect of enhancing the appearance feature retention of the target clothing f and making the generated features more delicate.

[0016] The present invention also provides a clothing customization method based on the Internet of Things, which adopts the cyanobacteria bloom meteorological forecasting and early warning system based on the hierarchical fusion algorithm according to any one of claims 1 to 7, characterized in that the clothing customization method comprises the following steps: S1: Collect target clothing images, target human body images and necessary size data; S2: Generate a virtual fitting image based on the target clothing image and the target human body image. The specific steps are as follows: S21: Matching the target clothing images output by the terminal with the target human body images one by one to generate a number of virtual fitting images; S22: performing aesthetic evaluation on the input virtual fitting pictures and sending the a virtual fitting pictures with the highest scores to the user; S23: The user adds personalized customization elements to a virtual fitting picture to generate a*b customized preview pictures; S24: performing aesthetic evaluation on a virtual fitting image and a*b customized preview images simultaneously and sending the image a with the highest score as the customized rendering to the customer; S3: preprocessing the input multi-view images of the target human body and necessary size data, constructing the target human body image according to the preprocessed data; and measuring the size data of each part of the target human body image by using the section method; S4: The user selects a favorite custom rendering and generates a custom clothing order based on the size data of various parts of the human body.

[0017] This method realizes online clothing customization based on virtual fitting and multi-dimensional aesthetic evaluation technology design, enabling users to choose clothing styles, colors, patterns and prints according to their preferences and needs, enhancing the user's personalized customization experience. This method provides online customization services to improve consumer satisfaction, while improving the customization and sales efficiency of clothing and reducing inventory and management costs.

[0018] In one embodiment, in step S21, the steps of matching and generating a virtual fitting picture are as follows: S211: extracting features from the input target clothing image and target human body image; S212: Input the human body representation feature r and the target clothing feature f, fuse the two features, and then obtain the TPS transformation parameter θ through the regression network, and then use θ to perform TPS transformation on f to obtain the distorted clothing f w ; S213: With f w With r as input, the U-shaped encoder-decoder network generates the rendered image Tr and the predicted fusion mask M, and Tr and M are fused together through the mask to obtain the final fitting result. .

[0019] In one embodiment, in step S22, the method of aesthetic evaluation is as follows: S221: Extracting aesthetic features of virtual fitting images; S222: Reduce the dimension of high-dimensional features through global average pooling, and then fuse multiple features after dimensionality reduction to generate fused features; the fusion method is horizontal splicing, and the calculation formula is as follows: F(I; )=Ω[GP( ), GP( ), GP( ), GP( )] In the formula, F(I;φ) represents the feature representation of the input image I; I represents the input image; φ represents all parameters of the entire feature extraction network; GP() represents global average pooling; Ω() represents the concatenation of features of multiple scales along the row direction. ; S223: Map the fused features to different output layers, and add a SoftMax layer after the output to generate the probability of each evaluation level. The process is described by the following formula: ) Where, t = 1, 2, 3, 4; Represents the distribution of aesthetic evaluation levels in different dimensions. There are four output layers in total, namely image details, image color, image color coordination, and overall aesthetic quality; Represents 4 different fully connected layer networks, and their network parameters are ; S224: Outputting aesthetic evaluation results through the quality regression model.

[0020] Compared with the prior art, the present invention has the following beneficial effects: 1. The CP-VTON virtual fitting network is used to extract features from the target clothing image and the target human body image, and the features of the two images are matched to generate a virtual fitting picture; by constructing the perceptual loss between the distorted image of the clothing and the real clothing image parsed from the target clothing image, not only the overall clothing alignment is ensured, but also the local distortion of the clothing is taken into account, so as to achieve the effect of improving the retention of the appearance features of the target clothing and make the generated features more delicate. The present invention provides users with highly personalized customization services, effectively improving the accuracy of online customization renderings and user satisfaction 2. The aesthetic quality of the virtual fitting pictures is evaluated by the aesthetic evaluation unit, and the virtual fitting pictures with the highest scores are sent to the terminal as customized preview pictures; the user can create customized preview pictures through the secondary creation unit. Figure 2 The creation generates a customized renderings, and the aesthetic evaluation unit simultaneously performs aesthetic evaluation on a customized rendering and a virtual fitting picture, and the image with the highest score is used as the customized rendering. The present invention is based on the design of virtual fitting and multi-dimensional aesthetic evaluation technology, which enables users to choose clothing styles, colors, patterns and prints according to their preferences and needs, thereby enhancing the personalized customization experience; the multi-dimensional aesthetic evaluation method is used to quickly and effectively recommend the most suitable customized design to users, and the aesthetic evaluation method is subjective, ensuring that it is easily recognized by the public.

[0021] 3. The invention provides users with a convenient, fast, interesting and innovative online customization service. At the same time, by driving the SMPL standard model to construct a target human body image for human body size measurement, the accuracy of size measurement data and the customization and sales efficiency of ready-made clothes are improved, and inventory and management costs are reduced. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 This is a structural schematic diagram of a clothing customization system based on the Internet of Things according to Embodiment 1 of the present invention; Figure 2 This is a step diagram of a clothing customization method based on the Internet of Things according to Embodiment 2 of the present invention; Figure 3 for Figure 1 Flowchart of the IoT-based clothing customization method. DETAILED DESCRIPTION

[0023] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0024] It should be noted that when a component is referred to as being "mounted on" another component, it may be directly on the other component or there may be a central component. When a component is considered to be "set on" another component, it may be directly set on the other component or there may be a central component at the same time. When a component is considered to be "fixed to" another component, it may be directly fixed on the other component or there may be a central component at the same time.

[0025] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art to which the present invention belongs. The terms used herein in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The term "or / and" used herein includes any and all combinations of one or more of the related listed items.

[0026] Example 1: Please refer to Figure 1 ,This embodiment provides a clothing customization system based on the Internet of Things, including a terminal, a size data generation module, a clothing customization module and a management module.

[0027] The terminal is used to register, log in and change personal information. Users upload multiple target clothing images, multi-view images of the target body and necessary size data through the terminal. The target body image includes front and side photos of the user wearing tight clothing, and the necessary size data includes the user's gender, height and weight. The terminal combines the input body size data and the customized effect map to generate a ready-made clothing customization order.

[0028] The size data generation module includes feature extraction network 1, fusion layer 1, fully connected layer 1, output layer 1 and output layer 2. Feature extraction network 1 is used to preprocess the multi-view images and necessary size data of the terminal output user and extract multi-view features. Feature extraction network 1 uses U-Net as the feature extraction network for semantic segmentation and uses ResNet as the encoder to extract deep features of the human body contour map. Fusion layer 1 is used to add the features of multiple views to obtain a fused feature vector; fully connected layer 1 is used to map the high-dimensional fused feature vector to the low-dimensional output space to obtain the morphological parameters of SMPL, and output layer 1 is used to drive the SMPL standard model with the morphological parameters of SMPL to transform the target human body images of different body shapes and postures.

[0029] The parameters of SMPL include morphological parameters and attitude parameters , morphological parameters , each element can expand or shrink the human body model along a certain direction; posture parameters , represents the rotation angle between the two parts connected by each joint. Through these two parameters, human models with different postures and poses are generated. The calculation steps of the SMPL model are as follows:

[0030]

[0031] In the formula, The skin weight representing the linear skin blend function, Indicates zero attitude 0 standard human body model, , Represents the morphological parameters and attitude parameters The impact on body shape, represents the position of the joint point, and P represents the weight of each joint point. It is a linear function that represents the changes in different human morphologies. The calculation formula is as follows:

[0032] In the formula =[ ], yes The number of parameters, Orthogonal principal component components representing shape displacement. Indicates the slight change of posture parameters to the morphology, and the calculation formula is as follows:

[0033] In the formula, function U transforms the representation of attitude transformation parameters from the form of Euler rotation angle to the form of rotation matrix vector, so U The length of is 23*9=207, Represents U The nth element of Indicates that the human body model is in 0 posture, This defines the effect of posture changes on vertex positions.

[0034] The SMPL model has built-in joint point positions. After the human body model morphology changes, the joint point positions can be calculated by weighting the positions of the human body vertices. The calculation formula is as follows:

[0035] Where I is a weight matrix that calculates the joint point coordinates through the vertex coordinates.

[0036] The second output layer uses the section method to measure the target human image and outputs the measurement data of each part; the section measurement method generates several closed curves for measurement by setting the section to intersect with the human body model. In the measurement of neck circumference, elbow circumference, and wrist circumference, the section to be used is determined by the joint point, and the intersection line of the section and the human body model is used for measurement; in addition, in the measurement of chest circumference, waist circumference, and hip circumference, multiple closed curves are generated by setting a horizontal section at intervals in the chest area, waist area, and hip area, and finally selecting the optimal curve for measurement.

[0037] This module extracts features from multi-view human images and fuses them, restores the target human image based on the parameters of the SMPL predicted by the fully connected layer 1, and achieves accurate measurement of user size data. This module effectively solves the problem of inconvenient user size measurement in online clothing customization, and more accurate size measurement data also improves clothing customization accuracy.

[0038] The clothing customization module includes a design unit, a secondary creation unit, and an aesthetic evaluation unit. The design unit includes a collection module, a geometric matching module, and a virtual fitting module. The collection module extracts features from the input clothing image and the target human body image. The geometric matching module includes an input layer 1, a fusion layer 3, a regression layer, and an output layer 4. The input human body representation feature r and the target clothing feature f are fused through the fusion layer 3 to generate a fusion feature. The regression layer is used to generate the TPS transformation parameter θ through the regression network. The output layer 4 performs a TPS transformation on the input θ and the target clothing feature f to obtain the distorted clothing f. w The virtual fitting module includes input layer 2, U-shaped network layer and output layer 5. The U-shaped network layer is based on f wWith r as input, the U-shaped encoder-decoder network is used to generate the rendered image Tr and the predicted fusion mask M. The output layer 5 fuses Ir and M through the mask to obtain the final fitting result. , try-on results The calculation formula is as follows:

[0039] In the formula represents the Hadamard product.

[0040] By constructing a distorted image of clothing f 𝑤 The real clothing image f parsed from the target image p The perceptual loss between and is used to achieve the effect of improving the appearance feature retention of the target clothing f. The penalty term calculation formula of the loss function of the geometric matching module is as follows:

[0041] In the formula, express , express ; Select VGG19 network as the pre-trained network to distort the clothing image f w The image f parsed from the target image p The perceptual loss between is calculated as follows:

[0042] In the formula represent Input the i-th layer feature map obtained by the pre-trained network, Represents the weight of the layer, and N represents the number of selected feature layers; The final loss function loss calculation formula of the geometric matching module is as follows

[0043] In the formula , is the hyperparameter of the model.

[0044] The aesthetic evaluation unit includes feature extraction network 3, fusion layer 2, fully connected layer 2, and output layer 3. Feature extraction network 3 selects the ResNet-50 feature extraction network to extract the aesthetic features of the virtual fitting pictures. The ResNet-50 feature extraction network includes 5 stages, which are used to extract aesthetic features of different dimensions. Fusion layer 2 reduces the dimension of high-dimensional features through global average pooling, and then fuses multiple features after dimensionality reduction to generate fusion features; the fusion method is horizontal splicing, and the calculation formula is as follows: F(I; )=Ω[GP( ), GP( ), GP( ), GP( )] Where F(I;φ) represents the feature representation of the input image I; I represents the input image; φ represents all parameters of the entire feature extraction network; GP() represents global average pooling; Ω() represents the concatenation of features of multiple scales along the row direction.

[0045] The fully connected layer 2 is used to map the fused features to different output layers. The SoftMax layer is added after the output of the fully connected layer 2 to generate the probability of each evaluation level. The process is described by the following formula: ) Where, t = 1, 2, 3, 4; Represents the distribution of aesthetic evaluation levels in different dimensions. There are four output layers in total, namely image details, image color, image color coordination, and overall aesthetic quality; Represents 4 different fully connected layer networks, and their network parameters are ; The output layer 3 is used to output the aesthetic evaluation results through the quality regression model.

[0046] The secondary creation unit is used to generate a customized preview image according to the personalized customization data provided by the user. The aesthetic evaluation unit simultaneously performs aesthetic evaluation on the virtual fitting image and the customized preview image and sends the image with the highest score to the terminal as the customized effect image.

[0047] The management module is used to manage orders placed by customers, import or delete algorithm models into the system, and manage basic information of anthropometric models.

[0048] This system is based on virtual fitting technology and multi-dimensional aesthetic evaluation technology to provide users with highly personalized customization services, effectively improving the accuracy of online customization renderings and user satisfaction.

[0049] Example 2: Please refer to Figure 2 and Figure 3 This embodiment provides a clothing customization method based on the Internet of Things, which can be applied to the clothing customization based on the Internet of Things in Example 1. The steps of the clothing customization method are as follows: S1: Collect target clothing images, target body images and necessary size data, where the target body images include front and side photos of the user wearing relatively tight clothing, and the necessary size data include the user's gender, height, and weight.

[0050] S2: Generate a virtual fitting image based on the target clothing image and the target human body image. The specific steps are as follows: S21: Match the target clothing images output by the terminal with the target human body images one by one to generate a number of virtual fitting images; the steps of matching and generating virtual fitting images are as follows: S211: extracting features from the input target clothing image and target human body image; S212: Input the human body representation feature r and the target clothing feature f, fuse the two features, and then obtain the TPS transformation parameter θ through the regression network, and then use θ to perform TPS transformation on f to obtain the distorted clothing f w ; S213: With f w With r as input, the U-shaped encoder-decoder network generates the rendered image Tr and the predicted fusion mask M, and Tr and M are fused together through the mask to obtain the final fitting result. ; Virtual fitting pictures The calculation formula is expressed as:

[0051] In the formula Represents the Hadamard product.

[0052] Among them, the calculation formula of the penalty term of the loss function of the regression network is as follows:

[0053] In the formula, express , express ; Select VGG19 network as the pre-trained network to distort the clothing image f w The image f parsed from the target image p The perceptual loss between is calculated as follows:

[0054] In the formula represent Input the i-th layer feature map obtained by the pre-trained network, Represents the weight of the layer, and N represents the number of selected feature layers; The final loss function loss calculation formula of the regression network is as follows

[0055] In the formula , is the hyperparameter of the model.

[0056] By constructing a distorted image of the garment The real clothing image parsed from the target image The perceptual loss between them not only ensures the overall clothing alignment, but also takes into account the local distortion of the clothing, thereby achieving the effect of enhancing the appearance feature retention of the target clothing f and making the generated features more delicate.

[0057] S22: Aesthetically evaluate the input virtual fitting pictures and send the a virtual fitting pictures with the highest scores to the user; the steps of aesthetic quality scoring are as follows: S221: extracting aesthetic features of the virtual fitting images; the feature extraction network selects the ResNet-50 feature extraction network; the ResNet-50 feature extraction network includes 5 stages, which are used to extract aesthetic features of different dimensions respectively.

[0058] S222: Reduce the dimension of high-dimensional features through global average pooling, and then fuse multiple features after dimensionality reduction to generate fused features; the fusion method is horizontal splicing, and the calculation formula is as follows: F(I; )=Ω[GP( ), GP( ), GP( ), GP( )] Where F(I;φ) represents the feature representation of the input image I; I represents the input image; φ represents all parameters of the entire feature extraction network; GP() represents global average pooling; Ω() represents the concatenation of features of multiple scales along the row direction.

[0059] S223: Map the fused features to different output layers, and add a SoftMax layer after the output to generate the probability of each evaluation level. The process is described by the following formula: ) Where, t = 1, 2, 3, 4; Represents the distribution of aesthetic evaluation levels in different dimensions. There are four output layers in total, namely image details, image color, image color coordination, and overall aesthetic quality; Represents 4 different fully connected layer networks, and their network parameters are ; S224: Outputting aesthetic evaluation results through the quality regression model.

[0060] S23: The user adds personalized customization elements to a virtual fitting picture to generate a*b customized preview pictures; the secondary design content of the clothing includes: the user can modify the color of the clothing, use the system's built-in stickers and pictures to add to the clothing, and customize personalized text on the clothing; in addition, the user can switch the front and back of the clothing for design.

[0061] S24: Perform aesthetic evaluation on a virtual fitting image and a customized preview image simultaneously and send the image with the highest score as the customized rendering to the customer.

[0062] S3: Preprocess the input multi-view image of the target human body and necessary size data, construct the target human body image according to the preprocessed data, and measure the size data of each part of the target human body image using the section method; the size measurement steps are as follows: S31: Based on the images uploaded by the user and the necessary size data, U-Net is used as the feature extraction network to extract multi-view features of the human body contour map.

[0063] S32: Add the features of multiple perspectives to obtain a fused feature vector; obtain the morphological parameters of SMPL by mapping the high-dimensional fused feature vector to the low-dimensional output space. The parameters of SMPL include morphological parameters and attitude parameters , morphological parameters , each element can expand or shrink the human body model along a certain direction; posture parameters , represents the rotation angle between the two parts connected by each joint. Through these two parameters, human body models with different postures and poses are generated.

[0064] S33: Drive the SMPL standard model transformation through the SMPL morphological parameters to form target human images of different body shapes and postures. The calculation steps of the SMPL model are as follows:

[0065]

[0066] In the formula, The skin weight representing the linear skin blend function, Indicates zero attitude 0 standard human body model, , Represents the morphological parameters and attitude parameters The impact on body shape, represents the position of the joint point, and P represents the weight of each joint point. It is a linear function that represents the changes in different human morphologies. The calculation formula is as follows:

[0067] In the formula =[ ], yes The number of parameters, Orthogonal principal component components representing shape displacement. Indicates the slight change of posture parameters to the morphology, and the calculation formula is as follows:

[0068] In the formula, function U transforms the representation of attitude transformation parameters from the form of Euler rotation angle to the form of rotation matrix vector, so U The length of is 23*9=207, Represents U The nth element of Indicates that the human body model is in 0 posture, This defines the effect of posture changes on vertex positions.

[0069] The SMPL model has built-in joint point positions. After the human body model morphology changes, the joint point positions can be calculated by weighting the positions of the human body vertices. The calculation formula is as follows:

[0070] Where I is a weight matrix that calculates the joint point coordinates through the vertex coordinates.

[0071] S34: Using the section method, a section is set to intersect with the human body model to generate several closed curves for measurement. In the measurement of neck circumference, elbow circumference, and wrist circumference, the section used is determined by the joint point, and the intersection line of the section and the human body model is used for measurement; in addition, in the measurement of chest circumference, waist circumference, and hip circumference, a horizontal section is set at intervals in the chest area, waist area, and hip area to generate multiple closed curves, and finally an optimal curve is selected for measurement, and the measurement data is sent to the user.

[0072] S4: The user selects the most favorite customized rendering on the terminal, and generates a customized clothing order based on the size data of various parts of the human body.

[0073] This method realizes online clothing customization based on virtual fitting and human body measurement technology, enabling users to choose clothing styles, colors, patterns and prints according to their preferences and needs, enhancing the user's personalized customization experience. This method provides online customization services to improve consumer satisfaction, while improving the customization and sales efficiency of clothing and reducing inventory and management costs.

[0074] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0075] The above embodiments only express several implementation methods of the present invention, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for those of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

Claims

1. A clothing customization system based on the Internet of Things, characterized in that: include: Terminal, used to register users, log in to accounts and change personal information; The user uploads multiple target clothing images, target human multi-view images and necessary size data through the terminal; The size data generation module includes a feature extraction network 1, a fusion layer 1, a fully connected layer 1, an output layer 1 and an output layer 2; the feature extraction network 1 is used to pre-process the multi-view images and necessary size data output by the terminal to extract multi-view features; The fusion layer 1 is used to add the features of multiple perspectives to obtain a fusion feature vector; the fully connected layer 1 is used to map the high-dimensional fusion feature vector to a low-dimensional output space to obtain the morphological parameters of SMPL; the output layer 1 is used to drive the SMPL standard model transformation through the morphological parameters of SMPL to form target human images of different body shapes and postures; the output layer 2 uses the section method to measure the target human image and output the size data of each part; The clothing customization module includes a design unit, a secondary creation unit, and an aesthetic evaluation unit; the design unit is used to match the target clothing picture output by the terminal with the target human body image to generate multiple virtual fitting pictures; the aesthetic evaluation unit is used to perform aesthetic evaluation on the virtual fitting pictures and send the top-scoring virtual fitting pictures to the secondary creation unit; the secondary creation unit is used to generate a customized preview picture according to the personalized customization data provided by the user; The aesthetic evaluation unit simultaneously performs aesthetic evaluation on the virtual fitting image and the customized preview image and sends the a images with the highest scores to the terminal as customized effect images; The management module is used to manage the orders placed by customers, import or delete algorithm models into the system, and manage the basic information of the anthropometric model.

2. The clothing customization system based on the Internet of Things according to claim 1 is characterized in that: The design unit includes a collection module, a geometric matching module and a virtual fitting module; The collection module extracts features from the input clothing pictures and target human images; the geometric matching module is used to distort the clothing according to the shape and posture features of the target human body; and the virtual fitting module is used to fuse the distorted clothing with the target human body features to obtain a virtual fitting image.

3. The clothing customization system based on the Internet of Things according to claim 2 is characterized in that: The aesthetic evaluation unit includes a feature extraction network three, a fusion layer two, a fully connected layer two, and an output layer three; the feature extraction network three selects a ResNet-50 feature extraction network to extract the aesthetic features of the virtual fitting image; the fusion layer two reduces the dimension of high-dimensional features through global average pooling, and then fuses the multiple features after dimensionality reduction to generate fused features; the fully connected layer two is used to map the fused features to different output layers three; the output layer three is used to output the aesthetic evaluation results through a quality regression model.

4. The clothing customization system based on the Internet of Things according to claim 1 is characterized in that: The feature extraction network 1 uses U2-Net as the feature extraction network for semantic segmentation, and uses ResNet18 as the encoder to extract deep features of the human body contour map.

5. The clothing customization system based on the Internet of Things according to claim 1 is characterized in that: The section method used in the output layer 2 is as follows: a section is intersected with the human body model, and the best curve is selected for measurement from several closed curves obtained by the intersection.

6. The clothing customization system based on the Internet of Things according to claim 2 is characterized in that: The geometric matching module includes an input layer 1, a fusion layer 3, a regression layer and an output layer 4. The fusion layer 3 fuses the input human body representation feature r and the target clothing feature f to generate a fusion feature. The regression layer is used to generate a TPS transformation parameter θ through a regression network for the fusion feature. The output layer 4 performs a TPS transformation on the input θ and the target clothing feature f to obtain a distorted clothing f. w ; The virtual fitting module includes an input layer 2, a U-shaped network layer and an output layer 5. The U-shaped network layer is based on f w With r as input, the U-shaped encoder-decoder network is used to generate the rendered image Tr and the predicted fusion mask M. The output layer 5 fuses Ir and M through the mask to obtain the final fitting result. , try-on results The calculation formula is as follows: In the formula Represents the Hadamard product.

7. The clothing customization system based on the Internet of Things according to claim 5 is characterized in that: The penalty term calculation formula of the loss function of the geometric matching module is as follows: In the formula, express , express ; The VGG19 network is selected as the pre-trained network, and the distorted clothing image f w The image f parsed from the target image p The perceptual loss between is calculated as follows: In the formula represent Input the i-th layer feature map obtained by the pre-trained network, Represents the weight of the layer, and N represents the number of selected feature layers; The final loss function loss calculation formula of the geometric matching module is as follows In the formula , is the hyperparameter of the model.

8. A clothing customization system based on the Internet of Things according to any one of claims 1 to 7, characterized in that: The IoT-based clothing customization method includes the following steps: S1: Collect target clothing images, target human body images and necessary size data; S2: Generate a virtual fitting image based on the target clothing image and the target human body image. The specific steps are as follows: S21: Matching the target clothing images output by the terminal with the target human body images one by one to generate a number of virtual fitting images; S22: performing aesthetic evaluation on the input virtual fitting pictures and sending the a virtual fitting pictures with the highest scores to the user; S23: The user adds personalized customization elements to a virtual fitting picture to generate a*b customized preview pictures; S24: performing aesthetic evaluation on a virtual fitting image and a*b customized preview images simultaneously and sending the image a with the highest score as the customized rendering to the customer; S3: preprocessing the input multi-view images of the target human body and necessary size data, constructing the target human body image according to the preprocessed data; and measuring the size data of each part of the target human body image by using the section method; S4: The user selects a favorite custom rendering and generates a custom clothing order based on the size data of various parts of the human body.

9. The clothing customization method based on the Internet of Things according to claim 8, characterized in that: In step S21, the method of matching and generating a virtual fitting picture is as follows: S211: extracting features from the input target clothing image and target human body image; S212: Input the human body representation feature r and the target clothing feature f, fuse the two features, and then obtain the TPS transformation parameter θ through the regression network, and then use θ to perform TPS transformation on f to obtain the distorted clothing f w ; S213: With f w With r as input, the U-shaped encoder-decoder network generates the rendered image Tr and the predicted fusion mask M, and Tr and M are fused together through the mask to obtain the final fitting result. .

10. The clothing customization method based on the Internet of Things according to claim 8, characterized in that: In step S22, the method of aesthetic evaluation is as follows: S221: Extracting aesthetic features of virtual fitting images; S222: Reduce the dimension of high-dimensional features through global average pooling, and then fuse multiple features after dimensionality reduction to generate fused features; the fusion method is horizontal splicing, and the calculation formula is as follows: F(I; )=Ω[GP( ),GP( ),GP( ),GP( )] In the formula, F(I;φ) represents the feature representation of the input image I; I represents the input image; φ represents all parameters of the entire feature extraction network; GP() represents global average pooling; Ω() represents the concatenation of features of multiple scales along the row direction. ; S223: Map the fused features to different output layers, and add a SoftMax layer after the output to generate the probability of each evaluation level. The process is described by the following formula: ) Where, t = 1, 2, 3, 4; Represents the distribution of aesthetic evaluation levels in different dimensions, with a total of four output layers, namely image details, image color, image color coordination and overall aesthetic quality; Represents 4 different fully connected layer networks, and their network parameters are ; S224: Outputting aesthetic evaluation results through the quality regression model.

Images