Clothing size measurement system and method based on convolutional neural network

Abstract

The present invention belongs to the technical field of automatic measurement of clothing dimensions, and discloses a T-shirt garment measurement system and method based on a convolutional neural network. An image import module is used for users to import T-shirt images to be detected; an image preprocessing module is connected to the image import module and is used to enhance the images imported by users; a key point recognition rule extraction module is connected to the image preprocessing module and is used to identify the key points of the T-shirt in the images imported by users; a data input module is used for users to input actual measurement data of the width of the neckline; a key size acquisition module is connected to the data input module and the key point recognition rule extraction module and is used to realize automatic measurement of the dimensions of other parts of the T-shirt; a size display module is connected to the key size acquisition module and is used to display the automatically measured T-shirt dimensions to users. The system of the present invention has high robustness and good generalization ability, the dimension measurement results meet the error range requirements, and the operation process is not interfered by external factors.

Description

Technical Field

[0001] The present invention belongs to the technical field of automatic measurement of clothing size, and in particular to a clothing size measurement system and method based on a convolutional neural network. Background Art

[0002] Clothing size is an important parameter to be considered in the clothing production process of enterprises. It is an indispensable indicator in the process of garment size quality inspection, garment shrinkage resistance inspection, and stretch performance inspection. Clothing is easy to deform, has a wide variety of categories, and has complex measurement parameters. At present, most clothing companies at home and abroad have been using inefficient manual measurement methods for size measurement. However, with the deepening of informatization and automation of the clothing industry and the emergence of large-scale batch customization, higher requirements have been put forward for clothing size measurement.

[0003] Judging from the current domestic clothing production situation, most companies still use the traditional manual measurement method in the clothing production process, which has high labor costs, low efficiency, and large errors in size measurement, which seriously affects the production efficiency of companies. In recent years, with the rise of e-commerce platforms such as Taobao and JD.com, more and more people buy clothes online, but after trying them on, they find that there are problems with the size of the clothes. The main reason is that there is no unified standard for clothing sizes and the errors caused by manual size measurement. Merchants have also received a lot of bad reviews, which seriously affects the sales of clothing.

[0004] Foreign research on convolutional neural networks is mainly focused on image detection, recognition and classification. Wicaksono et al. proposed a convolutional neural network model that can classify batik patterns, named IncRes, which merges the Inception module with the remaining network structure. Nawaz et al. proposed an Inception convolutional neural network model based on GoogleNet, which classified actual images of some traditional costumes in Bangladesh into pre-defined categories, and used models such as SGD to classify images in training sets and test sets. LiuZ et al. introduced the large-scale clothing dataset DeepFashion, and proposed the FashionNet convolutional neural network deep model to serve clothing recognition and detection, and optimized it through iteration.

[0005] Foreign studies on automatic measurement of clothing size are mostly based on the combination of clothing contour extraction algorithm and clothing image processing. Norton-Wayn proposed to use image processing technology to collect clothing images, detect corners and extract features, and finally complete the automatic measurement of its size. LICAO and ChenK et al. enhanced clothing images through image processing algorithms, detected the edges of clothing images with fast fuzzy edge detection, and used Freeman's new detection algorithm to locate corners, remove impulse noise and non-impulsive noise in clothing images, accurately detect the edges and corners of clothing images, and measure sizes close to actual sizes.

[0006] Jia Xiaojun and Deng Hongtao in China classified the texture of blue printed cloth by improving and optimizing the VGGNet16 convolutional neural network. The network was improved by adding convolution groups and adjusting network parameters. Li Zhi et al. constructed a clothing dataset B_DATClothing with 100,000 items of 16 attributes, and established a retrieval algorithm for attribute-based clothing attribute classification on this basis to improve the classification efficiency and retrieval speed of clothing. Qu Jinshan et al. proposed a face recognition algorithm based on YOLOv3, which uses the key point loss function to determine the key points of the face. YOLOv3 implements the output of the rectangular frame and the key points of the face detection in face detection. Chen Yuanyuan et al. proposed a method for locating the key points of clothing using a cascaded deep convolutional neural network. This method effectively combines high-level semantic information with low-level visual information of key parts. Wang Yiwen used a convolutional neural network model to detect the key points of Hanfu, combined the Hough transform circle detection algorithm and the K-Means clustering algorithm, eliminated invalid key points to obtain the coordinates of the key points of Hanfu, and then calculated the sizes of other parts of Hanfu through proportional mapping.

[0007] The existing research on automatic clothing size measurement in China can be roughly divided into three categories: combining image processing with corner detection algorithm, combining fuzzy edge detection algorithm with inflection point detection algorithm, and template matching algorithm to achieve clothing size measurement. These three research methods are all based on extracting clothing key points based on clothing contours, and then combining them with camera calibration coefficients to finally achieve automatic measurement of clothing size.

[0008] Convolutional neural networks are rarely used in the automated measurement of clothing sizes. There are very few literatures that mention the use of convolutional neural networks to measure the key sizes of Hanfu. However, no one has yet studied the application of convolutional neural networks in the automated measurement of ready-made clothing sizes. Therefore, there is still huge room for research in the automated measurement of ready-made clothing sizes based on convolutional neural networks.

[0009] Through the above analysis, the problems and defects of the existing technology are as follows: the currently known methods for automatic measurement of clothing size are mostly based on manually designed detection algorithms, which are relatively simple in sample feature extraction and are easily affected by the surrounding environment and clothing deformation. Summary of the invention

[0010] In view of the problems existing in the prior art, the present invention provides a T-shirt garment measurement system and method based on a convolutional neural network.

[0011] The present invention is implemented in this way: a clothing size measurement system based on a convolutional neural network, the clothing size measurement system based on a convolutional neural network comprises: a training image import module, an image preprocessing module, a key point recognition rule extraction module, a data input module, a key size acquisition module, a test image input module, and a size display module;

[0012] The training image import module is used to import clothing pictures of a specific category to be trained; the image data set is divided into three parts: training, verification, and testing. The training set is used to train the model, the verification set is used to observe the model performance, and the test set is used to verify the model effect.

[0013] The training image preprocessing module is connected to the training image importing module and is used to mark the key points corresponding to the specific test items on the training image using the image calibration tool according to the specific items of the clothing size test; the positions of the key points are determined by the specific items of the clothing size measurement.

[0014] A key point rule extraction module, connected to the training image preprocessing module, is used to establish key point recognition rules for specific categories of clothing based on training images imported by users;

[0015] A data input module is used for the user to input the real measurement size data of the benchmark part of the garment to be tested;

[0016] The measurement image import module is used by the user to import clothing pictures of the corresponding category to be detected;

[0017] The key size acquisition module is connected with the data input module and the key point recognition rule extraction module, and is used to realize the automatic measurement of the sizes of other parts of the garment to be measured;

[0018] The size display module is connected to the key size acquisition module and is used to display the automatically measured garment size to the user.

[0019] Furthermore, the key point recognition rule extraction module uses the DarkNet-53 convolutional neural network model to extract the key features of the garment, and then combines the Yolov3 target detection and K-means clustering algorithm to screen out the effective key point information.

[0020] Furthermore, Yolov3’s target detection algorithm detects the key points of clothing, and uses DarkNet-53 to perform feature fusion using the up and down sampling method to obtain the feature layer of the key points of clothing measurement. The feature layer is divided into three scales, which are used for targets at different scales.

[0021] Furthermore, the K-means clustering algorithm calculates the distance from the sample point to the cluster center point and uses the improved intersection-union ratio for clustering.

[0022] Furthermore, the key dimension acquisition module combines the actual dimension data of the benchmarked parts and realizes the automatic measurement of the dimensions of other parts of the clothing through the proportional mapping between the pixel distance of the key points and the actual dimensions.

[0023] The reference part must be located in the image of the garment to be tested. The reference part size is often the size of a standard part or a part with little or relatively minimal error in the image, or a known part size. For example, the size data of a fixed part on each garment to be tested, such as chest width, etc.; or the size of a standard part or part in a batch of garments of the same style, such as buttons, pocket width, etc. on the garment; or it can be a fixed-size mark attached or pasted on the garment, such as a fixed-size ruled sticker, etc.

[0024] Furthermore, the model training applies the YOLOv3 network model to the clothing key point dataset through transfer learning, extracts key point features through the backbone network DarkNet-53 convolutional neural network, obtains key point coordinates through the k-means clustering algorithm, and improves the key point recognition accuracy through repeated iterative training.

[0025] Furthermore, Python software is used to design the system GUI interface, and an acquisition system including image display, key point acquisition, and size display is established.

[0026] Another object of the present invention is to provide a convolutional neural network-based garment size measurement system and a convolutional neural network-based garment size measurement method, comprising:

[0027] (1) The user imports the training images produced by the clothing measurement rules through the training image import module, and the key points of the training images are marked through the training image preprocessing module;

[0028] (2) The DarkNet-53 convolutional neural network model is used to extract key features of garments, and then the Yolov3 target detection and K-means clustering algorithm are used to screen out effective key point information. Combined with the actual measured size of the reference part input by the user, the pixel distance of the key point is mapped to the actual size to achieve automatic measurement of the size of other parts of the garment;

[0029] (3) The size display module displays the sizes of other parts of the garment to be inspected to the user.

[0030] Another object of the present invention is to provide a convolutional neural network-based clothing size measurement system for application in clothing size measurement.

[0031] Furthermore, the user can directly obtain the size of the garment to be measured by taking a picture of the garment to be measured.

[0032] In combination with the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solutions to be protected by the present invention are as follows:

[0033] First, the present invention applies the convolutional neural network DarkNet-53 and Yolov3 target detection algorithm in computer deep learning to the automatic measurement of key dimensions of ready-made clothes, aiming to build a fully automatic ready-made clothing size measurement system, which can directly obtain the key dimensions of ready-made clothes through the photographed ready-made clothing pictures. The system has great potential value in the production and detection of ready-made clothes, and is of great help to the problems of low efficiency and large error of manual size measurement in the existing ready-made clothing production. Taking T-shirts as an example, the key point leakage recognition rate of T-shirt tops detected by the automatic measurement system for key dimensions of ready-made clothes is 0, the single point correct rate is 90.5%, and the recognition accuracy rate is 82.3%; the absolute error of T-shirt size is generally between 5-15mm, and the relative error is between 0.50%-4.00%, which meets the measurement error range given by the national T-shirt specification size deviation comparison range table.

[0034] The system of the present invention has high robustness and good generalization ability. The size measurement results meet the error range requirements, and the operation process is not affected by external factors. It only needs to input the pictures corresponding to the measured parts of the garment and the actual measured dimensions of the benchmark parts to quickly and accurately obtain the size mapping data of other key parts. The experiment was successful and the solution has practical application value.

[0035] Second, the system for acquiring key clothing dimensions based on convolutional neural networks can effectively solve the problems caused by manual measurement of clothing dimensions in the existing production process. On the one hand, workers can be relieved from heavy work and reduce work fatigue, and factories can also reduce industrial labor costs; on the other hand, the efficiency of clothing dimension measurement is improved, the subjectivity and uncertainty of traditional clothing dimension measurement methods are overcome, the accuracy of dimension measurement is improved, the measurement error is reduced, and the requirements of mass production measurement of enterprises are met.

[0036] Third, as auxiliary evidence of the inventiveness of the claims of the present invention, it is also reflected in the following important aspects:

[0037] (1) The expected benefits and commercial value of the technical solution of the present invention after transformation are:

[0038] After the technical solution of the present invention is converted, it can realize the automatic measurement of garment size by combining the hanging system and camera photography of the garment manufacturing factory. After the conversion, the technical solution can work 24 hours a day, with high measurement accuracy, and improve the work efficiency of garment quality inspection; combined with remote communication technology. The measurement results can be obtained remotely and in real time by measuring the size by the technical solution, which can improve the efficiency of quality management; at the same time, the technical solution replaces the work of garment size measurement personnel, saves the labor cost of garment size inspection, and has great commercial value.

[0039] (2) The technical solution of the present invention fills the technical gap in the industry at home and abroad:

[0040] The present invention proposes a clothing size acquisition system based on a convolutional neural network. First, a DarkNet-53 convolutional neural network model is built as the feature extraction backbone network of the YOLOv3 target detection algorithm. The efficiency of YOLOv3 network model training and the ability to extract key points of size measurement on clothing are improved through transfer learning. The coordinates of key points of size measurement are acquired by using the YOLOv3 target detection algorithm k-means clustering. The accuracy of key point recognition is improved through repeated iterative training. Finally, the pixel distance is obtained according to the key point coordinates, and then combined with the real measurement size data of the clothing benchmarking parts actively provided, the size data of other key parts are obtained through proportional mapping. This technical solution fills the gap in the automation of domestic clothing size measurement.

[0041] Fourth, the clothing size measurement system based on convolutional neural networks has achieved significant technical progress in many aspects:

[0042] 1. Automation and accuracy: Traditional clothing size measurement usually relies on manual operation, which is not only inefficient but also prone to errors. The system realizes automatic measurement through advanced convolutional neural network technology, which greatly improves the accuracy and consistency of measurement.

[0043] 2. Wide applicability: Regardless of different styles or types of clothing, as long as the corresponding training is carried out, the system can accurately measure and has strong versatility.

[0044] 3. Real-time: Traditional measurement methods take a long time, but this system can give measurement results in real time through automated processing, greatly shortening the time for dimension acquisition.

[0045] 4. Reduce the influence of human factors: Traditional measurement methods are greatly affected by the skills and experience of the operator, while this automated method reduces the interference of human factors and improves the objectivity of measurement.

[0046] 5. Data traceability: The system can store measured data and processes to facilitate subsequent data analysis, quality control and traceability.

[0047] 6. Advantages of integrating multiple algorithms: By combining DarkNet-53, Yolov3 target detection and K-means clustering algorithms, the system can not only accurately identify key points, but also effectively classify and screen these key points to ensure the accuracy of measurement.

[0048] 7. Easy to operate: Users only need to import the corresponding training and test pictures and enter a small amount of known dimensions, and the system can automatically complete the measurement. The operation is simple and convenient.

[0049] In general, this clothing size measurement system based on convolutional neural network has achieved significant technological progress in the field of clothing measurement, which not only improves the efficiency and accuracy of measurement, but also brings huge potential value to the clothing industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments of the present invention. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0051] Figure 1 is a structural diagram of a clothing size measurement system based on a convolutional neural network provided by an embodiment of the present invention;

[0052] Figure 2 is a flow chart of a clothing size measurement method based on a convolutional neural network provided by an embodiment of the present invention;

[0053] Figure 3 It is a relationship diagram between the YOLOv3 prediction box and the anchor provided by an embodiment of the present invention;

[0054] Figure 4 It is the specific size measurement item of T-shirt clothing.

[0055] Figure 5 These are the specific locations of 13 key points that need to be identified on the model corresponding to the dimensional measurement items.

[0056] Figure 6 It is a display diagram of some experimental image detection results provided by the embodiment of the present invention;

[0057] Figure 7 is a schematic diagram of manual measurement of T-shirt size provided by an embodiment of the present invention;

[0058] Figure 8A loss function curve diagram provided for an embodiment of the present invention. DETAILED DESCRIPTION

[0059] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0060] In view of the problems existing in the prior art, the present invention provides a clothing size measurement system and method based on convolutional neural network.

[0061] The present invention is achieved by: Figure 1 As shown, a clothing size measurement system based on a convolutional neural network, the clothing size measurement system based on a convolutional neural network includes: a training image import module, an image preprocessing module, a key point recognition rule extraction module, a data input module, a key size acquisition module, a test image input module, and a size display module;

[0062] The training image import module is used by users to import training images of a specific category of clothing for measurement purposes. It includes images used for key point recognition training and verification in the DarkNet-53 convolutional neural network model and images used to verify the model effect.

[0063] A training image preprocessing module, connected to the training image importing module, is used to mark key points of the training image;

[0064] A key point recognition rule extraction module, connected to the training image preprocessing module, is used to establish key point recognition rules for specific categories of clothing based on training images imported by users;

[0065] A data input module is used for the user to input the actual measurement size data of the benchmark part of the garment to be tested;

[0066] The measurement image import module is used by the user to import clothing pictures of the corresponding category to be detected;

[0067] The key size acquisition module is connected with the data input module and the key point recognition rule extraction module, and is used to realize the automatic measurement of the sizes of other parts of the garment to be measured;

[0068] The size display module is connected to the key size acquisition module and is used to display the automatically measured garment size to the user.

[0069] The key point recognition rule extraction module uses the DarkNet-53 convolutional neural network model to extract key features of clothing, and then combines Yolov3 target detection and K-means clustering algorithm to screen out effective key point information.

[0070] Yolov3’s target detection algorithm detects the key points of clothing, and uses DarkNet-53 to perform feature fusion using the up-and-down sampling method to obtain the feature layer of the key points of clothing measurement. It is divided into three scale feature layers, which are used for targets at different scales.

[0071] The K-means clustering algorithm calculates the distance from the sample point to the cluster center point and uses the improved intersection-union ratio for clustering.

[0072] The key dimension acquisition module combines the actual measurement data of the reference part and realizes the automatic measurement of the dimensions of other parts of the clothing through the proportional mapping between the pixel distance of the key points and the real measurement dimension data of the reference part.

[0073] The model training applies the YOLOv3 network model to the clothing key point dataset through transfer learning, extracts key point features through the backbone network DarkNet-53 convolutional neural network, obtains key point coordinates through the k-means clustering algorithm, and improves the key point recognition accuracy through repeated iterative training.

[0074] Use python software to design the system GUI interface and establish an acquisition system including image display, key point acquisition, and size display.

[0075] The clothing size measurement system based on convolutional neural network provided by the present invention mainly realizes automatic detection and measurement of clothing size through machine learning methods, especially YOLOv3 target detection and DarkNet-53 network in deep learning. The working principle of the system can be explained in detail according to the following steps:

[0076] Workflow:

[0077] 1) Training image import module:

[0078] The user imports clothing images for training into the system. These images should cover different types of clothing and annotate the key points of the clothing (such as cuffs, necklines, hems, etc.).

[0079] 2) Training image preprocessing module:

[0080] The imported training images are preprocessed, including marking the locations of key points, which is usually a manual annotation process.

[0081] 3) Key point recognition rule extraction module:

[0082] Using the preprocessed images and the DarkNet-53 network model, the features of the key points of clothing are extracted and recognition rules are established. This step is achieved through the training process of deep learning.

[0083] 4) Data input module:

[0084] The user enters the real-world measurements of a reference part of the garment (e.g., the standard length of a T-shirt). This data will be used to convert the pixel dimensions on the image into real-world dimensions.

[0085] 5) Measurement image import module:

[0086] Users import pictures of clothing to be measured. These pictures are new clothing instances, and the system needs to detect and measure their sizes.

[0087] 6) Key size acquisition module:

[0088] This module is responsible for converting the pixel distance of the identified key points into actual size. It combines the size data of the reference part in the data input module and the key point position obtained by the key point recognition rule extraction module.

[0089] 7) Size display module:

[0090] Finally, the measured dimensions will be displayed on the user interface, allowing users to intuitively see the size of each part.

[0091] DarkNet-53: is a deep convolutional neural network used to extract features from clothing images. It is the backbone network of the YOLOv3 object detection system and is responsible for generating feature maps for detection.

[0092] YOLOv3: is a real-time object detection algorithm used to locate key points in clothing images. It outputs predicted boxes that indicate the locations of key points.

[0093] K-means clustering algorithm: used to optimize the anchor boxes in the YOLOv3 algorithm. By calculating the distance from the sample point to the cluster center, the optimal prediction box size can be found to better match the actual key points in the data set.

[0094] Model training: Using transfer learning, the YOLOv3 network model is applied to the clothing key point dataset. Through repeated iterative training process, the network learns how to more accurately identify key points in new images.

[0095] System GUI interface: Use Python software for interface design to create a user-friendly operation interface, including image display, key point recognition results, and dimension measurement results display.

[0096] Figure 1 The overall structure of the system is shown, including the various modules and their interconnections.

[0097] Figure 2 A flowchart showing the clothing size measurement method, each step from image import to size display.

[0098] Figure 3 Describes the relationship between the prediction box and the anchor boxes in YOLOv3.

[0099] Figure 4 Shows the specific size measurements for T-shirt garments.

[0100] Figure 5 The specific locations of 13 key points that need to be identified in the model corresponding to the dimensional measurement project are shown.

[0101] Figure 6 The results of image detection in some experiments are shown, including examples of key point detection.

[0102] Figure 7 An example diagram showing manual measurement of clothing sizes.

[0103] Figure 8 The loss function curve during training is shown, reflecting the improvement of model performance during training.

[0104] Through the above steps and techniques, the clothing size measurement system of the present invention can automatically measure the size of clothing and provide accurate size information for clothing production and online retail.

[0105] The working principle of the clothing size measurement system based on convolutional neural network provided by the present invention is as follows:

[0106] 1. Training image import: Users first import clothing images of specific categories, such as T-shirts, pants, etc. through the training image import module. These images include not only images used for training, but also images used for subsequent verification of model effects.

[0107] 2. Image preprocessing: In the preprocessing module, the training images are processed to facilitate the marking of key points, which includes image cropping, resizing, normalization, etc., to ensure that each image can be effectively processed by the system.

[0108] 3. Key point recognition rule extraction: This module uses the DarkNet-53 convolutional neural network model to extract features from training images. DarkNet-53 is a deep neural network structure that can identify subtle features in images. These features, combined with the Yolov3 object detection algorithm, can help the system detect key points on clothing. The K-means clustering algorithm is used to classify these key points and filter out valid and meaningful key points.

[0109] 4. Data input: The user inputs a real measurement size data of the reference part, or a known size data, such as the standard part - the button diameter in the garment image to be measured, the pocket width that is difficult to produce size errors in garment production, etc. This data provides a reference benchmark for the system, enabling it to perform proportional mapping measurements of other parts based on this known size.

[0110] 5. Measurement image import: The user imports the pictures of the garment to be measured, which should belong to the same category as the previously imported training pictures.

[0111] 6. Key size acquisition: Based on the key point recognition rules obtained through previous training, the system identifies the key points on the garment image to be tested, and performs estimated size measurement based on the actual measurement size data of the reference parts provided by the user.

[0112] 7. Size display: Finally, the measured garment size will be displayed in the size display module, and the user can view it directly.

[0113] In short, this system can automatically and accurately measure clothing sizes through advanced convolutional neural networks and object detection technology. Users only need to provide a small number of known sizes and pictures of the clothing to be measured to obtain complete clothing size information.

[0114] Another object of the present invention is to provide a convolutional neural network-based garment size measurement system and a convolutional neural network-based garment size measurement method, comprising:

[0115] (1) The user imports the training images produced by the clothing measurement rules through the training image import module, and the key points of the training images are marked through the training image preprocessing module;

[0116] (2) The DarkNet-53 convolutional neural network model is used to extract the key features of the garment, and then the Yolov3 target detection and K-means clustering algorithm are used to screen out the effective key point information. The actual measurement data of the reference part input by the user is combined, and the pixel distance of the key point is mapped to the real size to achieve the automatic measurement of the size of other parts of the garment.

[0117] (3) The size display module displays the sizes of other parts of the garment to be inspected to the user.

[0118] Another object of the present invention is to provide a convolutional neural network-based clothing size measurement system for application in clothing size measurement.

[0119] Users can directly obtain the key dimensions of the garment by taking pictures of the garment to be tested.

[0120] In order to prove the creativity and technical value of the technical solution of the present invention, this section provides application examples of the claimed technical solution on specific products or related technologies.

[0121] Taking the T-shirt size measurement system as an example, according to national standards and the actual situation of T-shirt production, for T-shirt clothing, the key dimensions of clothing and the measurement method of the dimensions are shown in Table 1. The specific parts of the T-shirt measurement size are shown in Figure 4 .

[0122] Table 1 T-shirt key size measurement standards

[0123]

[0124] Based on the measurement parts in Table 1, there are 13 clothing key points corresponding to the measurement of these parts, see Figure 5 :

[0125] (1) Collect T-shirt images of the same style as the T-shirt to be tested, and eliminate images with the same background and clothing color, blurred images, or images that do not fully capture the corners of the T-shirt. Unify the format of the filtered images, change the image format to jpg, rename the images and reorder the order of the images to prepare for the key point marking later. Make sure there are no spaces or brackets in the image names. Finally, determine the number of model training images to be 1,634.

[0126] (2) Determine the T-shirt images that the model needs to recognize Figure 5 The 13 key points in the image are set as shown in Table 2. The key points of the T-shirt image are marked with the special image calibration tool labelimg, and the file is saved in XML format. The entire dataset is divided into three folders, namely Annotations, JPEGImages, and ImageSets. All the labeled XML files of the dataset are saved in the Annotation folder, the original images of all the datasets are saved in the JPEGImages folder, and the txt files are saved in the ImageSets folder. The txt files divide all the images in the dataset into various sets. The training image set is recorded in train.txt stored in the prediction folder, and the verification image set is recorded in val.txt.

[0127] Table 2 Key point preset label positions

[0128]

[0129]

[0130] (3) The dataset is divided into three parts: training, validation, and testing. The training set is used to train the model, the validation set is used to observe the model performance, and the test set is used to verify the model effect. Based on the experience of previous model training dataset division, the dataset images are divided into training set, validation set, and test set in a ratio of 9:1, where (training set and validation set): test set is 9:1, and training set: validation set is 9:1. The YOLOv3 target detection network is input for learning and training to obtain the trained parameter model.

[0131] (4) Based on the model transfer learning method, the Darknet-53 parameter model obtained from the large-scale dataset ImageNet is used as a pre-training model, and is applied to the T-shirt key point dataset for transfer training through transfer learning. The trained Darknet-53 parameter model is added to the Yolov3 network model structure.

[0132] (5) Mark the key point positions of the sizes to be measured on a large number of T-shirt clothing training images, and obtain the prediction results in the T-shirt key point feature map obtained by YOLOv3 from Darknet53.

[0133] (6) The initial setting of the learning rate is 0.001 according to the Yolov3 standard. Start with a relatively large learning rate. When the loss rate decreases slowly during the training process, adjust the learning rate to 0.0001 and continue training.

[0134] (7)YOLOv3 training loss function:

[0135]

[0136] The loss function is Figure 6 As shown in the figure, when the number of iterations (Epoch) is about 100, the loss curve begins to flatten out, and remains stable after Epoch 100, indicating that the model is gradually converging and the training effect is good. After transfer learning, a network parameter model that can accurately identify the key point features of the T-shirt is finally obtained for T-shirt feature extraction. Figure 6 This is a loss function curve graph provided by the present invention.

[0137] The Yolov3 target detection algorithm is used to detect the key points of the T-shirt. The entire image is used as the input of the network. The feature map of the key points of the T-shirt is obtained through the feature extraction network DarkNet-53. Multi-scale feature fusion is performed through upsampling and downsampling to obtain feature maps of three different scales. The sizes of the feature maps are 13×13, 26×26, and 52×52, which are used to extract large-scale, medium-scale, and small-scale targets, respectively.

[0138] Decode the prediction results, that is, calculate the coordinates bx, by, and the height and width bw, bh of the key point bounding box through offset, so as to obtain the position of the border. tx represents the offset from the x-axis to the upper left of the prediction grid, ty represents the offset from the y-axis to the upper left of the prediction grid, tw, th represent the scale scaling, and the network training will learn these parameters. Figure 3 As shown in the figure, the box represents the prediction box, bw and bh represent the prediction box height and width, the black dotted box represents the prior box, and Pw and Ph represent the width and height of the prior box. σ(tx) and σ(ty) represent the offset of the center point of the target relative to the grid responsible for detecting the target, and Cx and Cy represent the coordinates of the upper left corner of the grid where the center point is located. The final result is the position in the feature map, and the position in the original image should be increased by 32 times. The formula is shown in Table 3:

[0139] Table 3 Prediction box calculation in feature map and original image position

[0140]

[0141] (6) Sort the scores of the predicted bounding boxes and select the box with the highest score as the target box. Use the non-maximum suppression (NMS) method to multiply the classification information of each grid by the confidence of the bounding box prediction to obtain the score of each bounding box. Then sort all the bounding boxes in the network according to the score, take the predicted box with the highest score as the final bounding box, and eliminate the redundant predicted boxes on the same target to obtain the final detection result.

[0142] In order to test whether the Yolov3 network model after transfer learning can better identify the key points of T-shirts, 164 pictures in the T-shirt key point test set are input into the trained network model to detect the key points of T-shirts. The results are shown in Table 4. The rectangular area in the figure is the area where the key points of the output T-shirt are located. Each rectangular box represents the area where a key point is located. The center point of the rectangular box is the key point. A total of 13 clothing key points are generated. The key point name and the confidence of the prediction box are also displayed next to each key point.

[0143] Table 4 Key point detection results

[0144]

[0145] Note: Missed recognition rate: number of images without key points detected / total number of images; Single point accuracy: number of valid key points detected / total number of key points; Recognition accuracy: number of images with each key point detected / total number

[0146] Among the 164 test images, the number of samples with 13 key point labels detected is nearly 82.3%, and the number of samples with 7 to 12 key points detected accounts for 16.4% of the total sample size. After training on the T-shirt key point dataset, the model has a good effect on T-shirt key point recognition.

[0147] (7) Since the T-shirt uses the neckline width as the benchmark size for the mapping relationship, measure and inspect the neckline width of the T-shirt for input into the benchmark. Enter the neckline width size in the size display module interface, and the system will automatically display the sizes of other parts.

[0148] The following describes the implementation effect of the system in combination with the application of the technical solution in the T-shirt clothing size measurement test process.

[0149] In order to test the implementation effect of the present invention, 4 T-shirts were prepared for testing.

[0150] Iron the T-shirt before taking the photo to ensure that the surface of the T-shirt is flat. Keep the shooting angle parallel to the T-shirt; ensure sufficient light, but not overexposure; the shooting background should be a solid color different from the color of the T-shirt.

[0151] Before the automatic size measurement experiment, four T-shirts were manually measured to obtain the size information of the tested garments in order to compare the system and analyze the accuracy of the T-shirt acquisition system size measurement.

[0152] The present invention uses the neckline width as a mapping relationship and inputs it into the automatic T-shirt size measurement system. The following is an example of the measurement results of four T-shirts. The data order in the following table is as follows: Table 5 is the measurement data of T-shirt No. 1, Table 6 is the measurement data of T-shirt No. 2, Table 7 is the measurement data of T-shirt No. 3, and Table 8 is the measurement data of T-shirt No. 4.

[0153] Table 5

[0154]

[0155] Table 6

[0156]

[0157]

[0158] Table 7

[0159]

[0160] Table 8

[0161]

[0162] The key dimension deviations of length, chest circumference, sleeve length and collar size required for each grade of T-shirt quality are shown in Table 9:

[0163] Table 9

[0164]

[0165] According to the comparison of the sizes of various parts, the absolute error of the T-shirt size measured by the automatic T-shirt size measurement system is generally between 5-15mm, and the relative error is between 0.50%-4.00%, which meets the measurement error range given by the national T-shirt specification size deviation comparison table. It shows that the key point leakage recognition rate of the T-shirt detected by the model of the present invention is 0, the single point correct rate is 90.5%, and the recognition accuracy rate is 82.3%, which shows that the Yolov3 target detection model designed by the present invention has undergone transfer learning, and the model has high generalization ability and good key point detection effect.

[0166] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with the technical field within the technical scope disclosed by the present invention and within the spirit and principle of the present invention should be covered by the protection scope of the present invention.

Claims

1. A T-shirt garment measurement system based on convolutional neural network, characterized in that: include: The training image import module is used to import clothing pictures of a specific category to be trained; the image data set is divided into three parts: training, verification, and testing. The training set is used to train the model, the verification set is used to observe the model performance, and the test set is used to verify the model effect; The training image preprocessing module is connected to the training image importing module and is used to mark the key points corresponding to the specific test items on the training image using an image calibration tool according to the specific items of the clothing size test; the positions of the key points are determined by the specific items of the clothing size measurement; The key point recognition rule extraction module is connected to the training image preprocessing module and is used to recognize the key points of T-shirts in the images imported by the user; establish the key point recognition rules of specific categories of clothing based on the training images imported by the user; The measurement image import module is used for users to import the T-shirt images to be tested; A data input module, used for the user to input the actual measurement data of the neckline width; The key size acquisition module is connected with the data input module and the key point recognition rule extraction module, and is used to realize the automatic measurement of the sizes of other parts of the T-shirt; A size display module, connected to the key size acquisition module, is used to display the automatically measured T-shirt size to the user; The key point recognition rule extraction module uses the DarkNet-53 convolutional neural network model to extract the key features of the T-shirt, and then combines the Yolov3 target detection and K-means clustering algorithm to filter out effective key point information; Yolov3's object detection algorithm detects the key points of the T-shirt, and DarkNet-53 uses the up-and-down sampling method to perform feature fusion to obtain the feature layer of the T-shirt, which is divided into three scale feature layers, respectively used for objects at different scales; The K-means clustering algorithm calculates the distance from the sample point to the cluster center point and uses the improved intersection-union ratio for clustering; The key size acquisition module combines the actual measurement data of the neckline width and realizes the automatic measurement of other parts of the T-shirt by mapping the pixel distance of the key points with the real size. Model training applies the Yolov3 network model to the T-shirt key point dataset through transfer learning, extracts key point features through the backbone network DarkNet-53 convolutional neural network, obtains key point coordinates through the k-means clustering algorithm, and improves the key point recognition accuracy through repeated iterative training.

2. The T-shirt garment measurement system based on convolutional neural network according to claim 1, characterized in that: Use Python software to design the system GUI interface and establish an acquisition system including image display, key point acquisition, and size display.

3. A T-shirt garment measuring method based on a convolutional neural network applied to the T-shirt garment measuring system based on a convolutional neural network as claimed in any one of claims 1 to 2, characterized in that: The T-shirt garment measurement method based on convolutional neural network includes: (1) The user imports the T-shirt image to be tested through the measurement image import module; (2) The DarkNet-53 convolutional neural network model is used to extract the key features of the T-shirt, and then the Yolov3 target detection and K-means clustering algorithm are combined to screen out the effective key point information. Combined with the actual measurement data of the neckline width input by the user, the pixel distance of the key point is mapped to the ratio of the actual size to achieve automatic measurement of the size of other parts of the T-shirt; (3) The size display module shows the user the sizes of other parts of the T-shirt to be tested.

Images