A method and system for displaying clothing details

Abstract

The present invention belongs to the technical field of computer vision recognition, and specifically relates to a method and system for displaying clothing details. The method includes obtaining a to-be-detected picture and a display type, detecting the to-be-detected picture through a pre-trained type detection network to obtain a clothing picture of the details to be displayed; inputting the clothing picture into a pre-trained attribute recognition network, and performing recognition in combination with the display type to output corresponding clothing attribute values; inputting the clothing picture into a pre-trained anchor point prediction network, generating a clothing anchor point map in combination with the display type, and outputting a corresponding detail picture by cropping the clothing anchor point map; matching the clothing attribute values of the detail picture, obtaining a compliant display picture from the detail picture according to the display type, and displaying the display picture while displaying the matching clothing attribute values. By using a machine to recognize attributes and cut detail pictures, the efficiency is high, and through picture-text matching, the clothing is presented comprehensively and meticulously.

Description

Technical Field

[0001] The present invention belongs to the technical field of computer vision recognition, and particularly relates to a method and system for displaying clothing details. Background Art

[0002] Currently, selling clothing online has become a very important sales channel for merchants. Even for specific merchants, it has become the most important sales channel. Therefore, merchants are very eager to clearly display various details of clothing to consumers, such as collar types, sleeve types, front plackets, trouser types, skirt types, etc.

[0003] The existing manual photographed detail pictures or cut detail pictures are not only time-consuming but also laborious, and then they multiply the annotation work of clothing attributes. Even when using deep learning technology to identify clothing attributes, since the accuracy rate cannot reach 100%, manual review is ultimately required, which additionally increases the workload of reviewers. Summary of the Invention

[0004] The technical problem to be solved by the present invention is how to use deep learning technology to crop detail pictures of clothing and identify the attributes corresponding to the detail pictures to display the details of clothing.

[0005] To this end, according to the first aspect, an embodiment of the present invention discloses a method for displaying clothing details, including:

[0006] Obtain a picture to be detected and a display type, and detect the picture to be detected through a pre-trained type detection network to obtain a clothing picture of details to be displayed;

[0007] Input the clothing picture into a pre-trained attribute recognition network, and perform recognition in combination with the display type to output corresponding clothing attribute values;

[0008] Input the clothing picture into a pre-trained anchor point prediction network, generate a clothing anchor point map in combination with the display type, and crop the clothing anchor point map to output a corresponding detail picture;

[0009] Match the clothing attribute values of the detail picture, obtain a conforming display picture from the detail picture according to the display type, and display the display picture while showing the matched clothing attribute values.

[0010] Optionally, the training of the type detection network specifically includes: constructing a network mobileNet V3+SSD that has been trained on the Imagenet dataset, and configuring a first optimizer; inputting pre-obtained type annotation samples into the network mobileNet V3+SSD for iterative operation, optimizing the accuracy of type detection, calculating a first error function, and when the first error function converges, the training ends, obtaining a qualified type detection network and saving the first network parameters.

[0011] Optionally, the training of the attribute recognition network specifically includes: constructing the network InceptionV4 that has been trained on the Imagenet dataset, modifying the network into a multi-task output network and then configuring a second optimizer; inputting the pre-acquired attribute value annotation samples into the network InceptionV4 for iterative operations, optimizing the accuracy of the attribute output, calculating a second error function, and when the second error function converges, the training ends, obtaining a qualified attribute recognition network and saving the second network parameters.

[0012] Optionally, the training of the anchor prediction network includes: constructing the network inception-resnet-v2 that has been trained on the Imagenet dataset, configuring a third optimizer; inputting the pre-acquired anchor label samples into the network inception-resnet-v2 for iterative operations, optimizing the accuracy of the anchor prediction, calculating a third error function, and when the third error function converges, the training ends, obtaining a qualified anchor prediction network and saving the third network parameters.

[0013] Optionally, detecting the to-be-detected picture through a pre-trained type detection network to obtain a clothing picture of the details to be displayed specifically includes: constructing the type detection network according to the first network parameters stored during the training of the type detection network; preprocessing the to-be-detected picture and inputting it into the type detection network for detection, and outputting a clothing picture.

[0014] Optionally, preprocessing the to-be-detected picture and inputting it into the type detection network for detection to output a clothing picture and a display type specifically includes: performing convolution, pooling, normalization, and screening on the preprocessed to-be-detected picture through the type detection network to output multiple target bounding boxes in the to-be-detected picture; using the non-maximum suppression algorithm to eliminate similar bounding boxes of the same target to obtain the optimal target bounding box; cropping the to-be-detected picture according to the optimal target bounding box to obtain a clothing picture of the clothing details to be displayed.

[0015] Optionally, the attribute recognition network includes a collar type recognition network, a length recognition network, a sleeve-pant type recognition network, and a sleeve-front opening recognition network.

[0016] Optionally, inputting the clothing picture into the pre-trained attribute recognition network to output the corresponding clothing attribute value specifically includes: constructing the attribute recognition network according to the second network parameters stored during the training of the type detection network; calling the corresponding attribute recognition network according to the display type to perform convolution, pooling, dropout, and softmax operations on the clothing picture, and outputting the clothing attribute with the highest matching value.

[0017] Optionally, inputting the clothing picture into a pre-trained anchor prediction network to output a corresponding detail picture specifically includes: constructing the anchor prediction network according to the third network parameters stored when pre-training the type detection network; preprocessing the clothing picture to input it into the anchor prediction network, generating a clothing anchor map in combination with the display type, and outputting a corresponding detail picture by cropping the clothing anchor map.

[0018] According to a second aspect, an embodiment of the present invention provides a clothing detail display method and system, including:

[0019] A category detection module, configured to obtain a picture to be detected and a display type, and detect the picture to be detected through a pre-trained type detection network to obtain a clothing picture of details to be displayed;

[0020] An attribute recognition module, configured to input the clothing picture into a pre-trained attribute recognition network, and perform recognition in combination with the display type to output a corresponding clothing attribute value;

[0021] An anchor prediction module, configured to input the clothing picture into a pre-trained anchor prediction network, generate a clothing anchor map in combination with the display type, and output a corresponding detail picture by cropping the clothing anchor map;

[0022] A detail matching and display module, configured to match the clothing attribute value of the detail picture, obtain a conforming display picture from the detail pictures according to the display type, and display the display picture while displaying the matching clothing attribute value.

[0023] The beneficial effects of the present invention are as follows:

[0024] A clothing detail display method and system disclosed in an embodiment of the present invention are based on deep learning technology. By obtaining a picture to be detected and a display type, detecting the picture to be detected through a pre-trained type detection network to obtain a clothing picture of details to be displayed; inputting the clothing picture into a pre-trained attribute recognition network, performing recognition in combination with the display type to output a corresponding clothing attribute value; inputting the clothing picture into a pre-trained anchor prediction network, generating a clothing anchor map in combination with the display type, and outputting a corresponding detail picture by cropping the clothing anchor map; matching the clothing attribute value of the detail picture, obtaining a conforming display picture from the detail pictures according to the display type, and displaying the display picture while displaying the matching clothing attribute value. With this method, attributes can be recognized by a machine and detail pictures can be cut, with high efficiency. Moreover, as the business volume increases, the time cost and capital investment hardly change. The detail pictures of the clothing are described from both the text and picture aspects, presenting the clothing more comprehensively and meticulously. Description of the Drawings

[0025] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings required for the description of the specific embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those skilled in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0026] Figure 1 It is a flowchart of the clothing detail display method according to an embodiment of the present invention;

[0027] Figure 2 It is a reference anchor point map for drawing anchor point maps in the clothing detail display method according to an embodiment of the present invention;

[0028] Figure 3 It is the correspondence between the detail types and anchor point identifiers according to an embodiment of the present invention;

[0029] Figure 4 It is an effect diagram of clothing detail display according to an embodiment of the present invention;

[0030] Figure 5 It is a schematic diagram of the clothing detail display system according to an embodiment of the present invention. Specific Embodiments

[0031] The following will clearly and completely describe the technical solutions of the present invention in conjunction with the drawings. Obviously, the described embodiments are some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0032] The embodiments of the present invention also provide corresponding systems, which will be described in detail below respectively.

[0033] Figure 1 It shows a flowchart of the clothing detail display method provided by an embodiment of the present invention.

[0034] Refer to Figure 1 , the clothing detail display method includes:

[0035] S10. Obtain the image to be detected and the display type, and detect the image to be detected through a pre-trained type detection network to obtain a clothing image of the details to be displayed.

[0036] In this embodiment, the types of clothing are divided into 13 categories, namely short-sleeved, long-sleeved, short-sleeved coat, long-sleeved coat, vest, suspender, shorts, trousers, short skirt, short-sleeved dress, long-sleeved dress, vest dress, and suspender dress. Generally, the pictures for detection in the input may contain multiple detectable types. For example, if the picture is of a model, he may be wearing a short-sleeved top and trousers. Therefore, to make the detection more accurate and targeted, while obtaining the picture to be detected, the types to be displayed will also be obtained; and through a pre-trained type detection network, the clothing pictures matching the displayed types will be obtained.

[0037] Before using the type detection network to detect the picture to be detected, training is required. The training of the type detection network includes:

[0038] Construct a network mobileNet V3+SSD that has been trained on the Imagenet dataset and configure the first optimizer;

[0039] Input the pre-obtained type annotation samples into the network mobileNet V3+SSD for iterative calculation to optimize the accuracy of type detection, calculate the first error function. When the first error function converges, the training ends, and a qualified type detection network is obtained and the first network parameters are saved.

[0040] The type annotation samples are real data in the clothing industry obtained through web crawlers and have been verified by manual annotation. The annotation result format is: (x,y,w,h), where x and y are the upper left vertices of the smallest bounding rectangle of the clothing category in the picture, and w and h are the width and height of the smallest bounding rectangle. The network mainly includes: convolutional layer, pooling layer, dropout layer, and softmax layer. The optimizer is preferably a stochastic gradient descent optimizer with momentum. During training, the focus is on optimizing the network parameters, and the network parameters include weight parameters and offsets. The first error function is preferably a categorical cross-entropy function. During training, when the categorical cross-entropy function converges, the training is completed, and the weight parameters and offsets are saved.

[0041] Specifically, detecting the picture to be detected through the pre-trained type detection network to obtain the clothing pictures to be displayed in detail specifically includes:

[0042] Construct the type detection network according to the first network parameters stored during the training of the type detection network;

[0043] Preprocess the picture to be detected and input it into the type detection network for detection, and output clothing pictures.

[0044] In this embodiment, the acquired image to be detected is stored on the hard disk, and the relative path of the image is generated according to the current time and the base64 encoding of the image's url value. Since the category detection network limits the size of the image to be detected input to the network, before inputting it into the category detection network, the image is first processed to a size that meets the input requirements, specifically 640*640. Then, through the constructed type detection network, the pre-processed image to be detected is convolved, pooled, normalized, and screened by the type detection network to output multiple target bounding boxes in the image to be detected; the non-maximum suppression algorithm is used to eliminate similar bounding boxes of the same target to obtain the optimal target bounding box; the image to be detected is cropped according to the optimal target bounding box to obtain a clothing image of the clothing details to be displayed.

[0045] S20. Input the clothing image into a pre-trained attribute recognition network, and perform recognition in combination with the display type to output corresponding clothing attribute values.

[0046] Specifically, the training of the attribute recognition network specifically includes:

[0047] Construct the network InceptionV4 that has been trained on the Imagenet dataset, modify the network into a multi-task output network, and configure the second optimizer;

[0048] Input the pre-acquired attribute value annotation samples into the network InceptionV4 for iterative operations, optimize the accuracy of attribute output, calculate the second error function, and when the second error function converges, the training ends, and a qualified attribute recognition network is obtained and the second network parameters are saved.

[0049] The attribute value annotation samples are real data in the clothing industry obtained through web crawlers, specifically from real user uploads and promotional pictures on e-commerce websites. The attribute labels are based on the preliminary marking on e-commerce websites and the approval and modification by annotation engineers. The annotation result format is: attribute value, such as wide-leg pants in the trouser type, leg-of-mutton sleeves in the sleeve type, etc. The last average pooling layer, dropout layer, and fully connected layer of inceptionV4 are removed, and a randomly initialized fully connected layer is added. According to different types, there are 4 recognition networks in total, namely: collar type recognition network, length recognition network, sleeve type - trouser type recognition network, sleeve type - front opening of clothes recognition network. The second error function is the categorical cross-entropy function. The second optimizer is preferably the adam (adaptive moment estimation) optimizer in the early stage and preferably multiple stochastic gradient descent optimizers with gradually decreasing learning rates in the later stage. When the categorical cross-entropy function converges, the training is completed, and the second network parameters, that is, the second weight parameters and the second offsets, are saved.

[0050] Based on the obtained display type, select four pre-trained attribute recognition networks, and the corresponding relationships are as follows: for short-sleeved, long-sleeved, short-sleeved coats, long-sleeved coats, short-sleeved dresses, long-sleeved dresses, spaghetti strap dresses, and vest dresses, call the collar type recognition network; for short-sleeved, long-sleeved, short-sleeved coats, long-sleeved coats, short-sleeved dresses, long-sleeved dresses, spaghetti strap dresses, and vest dresses, call the sleeve type - front opening recognition network; for shorts, long pants, short skirts, short-sleeved dresses, long-sleeved dresses, spaghetti strap dresses, and vest dresses, call the pant type - skirt type recognition network. It should be noted that since any category has a length, the length recognition network needs to be called for any category. That is, the attribute recognition networks used for each type are not single, and even some categories will call all the attribute prediction networks to process clothing pictures. For example, various dresses have attributes such as collar type, sleeve type, front opening, and skirt type.

[0051] Specifically, the step of inputting the clothing picture into the pre-trained attribute recognition network to output the corresponding clothing attribute value specifically includes:

[0052] Construct the attribute recognition network according to the second network parameters stored when training the type detection network;

[0053] Call the corresponding attribute recognition network according to the display type to perform convolution, pooling, dropout, and softmax operations on the clothing picture, and output the clothing attribute with the highest matching value. For example, the pant type values output by the pant type - skirt type network are harem pants, tight pants, wide-leg pants, flared pants, straight-leg pants, and tapered pants; and the skirt type values are A-line skirt, lantern skirt, mermaid skirt, bias skirt, narrow skirt, and straight skirt. If the picture input into the category recognition network is a picture of a pair of pants, the system will select the value with the highest possibility for each of the pant type and the skirt type and output it.

[0054] S30: Input the clothing picture into the pre-trained anchor point prediction network, generate a clothing anchor point map in combination with the display type, and output the corresponding detail map by cropping the clothing anchor point map.

[0055] Specifically, the training of the anchor point prediction network includes:

[0056] Construct the network inception-resnet-v2 that has been trained on the Imagenet dataset, and configure the third optimizer;

[0057] Input the pre-obtained anchor point label samples into the network InceptionV4 for iterative operation, optimize the accuracy of attribute output, calculate the third error function, and when the third error function converges, the training ends, and a qualified anchor point prediction network is obtained and the third network parameters are saved.

[0058] The anchor label samples are obtained from the processing of an open-source dataset (DeepFashion2). The processing method is as follows: Use the key point coordinates in DeepFashion2 to generate a Gaussian distribution heatmap with a size of 56*56. The training error is the L2 value of the difference between the anchor map after the image is processed by the convolutional network and the heatmap with the same size as the anchor map generated from the key point data. The training uses the Adam (Adaptive Moment Estimation) optimizer, and the third error function is the categorical cross-entropy function. When the third error function converges, the training is completed, and the third network parameters, that is, the corresponding third weight parameters and third offsets, are saved.

[0059] Specifically, inputting the clothing image into the pre-trained anchor prediction network to output the corresponding detail map specifically includes:

[0060] Construct the anchor prediction network according to the third network parameters stored when pre-training the type detection network;

[0061] Preprocess the clothing image to input it into the anchor prediction network, generate a clothing anchor map in combination with the display type, and output the corresponding detail map by cropping the clothing anchor map.

[0062] Refer to Figure 2 , which is the reference benchmark anchor map called in the embodiment of the present invention and is all the anchor maps of 13 types. As can be seen from the figure, there are a total of 294 anchors in the anchor maps of 13 types of clothing. For the convenience of later matching types for drawing the anchor map, the 294 anchors are divided into different intervals. In this embodiment, the corresponding relationship between the type and the interval subscript is as follows: short-sleeved tops 0 to 24, long-sleeved tops 25 to 57, short-sleeved coats 58 to 88, long-sleeved coats 89 to 127, vests 128 to 142, suspenders 143 to 157, shorts 158 to 167, long pants 168 to 181, short skirts 182 to 189, short-sleeved dresses 190 to 218, long-sleeved dresses 219 to 255, vest dresses 256 to 274 (including the left and right boundary points). In this way, through the type, it can be obtained which interval the anchor belongs to, and then the clothing anchor map can be drawn according to the reference anchor map of the corresponding type. During the process of drawing the anchor map, the anchor identifiers of the detail maps of the same type are the same. For example, the anchor identifiers of the collar types are all 1, 2, 3, 4, 5, 6; the anchor marks of the waistbands of the pants are all 1, 2, 3.

[0063] Refer to Figure 3, then crop the detail images based on the anchor identifiers in the clothing anchors. Taking a short-sleeved top as an example: the anchor identifiers corresponding to the shoulder detail image of the short-sleeved top are 2, 7, and 8. Therefore, among the 25 anchor identifiers of the short-sleeved top returned, select the anchor identifiers 2, 7, and 8 as the anchor identifiers for the shoulders; the anchor identifiers corresponding to the neckline detail image are 1, 2, 3, 4, 5, and 6; the anchor identifiers corresponding to the cuff detail image are 9 and 10; the anchor identifiers corresponding to the waist detail image are 12, 13, 18, and 19; the anchor identifiers corresponding to the hem detail image are 15 and 16.

[0064] Crop the classified clothing anchor images. Select the minimum value x_min and the maximum value x_max of the anchor abscissa (x) from the anchor table identifiers corresponding to each detail in the clothing anchor image, and select the minimum value y_min and the maximum value y_max of the anchor ordinate (y). Use x_min and y_min as the upper left vertex of the detail image, and x_max and y_max as the lower right vertex of the detail image to perform cropping and output the corresponding detail image.

[0065] S40. Match the clothing attribute values of the detail image, obtain the conforming display images from the detail image according to the display type, and display the display images while showing the matched clothing attribute values. There is a corresponding relationship between the detail image and the clothing attributes. When the detail image is obtained, through the corresponding relationship between the two, perform matching. For the matched detail image, through the display image with attribute values required by the display type, the corresponding clothing attribute values are synchronously displayed during the display of the image, so as to display the clothing details from both the text and the image aspects. Specifically, refer to Figure 4 .

[0066] The present invention obtains the image to be detected and the display type, detects the image to be detected through a pre-trained type detection network to obtain a clothing image of the details to be displayed; inputs the clothing image into a pre-trained attribute recognition network, combines the display type for recognition to output the corresponding clothing attribute values; inputs the clothing image into a pre-trained anchor prediction network, combines the display type to generate a clothing anchor image, and outputs the corresponding detail image by cropping the clothing anchor image; matches the clothing attribute values of the detail image, obtains the conforming display images from the detail image according to the display type, and displays the display images while showing the matched clothing attribute values. Compared with the prior art, it has the following advantages:

[0067] 1. Through the type detection network and the input display type, a determined target clothing image is obtained, reducing the noise interference in the later detail detection;

[0068] 2. The method of multi-threaded data reading and distributed training, such as model pre-training, is adopted to complete the network tasks with a large amount of data (300,000 images), many iteration times, and many network parameters on limited machines within a short time (3 weeks).

[0069] 3. Visualize the training process to intuitively and quickly judge whether the model is running in the direction of convergence: Visualize the feature map of the model and the heatmap of the label coordinates to intuitively see whether the model is moving in a good direction.

[0070] 4. The currently popular heatmap in the industry is adopted, which greatly improves the accuracy of key points. Before predicting the key points, a category detection network is used to crop the clothing category image from the original image and input it into the attribute recognition network, reducing noise interference and improving the accuracy of key point prediction.

[0071] It should be understood that in the above embodiments, the magnitudes of the sequence numbers of the steps do not mean the order of execution. The order of execution of each step should be determined by its function and internal logic, and should not constitute any limitation to the implementation process of the embodiments of the present invention.

[0072] Figure 5 The schematic diagram of the clothing detail display system provided by the embodiments of the present invention is shown. For the sake of convenience of description, only the parts related to the embodiments of the present invention are shown.

[0073] In the embodiments of the present invention, the system is used to implement the above Figure 1 The clothing detail display method described in the embodiments can be a software unit, a hardware unit, or a unit combining software and hardware built into a computer or a server.

[0074] Refer to Figure 3 , the system includes:

[0075] The category detection module 10 is used to obtain the image to be detected and the display type, and detect the image to be detected through a pre-trained type detection network to obtain the clothing image of the details to be displayed;

[0076] The attribute recognition module 20 is used to input the clothing image into a pre-trained attribute recognition network, and perform recognition in combination with the display type to output the corresponding clothing attribute value;

[0077] The anchor point prediction module 30 is used to input the clothing image into a pre-trained anchor point prediction network, generate a clothing anchor point map in combination with the display type, and crop the clothing anchor point map to output the corresponding detail map;

[0078] The detail matching display module 40 is used to match the clothing attribute values of the detailed images, obtain the conforming display images from the detailed images according to the display type, and display the display images while showing the matched clothing attribute values.

[0079] It should be noted that the device in the embodiments of the present invention can be used to implement all the technical solutions in the above method embodiments. The functions of its respective functional modules can be specifically implemented according to the methods in the above method embodiments. The specific implementation process can refer to the relevant descriptions in the above examples and will not be elaborated here.

[0080] The above is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of changes or substitutions, which should all be covered by the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.

Claims

1. A method for displaying clothing details, characterized in that, Including: Obtain the picture to be detected and the display type, and detect the picture to be detected through a pre-trained type detection network to obtain a clothing picture of the details to be displayed; Input the clothing picture into a pre-trained attribute recognition network, and perform recognition in combination with the display type to output the corresponding clothing attribute value; Input the clothing picture into a pre-trained anchor prediction network, generate a clothing anchor map in combination with the display type, and output the corresponding detail map by cropping the clothing anchor map; match the clothing attribute value of the detail map, obtain the compliant display picture from the detail map according to the display type, and display the display picture while showing the matched clothing attribute value; The training of the anchor prediction network includes: Construct the network inception-resnet-v2 that has been trained on the Imagenet dataset, and configure the third optimizer; Input the pre-obtained anchor label samples into the network inception-resnet-v2 for iterative operations, optimize the accuracy of anchor prediction, calculate the third error function, and when the third error function converges, the training ends, obtain a qualified anchor prediction network and save the third network parameters; Inputting the clothing picture into the pre-trained anchor prediction network to output the corresponding detail map specifically includes: Construct the anchor prediction network according to the third network parameters stored when training the type detection network; Preprocess the clothing picture to input it into the anchor prediction network, generate a clothing anchor map in combination with the display type, and output the corresponding detail map by cropping the clothing anchor map.

2. The method for displaying clothing details according to claim 1, characterized in that, The training of the type detection network specifically includes: Construct the network mobileNet V3+SSD that has been trained on the Imagenet dataset, and configure the first optimizer; Input the pre-obtained type annotation samples into the network mobileNet V3+SSD for iterative operations, optimize the accuracy of type detection, calculate the first error function, and when the first error function converges, the training ends, obtain a qualified type detection network and save the first network parameters.

3. The method for displaying clothing details according to claim 1, characterized in that, The training of the attribute recognition network specifically includes: Construct the network InceptionV4 that has been trained on the Imagenet dataset, modify the network into a multi-task output network and then configure the second optimizer; Input the pre-obtained attribute value annotation samples into the network InceptionV4 for iterative operations, optimize the accuracy of attribute output, calculate the second error function, and when the second error function converges, the training ends, obtain a qualified attribute recognition network and save the second network parameters.

4. The method for displaying clothing details according to claim 2, characterized in that, Detecting the picture to be detected through a pre-trained type detection network to obtain a clothing picture of the details to be displayed specifically includes: Construct the type detection network according to the first network parameters stored when training the type detection network; Preprocess the picture to be detected to input it into the type detection network for detection, and output a clothing picture.

5. The method for displaying clothing details according to claim 4, characterized in that, Preprocess the picture to be detected to input it into the type detection network for detection to output a clothing picture and a display type specifically includes: The pre - processed image to be detected is subjected to convolution, pooling, normalization, and screening through a type - detection network, and multiple target bounding boxes in the image to be detected are output; The non - maximum suppression algorithm is used to eliminate similar bounding boxes of the same target, and the optimal target bounding box is obtained; the image to be detected is cropped according to the optimal target bounding box to obtain a clothing image of the clothing details to be displayed.

6. The method for displaying clothing details according to claim 3, characterized in that, The attribute recognition network includes a collar - type recognition network, a length recognition network, a sleeve - type / pant - type recognition network, and a sleeve - type / clothing - front - opening recognition network.

7. The method for displaying clothing details according to claim 6, characterized in that, Inputting the clothing image into a pre - trained attribute recognition network to output corresponding clothing attribute values specifically includes: Constructing the attribute recognition network according to the second network parameters stored when training the type - detection network; According to the display type, calling the corresponding attribute recognition network to perform convolution, pooling, dropout, and softmax operations on the clothing image, and outputting the clothing attribute with the highest matching value.

8. A system for the method of displaying clothing details, characterized in that, Including: A category - detection module, which is used to obtain the image to be detected and the display type, and detect the image to be detected through a pre - trained type - detection network to obtain a clothing image of the details to be displayed; An attribute - recognition module, which is used to input the clothing image into a pre - trained attribute recognition network, and perform recognition in combination with the display type to output corresponding clothing attribute values; An anchor - point prediction module, which is used to input the clothing image into a pre - trained anchor - point prediction network, generate a clothing anchor - point map in combination with the display type, and output a corresponding detail map by cropping the clothing anchor - point map; A detail - matching display module, which is used to match the clothing attribute values of the detail map, obtain a conforming display image from the detail map according to the display type, and display the display image while showing the matching clothing attribute values; The training of the anchor - point prediction network includes: Constructing the network inception - resnet - v2 that has been trained on the Imagenet dataset, and configuring the third optimizer; Inputting the pre - obtained anchor - point label samples into the network inception - resnet - v2 for iterative operations, optimizing the accuracy of anchor - point prediction, calculating the third error function, and when the third error function converges, the training ends, obtaining a qualified anchor - point prediction network and saving the third network parameters; Inputting the clothing image into a pre - trained anchor - point prediction network to output a corresponding detail map specifically includes: Constructing the anchor - point prediction network according to the third network parameters stored when pre - training the type - detection network; Pre - processing the clothing image to input it into the anchor - point prediction network, generating a clothing anchor - point map in combination with the display type, and outputting a corresponding detail map by cropping the clothing anchor - point map.

Images