KR20200084422A - Service providing apparatus and method for cleansing fashion related product information using neural network - Google Patents

Abstract

The present invention relates to a service providing device and a service providing method for cleansing fashion product-related information using a neural network. According to the present invention, a plurality of different neural networks learn category-related information for classification between a product and a product image of the fashion-related product and product property-related property information. Accordingly, support is provided so that cleansed product information can be generated that is capable of enhancing reliability during big data analysis using fashion-related product information. During the learning of the basic information by image and category and the information by property of the product related to the fashion-related product, even a user without fashion knowledge can easily grasp and specify product properties by keywords related to the product properties being provided for the user. As a result, accurate property-related information can be learned regarding the product. Categorization and automatic standardization with respect to a product image without product property-related information are performed based on the neural networks and cleansed product information is generated and provided. Accordingly, automatic product information generation is supported and product management can be facilitated. In addition, the cleansed product information is provided such that a highly accurate and reliable result can be provided during big data analysis based on the product information.

Description

Service providing apparatus and method for cleansing fashion related product information using neural network}

The present invention relates to a service providing apparatus and method for refining fashion product-related information using a neural network, and more specifically, related to category-related information and product attribute for product image and product classification of fashion-related products in a plurality of different neural networks. It relates to a service providing apparatus and method for purifying fashion product-related information using a neural network that supports to generate refined product information that can increase reliability when analyzing big data using fashion-related product information by learning attribute information.

Currently, along with the development of information and communication technology, the market of online markets such as online shopping malls is rapidly developing, and as a result, sellers are increasingly selling products through the online market.

In order to sell these products online, sellers create product information that includes various data such as product images related to the products to be sold, types of products to be sold, and targets of products to be sold, and register them in the online market. In addition, this product information is growing exponentially.

Accordingly, there are increasing attempts to utilize it for various purposes, such as grasping trends and targeting advertisements, through analyzing big data using a large number of product information forming big data.

In particular, the online market market related to fashion is sensitive to fashion, and changes in products that are demanded by the public are changing from time to time. In order to stand out in the fashion field, the use of such big data is more important than anything else. Compared to the field, since numerous products are continuously released in a short period of time, the amount of information for big data analysis is also satisfied.

However, in order for this product information to have a useful value as an analysis data, in addition to information on a category to which the product belongs, detailed information such as a material or a design of the product is required, but these product-related features are included in the product information and input. At the same time, users who input features according to their own subjective judgment for the same product or who have insufficient knowledge related to fashion may have an incorrect input when entering product information. There is a problem included in product information, and when this is used for big data analysis, there is a problem in that the reliability of the analysis result is poor.

Korean Patent Publication No. 10-2006-0056614

The present invention supports a user through a plurality of neural networks in relation to a fashion-related product, so that a user can generate product information in which attribute of a product is accurately reflected without misinputting attribute information related to a product attribute in addition to category-related information for product classification. , The purpose of this is to support the automatic generation of refined product information having a utilization value as information for big data analysis through a plurality of neural networks learned based on such product information.

In a method of providing a service for refining fashion product-related information using a neural network according to an embodiment of the present invention, the control unit is a low-level learning unit composed of a plurality of neural networks, and a product-related image region specified in a product image for each product and a plurality of preset categories. And a first learning step of matching and learning a plurality of different basic information related to each product, and in advance, in the order in which the number of overlapping times is high, based on a plurality of product-related information generated and pre-stored when the product is sold for each product. A set number of different product-related keywords are extracted and displayed through a user interface unit that receives user input, and the keywords selected for a plurality of preset attributes according to user input through the user interface unit are configured as a neural network and the plurality of attributes And a second learning step of learning a high-level learning unit corresponding to the selected keyword together with the image area matching the selected keyword among a plurality of high-level learning units corresponding to each, and the control unit is configured to receive a specific product image of a specific product. By applying the image area among the plurality of category-specific basic information and image areas related to the specific product obtained by applying to the low-level learning unit, the property information including attribute-specific data corresponding to the specific product is applied to the plurality of high-level learning units. The calculating step may include a refining step of generating and storing product information including the basic information and attribute information for each category generated through the output step in response to the specific product.

As an example related to the present invention, the product related information may include at least one of advertisement information of a product, product introduction information of a product, web page information of a product, and comment information of a product.

As an example related to the present invention, the plurality of categories may include a gender corresponding to a product, a body part where the product is worn, and a type of product.

As an example related to the present invention, the control unit displays the product image displaying the image area calculated through the first learning step in relation to the specific product through the user interface unit, and the control unit controls the user interface unit When receiving correction information according to the modification of the image area, the image area is updated from the specific product image to an area corresponding to the modification information, and the basic information for each of the categories corresponding to the specific product is updated. It may be characterized in that it further comprises the step of learning to match the image region and the low-level learning unit.

As an example related to the present invention, the first learning step includes a plurality of low-level learning units through the low-level learning unit so that the control unit identifies an image region of a learning target product matching the basic information for each of the plurality of categories in an image where one or more products are displayed. Further comprising the step of learning to a specific neural network of the neural network, the second learning step, the control unit from the product image to the image area from the product image through the image processing unit based on the learning target product-related image area obtained through the specific neural network. The method may further include generating a segmented image related to the product to be learned by clipping a corresponding region and then training the plurality of high-level learning units as the image region.

As an example related to the present invention, the refining step may further include the step of the controller including and storing at least one of a specific product image, an image area, and a divided image related to the specific product in the product information. have.

As an example related to the present invention, the control unit learns at least one of a brand and a manufacturing time corresponding to the product information by matching the product information to a design pattern learning unit composed of a neural network to learn the design pattern and manufacturing time for each brand. A third learning step that allows at least one of the design patterns to be learned, and input information received by the control unit through the user interface unit for at least one of a plurality of categories, a plurality of attributes, a brand, and a manufacturing time or the brand and a manufacturing time Applying the product information including at least one to the design pattern learning unit to provide a matching degree with the learned design pattern in the product information or error information related to the input error or the input information according to the matching degree It may be characterized in that it further comprises the step of providing a.

As an example related to the present invention, the plurality of attributes may include at least one of a material of a product, a design style of a product, a wearing season of a product, and a color of a product.

A service providing apparatus for refining fashion product-related information using a neural network according to an embodiment of the present invention is composed of a plurality of neural networks, and a product-related image region specified in a product image for each product and a plurality of preset categories, respectively. A low-level learning unit configured to match a plurality of related basic information to learn the plurality of neural networks, and the image area configured to consist of the plurality of neural networks and a separate first neural network to match selected keywords according to user input. A plurality of high-level learning units and low-level learning units for learning the first neural network are controlled, and a preset number of duplicates is sequentially set based on a plurality of product-related information generated and sold when each product is sold. After extracting different product-related keywords and displaying them through a user interface unit that receives the user input, a plurality of high-levels corresponding to the plurality of attributes and the keywords selected according to the user input through the user interface unit for a plurality of preset attributes. The plurality of categories related to the specific product obtained by learning the high-level learning unit corresponding to the selected keyword together with the image area matching the selected keyword among the learning units, and applying a specific product image of a specific product to the low-level learning unit By applying the basic information and the image area to each of the plurality of high-level learning unit to calculate the attribute information including the attribute-specific data corresponding to the specific product, the product including the attribute information and the plurality of category-specific basic information It may include a control unit for generating and storing information corresponding to the specific product.

The present invention provides keywords related to product attributes to users even if they have no knowledge of fashion in the process of learning the information related to fashion-related products and basic information for each category and various attributes of the product on multiple neural networks. By presenting and supporting the user to easily identify and specify the properties of the product, accurate property-related information about the product can be learned, and based on these multiple neural networks, there is no category classification and information related to the property of the product. By automatically generating and providing standardized and refined product information for product images, it supports the automatic creation of product information for products, thereby increasing the convenience of product management and providing refined product information, and based on the product information. It has the effect of supporting to provide results with high accuracy and reliability when analyzing big data.

In addition, the present invention enables the user to generate product information in which the attribute of the product is accurately reflected without incorrect input of attribute information related to the product, in addition to category-related information for product classification through a plurality of neural networks in relation to the fashion-related product. There is an effect of supporting the automatic generation of refined product information having a utilization value as information for big data analysis through a number of neural networks learned based on the product information.

In addition, the present invention supports the reliability and accuracy of product information generated in connection with fashion products by helping users who have no expertise in the fashion field to identify and correct product information generated in relation to products and matching degree between products. It has the effect of increasing.

1 is a configuration environment diagram of a service providing device for purifying information related to fashion products using a neural network according to an embodiment of the present invention.
2 and 3 are detailed configuration diagrams and operation examples of a service providing apparatus for refining fashion product-related information using a neural network according to an embodiment of the present invention.
FIG. 4 is an exemplary diagram related to modification of an image area, basic information, and attribute information of a service providing device for refining fashion product-related information using a neural network according to an embodiment of the present invention.
5 is an exemplary diagram illustrating an operation related to a design pattern analysis of a service providing device for refining information related to fashion products using a neural network according to an embodiment of the present invention.
6 is a flowchart of a service providing method for refining information related to fashion products using a neural network according to an embodiment of the present invention.

Hereinafter, exemplary embodiments of the present invention will be described with reference to the drawings.

1 is a configuration environment diagram of a service providing device (hereinafter referred to as a service providing device) 100 for purifying fashion product-related information using a neural network according to an embodiment of the present invention, and the service providing device 100 as illustrated It may be configured to include the holding DB 200, it may be connected to the user interface unit 10 for receiving user input and displaying various information.

In this case, the service providing apparatus 100 may be configured to include the user interface unit 10.

In addition, the service providing device 100 may be connected to the user interface unit 10 through a communication network to communicate with each other or through a communication line to communicate with each other.

At this time, the communication network may include a wired/wireless communication network, and examples of such a wireless communication network include a wireless LAN (WLAN), a Digital Living Network Alliance (DLNA), a wireless broadband (Wibro), and a WiMAX (World Interoperability for). Microwave Access: Wimax), Global System for Mobile Communication (GSM), Code Division Multi Access (CDMA), Code Division Multi Access 2000 (CDMA2000), Enhanced Voice-Data Optimized or Enhanced Voice-Data Only (EV-DO), WCDMA (Wideband CDMA), HSDPA (High Speed Downlink Packet Access), HSUPA (High Speed Uplink Packet Access), IEEE 802.16, Long Term Evolution (LTE), Long Term Evolution-Advanced (LTE-A), Broadband Wireless Mobile Mobile Broadband Service (WMBS), 5G mobile service, Bluetooth, LoRa (Long Range), Radio Frequency Identification (RFID), Infrared Data Association (IrDA), Ultra Wideband (UWB) , ZigBee, Near Field Communication (NFC), Ultra Sound Communication (USC), Visible Light Communication (VLC), Wi-Fi, Wi-Fi Direct (Wi -Fi Direct). In addition, wired communication networks include wired local area network (LAN), wired wide area network (WAN), power line communication (PLC), USB communication, Ethernet, serial communication, and optical/coaxial Cables and the like may be included.

In addition, the user interface unit 10 may be configured as a user terminal, the user terminal is a smart phone (Smart Phone), a portable terminal (Portable Terminal), a mobile terminal (Mobile Terminal), personal information terminal (Personal Digital Assistant) : PDA), PMP (Portable Multimedia Player) terminal, Telematics terminal, Navigation terminal, Personal computer, Notebook computer, Slate PC, Tablet PC, Ultrabook (ultrabook), wearable device (for example, a watch type terminal (smartwatch), glass type terminal (smart glass), head mounted display (HMD), etc.), Wibro terminal, IPTV (Internet Protocol Television) ) Terminal, Smart TV, Digital Broadcasting Terminal, Television, 3D Television, Home Theater System, AVN(Audio Video Navigation) Terminal, A/V(Audio/Video) System, Flexible Terminal It may include various terminals such as.

In addition, the user interface unit 10 includes a display module for displaying various information (for example, information received from the service providing device 100), an input module for receiving user input, and the service providing device 100 It may include a communication module for communication with, and a control module for performing an overall control function of the user interface unit 10 (for example, control of a display module, an input module, and a communication module), and the corresponding control module It can be configured to include a CPU, MCU (Micro Controller Unit). At this time, when the user interface unit 10 is configured to be included in the service providing device 100, the control module of the user interface unit 10 may be omitted, and a control unit configured in the service providing device 100 ( 160), overall control of the user interface unit 10 may be performed.

Meanwhile, various information related to a plurality of different fashion-related products may be previously stored in the DB 200, for example, product-related product images, product advertisement information, product product introduction information, Various information such as product web page information and product comment information may be stored.

At this time, the fashion described in the present invention may mean various products worn on the user's body, such as clothing and accessories, but is not limited thereto, and may mean various products related to fashion.

On the other hand, the neural network (or neural network model) described in the present invention may be composed of a deep learning algorithm consisting of an input layer, one or more hidden layers, and an output layer. Various types of neural networks such as Deep Neural Network (DNN), Recurrent Neural Network (RNN), Convolutional Neural Network (CNN), and Support Vector Machine (SVM) may be applied to the neural network model.

In this case, when the neural network is composed of CNNs, one or more convolutional layers, a consolidation layer, and a fully connected layer may be configured.

On the basis of the above-described configuration, the present invention is a standardization (standardization) as a product category in the process of learning the information related to the product to a plurality of neural networks so that the fashion-related products can be automatically classified and the product information is generated through a plurality of neural networks. ) In addition to the information, the definition of each product can be easily determined by the user for information related to various attributes of products that are difficult to formalize (standardize) and refine because many definitions are different for each seller, such as characteristics and characteristics of products. Product information that includes structured and refined information about product attributes as well as product image and category-related information by applying product-related images to multiple neural networks that have been learned and supported based on this. It can be supported to generate, it will be described in detail with reference to the drawings below.

2 and 3 are detailed configuration diagrams and operation examples of the service providing apparatus 100 according to an embodiment of the present invention.

As shown, the service providing apparatus 100 includes an image processing unit 110, a low-level learning unit 120, a reference information providing unit 130, and a plurality of different high-level learning units 140 and designs. It may be configured to include a pattern learning unit 150 and the control unit 160.

At this time, the image processing unit 110, the low-level learning unit 120, the reference information providing unit 130, and a plurality of different high-level learning units configured in the service providing apparatus 100 excluding the control unit 160 ( 140), and at least one of the design pattern learning unit 150 may be configured to be included in the control unit 160.

In addition, the control unit 160 executes the overall control function of the service providing apparatus 100 using pre-stored programs and data. The control unit 160 may include a RAM, ROM, CPU, GPU, bus, and RAM, ROM, CPU, GPU, etc. may be connected to each other through a bus. The CPU may access a separate storage unit configured in the service providing device 100 to boot using an operating system (O/S) stored in the storage unit, and various programs, contents, and data stored in the storage unit Various operations may be performed using the like.

In addition, the service providing apparatus 100 may be connected to the retention DB 200 or may be configured to include the retention DB 200, and the retention DB 200 may also be configured as the storage unit.

At this time, the holding DB 200 is a product DB storing a plurality of product images for each product related to fashion, and a plurality of preset basic information for product classification for each of the plurality of products identifies products. The basic information DB 220 stored in matching with information, and the product information refined by the sales-related DB 230 and the control unit 160 in which various product related information related to products are stored for each of the plurality of products are stored. Product DB 240 may be included.

Here, the basic information and the product image may include product identification information.

In addition, the plurality of categories are for classifying a plurality of products into a plurality of different product groups, and may include various categories such as a gender corresponding to a product, a body part where the product is worn, and a type of product.

In addition, the basic information includes male or female related information corresponding to the gender-related category, information such as top and bottom corresponding to the body part related category, and coat corresponding to the category related to the product type , Blouses, shirts, etc. may be included.

In addition, the product-related information may include information generated at the time of selling the product, such as product advertisement information, product introduction information, product web page information, product comment information, and the like.

First, the low-level learning unit 120 may be composed of a plurality of neural networks or may include the plurality of neural networks, and among the plurality of products, a learning target product-related image area and a preset product-specific image area for each learning target product are preset. After extracting from the image DB 210 and the basic information DB 220, a plurality of different basic information related to a learning target product corresponding to a plurality of categories and matching with each other can be learned in the plurality of neural networks.

At this time, the control unit 160 is a low-level learning unit 120 composed of a plurality of neural networks, a learning target product-related image area specified in a product image for each learning target product, and a plurality of preset product-related plurality respectively corresponding to a predetermined category. You can learn by matching different basic information of.

To this end, the control unit 160, when extracting a product image for learning from the image DB 210 to the low-level learning unit 120, is matched to the product image based on product identification information included in the product image Basic information for each category of can be extracted from the basic information DB 220, through which a product image extracted from the image DB 210 for a product to be learned and a plurality of categories extracted from the basic information DB 220 By matching the basic information with each other, the plurality of neural networks may be trained through the low-level learning unit 120.

In addition, the control unit 160 in order to learn a specific neural network of any one of a plurality of neural networks configured in the low-level learning unit 120, the product-related image area in which the product to be learned is displayed in the product image in which one or more products are displayed. The product image may be displayed on the user interface unit 10, and a selected region is set as an image region according to a user input through the user interface unit 10 from the product image, and the low-level learning unit 120 is set. Through this, the segmented image corresponding to the image area in the product image can be learned in the specific neural network.

At this time, the controller 160 may provide a product image and an image area of the product to be learned to the image processing unit 110 when setting an image area in the product image for the product to be learned according to the user input, and the image The processor 110 generates a split image in which the image of the product to be learned occupies most of the area by clipping (or cutting or cropping) the image area from the product image, and then the low-level learning unit 120 It can be transmitted to the low-level learning unit 120 along with the basic information for each category of the learning target product in relation to the learning target product.

That is, the control unit 160 generates an image region to be learned by the low-level learning unit 120 as a divided image through the image processing unit 110 and converts the divided image into an image area through the low-level learning unit 120. The plurality of neural networks can be trained.

In addition, when the number of product-related image targets related to a learning target product including the image region related to the product image learned by the low-level learning unit 120 and the plurality of category-specific basic information is greater than or equal to a preset reference value, the controller 160 When only a specific product image of a specific product is input to the low-level learning unit 120, the image area corresponding to the specific product image and basic information for each category corresponding to the specific product image are output through the low-level learning unit 120 ( Output).

At this time, as shown in Figure 4, the control unit 160 corresponds to the specific product image, the image area output through a plurality of neural networks configured in the low-level learning unit 120 and the basic information for each category of the user It can be provided to the interface unit 10 to be displayed.

In addition, the control module of the user interface unit 10 may display the image area and basic information for each category through a display module, and display the image area in the state where the image area is set in the specific product image or the The divided image corresponding to the image area may be displayed.

In addition, the control module resets the image area in the specific product image based on user input through the user interface unit 10 when an image area is incorrectly set in the specific product image or an error occurs in the basic information for each category. Or, it is possible to provide the correction information generated by correcting the basic information in which the error occurred to the control unit 160 of the service providing apparatus 100.

Through this, the control unit 160 updates the image area from the specific product image to a region corresponding to the correction information when the correction information is received, and updates the basic information for each of the categories corresponding to the specific product. Matched with the image area, it is possible to learn the plurality of neural networks through the low-level learning unit 120.

In addition, when the correction information is an error correction of the basic information, the control unit 160 matches the image area corresponding to the specific product image and the basic information for each category in which the error is corrected, through the low-level learning unit 120. The plurality of neural networks can be trained.

Through this, the control unit 160 may increase the accuracy of information output from a plurality of neural networks constituting the low-level learning unit 120 or the accuracy of the low-level learning unit 120.

Meanwhile, each of the plurality of high-level learning units 140 may be configured as a first neural network that is a separate neural network separated from the plurality of neural networks, or may include the first neural network.

In addition, the reference information providing unit 130 extracts a plurality of pre-stored product-related information generated when a product is sold in relation to a learning target product from the sales-related DB 230, and based on the plurality of product-related information In the order in which the number of overlapping times is high, a preset number of keywords related to different learning target products may be extracted and provided to the controller 160.

At this time, the product-related information may include product identification information of the product to be learned.

Accordingly, as illustrated in FIG. 4, the control unit 160 may cause the predetermined number of different keywords related to the learning target product to be displayed through the user interface unit 10 in relation to the learning target product. , Based on user input through the user interface unit 10, keyword information including keywords selected by a user for each of a plurality of preset attributes may be generated.

At this time, the plurality of attributes may include properties or characteristics of the product to be studied, such as the material of the product, the design style of the product, the season of wearing the product, and the color of the product.

In addition, the control unit 160 may generate keyword information generated in relation to the learning target product and an image region learned in the low level learning unit 120 in connection with the learning target product through the plurality of high level learning units 140. The first neural network configured in each of the plurality of high-level learning units 140 may be trained.

That is, the plurality of high-level learning units 140 are respectively configured in the service providing apparatus 100 in correspondence with the plurality of attributes, and the control unit 160 respectively controls the image area of the plurality of high-level learning units 140. When learning each of the first neural networks of the keyword, a keyword corresponding to any one attribute among a plurality of attributes included in keyword information matching the image region may be matched with the image region to be learned.

For example, the controller 160 may include a plurality of high-level learning units 140 when a specific keyword selected by a user included in keyword information generated in association with a product to be learned is an attribute value of a first attribute among the plurality of attributes. Among them, the first high-level learning unit corresponding to the corresponding first attribute may be trained by matching the attribute value with the image region of the learning target product.

In addition, if the other keyword selected by the user included in the keyword information generated in association with the product to be learned is the attribute value of the second attribute among the plurality of attributes, the control unit 160 of the plurality of high-level learning units 140 The second high-level learning unit corresponding to the second attribute may be trained by matching the attribute value of the second attribute with the image region of the learning target product.

Through this, each of the plurality of high-level learning units 140 can train a correlation between the image region and the attribute corresponding to itself among the plurality of attributes to the first neural network, and accordingly, the plurality of high-level learning units ( 140) The correlations learned in each may be different from each other.

At this time, the control unit 160 corresponds to the image area learned by the low-level learning unit 120, and the divided image extracted from the product image through the image processing unit 110 as the image area is the plurality of high-level learning units ( 140) Each can be learned.

Through this, the control unit 160 may learn the keyword information and the image area for each of a plurality of different learning target products in the first neural network of the high-level learning unit 140, thereby distinguishing products with similar attributes. The first neural network of the high-level learning unit 140 may be trained to be made.

That is, the control unit 160, in conjunction with the reference information providing unit 130, the seller related to the product to be learned has a specific characteristic and property-related keyword, and consumers who purchase the product to be studied have the same comments as the product. By synthesizing keywords generated from information and presenting keywords with high frequency of overlapping to the product to be studied, users can provide common attributes that the public thinks about the product to be learned, and through this, The subjective judgment of the user who inputs the attribute does not interfere with the attribute-related information, and objective information that matches the public's (including the seller's) thoughts about the product to be learned is input as the attribute-related information, thereby Accurate and objective attribute-related information may be learned by the plurality of high-level learning units 140.

Through such an operation configuration, the present invention can secure the accuracy of product information generated through a plurality of neural networks configured in the service providing apparatus 100 in relation to a specific product to be described later, and unnecessary information is included in the product information Can be prevented.

According to the learning of the low-level learning unit 120 and the high-level learning unit 140 as described above, the controller 160 provides standardized product information for analyzing big data on a specific product image of a specific product. It can be, it will be described in detail with reference to Figures 2 and 3.

First, the controller 160 may apply a specific product image of a specific product to the low-level learning unit 120, and the low-level learning unit 120 may apply the specific product to the specific product image through a plurality of neural networks. An image area for each of one or more products in the image may be identified, and basic information for each category highly related to the image area of the specific product may be generated and output (calculated).

In one example, the low-level learning unit 120 applies the specific product image to a plurality of neural networks when receiving a specific product image in which a specific product related to a male shirt is displayed by the control unit 160, and through the plurality of neural networks, The image area corresponding to the specific product related to the men's shirt is identified from a specific product image, and a'male' is displayed as basic information as a gender-related category corresponding to the corresponding image area related to the male shirt, and the category corresponding to the body part-related category. The'top' may be output as basic information, and the'shirt' may be output as basic information corresponding to the category related to the product type.

In this case, when a plurality of products are included in the specific product image, the low-level learning unit 120 may generate and output image area and category basic information for each product included in the specific product image. .

In addition, the control unit 160 provides an image region and a specific product image obtained corresponding to the specific product through the low-level learning unit 120 to the image processing unit 110, and the identification from the image processing unit 110 A segmented image generated by clipping the image area in a product image may be generated.

In the above-described configuration, the control unit 160 may display the product image in which the image area calculated through the low-level learning unit 120 in relation to the specific product is displayed through the user interface unit 10, When receiving the correction information according to the modification of the image area through the user interface unit 10, the image area is updated from the product image to an area corresponding to the modification information, and the plurality of areas corresponding to the specific product are updated. By matching the basic information for each category with the updated image area, the low-level learning unit 120 may be trained.

In addition, the controller 160 applies the image area related to the specific product to the first neural network of each of the plurality of high-level learning units 140 through the plurality of high-level learning units 140 or the plurality of high-level learning units. (140).

In this case, the controller 160 may apply the specific product-related segmented image obtained through the image processing unit 110 as the specific product-related image area to the first neural network for each of the plurality of high-level learning units 140.

In addition, when receiving the correction information for the image area, the controller 160 applies an updated image area corresponding to the correction information to the first neural network of each of the plurality of high-level learning units 140 or the updated image area. The split image corresponding to may be applied to the first neural network of each of the plurality of high-level learning units 140.

In addition, each of the plurality of high-level learning units 140 may output (calculate) data (or parameters) of a specific attribute having high relevance to the image region according to a correlation previously learned in its first neural network.

In this case, data of the specific attribute calculated by the high-level learning unit 140 may be a keyword.

For example, the high-level learning unit 140 corresponding to a material-related attribute among a plurality of high-level learning units 140 applies the image area of the specific product to its first neural network to analyze the specific product through analysis of the first neural network. A specific material (eg, silk) that is highly related to can be output as data of material-related properties.

In addition, among the plurality of high-level learning units 140, the high-level learning unit 140 corresponding to the design style related attribute applies the image area of the specific product to its first neural network to analyze the specific product through analysis of the first neural network. A design style (eg, a wave pattern) that is highly related to can be output as data of properties related to the design style.

Accordingly, the control unit 160 may apply the image area of the specific product to the plurality of high-level learning units 140 to obtain data for each of the plurality of attributes from the plurality of high-level learning units 140, and Attribute information including a plurality of attribute-specific data may be generated.

That is, the controller 160 may apply the image area of the specific product to the plurality of high-level learning units 140 to provide a prediction result for the attribute of the specific product.

In addition, the controller 160 may generate product information including basic information for each category and the attribute information obtained through the low-level learning unit 120 in correspondence with a specific product image of the specific product.

That is, the control unit 160 of the low-level learning unit 120 and the high-level learning unit 140 only by recognizing an image of a specific product through the low-level learning unit 120 and the high-level learning unit 140 composed of a neural network. Through neural network analysis, product information including attribute-related information including characteristics and properties of a specific product as well as category-related information for classification of a specific product can be generated. Refined product information can be generated.

In addition, the control unit 160 may store the product information in the product DB 240 when the product information is generated, thereby providing standardized and refined product information, thereby providing high accuracy through corresponding product information when analyzing big data. We can help you get results.

In addition, the control unit 160 may include at least one of a specific product image, an image area, and a divided image related to the specific product in the product information and store it in the product DB 240, through which the category included in the product information Images of specific products corresponding to the related basic information and attribute information may also be included in the product information in a refined state.

In addition, when the product information is automatically generated corresponding to the specific product, the control unit 160 may display the corresponding product information through the user interface unit 10, in accordance with the user input through the user interface unit 10 When correction input information for product information is received, the product information may be updated based on the correction input information and stored in the product DB 240.

Through the above-described configuration, the present invention relates to a fashion-related product, even if the user has no knowledge of fashion in the process of learning images and category-specific basic information and information on various attributes of the product on multiple neural networks. By presenting keywords related to the user, the user can easily identify and specify the properties of the product so that accurate property-related information about the product can be learned. By automatically generating and providing standardized and refined product information for product images with no attribute-related information, it supports automatic generation of product information for products, thereby enhancing convenience in product management and refined product information as described above. By providing, it is possible to assist in providing a result with high accuracy and reliability when analyzing big data based on the product information.

Meanwhile, as illustrated in FIGS. 2 and 5, the control unit 160 corresponds to the product information stored in the product DB 240, such as a brand and a manufacturing time (eg, manufacturing year, manufacturing year, month) , Manufacturing date, etc.) may be learned by matching the product information to a design pattern learning unit 150 composed of a plurality of neural networks and a second neural network that is a separate neural network separated from the first neural network.

At this time, the control unit 160 receives the brand and manufacturing time related information corresponding to the product information based on user input through the user interface unit 10, and receives the brand and manufacturing time related information in the product information. It can be included to be trained in the second neural network of the design pattern learning unit 150 or the design pattern learning unit 150.

Alternatively, the controller 160 matches the product identification information included in a specific product image of a specific product corresponding to the product information, and provides the reference information to the brand and manufacturing time information stored in the sales-related DB 230 Received from (130), corresponding to the specific product, the reference information providing unit 130 includes the brand and manufacturing time extracted from the sales-related DB (230) in the product information or matches the product information By doing so, the second neural network of the design pattern learning unit 150 may be trained.

In addition, the control unit 160 may learn the product pattern, brand, and manufacturing time generated for each of a plurality of products in the design pattern learning unit 150.

Accordingly, the design pattern learning unit 150 may learn the plurality of product information, the brand, and the manufacturing time in the second neural network so that at least one of the design pattern for each brand and the design pattern for each manufacturing time is learned.

At this time, the control unit 160, if the design pattern learning unit 150 includes at least one of the product image, the image area and the divided image included in the product information, except for this, the design pattern learning unit 150 Can be learned.

That is, the design pattern learning unit 150 in which a plurality of product information is learned along with the brand and the manufacturing time may learn the design pattern for each brand and each manufacturing time.

At this time, the design pattern learned in the second neural network of the design pattern learning unit 150 may mean a range value of the plurality of attributes for a product group having a specific value for a plurality of categories at a specific manufacturing time of a specific brand, , It may mean the design type of each product group produced by the brand by manufacturing time.

Accordingly, the control unit 160 may input information received for at least one of a plurality of categories, a plurality of attributes, a brand, and a manufacturing time through the user interface unit 10 in relation to a product to be analyzed or the brand and the manufacturing. The product information including at least one of the periods is applied to the design pattern learning unit 150 in which the plurality of product information is learned to provide a matching degree between the product to be analyzed and the design pattern learned in the product information. Can be.

At this time, the design pattern learning unit 150 applies the product information of the product to be analyzed to the second neural network to design the design pattern of the product to be analyzed according to criteria related to the classification of the design pattern learned in the second neural network. Can be classified (classified) and defined (or identified).

That is, the controller 160 directly inputs at least one of the plurality of categories, a plurality of attributes, a brand, and a manufacturing time in relation to an analysis target product to analyze a design through the design pattern learning unit 150 A product in which an analysis target product corresponding to the input information existed based on the matching degree obtained through the second neural network of the design pattern learning unit 150 by applying one input information to the design pattern learning unit 150 It is possible to provide the user with a degree of difference from the design pattern.

For example, if the matching degree is high, the user can recognize that the design of the product to be analyzed is a safe product with a design similar to the existing products. If the matching degree is low, the unique design pattern does not exist in the existing products. By having it, the user can grasp it.

In addition, if the product to be analyzed is specified by the brand and manufacturing time, the matching degree can be used to identify the degree of difficulty (degree of pandemic) or originality of the product to be analyzed in relation to a specific brand at a specific manufacturing time.

Through this, the present invention can assist the designer in making a fashion-related product by assisting the designer in determining the originality or fashionability of the product when the designer makes the product.

In addition, the control unit 160 corresponds to a specific product image of a specific product and adds brand and manufacturing time related data to product information generated through the low-level learning unit 120 and the high-level learning unit 140, and then the design pattern A design pattern produced at a time adjacent to the manufacturing time of the corresponding product information in a brand corresponding to the product information through the design pattern learning unit 150 (or the second neural network) by applying it to the second neural network of the learning unit 150 Matching degree can be provided, and through this, the user can determine that there is an incorrect input in the product information when the matching degree is low.

That is, according to the present invention, a matching degree between a design pattern according to product information and a design pattern learned by the design pattern learning unit 150 through a matching degree provided by the user through the design pattern learning unit 150 corresponding to the product information In the low case, the brand corresponding to the product information can be assisted to understand that a design pattern based on the information entered in the product information has not been produced at the time of manufacture corresponding to the product information. Accordingly, an error occurs in the product information. It can be identified and corrected to support more accurate product information.

Meanwhile, the control unit 160 may provide a matching degree corresponding to each of the plurality of categories and a plurality of attributes through the design pattern learning unit 150 in response to the product information, and the plurality of categories and the plurality of If there is an item whose matching degree is lower than a preset reference value among a plurality of items including the attribute of, a corresponding item is designated as an error item, and it is determined that an input error for the product information has occurred in the error item and related to the error item. After generating error information and providing the error information to the user interface unit 10, the error information may be displayed through the user interface unit 10 to assist the user in confirming an incorrect entry.

In addition, when the controller 160 receives correction request information according to user input from the user interface unit 10 with respect to the mistyped item in relation to the product information, the product information is based on the modification request information. Or modify the data according to the modification request information in relation to the mistyped item of the product information and store it in the product DB 240.

Through the above-described configuration, the present invention provides product information when a design pattern of product information generated through a plurality of neural networks based on a product image of a specific product is largely different from an existing design pattern for a product group (product group) similar to product information. It is possible to inform the user that an incorrect input has occurred, and through this, the user can modify the product information to create highly reliable product information, and accordingly, reliability of the analysis result when analyzing big data using the product information. To improve.

According to the above-described configuration, the present invention provides product information in which the attribute of the product is accurately reflected without incorrect input of attribute information related to the product in addition to category-related information for the user to classify the product through multiple neural networks in relation to the fashion-related product. It is possible to support creation, and to automatically generate refined product information having utilization value as information for big data analysis through multiple neural networks learned based on such product information.

6 is a flowchart of a service providing method for refining information related to fashion products using a neural network of the service providing apparatus 100 according to an embodiment of the present invention.

As illustrated, the control unit 160 is associated with a product corresponding to a product-related image region specified in a product image and a plurality of preset categories, respectively, in a product image (by product for learning) in the low-level learning unit 120 composed of a plurality of neural networks. A plurality of different basic information may be matched and learned (S1).

In addition, the controller 160 extracts a predetermined number of different product-related keywords in order of a large number of duplicates based on a plurality of pre-stored product-related information generated when goods are sold for each product, and receives a user input. A plurality of high-level learning units displayed through the unit 10 and configured with a neural network for the keywords selected for a plurality of preset attributes according to user input through the user interface unit 10 and corresponding to the plurality of attributes, respectively 140), the high-level learning unit 140 corresponding to the keyword may be trained together with the image area matching the keyword (S2).

In addition, the controller 160 learns the image area from among the plurality of category-specific basic information and image areas related to the specific product obtained by applying a specific product image of a specific product to the low-level learning unit 120. It can be applied to the unit 140 to calculate attribute information including attribute-specific data corresponding to the specific product (S3, S4).

In this case, the controller 160 specifies the specific product when the number (or summation capacity) of all the learning-related product-related information learned by the low-level learning unit 120 and the high-level learning unit 140 is greater than or equal to a preset reference value. Corresponding to a product image, it is possible to obtain basic information and attribute information for each image region and category through the low-level learning unit 120 and the high-level learning unit 140.

In addition, the control unit 160 may generate and store product information including basic information and attribute information for each category generated in correspondence with the specific product (S5).

As described above, according to the present invention, basic information held in relation to a fashion-related product is learned in a low-level neural network to enable product classification using only images, and a user can easily identify and specify characteristic information related to a specific product. Support, and by linking these feature information with basic information and product images to learn in a high-level neural network, support not only information for classifying products by product images but also product information including feature-related feature information of products. It can support product information that can be used for big data analysis.

The various devices and components described herein can be implemented by hardware circuitry (eg, CMOS-based logic circuitry), firmware, software, or a combination thereof. For example, it may be implemented using transistors, logic gates, and electronic circuits in the form of various electrical structures.

The above-described contents may be modified and modified without departing from the essential characteristics of the present invention to those skilled in the art to which the present invention pertains. Therefore, the embodiments disclosed in the present invention are not intended to limit the technical spirit of the present invention, but to explain, and the scope of the technical spirit of the present invention is not limited by these embodiments. The scope of protection of the present invention should be interpreted by the claims below, and all technical spirits within the scope equivalent thereto should be interpreted as being included in the scope of the present invention.

10: user interface unit 100: service providing device
110: image processing unit 120: low-level learning unit
130: reference information providing unit 140: high-level learning unit
150: design pattern learning unit 160: control unit
200: Reserved DB

Claims (9)

A first learning step in which the control unit matches and learns a product-related image region specified in a product image for each product and a plurality of preset categories related to a plurality of different basic information related to each product corresponding to a low-level learning unit composed of a plurality of neural networks;
The control unit extracts a preset number of different product-related keywords in order of a large number of duplicates based on a plurality of pre-stored product related information generated when goods are sold for each product, and displays them through a user interface unit receiving user input, The keyword selected by a plurality of preset attributes according to a user input through the user interface unit is composed of a neural network, and the selected image is selected together with the image area matched with the selected keyword among a plurality of high-level learning units corresponding to the plurality of attributes. A second learning step of learning the high level learning unit corresponding to the keyword;
The controller applies the image area among the plurality of category-specific basic information and image areas related to the specific product obtained by applying a specific product image of a specific product to the low-level learning unit, and applies the image area to the specific product. An output step of calculating attribute information including corresponding attribute-specific data; And
A refinement step in which the controller generates and stores product information including basic information and attribute information for each category generated through the output step in response to the specific product.
Method for providing a service for refining information related to fashion products using a neural network including a.
The method according to claim 1,
The product-related information includes at least one of product advertisement information, product product introduction information, product web page information, and product comment information, and a service providing method for purifying fashion product-related information using a neural network.
The method according to claim 1,
The plurality of categories includes a gender corresponding to the product, a body part where the product is worn, and a service providing method for refining fashion product-related information using a neural network, characterized in that it comprises a product type.
The method according to claim 1,
Displaying, by the control unit, the product image on which the image area calculated through the first learning step is displayed in relation to the specific product through the user interface unit; And
When the control unit receives correction information according to the modification of the image area through the user interface unit, the image area is updated from the specific product image to an area corresponding to the modification information, and the plurality of categories correspond to the specific product. And matching the basic information of each star with the updated image area to train the low-level learning unit.
The method according to claim 1,
In the first learning step, the controller learns a specific neural network from among a plurality of neural networks through the low-level learning unit so as to identify an image area of a product to be learned that matches the basic information for each of the categories in an image in which one or more products are displayed. Further comprising,
In the second learning step, the control unit clips an area corresponding to the image area from the product image through an image processing unit based on the image area related to the product to be learned obtained through the specific neural network, thereby dividing the image related to the product to be learned. After generating the, further comprising the step of learning as the image area in the plurality of high-level learning unit Service providing method for purifying information related to fashion products using a neural network.
The method according to claim 5,
In the refining step, the control unit further includes refining the fashion product-related information using a neural network, wherein the product information includes at least one of a specific product image, an image area, and a segmented image related to the specific product. How to provide services.
The method according to claim 1,
The control unit learns by matching and learning at least one of a brand and a manufacturing time corresponding to the product information by matching the product information with a design pattern learning unit composed of a neural network so that at least one of the design pattern by each brand and a design pattern by manufacturing time is learned. A third learning step; And
Learning the design pattern through the user interface through the user interface, input information received for at least one of a plurality of categories, a plurality of attributes, and a brand and a manufacturing time or the product information including at least one of the brand and a manufacturing time. A neural network characterized in that it further comprises the step of providing a matching degree with the learned design pattern to the product information by applying to the unit or providing error information related to an input error for the input information or product information according to the matching degree. A method of providing services for refining information related to used fashion products.
The method according to claim 1,
The plurality of attributes include at least one of a material of a product, a design style of a product, a wearing season of a product, and a color of a product. A service providing method for refining information related to fashion products using a neural network.
A low-level learning unit configured to consist of a plurality of neural networks and match a product-related image region specified in a product image for each product and a plurality of predetermined basic information related to a product corresponding to a plurality of preset categories, to train the plurality of neural networks;
A plurality of high-level learning units configured to consist of the plurality of neural networks and a separate first neural network to train the first neural network with the image regions matching selected keywords according to user input; And
Control the learning of the low-level learning unit, extract the preset number of different product-related keywords in the order of the number of duplicates based on a plurality of product-related information generated and pre-stored when the product is sold for each product, and receive the user input And the image area matched with the selected keyword among a plurality of high-level learning units respectively corresponding to the plurality of attributes by displaying the keyword selected according to a user input through the user interface unit for a plurality of preset attributes displayed through the user interface unit. Together, the high level learning unit corresponding to the selected keyword is trained, and the plurality of category-specific basic information and image regions related to the specific product obtained by applying a specific product image of a specific product to the low level learning unit is the plurality of high level learning units. A control unit for generating attribute information including the attribute-specific data corresponding to the specific product, and generating and storing product information including the attribute information and the plurality of category-specific basic information corresponding to the specific product
A service providing device for refining fashion product-related information using a neural network including a.

Images