A method for processing a commodity in a picture

Abstract

The invention discloses a processing method for commodities in images. By adopting machine vision technology based on deep learning, commodity objects in images can be identified, segmented, automatically classified, and the position information, range information and the like of commodities can be marked with highlighting. Commodities contained in any image are managed through commodity SKU objectification, and each identified and marked commodity SKU object can be further subjected to online ordering processing. The invention separates commodity objects from massive live pictures first, allowing procurement-side users to select commodities on a broad scale just like being personally on the spot in a physical spot warehouse, and provides an interactive transaction mode in which procurement-side users select commodities one by one for batch procurement. It truly reproduces the interactive experience online that users can enter a large centralized warehouse for wholesale procurement on site anywhere, and at the same time simplifies the workload of suppliers maintaining commodity SKUs and inventory quantities in the warehouse, and strives to achieve the use experience of wholesale ordering that users can get what they point to buy at large warehouse exhibition halls without going out.

Description

Technical Field

[0001] This invention relates to the field of e-commerce technology, and in particular to a method for processing products in images. Background Technology

[0002] In various e-commerce systems, such as Taobao, JD.com, Alibaba, and Dewu, whether targeting enterprise users or consumers, product images are typically processed one image per product for procurement. It's rare to find an image containing multiple products, each available for individual sale. This is because purchasing decisions for any given product generally require numerous images accompanied by extensive text descriptions, showcasing the product from visual and informational perspectives to gain the buyer's understanding and approval, thus securing a sale. Therefore, there is no industry demand or practical significance for simultaneously displaying multiple products within a single image and allowing each to be sold independently.

[0003] Furthermore, for any product listed and sold on an e-commerce system, it is necessary to catalog and manage the inventory in terms of Stock Keeping Unit (SKU) in advance. Therefore, even if multiple products can be listed and sold in one image, each product must have a pre-cataloged SKU for organization and standardization. Then, the inventory quantity of the product's SKU needs to be maintained before it can be used for sales in the e-commerce system. Therefore, when multiple products in a certain image need to be listed and sold, the common practice of e-commerce systems in various industries is to create an SKU for each product and provide an independent display image for that product. The operations staff maintains the inventory data of each product for display and sales in the e-commerce system.

[0004] The design principles of several generations of e-commerce systems, such as Taobao, Alibaba, Pinduoduo, and Dewu, have been regarded as industry common sense, influencing practitioners across various industries and corresponding technology suppliers. However, years of industry practice have proven that this design of e-commerce systems is ill-suited to the actual needs of the precious metals and jewelry wholesale industry, resulting in long-term stagnation in the digitalization process of the precious metals wholesale industry. The causes of this problem are explained from several perspectives below:

[0005] 1. Industry Status and Background

[0006] The wholesale of precious metal jewelry (materials such as gold, karat gold, and platinum) is a B2B transaction scenario involving thousands of suppliers and hundreds of thousands of buyers. On the supplier side, there are numerous jewelry processing factories (thousands of large-scale factories in China). These suppliers are relatively small in scale, and their level of informatization is limited by their enterprise size. Most of them do not have complete enterprise resource planning (ERP) and product data management (PDM) systems. Most of these companies cannot provide downstream wholesalers and retailers with complete, timely, and effective e-commerce system data packages (including product images, specifications, and current inventory).

[0007] Currently, most suppliers can provide an electronic order catalog or even an online ordering system, but they lack accurate inventory data (the reasons will be explained in point 2), and do not offer a commitment to the order production cycle. The supplier's sales channels either produce and deliver according to orders, or raise funds to produce a batch of best-selling spot goods based on the supplier's own market forecasts, and sell them to various wholesale markets for downstream buyers to see the physical samples and purchase in bulk.

[0008] Because downstream buyers are located throughout the country, regularly visiting jewelry wholesale markets to view and purchase physical goods from various jewelry showrooms is the most mainstream supply chain channel. Even though a few leading suppliers have invested in developing and operating complete online ordering systems, downstream buyers, constrained by various factors, still rely on offline showroom purchases as their primary procurement channel. Currently, bulk purchases of physical goods through offline physical viewing account for over 90% of the wholesale trade volume in the precious metals and jewelry industry.

[0009] 2. Inventory Settlement Method

[0010] The aforementioned e-commerce systems are all designed with a single item having a unique code. However, the precious metals and jewelry industry presents a unique challenge: wholesalers struggle to manage each item with a unique code at low cost. Daily inventory checks, inter-company transactions, and settlements within the industry all rely on weighted counting, failing to manage inventory and process orders precisely on a per-item basis. The main reasons for this are as follows:

[0011] A. The cost of managing each item with a unique code is too high. The gross profit margin in the precious metals and jewelry wholesale industry is low. After adopting this system, downstream buyers do not accept the premium caused by this cost.

[0012] B. The prices of precious metal materials fluctuate frequently and significantly. Companies in the industry must closely monitor their net holdings by weight every day. Each company must conduct daily inventory checks, clearing, and settlement of its holdings in order to control the operational risks of holding precious metal materials.

[0013] C. For enterprises of a certain size, the daily volume of goods received into the warehouse, categorized and displayed in the showroom, and sold out based on samples involves a large number of styles and items. Furthermore, the operations at both ends, such as buyers viewing and selecting samples and suppliers settling labor costs according to procurement requirements, as well as packaging and shipping, all rely on professional manual processing. Under the premise of ensuring that the weight of all goods in the inventory must be accurate, counting individual items is neither absolutely necessary nor conducive to increasing operating costs. This is a process step that has been optimized and eliminated after years of practice. Therefore, the settlement packing documents in the precious metal and jewelry wholesale market rarely have accurate item count records.

[0014] D. Wholesale businesses have high product turnover. In the showrooms of mainstream offline wholesale businesses, many new styles arrive and are put on display every day, while many existing styles are sold out. Some items are even snapped up before they even hit the shelves. Year after year, styles are updated and new styles emerge daily. Even if some companies spare no expense to build SKU databases for a considerable number of styles on sale, and create SKU files for all styles within a certain period or range, and manage inventory down to the individual item level, as older styles sell out and are discontinued, new styles constantly appear, and situations where they quickly sell out and there is no follow-up production frequently occur. Therefore, maintaining the accuracy of SKU listings and inventory quantities becomes a costly, low-profit, and inefficient operating method, which is generally unsustainable.

[0015] E. Precious metal jewelry has numerous subtle changes between old and new styles, and it is constantly iterating and upgrading at a very fast pace. A product may seem to be sold every year and every month, but different batches at different times will have various differences. Manually cataloging these styles into SKUs is difficult without the upstream supplier's PDM database being strictly provided to downstream buyers along with production and supply. It is difficult to classify and code these very similar styles in a long-term and rigorous manner. Imprecise classification will result in the creation of a large number of similar SKU categories, making it difficult to identify and enter new stock into the correct inventory, which objectively increases the management complexity of one item per code.

[0016] The above reasons summarize the current situation of the precious metals and jewelry industry and the constraints of digital operation. Objectively, it is difficult to fully promote and implement conventional e-commerce systems based on the one-item-one-code management system in this industry. Enterprises working in this field are all constrained by the above-mentioned rules and find it difficult to break through and develop. Summary of the Invention

[0017] To address the aforementioned technical problems, this invention provides a method for processing goods in an image.

[0018] To achieve the above objectives, the present invention provides the following technical solution:

[0019] A method for processing products in an image includes the following steps:

[0020] S1. A machine vision model trained with a large number of manually labeled product images is used to perform deep learning and machine vision recognition on the target (hereinafter referred to as: real-time) image (the present invention uses machine vision models such as R-CNN / YOLO);

[0021] S2. Based on the results of deep learning and machine vision recognition, the suspected goods identified (hereinafter referred to as: Live Goods, abbreviated as LG) are identified in terms of location and classification. All judgment results are output and a set of Live Goods Dataset (abbreviated as: LGD) is established. The Live Goods Dataset includes the product number, location, range, machine vision classification judgment, etc. The data in the Live Goods Dataset are all non-human processing results provided by machine vision.

[0022] S3. Based on the aggregated Live Goods dataset from S2, create uniquely coded Live Goods Records (LGRs) for all historical Live Goods. All Live Goods Records (LGRs) are aggregated into a Live Goods Database (LGDB). This database not only includes the non-human-processed information contained in the LGD, but also adds Homogenization Information (HI), which standardizes all Live Goods data records to ensure uniformity. HI includes background information for each Live Goods image, such as the batch, time, location, owner, and contact person of the item in the image, as well as individual information for each Live Goods item, such as condition, category, style, size, and specifications. The Live Goods Database (LGDB) is a fully redundant database that stores all relevant information about Live Goods automatically identified by machine vision and remains open and editable, allowing for additions and revisions at any time.

[0023] S4. When displaying the live image to the end user, the live image must be redrawn. A new layer of the same size as the live image, allowing for user interaction, must be added. For each live item (LG) in the live image, a highlighted polygonal outline is drawn according to its spatial location and scope. Each polygonal outline corresponds to a live item record (LGR), and the information for each LGR comes from the LGDB database created in step S3. These polygonal outlines corresponding to live items (LG) are further named Live Goods Objects (LGOs), allowing for user-friendly interface interactions including, but not limited to, adding to favorites, adding to cart, locking payment, and placing purchase orders. The Live Goods Object (LGO) is, to some extent, a "digital twin" of the physical goods within the live image.

[0024] S5. Establish a Live Goods Object Tracking and Draft (LGOTD) archive database for recording dialogue tracking and saving real-time operations for all end users' interface interaction operations on Live Goods Objects (LGOs). Record and save the actions of each end user on each Live Goods Object (LGO).

[0025] S6. When an end user performs operations related to generating a purchase order, such as adding to favorites, shopping cart, payment lock, or purchase order, a Live StockKeeping Unit (LSKU) is created. The data for the LSKU comes from the Live Stock Database (LGDB). The Live Stock Database is required to pre-store all the information needed to create the LSKU.

[0026] Since S6 creates some duplicate or similar LSKUs, which affects system efficiency, the system periodically (e.g., every time payment is locked, every time an order is accepted, daily, weekly, monthly, yearly, etc.) based on models using deep learning algorithms (GNN, K-means, DBSCAN, HDBSCAN, etc.) clusters, removes duplicates, and merges similar LSKUs.

[0027] Since S6 creates some LSKUs but these LSKUs do not ultimately generate orders, which affects system efficiency, the system regularly cleans up and deletes LSKUs that have expired and have no order submission operations.

[0028] Furthermore, in step S1, the live images are digital photos, video files, live streaming data, or 3D real-time modeling data.

[0029] Furthermore, in step S1, the information identified by the machine vision model includes, but is not limited to, category, quality, style, specifications, process, manufacturer, inlay, etc.

[0030] Furthermore, the purity levels include: 999, 3D, 5G, 18K, platinum, etc.

[0031] Furthermore, the categories include: pendants, rings, bracelets, earrings, bracelets and anklets, necklaces, beaded bracelets, necklace sets, etc.

[0032] Furthermore, in step S2, the product location information includes the relative position and boundary range information of the product in the image.

[0033] Furthermore, in step S4, the dynamic graphics are highlighted using a polygonal outline.

[0034] Furthermore, step S4 includes various interactive operation states for the Live Goods Object (LGO), such as: selectable state, selected state, currently selected and focused state, payment locked state, sold state, etc. These operation states can be distinguished by using different colored outlines. Each end user's operation on each Live Goods Object is recorded in the Live Goods Draft Archive Database (LGOTD) provided in S5.

[0035] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0036] Unlike common e-commerce systems that require users to first create a catalog of product SKUs before placing orders, this invention uses machine vision and artificial intelligence to automatically identify and classify product objects from a massive database of real-time images. This allows purchasing users to browse products as if they were physically in a warehouse, and provides an interactive transaction model where users can pick and choose items one by one for bulk purchasing. It realistically recreates the experience of entering a large, centralized warehouse for wholesale purchasing from anywhere online. At the same time, it simplifies the workload for suppliers in maintaining warehouse SKUs and inventory levels, striving to provide a user experience where they can buy whatever they want without leaving their home, just like at a large warehouse showroom or trade show.

[0037] The advantages of this invention are as follows:

[0038] 1. Reduce spatial distance: Buyers can view the latest or even immediate inventory and make purchases without having to visit the supplier's warehouse or showroom in person, saving travel costs.

[0039] 2. Time-saving: Because artificial intelligence machine vision automatically classifies products, such as the purity, category, style, specifications, and even which existing SKU the product is approximately matched with in the jewelry industry, all of these can be pre-classified and archived. Therefore, buyers can not only easily browse all the supplier's inventory like visiting a showroom or warehouse, but also filter by product category like using e-commerce platforms such as Taobao and JD.com, quickly locate the purchase target, and filter out unnecessary products. It is even easier to find specific products than to go to the store in person.

[0040] 3. Timeliness value: This kind of live product delivery can achieve the ultimate in timeliness with a live camera, or it can sacrifice some timeliness by updating daily or periodically, and shooting clear and exquisite photos of individual items in different areas.

[0041] 4. Significantly reduces the data operation and maintenance costs of e-commerce platforms. Suppliers do not need to spend a lot of manpower, capital, and technical operation and management experience to maintain an e-commerce backend system that is very similar to offline wholesale operations. This allows their customers to remotely purchase goods as if they were on-site. The biggest saving is that there is no need to maintain online inventory addition and subtraction data for each SKU, which is a major burden of management costs for operators of conventional e-commerce systems. Attached Figure Description

[0042] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0043] Figure 1 This is a flowchart illustrating the method for processing goods in images provided in an embodiment of the present invention.

[0044] Figure 2 This invention provides an embodiment of the method for performing deep learning and machine vision recognition on real-time images.

[0045] Figure 3 and Figure 4 This image, provided as an embodiment of the present invention, shows the collection of pseudo-SKU information for all products in the real-world image.

[0046] Figure 5 The images provided in this embodiment of the invention are archived images of all the real-world products in the figure.

[0047] Figure 6 Image showing the location information of a pendant provided in an embodiment of the present invention.

[0048] Figure 7Images of a real-time product operation interface provided for embodiments of the present invention.

[0049] Figure 8 The product image is shown in the product image provided in the embodiments of the present invention for editing and operation.

[0050] Figure 9 These are images provided in an embodiment of the present invention for the process of adding real-time products to favorites.

[0051] Figure 10 These are images showing the process of adding real-time products to a shopping cart, as provided in an embodiment of the present invention.

[0052] Figure 11 These are images showing the process of removing real-time items from a shopping cart, as provided in an embodiment of the present invention.

[0053] Figure 12 The S2 real-time commodity dataset LGD table style provided in the embodiments of the present invention.

[0054] Figure 13 The S2 real-time commodity dataset LGD table data is provided in the embodiments of the present invention.

[0055] Figure 14 The LGR table structure for the real-time commodity dataset provided in this embodiment of the invention.

[0056] Figure 15 The LGOTD table structure is provided for the real-time product draft archive in the embodiments of the present invention.

[0057] Figure 16 The actual minimum inventory unit (LSKU) table structure provided in this embodiment of the invention.

[0058] Figure 17 This is an LSKU aggregation judgment and analysis diagram provided in an embodiment of the present invention.

[0059] Figure 18 19 is a graph showing the aggregation judgment conclusion provided in an embodiment of the present invention.

[0060] Figure 20 This refers to unaggregated LGR data records provided in embodiments of the present invention.

[0061] Figure 21 This refers to the aggregated LGR data record provided in this embodiment of the invention.

[0062] Figure 22 This is a pseudo-query statement for deduplicated LGR data and its return result provided in an embodiment of the present invention. Detailed Implementation

[0063] To enable those skilled in the art to better understand the technical solutions of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0064] The present invention provides a method for processing goods in images, including technical methods for locating, identifying, marking, and providing procurement services for the main body of the goods in the image. Based on any image (including frame-by-frame images in live streaming media, hereinafter referred to as: live images), artificial intelligence technology is used to identify the goods in the image, marking their information and location within the image. Based on the identification results, corresponding product classifications and online display positions are selected for different identification results. Each product in the live image (hereinafter referred to as: live product LG) has many characteristic information that will be used later; therefore, each product is individually marked, an information database is established, and all information is saved. Since the characteristic information of many products in the live image is homogeneous (i.e., products in the same live image have similar or identical information such as category, condition, style, manufacturer, and warehouse location), this method merges the identification results of each product in the live image (hereinafter referred to as: live product dataset LGD) and the homogeneous information obtained when acquiring the live image (hereinafter referred to as: homogeneous information HI) into a live product data record LGR, plus the product information... The relative position and boundary range of the product in the image are marked as the product's location (referred to as: product location information). All of the above information is pre-stored in a database that stores all live images and the live products contained therein (referred to as: live product database LGDB). At this point, each live product has all the necessary information to generate the minimum stock unit (SKU) (referred to as: pseudo-SKU information). Once collected, the pseudo-SKU information is permanently saved for extraction and use in subsequent functions. Subsequently, in the live product display area, each live product is redrawn as an operable transparent polymorphic box object (referred to as: LGO) and prominently displayed, with its location range clearly marked for selection. Users can freely select the LGO corresponding to each live product to complete further product transaction operations such as selection, favorite, adding to cart, payment lock, and placing an order. The previously pre-stored pseudo-SKU information is extracted as the necessary information for filling in the order and used to generate the live stock unit (LSKU) for the order. It is also created simultaneously with the order before the first order is submitted.

[0065] This invention provides a method for processing products in an image, involving product image recognition, location positioning of multiple product entities, and determining when to generate a product's LSKU based on user browsing and other operations. Figure 1 As shown, the specific implementation process is as follows:

[0066] Process S1:

[0067] S101. Use a pre-trained product image model to perform deep learning and machine vision recognition on real-world images; for example... Figure 2 As shown, the automatically labeled boxes from artificial intelligence machine vision accurately selected the location and range information of multiple items in the wholesale market showroom and inventory. The text outside the boxes contains information on the quality, category, and style of the goods as determined by machine vision. Different colors represent a new combination of quality, category, and style. Basic product information includes: category, quality, style, specifications, craftsmanship, manufacturer information, etc. Product quality includes: 999, 3D, 5G, 18K, and platinum; product categories include: pendants, rings, bracelets, earrings, anklets, necklaces, and beaded bracelets.

[0068] Process S2:

[0069] S201. Based on the results of deep learning and machine vision recognition, collect LGD information for all items in the live image, and record the information in the following format: Figure 12 LGD table structure design Figure 13 The LGD table data example is shown below; however, the LGD information for each live-action product is not verified for accuracy and is entirely based on and recorded manually entered by the person collecting the live-action photos. Figure 4 As shown, the recognition feedback results of the deep learning and machine vision recognition training models are as follows: Figure 3 As shown, the images include location information and classification results. The bounding boxes automatically labeled by AI machine vision select the location and range information of multiple items in the wholesale market showroom and inventory. The text outside the boxes contains the product's quality, category, and style information as determined by machine vision. Different colors represent new combinations of quality, category, and style. During image acquisition, the person collecting the images is required to provide homogeneous information (HI) as comprehensively and accurately as possible, such as... Figure 4 As shown.

[0070] Process S3:

[0071] S301. Based on the summarized live commodity dataset from S2, create uniquely coded live commodity records (LGRs) for all historical live commodities, such as... Figure 14 As shown, all Live Goods Records (LGR) are aggregated into the Live Goods Database (LGDB). This database not only includes the non-manually processed information contained in the LGD, but also adds homogeneous information (HI) that should be manually entered and supplemented, such as... Figure 4As shown. The Live Product Database (LGDB), composed of all LGR records, is a fully redundant database. It stores all relevant information about all live products automatically identified by machine vision and remains open and editable, allowing for additions and revisions at any time. After integrating all live products, each LGR in the LGDB contains a pseudo-SKU information record (the LGR is not deduplicated and contains a large amount of duplicate data). The pseudo-SKU serves as the information source for subsequently establishing LSKUs for various live products.

[0072] Process S4:

[0073] S401 Figure 5 As shown, the AI ​​machine vision-generated, automatically labeled operable objects (LGOs) contain not only the images shown but also the previously collected pseudo-SKU information. Each operable product corresponds to one object, displayed by number. Product position information is such as: left, top, width, height ('0.3645” 0.4885” 0.095” 0.103') or the polygon label is the image path [space] x1, y1, x2, y2, x3, y3, x4, y4, label [space]

[0074] x1, y1 represent the x and y coordinates of the first point. Below is an example:

[0075] 71.7312,201.002,76.7312,240.002,31.7312,295.002,21.7312,345.002,38.7312,391.002,78.7312,414.002,135.731,406.002,169.731,351.002,151.731,288.002,102.731,238.002,108.731,201.002,100.731,186.002,75.7312,191.002,73,195, Pendant. (Example) Figure 6 As shown.

[0076] S402. When displaying the live image to the end user (consumer), dynamically draw graphics on the LGO (Local Object Item). The drawing information comes from the product location information collected in step S401. Highlight the main position of each live product in the image to allow the end user to further interact with all the live products in the live image. The highlighting method described here is a rectangular outline (or polygon, for example, a rectangle). This shape is drawn with (left, top, width, height) at distances from the left, right, top, and bottom. These four points are dynamically drawn as graphics, and several operation states are defined, such as: selectable state, selected state, currently selected focus state, paid lock state, sold out state, etc. Different colors are used to draw the rectangular outlines, for example, red for selectable, green for selected, and white for currently focused. The operable object range is determined based on the presented rectangular outlines. Within each rectangular outline, the user can manually touch or slide the mouse to select and focus on the live product object for further interaction. The user then performs the next operation on the currently selected object. Figure 7 As shown.

[0077] The following is sample code for drawing a Live Item Object (LGO):

[0078]

[0079]

[0080] Process S5:

[0081] S501, based on the real-time product operation interface presented by S402, establishes a dialogue tracking and real-time draft archive database for all user operations (e.g., Figure 15 As shown), users can edit and manipulate product objects in live product images at any time. For example... Figure 8 As shown.

[0082] The following is an example code for LGO operations:

[0083]

[0084]

[0085] Process S6:

[0086] S601. When an end user browses the live product operation interface of the live image, and performs operations related to generating purchase orders, such as adding to favorites, adding to cart, locking payment, or generating purchase orders, an LSKU is created for that live product. The LSKU is only used for long-term monitoring and tracking by the end user in their favorites. Therefore, when the pseudo-SKU information of the live product changes for any reason (such as when the maintenance team deletes, removes, or edits the information), the LSKU in a particular end user's favorites will change accordingly. Figure 9 , Figure 10 and Figure 11 When an end user selects a live product and performs actions related to order generation, such as adding it to their favorites, shopping cart, payment lock, or generating a purchase order, an LSKU (Local Stock Unit) is generated for that live product object (LGO). The LSKU is generated simultaneously with these actions. The LSKU function is the same as a typical e-commerce SKU. The reason for creating the LSKU only after an end user action is that the live product database (LGDB) is very large, and most of the data does not generate actual order demand. To reduce unnecessary resource consumption and performance pressure caused by an increase in the number of SKUs, the LSKU is created only after an end user action for subsequent functionality implementation in the e-commerce system. The scope of each live product entity is as follows: Figure 2 , Figure 3 As shown, each product within the wireframe is a product entity capable of independent actions (its labeling information is the descriptive text on the wireframe, but is not limited to this; more information is stored in the product image library).

[0087] The dynamic graphics used to distinguish, mark, and select each live product in the live product interface can be drawn using the most basic canvas. However, this method is not logically limited to display and interactive operation technologies. It includes any method that can use other image and graphics frameworks for drawing, marking, and selection, such as VR virtual reality or XR augmented reality implementations using 3D engines. All of these can implement the logic described in this method.

[0088] Process S7:

[0089] S701. Because S6 creates some duplicate or similar LSKUs, which affects system efficiency, the system periodically (e.g., each time payment is locked, each time an order is processed, daily, weekly, monthly, yearly, etc.) based on a deep learning algorithm-based model (GNN, K-means, DBSCAN, HDBSCAN, etc.) clusters and deduplicates the actual minimum inventory unit (LSKU), merging similar LSKUs. The clustering algorithm calculates the similarity between different LGOs based on the LGO images in S6. LGOs with similarity higher than a set threshold are clustered and deduplicated, then associated and merged. Aggregated Product 0 and Aggregated Product 1 are the aggregation result images for those with similarity higher than the threshold. LSKU aggregation judgment analysis, such as... Figure 17 Aggregate judgment conclusions, such as Figure 18 , Figure 19 Unaggregated LGR data records, such as Figure 20 After aggregation, LGR data records, such as Figure 21 The pseudo-query statement and returned results of LGR data after deduplication, such as Figure 22 The content shown.

[0090] Process S8:

[0091] S801. Since some LSKUs were created by S6 but no orders were ultimately generated by these LSKUs, which would affect system efficiency, the system regularly cleans up and deletes LSKUs that have expired and have no order submission operations.

[0092] Summary of Implementation Methods

[0093] This invention provides a method for processing products in images. It designs a low-level e-commerce architecture and employs deep learning-based machine vision technology to identify, segment, and automatically classify product objects in images, highlighting their location and range information. It manages the products contained in any image as SKU objects, and each identified and labeled SKU object can be further processed for online ordering.

[0094] This invention is a general method that creates operable product objects for labeled products in any image, allowing operators who view the image to select methods from the following list to further process the product object. Specific operations include: selection, deselection, screenshotting, numbering, and creation as a Real-Time Minimum Inventory Unit (LSKU). Figure 16As shown), users can add items to the cart, remove items from the cart, add items to favorites, remove items from favorites, edit names, categories, specifications, and notes, etc. Without needing to manage each item's SKU individually, this system fulfills the bulk selection and purchasing needs of buyers. It transforms the vast quantities of goods in the wholesale market—which don't require rigorous coding and warehousing—into a visual, operable, and orderable e-commerce platform. Products are first displayed and promoted, then buyers select and place orders. Only when there is a genuine order demand are generated are typical e-commerce SKUs and e-commerce orders generated. This allows suppliers to simultaneously conduct on-site physical wholesale activities and receive remote electronic orders from buyers. This process eliminates the need for individual item cataloging and shelving, as well as maintaining real-time inventory levels. As long as there is stock, images (or live streaming, essentially real-time image updates) can be uploaded and sold. During the sales process, there's no need to maintain dynamic online and offline inventory. A single inventory system caters to the remote and local sales needs of both offline and online wholesale customers.

[0095] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. However, these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method of processing a product in a picture, characterized by, Includes the following steps: S1. Use a pre-trained machine vision model to perform deep learning and machine vision recognition on live images; live images are digital photos, video files, live streaming data, or 3D modeling data. The information recognized by the machine vision model includes the jewelry's category, purity, style, specifications, craftsmanship, manufacturer, and setting; purity includes: 999, 3D, 5G, 18K, and platinum; categories include: pendants, rings, bracelets, earrings, bracelets and anklets, necklaces, beaded bracelets, and necklace sets; S2. Based on the results of deep learning and machine vision recognition, perform location identification and classification judgment on all products in the live images, output all judgment results and build a set of live product datasets; the live product dataset includes product serial number, location, range, and machine vision classification judgment; S3. Based on the results of deep learning and machine vision recognition, create uniquely coded real-time product records for all historical real-time products in the aggregated real-time product dataset of S2, and aggregate all real-time product records into a real-time product database. S4. When displaying the live image to the end user, dynamically draw the live image and add a layer of the same size as the live image for the end user to interact with. For each live product in the live image, draw a highlighted polygonal outline of the product according to its spatial position and range. Each polygonal outline corresponds to a live product record. The information of each live product record comes from the live product database created in step S3. The polygonal outline is further named a live product object for the end user to perform interface interaction. S5. Establish a real-time product draft archive database for recording dialogue tracking and saving real-time operations for all end-users' interface interaction operations on real-time product objects. Record and save the actions of each end-user on each real-time product object independently. S6. When an end user performs an operation related to generating a purchase order on the Live Goods Object (LGO), a Live Minimum Inventory Unit is created. The data for the minimum inventory unit comes from the Live Goods database.

2. The method of claim 1, wherein, In step S2, the product location information includes the relative position and boundary range information of the product in the image.

3. The method for processing goods in an image according to claim 1, characterized in that, In step S4, the dynamic graphics are highlighted using a polygonal outline.

4. The method for processing goods in an image according to claim 1, characterized in that, In step S4, the interactive operation states of the real-time product object include: selectable state, selected state, current selection focused state, paid lock state, and sold state. Different interactive operation states are drawn with different colored display frames to distinguish them.

5. The method for processing goods in an image according to claim 1, characterized in that, Using a model based on deep learning algorithms, the actual minimum inventory units in step S6 are periodically clustered, deduplicated, and statistically combined to identify similar actual minimum inventory units.

6. The method for processing goods in an image according to claim 1, characterized in that, Using a model based on deep learning algorithms, the smallest real-time inventory units that have expired in step S6 and have no order submissions are periodically cleaned up and deleted.

Images