Method and system for barcode prediction in self-checkout

Abstract

This invention provides a method and system for predicting a complete barcode from a damaged or partially unreadable barcode in a self-checkout (SCO) terminal. [Solution] A hybrid machine learning architecture is used, combining array processing of partial barcode data with metadata-based prediction using product attributes such as weight and category information. When a damaged barcode is scanned, available data is extracted and combined with contextual information to predict the most likely complete barcode or candidate barcode. Details of the predicted product are displayed for customer verification, and the customer is given the option to accept or reject the prediction.

Description

TECHNICAL FIELD

[0001] The present invention relates to self-checkout systems, and in particular, to a method and system using machine learning for predicting a complete barcode from a damaged or partially unreadable barcode.

SUMMARY OF THE INVENTION

PROBLEMS TO BE SOLVED BY THE INVENTION

[0002] Self-checkout (SCO) terminals are becoming increasingly prevalent in retail environments, providing convenience to customers and operational efficiency to retailers. However, these SCO terminals face significant challenges when barcodes become damaged or partially unreadable due to various factors during manufacturing, printing, processing, logistics, and handling. If the barcode cannot be fully scanned, the SCO process is interrupted, delays occur, and manual intervention by store employees is required. Store employees must then assist in the identification and processing of the goods, which reduces the overall value proposition of SCO technology and increases the operating costs of retailers. This not only causes customer dissatisfaction but also reduces the efficiency benefits provided by SCO terminals.

[0003] The present invention solves the above problems, and aims to enable reliable product identification even for damaged barcodes, minimize interruptions during SCO transactions, and reduce the need for employee intervention.

MEANS FOR SOLVING THE PROBLEMS

[0004] To solve the above problems, a method according to one aspect of the present invention includes: a processor receiving a partial barcode sequence from an SCO terminal; a processor obtaining metadata associated with a product having a partial barcode sequence; a processor processing the partial barcode sequence with a first machine learning model (MLM) to generate a first prediction of a complete barcode; a processor processing the metadata with a second MLM to generate a second prediction of a complete barcode; a processor combining the first and second predictions to determine a candidate barcode; and a processor providing the candidate barcode to the SCO terminal for customer confirmation during an SCO transaction.

[0005] Another aspect of the present invention includes: detecting scan data of a damaged barcode of a product in transaction using the processor of an SCO terminal; measuring the weight of the product using a scale of the SCO terminal; determining the category of the product using the processor; extracting a partial barcode sequence from the damaged barcode using the processor; inputting the partial barcode sequence into an array-based MLM to generate an array-based prediction using the processor; inputting the weight and category into a metadatabase MLM to generate a metadata-based prediction using the processor; combining the array-based prediction and the metadata-based prediction to identify a candidate barcode using the processor; and displaying product information associated with the candidate barcode on an interface for customer confirmation using the processor.

[0006] Furthermore, a system according to another aspect of the present invention includes an SCO terminal including a scanner, a scale, and a display, a processor, and memory, the memory including instructions that, when executed by the processor, cause the processor to perform an operation, the operation including receiving a partial barcode sequence from the scanner if a complete barcode cannot be scanned, obtaining a weight measurement from the scale, processing the partial barcode sequence with a recurrent neural network (RNN) MLM to generate a first set of barcode predictions, processing the weight measurement with a gradient boosting MLM (GBM) to generate a second set of barcode predictions, combining the first set of barcode predictions and the second set of barcode predictions to determine a candidate barcode, and presenting product information associated with the candidate barcode on the display for customer confirmation. [Effects of the Invention]

[0007] According to this invention, interruptions during SCO transactions are minimized, the need for employee intervention is reduced, and reliable product identification is guaranteed even for damaged barcodes. This makes SCO terminals more robust and user-friendly while giving retailers the impression of being technologically advanced and customer-oriented. Furthermore, continuous learning through a feedback loop improves predictive accuracy over time, enhancing the overall SCO experience. [Brief explanation of the drawing]

[0008] [Figure 1] This is a diagram of a system for barcode prediction in SCO, according to an exemplary embodiment. [Figure 2] This is a flowchart of a method for predicting product barcodes during an SCO transaction in an SCO terminal, according to an exemplary embodiment. [Figure 3] This is a flowchart of another method for predicting product barcodes during an SCO transaction in an SCO terminal, according to an exemplary embodiment. [Modes for carrying out the invention]

[0009] SCO terminals and systems have transformed the retail environment by offering greater autonomy to customers and reducing checkout wait times. These systems rely heavily on barcode scanning technology to accurately and efficiently identify products. However, the effectiveness of these systems is significantly impaired if barcodes are damaged, partially illegible, or otherwise unreadable. Barcode damage can occur at various stages of the product lifecycle, including manufacturing, shipping, inventory replenishment, and customer handling. If barcodes cannot be scanned properly, the SCO process is interrupted, requiring intervention from store staff, and causing frustration for customers who have chosen the SCO experience for its speed and convenience.

[0010] Barcode scanning failures impact not only the customer experience but also retail operations. Store staff must be distracted from other tasks to address barcode issues, reducing overall operational efficiency. Often, products with unreadable barcodes must be processed as general departmental sales, which can lead to inaccurate inventory tracking and potential revenue losses. As SCO adoption continues to increase, the frequency of such interruptions is increasing proportionally, creating significant problems for both retailers and customers and undermining the value proposition of SCO technology.

[0011] The technical challenge addressed by the technologies disclosed herein is that current SCO terminals and systems are unable to process and identify products when barcodes are damaged or partially unreadable. This limitation creates a reliance on manual intervention, interrupting the automated checkout process and negating many of the efficiency benefits offered by SCO terminals and systems. This technical challenge is further exacerbated by the diverse factors that can damage barcodes and the existence of a vast variety of products with different characteristics that must be accurately identified during a transaction.

[0012] Embodiments of the technology disclosed herein provide a technical solution to the aforementioned technical challenges by leveraging machine learning models (MLMs) to combine metadata-based prediction using product or item attributes with array processing on partial barcode data. This hybrid architecture employs a recurrent neural network (RNN) MLM to process and predict complete barcodes from partial arrays, while simultaneously utilizing a gradient boosting MLM (GBM) to analyze metadata such as weight and product category. The outputs from both MLMs are combined to generate highly accurate predictions of complete barcodes. The disclosed technology seamlessly integrates with existing SCO infrastructure, presents predictions to customers for validation, and continuously improves MLM predictions through a feedback mechanism.

[0013] In one embodiment, the training data includes normalized weight data and encoded categorical information, along with various forms of partial barcodes, such as sequences partially replaced with placeholder characters (e.g., "1234XXXXXX"). This diverse training dataset enables the MLM to recognize patterns in damaged barcodes and associate them with product metadata.

[0014] In one embodiment, the trained MLM is deployed on an SCO terminal using an appropriate framework. In complex cases where edge processing is insufficient, the method and system may utilize a cloud-based API service as a fallback, thereby ensuring that predictive capabilities remain available even under challenging circumstances.

[0015] In one embodiment, the method and system implement a continuous learning process in which customer feedback on predictions is logged and used to improve the MLM. This feedback loop allows the system to adapt over time to new products, fluctuating inventory, and changing barcode damage patterns.

[0016] In one embodiment, the method and system employ a comprehensive data acquisition and preparation process. Product data, including complete barcodes, weight, category, inventory details, and pricing, is collected and stored in a central database. This data is augmented with examples of damaged barcodes and partial barcode sequences to create a robust training dataset. The method and system normalizes weights, encodes categories, and generates partial barcode sequences to prepare data for MLM training.

[0017] In one embodiment, the prediction process is initiated when a customer scans a damaged or partially readable barcode at an SCO terminal. The method and system extracts the available partial barcode sequence and combines it with contextual information such as product weight and product category data from the SCO terminal's scale. The partial barcode sequence is processed by an RNN-MLM, which generates predictions based on the sequence pattern. Simultaneously, the metadata is processed by a GBM, which generates predictions based on the weight and category relevance. In an exemplary embodiment, these two sets of predictions are then combined using an averaging approach to determine the most likely complete barcode or candidate barcode.

[0018] In one embodiment, the product barcode prediction unit employs multiple approaches in addition to an averaging approach as a method for combining the outputs from the first MLM and the second MLM. These approaches include: (1) a weighted ensemble method in which predictions from each MLM are assigned different weights based on confidence scores or historical accuracy for a particular product category; (2) a sequential filtering approach in which the output from the first MLM 113 generates a set of candidate barcodes that are then improved by the second MLM 114 based on metadata consistency; (3) a decision tree approach that selects an appropriate combination strategy based on transaction context factors such as time, store department, or customer loyalty status; (4) a confidence-based threshold approach in which a prediction is accepted only if both MLMs exceed a predetermined confidence level, and an additional validation step is triggered otherwise; and / or (5) a reinforcement learning approach that dynamically adjusts the combination strategy over time based on feedback from successful and failed predictions. The method and system may also implement a hybrid approach that selects the optimal combination method based on specific characteristics of the damaged barcode and available metadata, further improving prediction accuracy across diverse retail environments and product types.

[0019] In one embodiment, the method and system display details of predicted products on the SCO terminal interface for customer verification. The customer is presented with acknowledgment options to accept or reject the predicted products. This interaction is logged regardless of the outcome, providing useful data for continuous improvement and learning of the MLM. If the prediction is incorrect, the method and system can generate an intervention request for assistance from store staff. This feedback loop ensures that prediction accuracy improves over time, further reducing the need for manual intervention and improving the overall SCO experience.

[0020] In one embodiment, the method and system address potential ticket swapping scenarios where a customer may swap the barcode of a low-value item for a high-value item, particularly when the barcode is partially damaged and triggers the prediction process. When the product barcode prediction unit generates a prediction for product identification, the method and system performs a multi-factor validation process to verify the prediction against other product attributes. This validation includes comparing the weight measured by the SCO terminal scale with the predicted weight range of the predicted product, analyzing the correlation between product characteristics and the predicted barcode using the metadatabase MLM, and matching it with inventory data from the retailer's transaction system. If a discrepancy is detected, for example, if the predicted low-value product code corresponds to a product with a weight that differs significantly from the predicted weight range in the database, the method and system flag this as a potential ticket swapping attempt. Upon detecting such conflicting data, the method and system automatically triggers an intervention request for assistance from store staff. The method and system may display alternative predictions on the SCO terminal UI or request customer validation before proceeding with the transaction. This multi-layered approach leverages both sequence-based prediction using a first MLM and metadata-based prediction using a second MLM to ensure that prediction results are not misused for malicious purposes, even when processing damaged barcodes.

[0021] In one embodiment, the barcode prediction presented herein minimizes interruptions during SCO transactions, reduces the need for employee intervention, and ensures reliable product identification even for damaged barcodes. This makes SCO terminals more robust and user-friendly while giving retailers the impression of being technologically advanced and customer-oriented.

[0022] In this specification, the terms "merchandise" and "product" may be used synonymously and interchangeably. These refer to barcode-affixed articles that are the subject of purchase in SCO transactions at SCO terminals. In this specification, the terms "user" and "customer" may be used synonymously and interchangeably. These refer to individuals engaged in SCO transactions at SCO terminals.

[0023] The following detailed description relating to Figures 1 to 3 illustrates the components, data flows, and processes involved in predicting a complete barcode from a damaged or partially unreadable barcode in an SCO transaction. Each figure shows how the RNN-MLM and GBM work together to generate accurate predictions, and how customer feedback is incorporated to continuously improve the predictions of the methods and systems presented in this specification.

[0024] Figure 1 is a diagram of an exemplary system for barcode prediction in self-checkout (SCO) according to an embodiment. It should be noted that each component is schematically illustrated in a greatly simplified form, and only the components relevant to the understanding of the embodiment are shown.

[0025] Furthermore, various components (identified in system 100) are illustrated, and the arrangement of components is presented for illustrative purposes only. It should be noted that other arrangements with more or fewer components are possible without departing from the teachings for product barcode prediction during SCO transactions at SCO terminals presented in this specification and hereinafter.

[0026] System 100 includes a cloud 110 or server, one or more SCO terminals 120, and one or more retailer servers 130. The cloud 110 includes at least one processor 111 and a non-temporary computer-readable storage medium 112 (medium), the non-temporary computer-readable storage medium 112 includes instructions for a first MLM 113, a second MLM 114, a product barcode prediction unit 115, and an application programming interface (API) 116. When these instructions are executed by the processor 111, the processor 111 is caused to perform the operations described herein and below with respect to 113-116.

[0027] Each SCO terminal 120 includes at least one processor 121 and a medium 122, the medium 122 containing instructions for a transaction manager 123 and a user interface (UI) 124. When these instructions are executed by the processor 121, the processor 121 is caused to perform the operations described herein and below with respect to 123-124. Each SCO terminal 120 also includes a scanner 125, a scale 126, and a display 127.

[0028] Each retailer server 130 includes at least one processor 131 and a medium 132, the medium 132 containing instructions for a transaction system 133. When these instructions are executed by the processor 131, the processor 131 is caused to perform the operations described herein and below with respect to 133.

[0029] In one embodiment, the first MLM113 is implemented as a recurrent neural network (RNN) MLM specifically trained to process partial barcode sequences and predict complete barcodes. During training, the RNN-MLM113 is provided with various forms of partial barcode sequences, such as "12345XXXXX", where parts of the complete barcode are replaced with placeholder characters. This training enables the RNN-MLM113 to recognize patterns in damaged or partially readable barcodes and generate probability distributions for possible complete barcodes.

[0030] In one embodiment, the second MLM114 is implemented as a gradient boosting MLM (GBM) trained to analyze product metadata such as weight and category information. During training, the GBM114 is provided with normalized weight data and encoded category information paired with corresponding complete barcodes. This training enables the GBM114 to associate product characteristics with specific products and generate a probability distribution for possible complete barcodes based solely on these product characteristics.

[0031] The product barcode prediction unit 115 functions as a central coordinating component, receiving input from both MLMs 113 and 114 and providing input to MLMs 113 and 114. During an SCO transaction, the product barcode prediction unit 115 receives a partial barcode from the transaction manager 123 and sends it as input to the first MLM 113. Similarly, the product barcode prediction unit 115 receives product weight data (originally obtained by the scale 126) from the transaction manager 123 and sends it as input to the second MLM 114. The product barcode prediction unit 115 then combines the outputs from MLMs 113 and 114 using an averaging approach to determine a complete barcode that is a candidate for the product being processed.

[0032] The transaction manager 123 manages the overall SCO transaction process through interaction with the corresponding transaction system 133. The transaction manager 123 also communicates with the product barcode prediction unit 115 via API 116 if it encounters a damaged or partially readable barcode. The scanner 125 retrieves barcode data from the product, and the scale 126 measures the product weight. The UI 124 and display 127 provide an interface for the customer to interact with the SCO terminal 120 and verify the predicted barcode. The transaction manager 123 controls the state and the UI screens presented during the SCO transaction.

[0033] The transaction system 133 holds the retailer's inventory, pricing, and product databases. These can be accessed by the product barcode prediction unit 115 via API 116 to retrieve additional metadata about products as needed. During the operation of system 100, the product barcode prediction unit 115 provides a partial barcode to the first MLM 113 and metadata retrieved from the product database to the second MLM 114. The barcode predictions provided by MLMs 113 and 114 are analyzed, and one or more of the most likely complete barcodes or candidate barcodes are selected and returned to the transaction manager 123 for presentation to the user within the user UI 124.

[0034] A typical user journey in system 100 begins when a customer scans an item with a damaged barcode using the SCO terminal 120. The scanner 125 attempts to read the barcode but can only obtain a partial sequence due to damage or occlusion. The transaction manager 123 recognizes that it could not obtain a complete barcode and, via API 116, transfers the partial barcode sequence to the product barcode prediction unit 115. Simultaneously, the scale 126 measures the weight of the item, which is also transferred to the product barcode prediction unit 115 via the transaction manager 123 using API 116.

[0035] Once a candidate complete barcode is determined through analysis of the outputs from the first MLM 113 and the second MLM 114, the product barcode prediction unit 115 sends this information back to the transaction manager 123. The transaction manager 123 displays the details of the corresponding product on the display 127 via the UI 124. The customer is then presented with the option to approve or reject the predicted product. This feedback is obtained via the UI 124 and sent back to the product barcode prediction unit 115. The product barcode prediction unit 115 logs this for future retraining of the MLMs.

[0036] If the customer approves the prediction, the transaction proceeds successfully. If the customer rejects the prediction, system 100 can offer an alternative prediction or generate an intervention request for assistance from store staff. In either case, the feedback is logged and used to retrain both the first MLM 113 and the second MLM 114 to continuously improve their prediction accuracy over time. This retraining process involves periodically updating the MLMs with new feedback data, enabling them to adapt to fluctuating product inventory and changing barcode damage patterns.

[0037] API 116 provides a cloud-based fallback mechanism in case the edge processing capabilities of the SCO terminal 120 are insufficient to determine the most likely complete barcode or candidate barcode. In such a scenario, the product barcode prediction unit 115 can utilize API 116 or additional APIs to access additional computing resources or a more comprehensive product database to improve prediction accuracy.

[0038] The product barcode prediction unit 115 initiates the prediction process by sending a partial barcode sequence to the first MLM 113 and product or product metadata (e.g., weight, product classification, etc.) to the second MLM 114. The first MLM 113 processes the partial sequence and generates a probability distribution for possible complete barcodes. Simultaneously, the second MLM 114 analyzes the metadata and generates its own probability distribution for possible complete barcodes. The product barcode prediction unit 115 combines these distributions, typically by averaging them, to determine candidate complete barcodes for a product.

[0039] In one embodiment, the first MLM 113, the second MLM 114, and the product barcode prediction unit 115 are hosted on a retail server 130. In another embodiment, the first MLM 113, the second MLM 114, and the product barcode prediction unit 115 are processed on an SCO terminal 120.

[0040] In one embodiment, the first MLM 113 and the second MLM 114 are deployed on the SCO terminal 120 using TensorFlow (Registered Trademark) Lite or a similar framework optimized for edge processing. This deployment approach enables real-time barcode prediction without requiring a constant network connection, while still allowing cloud-based processing via API 116 when needed for more complex predictions.

[0041] Figure 2 is a flowchart of Method 200 for predicting product barcodes during an SCO transaction in an SCO terminal, according to an exemplary embodiment. The software module implementing Method 200 is called the "Product Barcode Prediction Unit". The Product Barcode Prediction Unit is implemented as executable instructions programmed and resident in memory and / or non-temporary computer-readable (processor-readable) storage medium and executed by one or more processors in one or more devices. The processor of the device executing the Product Barcode Prediction Unit is configured and programmed specifically to process the Product Barcode Prediction Unit. The Product Barcode Prediction Unit may have a network connection, including wired, wireless, or a combination of wired and wireless.

[0042] In one embodiment, the device running the product barcode prediction unit is a cloud 110, an SCO terminal 120, or a retailer server 130. In one embodiment, the product barcode prediction unit is a combination of all or some of components 113, 114, 115, and / or 116.

[0043] In step 210, the product barcode prediction unit receives a partial barcode sequence from the SCO terminal 120. In one embodiment, in step 211, the product barcode prediction unit receives scan data of a damaged barcode from the scanner 125 of the SCO terminal 120.

[0044] In 220, the product barcode prediction unit obtains metadata associated with a product having a partial barcode sequence. In one embodiment, in 221, the product barcode prediction unit obtains at least one of the following: the weight of the product from the scale 126 of the SCO terminal 120, the product category, or the price range of the product.

[0045] In 230, the product barcode prediction unit processes a partial barcode sequence in the first MLM 113 to generate a first prediction of a complete barcode. In one embodiment, in 231, the first MLM 113 is an RNN-MLM trained on a partial barcode sequence to provide a prediction of the barcode sequence.

[0046] In 240, the product barcode prediction unit processes metadata in the second MLM 114 to generate a second prediction of the complete barcode. In one embodiment, in 241, the second MLM 114 is a GBM trained with product metadata.

[0047] At 250, the product barcode prediction unit combines the first prediction and the second prediction to determine a candidate complete barcode. In one embodiment, at 251, the product barcode prediction unit averages the first probability distribution output by the first MLM 113 and the second probability distribution output by the second MLM 114.

[0048] In step 260, the product barcode prediction unit provides candidate barcodes to the SCO terminal 120 for customer verification during an SCO transaction. In one embodiment, in step 261, the product barcode prediction unit provides product details associated with the candidate barcodes for display on the interface (e.g., UI 124) of the SCO terminal 120.

[0049] In one embodiment, in 270, the product barcode prediction unit receives customer feedback indicating whether a candidate barcode is correct or not. The product barcode prediction unit logs the customer feedback for retraining of at least one of the first MLM 113 and / or the second MLM 114.

[0050] In one embodiment, at 280, the product barcode prediction unit generates an intervention request for store staff assistance if customer feedback indicates that the candidate barcode is incorrect. In one embodiment, at 290, the product barcode prediction unit deploys the first MLM 113 and the second MLM 114 on the SCO terminal 120 using a machine learning framework optimized for edge processing. In one embodiment, at 295, the product barcode prediction unit utilizes a cloud-based API 116 as a fallback if edge processing is insufficient to determine the complete candidate barcode.

[0051] Figure 3 is a flowchart of another method 300 for predicting product barcodes during an SCO transaction in an SCO terminal, according to an exemplary embodiment. The software module implementing method 300 is called the “SCO Barcode Prediction Unit.” The SCO Barcode Prediction Unit is implemented as executable instructions programmed and resident in memory and / or a non-temporary computer-readable storage medium and executed by one or more processors in the device. The processor executing the SCO Barcode Prediction Unit is configured and programmed specifically to process the SCO Barcode Prediction Unit. The SCO Barcode Prediction Unit may have a network connection, including wired, wireless, or a combination of wired and wireless.

[0052] In one embodiment, the device running the SCO barcode prediction unit is a cloud 110, an SCO terminal 120, or a retailer server 130. In one embodiment, the SCO barcode prediction unit is a combination of all or part of components 113, 114, 115, 116, 123, 124, 133, and / or method 200. The SCO barcode prediction unit presents a different, and in some respects enhanced, processing perspective from that described above in method 200 of Figure 2.

[0053] In 310, the SCO barcode prediction unit detects scanned data of a damaged barcode of an item in a transaction. In one embodiment, in 311, the SCO barcode prediction unit identifies that the scanned barcode includes an unreadable portion.

[0054] In step 320, the SCO barcode prediction unit measures the weight of the product via the scale 126 of the SCO terminal 120. In step 330, the SCO barcode prediction unit determines the category of the product. In one embodiment, in step 331, the SCO barcode prediction unit retrieves the category from the data store based on the available characters in the partial barcode sequence.

[0055] In step 340, the SCO barcode prediction unit extracts a partial barcode sequence from the damaged barcode. In step 350, the SCO barcode prediction unit inputs the partial barcode sequence into the sequence-based MLM 113 to generate a sequence-based prediction. In one embodiment, in step 351, the SCO barcode prediction unit preprocesses the partial barcode sequence by replacing unreadable portions with placeholder characters.

[0056] In step 360, the SCO barcode prediction unit inputs weight and category into the metadatabase MLM114 to generate metadata-based predictions. In step 370, the SCO barcode prediction unit combines sequence-based predictions and metadata-based predictions to identify candidate barcodes for a product. In one embodiment, in step 371, the SCO barcode prediction unit selects the barcode with the highest probability from a combined probability distribution based on sequence-based predictions and metadata-based predictions.

[0057] In 380, the SCO barcode prediction unit displays product information associated with candidate barcodes on an interface (e.g., UI124) for customer verification. In one embodiment, in 390, the SCO barcode prediction unit stores customer verification results as training data to enable improved accuracy in the array-based MLM113 and the metadatabase MLM114.

[0058] In one embodiment, in 395, the SCO barcode prediction unit generates a plurality of candidate barcodes ranked by probability. If a candidate barcode is rejected through the UI 124 of the SCO terminal 120, the SCO barcode prediction unit allows the customer to make a selection from the plurality of candidate barcodes.

[0059] When software is described in a specific form (such as components or modules), this is merely to aid understanding and is not intended to limit how the software implementing those functions may be architectured or structured. For example, modules may be illustrated as separate modules, but they may be implemented as uniform code, as individual components, as some (but not all) of these modules combined, or as software structured in any other convenient form.

[0060] Furthermore, although the software module is illustrated as running on a single piece of hardware, the software may be distributed across multiple processors or in any other convenient form.

[0061] The above description is illustrative and not limiting. By examining the above description, many other embodiments will become apparent to those skilled in the art. Therefore, the scope of each embodiment should be determined by reference to the appended claims and the entire scope of equivalents to be given thereto.

[0062] In the above description of the embodiments, various features are grouped into a single embodiment for the purpose of streamlining the disclosure. This method of disclosure should not be interpreted as reflecting that the above embodiments have more features than are explicitly described in each claim. Rather, as reflected in the following claims, the subject matter of the invention lies in fewer features than all the features of the single disclosed embodiment combined. Accordingly, the following claims are incorporated herein into the description of the embodiments, and each claim stands as an independent exemplary embodiment.

Claims

1. The processor receives a partial barcode sequence from the self-checkout (SCO) terminal, The processor obtains metadata associated with the product having the partial barcode sequence, The processor processes the partial barcode sequence with a first machine learning model (MLM) to generate a first prediction of a complete barcode. The processor processes the metadata in a second MLM in order to generate a second prediction of the complete barcode. The processor combines the first prediction and the second prediction to determine a candidate barcode. The processor provides the candidate barcode to the SCO terminal for customer verification during an SCO transaction. Methods that include...

2. Receiving the aforementioned partial barcode sequence is equivalent to receiving scan data of a damaged barcode from the scanner of the SCO terminal. The method according to claim 1, further comprising:

3. Obtaining the aforementioned metadata means obtaining at least one of the following from the scale of the SCO terminal: the weight of the product, the category of the product, or the price range of the product. The method according to claim 1, further comprising:

4. The first MLM includes a recurrent neural network (RNN) MLM trained on a partial barcode sequence. The method according to claim 1.

5. The second MLM includes a gradient boosting MLM (GBM) trained with product metadata, The method according to claim 1.

6. Combining the first prediction and the second prediction is equivalent to averaging the first probability distribution output by the first MLM and the second probability distribution output by the second MLM. The method according to claim 1, further comprising:

7. Providing the candidate barcodes means providing product details associated with the candidate barcodes for display on the interface of the SCO terminal. The method according to claim 1, further comprising:

8. The processor receives customer feedback indicating whether the candidate barcode is correct or not. The processor logs the customer feedback in order to retrain at least one of the first MLM or the second MLM. The method according to claim 1, further comprising:

9. The processor generates an intervention request for store staff assistance if the customer feedback indicates that the candidate barcode is incorrect. The method according to claim 8, further comprising:

10. The processor deploys the first MLM and the second MLM on the SCO terminal using a machine learning framework optimized for edge processing. The method according to claim 1, further comprising:

11. If edge processing is insufficient to determine the candidate barcode, the processor may use a cloud-based application programming interface (API) as a fallback. The method according to claim 1, further comprising:

12. The processor in the self-checkout (SCO) terminal detects scanned data of damaged barcodes on items during a transaction, The weight of the product is measured using the scale of the SCO terminal, The processor determines the category of the product, The processor extracts a partial barcode sequence from the damaged barcode, The processor inputs the partial barcode sequence into a sequence-based machine learning model (MLM) in order to generate sequence-based predictions. The processor inputs the weight and the category into the metadatabase MLM in order to generate metadata-based predictions. The processor combines the sequence-based prediction and the metadata-based prediction to identify candidate barcodes. The aforementioned processor displays product information associated with the candidate barcode on the interface for customer verification. Methods that include...

13. Detecting the scanned data of the damaged barcode means identifying that the scanned barcode includes an unreadable portion. The method according to claim 12, further comprising:

14. Determining the category of the aforementioned product involves searching for the category in the data store based on the available characters in the partial barcode sequence. The method according to claim 12, further comprising:

15. Inputting the aforementioned partial barcode sequence into the sequence-based MLM, pre-processing the aforementioned partial barcode sequence by replacing unreadable parts with placeholder characters, The method according to claim 12, further comprising:

16. Combining the sequence-based prediction and the metadata-based prediction allows for the selection of the barcode with the highest probability from the combined probability distribution. The method according to claim 12, further comprising:

17. The processor stores the results of customer verification as training data to improve the accuracy of the sequence-based MLM and the metadatabase MLM. The method according to claim 12, further comprising:

18. The aforementioned processor generates multiple candidate barcodes ranked by probability, If the candidate barcode is rejected, the processor allows the customer to select from the multiple candidate barcodes. The method according to claim 12, further comprising:

19. A self-checkout (SCO) terminal including a scanner, scale, and display, Processor and Memory and A system that includes, The memory, when executed by the processor, includes instructions that cause the processor to perform an operation including: If the complete barcode cannot be scanned, the scanner receives a partial barcode sequence, Obtaining weight measurements from the aforementioned scale, To generate a first set of barcode predictions, the partial barcode sequence is processed using a recurrent neural network (RNN) machine learning model (MLM), To generate a second set of barcode predictions, the weight measurements are processed using gradient boosting MLM (GBM), In order to determine a candidate barcode, the first set of barcode predictions and the second set of barcode predictions are combined, To allow customer verification, product information associated with the candidate barcode is displayed on the display.

20. The aforementioned operation, Receiving customer feedback through the aforementioned SCO terminal, The aforementioned customer feedback will be recorded in a log, To improve prediction accuracy, the RNN-MLM and GBM are retrained using the customer feedback recorded in the logs, The system according to claim 19, further comprising:

Images