Method and system for intelligent scoring of merchandise planogram images based on dynamic rule configuration

Abstract

This disclosure provides a method and system for intelligent scoring of product display images based on dynamic rule configuration. The method includes recognizing and performing basic structured modeling on the collected shelf images to generate a structured data set; identifying a cross-category and spatially continuous set of products to construct a combined display semantic unit with business semantics; semantically reconstructing the original scoring rules through a dynamic rule engine, transforming the single-product rule system into a two-layer rule structure consisting of combination priority rules and single-product supplementary rules; calculating the overall combined score of the combined display semantic unit according to the mapping relationship, performing single-product supplementary scoring on products not covered by the combination priority rules, and merging to generate the shelf scoring result; and re-executing the dynamic rule reconstruction and scoring calculation based on the archived combined display semantic units and structured data set in response to changes in scoring rules. This method solves the problems of slow rule adjustment response, single scoring dimension, and difficulty in reusing historical data.

Description

Technical Field

[0001] This disclosure belongs to the field of product display image recognition and intelligent scoring technology, and more specifically, relates to a product display image intelligent scoring method and system based on dynamic rule configuration. Background Technology

[0002] The quality of merchandise display is crucial to sales performance in the retail industry, and various automated evaluation technologies for shelf displays have been explored. Existing solutions, such as image recognition-based shelf display detection methods, determine shelf quality by capturing display videos, extracting keyframes, and matching them with preset image samples. While these methods can partially replace manual labor, they are essentially static comparisons, struggling to handle complex and ever-changing display rules, and have limited resistance to interference. There are also AI-based display item review methods that perform multi-level reviews of displays using primary and secondary standards, generating adjustment strategies when displays are deemed unsuitable. This method introduces a deeper level of rationality judgment, but its review rules are relatively fixed and integrated within the system, making it difficult to quickly adjust to changes in marketing strategies.

[0003] It's not difficult to see that existing technical solutions suffer from the following main drawbacks: Firstly, evaluation rules are mostly hard-coded into the system logic. When business needs change, such as adjusting scoring weights or adding new scoring dimensions, the program code must be modified and redeployed, resulting in slow response times and an inability to support rapid marketing campaign iterations—in other words, rigid rules with poor adaptability. Secondly, many solutions focus on compliance checks or single indicators, lacking comprehensive quantitative evaluation of multi-dimensional marketing indicators such as display occupancy and prime shelf scores. This makes the scoring results unable to fully reflect the actual impact of displays on sales—in other words, the scoring dimensions are singular and lack comprehensiveness. Furthermore, traditional solutions typically do not structure and store the intermediate results after image recognition and scoring. When scoring rules are updated, re-scoring historical display images according to the new rules requires re-calling the AI ​​model for recognition, resulting in high computational costs and hindering efficient historical data retrospective analysis and trend comparison—in other words, data cannot be reused, and historical analysis costs are high.

[0004] Therefore, there is an urgent need for an intelligent scoring method for product display images that can support dynamic rule configuration, achieve multi-dimensional comprehensive scoring, and perform low-cost and rapid re-scoring of historical data, in order to solve the problems of slow rule adjustment response, single scoring dimension, and difficulty in reusing historical data in existing technologies. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the present invention aims to resolve the aforementioned deficiencies and propose an intelligent scoring method for product display images based on dynamic rule configuration.

[0006] The present invention adopts the following technical solution.

[0007] The first aspect of this invention discloses an intelligent scoring method for product display images based on dynamic rule configuration, the method comprising: S1. Recognize and perform basic structured modeling on the collected shelf images to generate a structured data set; S2. Based on the structured data set, identify a collection of products that are cross-category and spatially continuous, and construct a combined display semantic unit with business semantics; S3. The original scoring rules are semantically reconstructed through a dynamic rule engine, transforming the flat single-item rule system into a two-layer rule structure consisting of combination priority rules and single-item supplementary rules, and establishing a mapping relationship between each of the combined display semantic units and the combination priority rules. S4. Under the constraints of the two-layer rule structure, the combined display semantic unit is calculated according to the mapping relationship, and the individual product supplementary score is performed on the products not covered by the combined priority rule, and the shelf score result is generated by fusion. S5. The shelf scoring results are jointly archived with the combined display semantic unit and the structured data set. In response to changes in the scoring rules, S3 and S4 are re-executed based on the archived combined display semantic unit and the structured data set to achieve rapid re-scoring.

[0008] Optionally, S1 includes: The shelf images are preprocessed to establish a unified spatial reference, and a deep learning object detection network is used to identify the goods and calculate the comprehensive credibility and actual projected area. Based on the product bounding box, determine the layer number and column area, and generate a basic structured data set. Each product record includes product identifier, product category, product brand, center coordinates, layer number, column area number, and projected area. Calculate the spatial adjacency strength between any two products, merge products that meet the adjacency conditions into an initial spatial clustering unit, and record the cluster number, cover layer number, starting column area, ending column area, product quantity and brand composition for each clustering unit; The structured dataset includes the basic structured dataset and the initial spatial clustering unit.

[0009] Optionally, S2 includes: Based on the initial spatial clustering units, a region growing merging algorithm is used to identify cross-category and spatially continuous candidate combinations; By combining fill rate, category diversity and inter-layer span, the cross-category effectiveness strength is calculated, and effective combinations are selected from the candidate combinations. A unique identifier is assigned to each valid combination, the member product list and spatial boundaries are recorded, the semantic strength and the number of coverage layers are calculated, and the basic information of the combination is formed. Based on the basic information of the combination, according to the category and brand composition of the effective combination, a predefined template library is matched to generate a combination type identifier. The coverage area is calculated according to the spatial boundary, and the internal product proportion structure is calculated according to the member product list. The combination display semantic unit containing the combination identifier, member product list, spatial boundary, number of coverage layers, coverage area, internal product proportion structure and combination type identifier is constructed.

[0010] Optionally, S3 includes: The original scoring rules are broken down into rule objects, conditions, scoring factors, score ranges, and applicable scope; Based on whether the rule conditions involve multiple categories co-occurring or combined attributes, the rules are divided into combination priority rules and single-item supplementary rules, thus constructing a two-layer rule framework; Scan each of the combination priority rules one by one, establish a mapping relationship between each combination display semantic unit and the combination priority rule, and generate a scoring expression unit that includes a combination identifier, a hit rule identifier, and a score limit. Based on the scoring expression unit, the coverage and exclusion relationships of the combination priority rule on the single item supplement rule are determined by the difference between the condition coverage rate and the scoring contribution, and the single item supplement rule is marked as exclusion, reduction or retention state according to the determination result. Based on the mapping relationship, the coverage relationship, and the exclusive relationship, an executable rule expression result containing the rule version number, execution sequence number, and current valid state is generated, forming a stable two-layer rule structure.

[0011] Optionally, determining the coverage and exclusion relationship between the combination priority rule and the single-item supplement rule includes: Calculate the condition coverage of the single-item replenishment rule, and the exclusivity strength between the combination priority rule and the single-item replenishment rule; When the condition coverage rate is greater than or equal to the condition coverage rate threshold and the exclusion strength is greater than or equal to the score contribution difference threshold, the single item supplement rule is marked as exclusionary. When the condition coverage rate is greater than or equal to the condition coverage rate threshold and the exclusivity intensity is less than the score contribution difference threshold, the single item supplement rule is marked as a reduction state. When the condition coverage rate is less than the condition coverage rate threshold, the single-item supplement rule is retained as a valid rule.

[0012] Optionally, S4 includes: For each of the combined display semantic units, verify the type, number of layers, area and proportion conditions in the mapped combination priority rules, and generate a valid execution set containing comprehensive hit rate; Based on the effective execution set, the theoretical score of a single rule is calculated by combining the full score of each rule, the comprehensive hit rate, and the execution weight. The scores of multiple rules in the same combination are then compressed and merged to obtain the overall combined score. Based on the determination results of the coverage relationship and the exclusivity relationship, the products in the combination are assigned, and the individual product rule conditions that have been absorbed by the combination priority rule are trimmed to obtain the individual product residual condition set. Based on the set of residual conditions for individual items, supplementary scoring is performed on the products within the combined display semantic unit, and products that do not belong to any combined display semantic unit are directly scored according to the complete individual item rules to obtain a set of supplementary individual item scores. The overall score of the combination and the supplementary scores of the individual items are summed, and the shelf score result is output after applying the upper limit constraint.

[0013] Optionally, S5 includes: The shelf rating results, the combined display semantic units, and the structured data set are jointly archived to construct a historical rating master record that includes a master record table, a combination detail table, and an individual item detail table. Based on the aforementioned historical rating master record, a reusable index is established according to the shelf dimension, version dimension, and time dimension; When a new rule version is released, the stored combined display semantic units and the structured data set are located based on the reusable index, and the rule reconstruction and score calculation are directly read and re-executed. The recalculated scoring results are archived according to the new rule version and stored in parallel with the original historical records to form a multi-version scoring record.

[0014] Secondly, this disclosure provides an intelligent scoring system for product display images based on dynamic rule configuration, applied to the method in the first aspect, the system comprising: The image recognition and modeling module is used to recognize and perform basic structured modeling on the collected shelf images, generating a structured dataset. The combined semantic unit construction module is used to identify a collection of products that are cross-category and spatially continuous based on the structured data set, and to construct a combined display semantic unit with business semantics. The rule dynamic reconstruction module is used to semantically reconstruct the original scoring rules through a dynamic rule engine, transforming the flat single-item rule system into a two-layer rule structure consisting of combination priority rules and single-item supplementary rules, and establishing a mapping relationship between each of the combined display semantic units and the combination priority rules. The scoring calculation module is used to perform overall score calculation on the combined display semantic unit according to the mapping relationship under the constraints of the two-layer rule structure, and to perform supplementary individual scores on products not covered by the combined priority rule, and to generate shelf score results by merging them. The archiving and re-scoring module is used to jointly archive the shelf scoring results with the combined display semantic units and the structured data set, and in response to changes in scoring rules, to re-trigger the rule dynamic reconstruction module and the scoring calculation module based on the archived combined display semantic units and the structured data set to achieve rapid re-scoring.

[0015] Thirdly, this disclosure provides an electronic device including a memory and at least one processor, the memory storing a computer program, and the processor executing the computer program to implement the method of the first aspect described above.

[0016] Fourthly, this disclosure provides a computer storage medium storing a computer program that, when executed, implements the method described in the first aspect.

[0017] The beneficial effects of the present invention are as follows: Compared with the prior art, the present invention has the following advantages: (1) This invention uses a dynamic rule engine to semantically reconstruct the original scoring rules, transforming the flat single-item rule system into a two-layer structure consisting of combination priority rules and single-item supplementary rules. It also enables hot updates when rules change without requiring modification of program code or redeployment of the system. This technical solution solves the problems of rule rigidity and poor adaptability caused by hard-coded evaluation rules in existing technologies, enabling business personnel to quickly adjust scoring weights and add or remove scoring dimensions according to marketing strategies, significantly improving the system's responsiveness to complex and ever-changing retail environments.

[0018] (2) This invention organizes discrete product objects into business-meaning combinations through product structured modeling and semantic unit construction of combined displays. Furthermore, it integrates multi-dimensional marketing indicators such as combination coverage, golden shelf score, internal proportion structure of the combination, and cross-category effectiveness strength during the scoring process, achieving a leap from single-product compliance checks to comprehensive evaluation of combined marketing value. This technical solution overcomes the shortcomings of existing technologies, such as single scoring dimensions and lack of comprehensiveness, enabling the scoring results to fully reflect the actual impact of displays on sales.

[0019] (3) This invention constructs a historical record system with rule versioning and structure reuse by jointly archiving product identification results, combined semantic units, and scoring results. When the scoring rules are updated, the system can quickly re-score based on the stored structured data without having to call the image recognition model again. This technical solution solves the problems of unusable data and high historical analysis costs in the prior art, significantly reduces the computational cost of rule iteration, and provides efficient and feasible technical support for historical trend comparison and marketing effect retrospective. Attached Figure Description

[0020] The accompanying drawings, which are incorporated in and form a part of this specification, illustrate embodiments consistent with this disclosure and, together with the description, serve to explain the principles of this disclosure.

[0021] Figure 1 A flowchart of the intelligent scoring method for product display images based on dynamic rule configuration provided in an embodiment of this disclosure is shown. Figure 2 This illustration shows a flowchart of the rule semantic reconstruction and two-layer rule structure generation provided in an embodiment of this disclosure; Figure 3 A schematic diagram of the structure of a product display image intelligent scoring system based on dynamic rule configuration provided in an embodiment of this disclosure is shown.

[0022] The accompanying drawings have illustrated specific embodiments of this disclosure, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concepts of this disclosure to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0023] The present disclosure will be further described below with reference to the accompanying drawings. The following embodiments are only used to illustrate the technical solutions of the present disclosure more clearly, and should not be used to limit the scope of protection of the present disclosure.

[0024] The components of the embodiments of the invention described and illustrated herein can typically be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0025] In the following, the terms “comprising,” “having,” and their cognates, which may be used in various embodiments of the invention, are intended only to indicate a particular feature, number, step, operation, element, component, or combination thereof, and should not be construed as excluding, firstly, the presence of one or more other features, numbers, steps, operations, elements, components, or combinations thereof, or adding the possibility of one or more features, numbers, steps, operations, elements, components, or combinations thereof.

[0026] Unless otherwise specified, all terms used herein (including technical and scientific terms) shall have the same meaning as commonly understood by one of ordinary skill in the art to which the various embodiments of the invention pertain. Terms (such as those defined in commonly used dictionaries) shall be interpreted as having the same meaning as in their contextual meaning in the relevant technical field and shall not be interpreted as having an idealized or overly formal meaning, unless clearly defined in the various embodiments of the invention.

[0027] Figure 1 A flowchart of the intelligent scoring method for product display images based on dynamic rule configuration provided in this disclosure embodiment is shown below. Figure 1 As shown, the method may include the following steps: S1 identifies and performs basic structured modeling on the collected shelf images, generating a structured data set containing product categories, brands, locations, areas, shelf numbers, and spatial adjacency relationships.

[0028] The acquired shelf images are systematically identified and modeled, transforming the visual information in the images into structured data containing product categories, brands, locations, areas, shelf numbers, and spatial adjacency relationships. This step provides a unified spatial benchmark and reusable foundational data for subsequent combined identification and rule mapping. This is achieved through the following sub-steps.

[0029] S1.1: Perform distortion correction, brightness normalization, and cropping of the main body of the shelf on the original shelf image, and calculate the horizontal and vertical pixel scale parameters to establish a unified spatial reference.

[0030] After acquiring the raw shelf images, the images are preprocessed to eliminate distortions and lighting differences introduced during the acquisition process. Specifically, distortion correction is performed on the raw shelf images, and geometric distortions in the images are corrected using lens calibration parameters.

[0031] Next, brightness normalization processing is performed to adjust the image brightness to a standard range, eliminating the impact of lighting differences under different shooting conditions on the recognition results. Then, the main body of the shelf is cropped, and the boundaries of the shelf area are identified using an edge detection algorithm. Background parts that do not belong to the shelf area are removed from the image, resulting in a standardized shelf image. Based on this, effective areas are extracted from the standardized shelf image. Image segmentation technology is used to identify structural elements such as shelf shelves and back panels, removing the floor, ceiling, price tag hanging areas, and non-display background areas, retaining only the effective areas for product display.

[0032] At the same time, based on the calibration parameters recorded during shooting, the number of pixels per centimeter horizontally is calculated. And the number of pixels per centimeter vertically The value typically ranges from 20 to 120 pixels per centimeter. These two parameters serve as a unified spatial reference for subsequent calculations of the actual projected area of ​​the goods, lateral distances, and shelf divisions, ensuring that all positional and size-related calculations are performed within the same coordinate system.

[0033] S1.2: A deep learning object detection network is used to identify product categories and brands. The overall credibility is calculated by combining boundary integrity and recognition probability, and the actual projected area of ​​the product is converted.

[0034] After image preprocessing, product object detection is performed on standardized shelf images. A deep learning-based object detection network is preferably used, with an improved convolutional detection model as the backbone. Through training on a large number of product samples, this network can simultaneously identify product category and brand. During detection, the confidence threshold is set to 0.55 to 0.80, and the non-maximum suppression threshold is set to 0.40 to 0.65 to balance accuracy and recall. The detection output includes the left boundary of each candidate product. Upper boundary Right boundary Lower boundary and the corresponding category recognition probability Brand recognition probability ,in, and The values ​​range from 0.50 to 1.00.

[0035] To mitigate the impact of incomplete or occluded targets on subsequent scoring, a secondary screening of the detection results is necessary. First, the boundary integrity of each candidate item is calculated. This parameter is defined as the ratio of the effective boundary length to the theoretical boundary length, and its value ranges from 0.60 to 1.00. The effective boundary refers to the portion of the product boundary that is not truncated by image edges or occlusions. Subsequently, the overall credibility of each candidate product is calculated. This value is calculated using a weighted geometric mean: ; in, , , Categories ,brand and boundary integrity The recommended weights are 2 to 5, 1 to 4, and 1 to 3, respectively. This calculation method can effectively prevent situations where an extremely low value for one indicator is masked by averaging when other indicators are high, ensuring that only targets with high overall credibility can proceed to subsequent processing.

[0036] In addition, to calculate the area ratio and eliminate false targets that are too small, it is necessary to calculate the actual projected area of ​​each candidate product. Based on the pixel coordinates of the product bounding box and the aforementioned pixel scale parameters, the area of ​​the pixel box is converted into the projected area of ​​the shelf plane: ; S1.3: Calculate the center coordinates based on the product bounding box, determine the layer number and column area based on the shelf boundary line, eliminate false targets with low credibility and too small area, and form a basic structured data set.

[0037] Based on the candidate product bounding boxes obtained in sub-step S1.2, the center position of each product is calculated for subsequent layer number determination and adjacency relationship determination. The horizontal coordinates of the product center are... and vertical coordinates The values ​​were calculated from the averages of the left and right boundaries and the top and bottom boundaries of the bounding box, respectively. ; ; To determine the shelf number of each product, the longitudinal coordinates of the shelf's shelf boundaries must first be obtained through lateral edge density peak detection. Where m is the shelf division line number. Each shelf division line is determined by the edge position of the transverse support members in the shelf structure. Then, based on the longitudinal coordinates of the product center... Boundary lines between each layer Based on the location relationship, calculate the floor number of the goods. : ;

[0038] The formula sums all shelf boundaries. If the center of a product is below a shelf boundary, the corresponding term is counted as 1; if it is above, it is counted as 0. The final shelf number is the cumulative value plus 1. In this way, continuous vertical positions are mapped to discrete shelf numbers, making the shelf number determination correspond one-to-one with the physical shelves.

[0039] While determining the shelf number, each product also needs to be assigned column information. Columns can be divided equally according to the effective width of the shelf, or unequally according to the actual guide rail spacing; the specific division method can be preset according to the shelf structure characteristics. For candidate products with an overall credibility score below the preset threshold and a projected area smaller than the minimum display area, they are removed from the candidate target set to avoid false targets interfering with subsequent scoring.

[0040] In the final structured dataset, each product record includes a product identifier, product category, product brand, and center coordinates. and Floor number Column number and projected area .

[0041] S1.4: Calculate the spatial adjacency strength between any two products, merge products that meet the adjacency conditions into an initial spatial clustering unit, and record the cluster number, cover layer number, number of products and brand composition.

[0042] After obtaining the basic structured data set, it is necessary to further establish the spatial relationships between products to provide a basis for subsequent combination display identification.

[0043] This sub-step calculates the spatial adjacency strength between any two items, taking into account three dimensions: lateral distance, size consistency, and layer number consistency. Specifically, for the i-th item and the j-th item, the lateral distance between their centers is first calculated. This distance is obtained by dividing the difference in the center horizontal coordinates by the horizontal pixel scale. The average width of the two items is also calculated. The values ​​are calculated from the widths of their respective bounding boxes. Then, the projected areas of the two items are calculated. and and floor number and Spatial adjacency strength The calculation formula is: ; In this formula, the first term The term represents the relative distance in the horizontal direction. When the horizontal distance is less than the average width of the product, this term approaches 1; as the distance increases, this term rapidly decreases. The second term... This term represents dimensional consistency; it approaches 1 when two items have similar areas, and decreases as the area difference increases. (The third term...) This represents the consistency of the tier number; when two products are in the same tier, this item is 1, and the greater the difference in tier numbers, the smaller this item is. The product of these three items... This comprehensively reflects the possibility of the two products forming a continuous display relationship in space. When The spatial adjacency strength exceeds the preset threshold, which is recommended to be between 0.55 and 0.85, and the lateral center distance between the two items is also greater. If the horizontal distance is less than the preset threshold, the two products are considered to be in a continuous display relationship. It is recommended that the horizontal distance threshold be set to 0.15 to 0.40 times the average width of the products, based on the width of a single shelf unit.

[0044] Based on the adjacency determination results, multiple products that meet the adjacency conditions are merged into initial spatial clustering units. Each clustering unit records the cluster number, cover layer number, starting column area, ending column area, product quantity, and brand composition. These initial spatial clustering units constitute the basic input for subsequent identification of combined display semantic units, enabling step S2 to perform higher-level combined identification based on this.

[0045] In the technical solution of this disclosure, the original shelf image is transformed into structured data containing product category, brand, location, area, shelf number, and spatial clustering information through image preprocessing, product target detection, comprehensive credibility assessment, and spatial adjacency relationship construction. This step establishes a unified physical spatial benchmark, making images acquired under different shooting conditions comparable. At the same time, the interference of incomplete edge targets and occluded targets is effectively suppressed through boundary integrity and weighted geometric mean calculation, providing accurate and reusable basic data for subsequent combined display recognition.

[0046] S2: Based on structured data, identify cross-category and spatially continuous product sets, construct combined display units with business semantics, and generate combined-level attributes.

[0047] Based on the structured data output by S1, product sets that meet the conditions of cross-category coexistence and spatial continuity are identified and constructed into combined display units with business semantics. This step organizes discrete product objects into holistic combined objects, providing a semantic carrier for mapping scoring rules from the single-item level to the combined level. This is specifically achieved through the following sub-steps.

[0048] S2.1: Based on the initial spatial clustering units, a region growing merging algorithm is used to identify cross-category and spatially continuous candidate combinations, and the continuous aggregation strength and the area of ​​the outer boundary are calculated.

[0049] Read the basic structured data set and initial spatial clustering units output in step S1, and perform candidate combination search based on the initial spatial clustering units. Unlike traditional single-item level identification, the identification of candidate combinations no longer uses a single product as the evaluation object, but uses cross-category coexistence, spatial continuity, and display relevance as constraints to perform a secondary merging of multiple initial spatial clustering units.

[0050] In practice, the product category, brand, shelf number, column area, center position, projected area, and initial spatial clustering unit of each product object are first obtained. Then, based on spatial adjacency, a region-growing merging algorithm is used for combination search: first, the initial spatial clustering unit with a larger area and more adjacent products is selected as the seed unit, and then expanded to the left and right adjacent areas and adjacent shelf areas. When a new product object meets the cross-category constraint and spatial continuity constraint, it is merged into the current candidate combination. During the expansion process, the cross-category constraint requires that the total number of product categories in the combination is not less than the preset lower limit of the number of categories, which is recommended to be 2 to 4, meaning that only when two or more types of products appear can it be considered a combination display; the spatial continuity constraint requires that the horizontal center distance between the new product and the existing products in the combination does not exceed the horizontal continuity interval threshold, which is recommended to be 0.20 to 0.45 times the average width of a single product, and the shelf number difference does not exceed the shelf number difference threshold, which is recommended to be 1 to 2 layers, to avoid misjudging independent displays that are too far apart as the same combination.

[0051] To quantify the degree of continuous aggregation of candidate combinations, it is necessary to calculate the continuous aggregation strength of each candidate combination. This value is obtained by dividing the sum of the products of the spatial adjacency strength and the class difference of all product pairs within the combination by the square of the third of the number of products in the combination: ; in, Let be the spatial adjacency strength between the i-th and j-th items in step S1, which is dimensionless; This value indicates the category difference between the i-th and j-th products. It is set to 1 when the two products are in different categories and belong to a preset allowed combination relationship; otherwise, it is set to 0. The allowed combination relationship refers to a set of category pairs that are predefined according to business marketing strategies and encouraged for combined display, such as bread and milk, coffee and coffee creamer, potato chips and carbonated drinks, etc. This set of relationships is stored in the system in the form of a configuration file or data table and can be dynamically added or deleted by business personnel. Let be the number of product objects within the u-th candidate combination, ranging from 2 to 50. This formula emphasizes both spatial proximity and category difference within the candidate combination. The denominator uses the square of the third of the quantity to strengthen the penalty, preventing large, ineffective areas from being incorrectly identified as combined displays simply due to an excessive number of products. The calculated continuous aggregation intensity... As an indicator of the aggregation degree of candidate combinations, it is stored in the corresponding attributes of the candidate combinations for subsequent semantic strength evaluation.

[0052] Simultaneously, a unified spatial boundary is established for each candidate combination, and its pixel outer boundary area is calculated. : ; in, The leftmost horizontal pixel coordinate of the candidate combination. The rightmost horizontal pixel coordinate of the candidate combination. The vertical pixel coordinates of the uppermost boundary of the candidate combination. The vertical pixel coordinates of the lowest boundary of the candidate combination are used. The coordinates of each boundary are obtained by statistically analyzing the extreme values ​​of the bounding boxes of the member items in the candidate combination. This boundary area provides a common spatial benchmark for subsequent judgments on the compactness, number of coverage layers, and coverage area of ​​the combination, and is also stored in the candidate combination attributes.

[0053] Through this step, the structured results of a single product are transformed into a candidate combination display set containing information on continuous aggregation intensity and spatial boundaries, enabling subsequent steps to focus on the combined semantic objects rather than individual products.

[0054] S2.2: Calculate the cross-category effectiveness strength by combining fill rate, category diversity and inter-layer span, and select effective combination display sets for real joint display.

[0055] The candidate combination display set output from sub-step S2.1 is filtered for validity, eliminating pseudo-combination areas that, although spanning multiple categories, are not actually jointly displayed. The filtering logic includes three dimensions: the spatial compactness of the products within the combination, the category diversity within the combination, and the degree of hopping between layers of products within the combination.

[0056] In practice, the sum of the projected areas of all goods within each candidate combination is first calculated, and then compared with the area of ​​the outer boundary of the candidate combination to obtain the combination fill rate. : ; in, The sum of the actual projected areas of all items in the u-th candidate combination is calculated by step S1; Let be the area of ​​the pixel boundary of the u-th candidate combination; and These represent the number of pixels per centimeter horizontally and vertically, respectively. The denominator represents the actual area corresponding to the outer boundary of the u-th candidate combination. This formula uses the ratio of the actual area occupied by the items within the combination to the overall envelope area of ​​the combination to measure whether the combination is displayed continuously rather than being spatially adjacent. A fill rate threshold of 0.35 to 0.80 is recommended.

[0057] Then, the number of different product categories within the combination was counted. With different number of brands The interlayer span is obtained by calculating the difference between the maximum and minimum values ​​of the layer numbers to which the products in the combination belong. The values ​​range from 0 to 5. Based on these parameters, the cross-category effectiveness strength of candidate combinations is calculated. : ; In this formula, the more categories and brands there are, the more likely a genuine promotional mix will be formed; the larger the span between layers, the more likely it is to be a non-coordinated display. Therefore, the denominator is used to suppress pseudo-mixes with excessively large spans between layers. It is recommended to set the diversity intensity threshold to 1.20 to 3.50, and the span between layers threshold to no more than 2 layers.

[0058] Only when the fill rate, diversity intensity, and interlayer span all meet the threshold conditions will the candidate combination be retained as a valid combination display set; otherwise, it will be rejected.

[0059] S2.3: Assign a unique identifier to each valid combination, record the list of member products and spatial boundaries, calculate semantic strength and determine the number of coverage layers, and construct basic information about the combination.

[0060] Semantic units are constructed from the set of valid combinations output by sub-step S2.2, and a unified object is created for each combination that can be directly called by the subsequent rule engine.

[0061] Each combination is assigned a unique combination identifier, and a list of product identifiers for each combination member is recorded. The left, right, top, and bottom boundaries of the combination are calculated, and the starting and ending layer numbers, starting and ending column areas are recorded. The product category and brand composition within the combination are also labeled. During the construction process, a threshold-based rule template matching method is used to initially merge combinations. For example, when a combination contains food, dairy products, and beverages and is displayed consecutively in adjacent layers or two layers, it is prioritized as a breakfast combination candidate. When a combination contains areas with prominent price tags, complimentary items, or promotional endplates, it is prioritized as a promotional combination candidate.

[0062] To avoid the same product belonging to multiple semantic units simultaneously, the principle of assigning based on maximum semantic strength is adopted: if a product is included in multiple valid combinations, it is assigned to the combination with the highest semantic strength; the remaining combinations remove the product and re-adjust their boundaries. Semantic strength The calculation comprehensively considers continuous polymerization strength, filler content, and cross-category effectiveness strength: ; in, Let be the continuous polymerization intensity of the u-th candidate combination. Let be the fill rate of the u-th candidate combination. Let be the cross-category effectiveness strength of the u-th candidate combination. , , These are the weighting coefficients for continuous aggregation strength, fill rate, and cross-category effectiveness strength, respectively, with suggested values ​​ranging from 2 to 5, 1 to 4, and 1 to 4. This formula simultaneously constrains three conditions—close clustering, full placement, and genuine category mixing—using a weighted geometric mean. Too low a value for any dimension will significantly reduce semantic strength, thus ensuring strong physical and business consistency of the combined semantic units.

[0063] Simultaneously calculate the number of coverage layers for each combined display semantic unit. : ; in, This is the lowest layer number in this combination. This represents the maximum layer number of the combination, which is obtained by statistically analyzing the layer numbers of the combination's member products. This formula maps the discrete product set to a single coverage layer attribute, providing direct input for the combination coverage layer score in subsequent combination-level rules. At this point, each valid combination possesses a combination identifier, a list of member products, spatial boundaries, semantic strength, and coverage layer number, forming the basic information of the combination.

[0064] S2.4: Match the predefined template library with the category composition and brand composition of the combination, generate combination type identifier, calculate the coverage area and internal product proportion structure, generate combination type identifier and complete combination-level attribute archiving, and construct a complete combination display semantic unit.

[0065] Based on the combined basic information output in step S2.3, combined-level attributes are further calculated for each combined semantic unit to provide rich business features for subsequent rule mapping.

[0066] Based on the category composition, brand composition, coverage layer, and spatial boundary characteristics of the bundled members, a bundle type identifier is generated. This identifier can use business definition tags such as breakfast bundle, promotional bundle, cross-selling bundle, and gift box bundle. The type identifier is generated by comparing it against a predefined bundle type rule template library. Each template in the library consists of a condition combination and a corresponding bundle type tag. The condition combination can include category composition requirements, brand composition requirements, coverage layer range, and spatial boundary characteristics. The system compares the current bundle's category composition, brand composition, coverage layer, and spatial boundary attributes with the condition combinations in the template library one by one. When there is a complete match, the bundle is assigned the corresponding type tag. For example, if the bundle contains food, dairy products, and beverages, and the coverage layer does not exceed two, it corresponds to a breakfast bundle; if the bundle contains small-area gift items and multiple brands appear together, it corresponds to a promotional bundle.

[0067] Then calculate the combined coverage area. : ; in, Let be the area of ​​the bounding boundary of the pixels in the u-th combination. and These are the number of pixels per centimeter horizontally and vertically, respectively. This formula transforms the combination boundary from pixel space to the actual shelf space, which is used for subsequent determination of the combination area ratio, promotion coverage, and historical recalculation.

[0068] Next, the quantity and area ratio of each type of product within the combination are calculated to form the product proportion structure within the combination. Proportion calculated by quantity. We obtain it from the following formula: ; in, Let be the number of product objects of category c in the u-th combination, and the denominator be the sum of the number of product objects of all categories in the u-th combination. If calculated by area percentage, then... Replace with the total area of ​​the corresponding product category. The percentage structure stores both percentages calculated by quantity and percentages calculated by area, allowing for flexible use of subsequent scoring rules based on business priorities. If a combination, after correction, exhibits insufficient categories, insufficient coverage area, or severely unbalanced percentages, it will be relegated to the invalid combination set to prevent pseudo-combinations from entering the rule engine.

[0069] In the final output set of combined display semantic units, each combination carries complete attributes such as combination identifier, combination type identifier, member product list, spatial boundary, number of coverage layers, coverage area, and internal product proportion structure, providing a structured combination object foundation for the semantic reconstruction of execution rules and combination-level rule mapping in step S3.

[0070] In the technical solution of this disclosure embodiment, based on structured data, discrete product objects are constructed into semantic units of combined display with business significance through region growth and merging, compactness evaluation, and semantic strength calculation. This step solves the problem that traditional solutions cannot identify cross-category combined displays, enabling the scoring objects to be upgraded from the single-item layer to the combined layer. At the same time, through multi-dimensional screening of fill rate, cross-category effectiveness strength, and semantic strength, pseudo-combinations that are spatially close but not truly jointly displayed are eliminated, providing the rule engine with semantically clear and well-defined combined objects.

[0071] S3: The original scoring rules are semantically reconstructed through a dynamic rule engine, transforming the flat single-item rule system into a two-layer rule structure that prioritizes combinations and supplements single items.

[0072] Based on the combined display semantic units output by S2, the dynamic rule engine semantically reconstructs the original scoring rules, transforming the flat single-item rule system into a two-layer rule structure consisting of combined priority rules and single-item supplementary rules. This step resolves the conflict problem of multiple rules hitting repeatedly in combined display scenarios, achieving decoupling between rules and recognition results. Figure 2 The flowchart illustrating the rule semantic reconstruction and two-layer rule structure generation process provided in this disclosure embodiment is shown, as follows: Figure 2 As shown, this is achieved through the following sub-steps.

[0073] S3.1: Read the original scoring rules and break them down into five categories of elements: rule object, condition, scoring factor, score range and scope of application. Calculate the clarity of object attribution and the standardized weight value.

[0074] The dynamic rule engine first reads the original set of scoring rules. These rules are typically defined by business personnel through a visual configuration interface and stored in a structured format in the database. The engine breaks down each original scoring rule into five basic elements: rule object, rule condition, scoring factor, score range, and scope of application. During the breakdown process, the rule object does not directly use the original single-item concept. Instead, it first determines whether the rule applies to the product brand, product category, combination type, number of combination coverage layers, combination proportion structure, or combination space boundary. If the original rule contains expressions such as a specified brand must be located in a certain layer, a specified category proportion reaches a certain threshold, or a specified combination of products must appear together, these must be uniformly broken down into computable condition fragments and mapped to the combined display semantic units already formed in step S2.

[0075] The specific decomposition adopts a segmented semantic parsing process: the first stage identifies object keywords in the rules to form a candidate set of object affiliations; the second stage identifies positional constraints, proportion constraints, hierarchical constraints, and parallel constraints to form a set of conditions; the third stage extracts scoring values ​​and threshold ranges to form a set of scoring factors. For cases where multiple conditions exist within the same rule, the original logical relationships must be preserved, such as simultaneously satisfying one condition, satisfying two conditions but not the third, etc. To facilitate subsequent unified reconstruction, all rule conditions should be standardized into numerical or Boolean conditions, and each rule should retain only one main scoring factor.

[0076] During the decomposition process, the rule parsing confidence threshold is set to 0.70 to 0.95 to determine the reliability of the rule element identification results; elements below this threshold are not retained. The minimum number of condition fragments to retain is set to 2 to 8 to filter rules with overly simple conditions. The maximum nesting level of a single rule is no more than 4 levels to avoid excessive execution costs for the rule engine. If an original scoring rule cannot be clearly assigned to a combination object or a single item object, it is first entered into the set of rules to be decided, and its assignment is corrected in subsequent sub-steps.

[0077] To quantify the reconfigurability of each original scoring rule, the clarity of object attribution is calculated. : ; in, This represents the number of object keywords accurately identified in the rule, with a value ranging from 1 to 20. This represents the number of condition fragments that are successfully decomposed in the rule, ranging from 1 to 30. This represents the number of ambiguous condition fragments in the rule, ranging from 0 to 10. This formula measures whether an original rule is sufficiently refactorable; the more explicit the object keywords, the more complete the condition fragments, and the fewer the ambiguities, the more suitable it is for subsequent combination-priority semantic reconstruction. The calculated object attribution clarity... The results of the rule element decomposition are stored in the set of rule element decomposition results, which serve as an indicator of the reconfigurability of the rule.

[0078] Simultaneously calculate the standardized weight value for each original scoring rule. : ; in, This is the local weight corresponding to the p-th condition fragment of the rule, with a value ranging from 0.10 to 1.00, determined jointly by the rule level, historical activation frequency, and marketing priority configured on the business side; The number of condition fragments for this rule, ranging from 1 to 30. This formula first compresses complex rules into a unified weight benchmark, facilitating subsequent comparisons of whether combination rules or individual rules should take precedence. The calculated standardized weight values... Similarly, it is stored in the rule element decomposition result set as a quantitative basis for the importance of the rule.

[0079] As an optional implementation, the dynamic rule engine loads the original set of scoring rules from an external configuration center or database upon startup and listens for rule change events. When business users modify the rules through a visual interface, such as adjusting scoring weights, adding or removing condition fragments, or modifying score ranges, the configuration center pushes a change notification to the engine. The engine only needs to re-execute the current steps S3.1 to S3.5 without restarting the service or modifying the underlying code. This hot update mechanism shortens the iteration cycle of marketing strategies from several days to minutes.

[0080] This step completes the first-level transformation from the original natural business rules to computable rule elements that include reconfigurability and weighting benchmarks, laying the data foundation for subsequent combination-priority reconfiguration.

[0081] S3.2: Based on whether the rule conditions involve multiple categories co-occurring or combined attributes, the rules are divided into combination priority rules and single-item supplementary rules to construct a two-layer rule framework.

[0082] After rule decomposition is completed in sub-step S3.1, the dynamic rule engine begins to perform rule semantic reconstruction, rewriting the original flat structure where all rules directly apply to product objects into a two-layer rule structure consisting of combination priority rules and single-item supplementary rules.

[0083] In practice, the set of objects corresponding to each rule is first read to determine whether it can be directly applied to the semantic unit of the combined display. Rules involving multiple categories appearing together, multiple brands jointly displayed, the number of layers covered by the combination, the area covered by the combination, or the internal proportion structure of the combination are prioritized and assigned to the combination priority rule set. Rules involving only a single product, brand, single item position, or the independent proportion of a single category are assigned to the single item supplementary rule set.

[0084] In constructing combination priority rules, the integrity of the combination object must be preserved. That is, the judgment object of the rule is no longer any single item within the combination, but the entire semantic unit of the combination display. For example, the requirement in the original rule that bread, milk, and coffee appear together and are located on the gold layer to score 20 points is no longer broken into three single-item rules after reconstruction. Instead, a combination priority rule is generated, directly using the semantic unit of the breakfast combination as the scoring object. At the same time, if the original rule also includes the requirement that milk brand A must be on the upper layer, this rule is retained as a supplementary single-item rule, but its effectiveness is premised on that single item not being completely covered by the combination rule.

[0085] This approach employs a rule reconstruction method that prioritizes the main object and then strips away the remaining conditions. First, a main object is identified for each rule. If the main object is a combination, the rule's main scoring weight and main conditions are assigned to the combination priority rule. Then, the remaining conditions that haven't yet been absorbed by the combination object are stripped out to form corresponding supplementary rules for individual items. The main object priority threshold is set between 0.60 and 0.85, and the remaining condition stripping ratio is controlled between 10% and 40% to ensure sufficient completeness of the combination rules.

[0086] To quantify the fit of each original rule reconstructed into a combination-priority rule, firstly, from all the combined display semantic units output in step S2, the combination with the most overlap with the categories or brands mentioned in the current rule is selected as the best-matching combination. If multiple combinations are tied, the one with the highest semantic strength is selected. Calculate the fit strength : ; in, For the clarity of object attribution in the r-th original scoring rule, Let r be the standardized weight value of the r-th original scoring rule. The semantic strength of the u-th combined semantic unit that best matches the rule in step S2 is given. , , These are the weighting coefficients for clarity of attribution, rule weight, and semantic strength of combination, with suggested values ​​ranging from 2 to 5, 1 to 4, and 2 to 5, respectively. This formula uses the clarity of the rule itself, the importance of the rule, and the maturity of the combination objects to jointly determine whether a rule should be elevated to a priority combination rule.

[0087] Simultaneously, calculate the retained weight value of each original scoring rule for the individual product supplementary rule layer after reconstruction. : ; In this formula, the stronger the combination fit, the smaller the weight of the single-item supplement layer; if the combination fit is weak, the rule is retained more in the single-item layer to prevent the rule semantics from being excessively emptied.

[0088] This step completes the transformation from a flat rule structure to a two-layer rule skeleton, providing a clear rule hierarchy for the subsequent establishment of composite-level rule mapping.

[0089] S3.3: Scan the combination priority rules one by one, establish a rule mapping relationship for each combination semantic unit, and generate a scoring expression unit containing the combination identifier, the hit rule identifier, and the upper limit of the score.

[0090] After the two-layer rule skeleton is completed, a rule mapping relationship is established for each combined display semantic unit. The rule mapping relationship refers to clarifying which combination priority rules should match a certain combined semantic unit, which condition in these rules corresponds to the combination type identifier, which condition corresponds to the number of combination coverage layers, which condition corresponds to the combination coverage area, and which condition corresponds to the product proportion structure within the combination. Finally, the rules and scoring factors are bound one by one to form a combination-level scoring expression unit.

[0091] In practice, the set of combined display semantic units and the set of combined attributes output in step S2 are read first, and then the set of combined priority rules is scanned one by one. If a combined priority rule specifies a combined type identifier, it is first determined whether the combined type identifier of the combined semantic unit matches it; if it matches or the rule does not specify a type identifier, conditions such as the number of coverage layers, coverage area, and percentage structure are further checked. If the combined type, number of coverage layers, coverage area, and percentage structure required by a combined priority rule match a combined semantic unit, a mapping relationship is established.

[0092] After establishing the mapping relationship, the multi-condition expressions in the original rules need to be converted into a unified combination-level scoring expression unit. This scoring expression unit should at least include a combination identifier, a matching rule identifier, the number of matching conditions, the total number of conditions, the coverage level, the score cap, and the currently computable score. For cases where a combination may simultaneously match multiple combination priority rules, a rule binding decision should be executed. During the decision, lower-priority rules are not directly deleted; instead, all binding relationships are retained first, and then precise pruning is performed in subsequent sub-steps based on coverage and exclusivity relationships. The focus of the rule binding stage is to minimize missed bindings; therefore, the hit threshold is set relatively loosely, and the combination condition matching rate threshold is set to 0.55 to 0.90.

[0093] To quantify the mapping strength between each combination and each combination priority rule, the mapping strength is calculated. : ; in, This represents the number of conditions that the combination and the rule actually matched. Let r be the total number of conditions for the r-th combination priority rule. Let u be the actual number of coverage layers for the u-th combination. The target coverage layer configured in the r-th combination priority rule. Let u be the actual coverage area of ​​the u-th combination. The target coverage area configured in the r-th combination priority rule. The combination type matching indicator value is set to 1 if the combination type identifier of the u-th combination matches the combination type required by the r-th rule, otherwise it is set to 0. If the rule does not specify a combination type requirement, then... Always set to 1. This formula combines the number of conditions hit and the deviation of key attributes into a single mapping strength value, considering both the number of matches and the similarity of key physical attributes.

[0094] The theoretical score contribution of each combination under each combination priority rule is calculated based on the mapping intensity. : ; in, Let r be the full score for the r-th rule, with a suggested range of 1 to 100. This formula first establishes the theoretical upper limit of the combined score, providing a quantitative basis for the next sub-step to determine which rules should be retained and which should be excluded.

[0095] This step truly binds the rules and the combined objects together, and the output set of combined-level scoring expression units becomes the direct basis for subsequent determination of coverage and exclusivity relationships.

[0096] S3.4: Determine the coverage and exclusion relationships of the combined rules to the individual rules by using the difference between the condition coverage rate and the score contribution, calculate the exclusion strength, and mark the exclusion or reduction status.

[0097] This sub-step addresses the core question of whether a product can continue to participate in brand rules or category rules when it is within a combination. First, it reads the already bound combination priority rules and related individual product supplementary rules for each combination display semantic unit, and then determines the overlap and exclusivity relationships between these two types of rules.

[0098] Coverage refers to a situation where a priority rule for a combination has already absorbed the conditions for a portion of the individual items within the combination, meaning that the corresponding individual item rules should not be applied again. For example, if the breakfast combination rule has already scored bread, milk, and coffee together in the gold layer, then these three types of products within the combination should not be scored separately according to the brand co-occurrence rule. Exclusivity refers to a situation where, once one rule takes effect, it directly prohibits another rule from taking effect simultaneously within the same scope. For example, if a promotional combination rule requires the combination percentage to reach a specific threshold, then the percentage rules for each product within that combination can only be used as supplementary adjustments and cannot be independently scored at the original maximum score.

[0099] At the execution level, the condition coverage rate of each individual supplementary rule is first calculated based on the hit condition distribution in the combined-level scoring expression unit. and exclusivity strength The condition coverage This indicates the proportion of the number of conditions absorbed by the combination rule in the single-item rule out of the total number of conditions; the exclusivity strength. It is calculated by the ratio of the theoretical score contribution value of the combined priority rule to the retention weight value of the single item supplement rule.

[0100] The final state of the single-item replenishment rule is determined according to the following dual-threshold determination rule: like Greater than or equal to the conditional coverage threshold, which ranges from 0.70 to 1.00, and If the difference in scoring contribution is greater than or equal to the threshold, which ranges from 5% to 30%, then the rule for that single item is determined to be completely covered by the combined rule and is marked as excluded. like Greater than or equal to the condition coverage threshold, but If the difference in scoring contribution is less than the threshold, the rule for that item is determined to be partially covered, marked as a reduction state, and its retention weight is adjusted according to the proportion of the remaining conditions. like If the coverage rate is less than the condition coverage threshold, the rule for that single item is determined to have not been effectively absorbed by the combined rule and is retained as a supplementary rule for that single item.

[0101] This system employs a dual-judgment mechanism based on condition coverage and score contribution difference to avoid false positives caused by simple priority. The condition coverage threshold is set to 0.70 to 1.00, and the score contribution difference threshold is set to 5% to 30%.

[0102] To quantify the extent to which the supplementary rules for individual items are covered by the combination priority rules, calculate the condition coverage rate. : ; in, This represents the number of conditions in the s-th single-item supplementary rule that have been absorbed by the r-th combination priority rule. This formula represents the total number of conditions for supplementing the rules for the s-th item. It quantifies how many item rules' semantics the combined rules cover, providing a direct basis for subsequent decisions on exclusion, reduction, or retention. The calculated... Store it in the coverage relationship set as a quantitative indicator of the degree to which the rule for that single item is covered.

[0103] Simultaneously calculate the exclusivity strength of each individual item supplement rule when facing the combination priority rule. : ; in, Let u be the theoretical score contribution value of the u-th combination under the r-th combination priority rule. This is a dimensionless weight value for the retained weight of the s-th supplementary rule for a single item. In this formula, the higher the contribution value of the combined rule and the lower the remaining retained weight of the single item rule, the more likely that single item rule should be excluded to avoid double scoring. The calculated... Stored in the set of exclusive relationships as the basis for determining the strength of repulsion.

[0104] As an optimized implementation method, the adjudication results for coverage and exclusivity relationships are not static. The system allows business personnel to set dynamic priority labels for different rules; for example, "promotional rules" have higher priority than "routine rules." When the same combination simultaneously hits multiple combination priority rules and there is a mutual exclusion relationship, the engine dynamically adjusts the adjudication result based on the currently effective priority label, rather than relying on a fixed rule number. This design allows the same scoring logic to adapt to rapid switching between various marketing scenarios.

[0105] This step quantifies the adjudication of covering and exclusive relationships, and stores the adjudication results in the corresponding sets of covering and exclusive relationships, providing a basis for conflict resolution in generating the final executable rule expression.

[0106] S3.5: Generate the final executable rule expression result, including the rule version number, execution sequence number and current valid status, forming a stable two-layer scoring framework.

[0107] After determining the coverage and exclusivity relationships, the dynamic rule engine generates the final executable rule expression result.

[0108] In practice, the combination priority rule is first executed on each combination display semantic unit at the combination layer, and then the single item supplement rule is executed on the single item objects that are not fully covered or excluded, thus forming a stable two-layer scoring framework.

[0109] When generating the final executable rule expression, the following information must be saved for each combined display semantic unit: combined identifier, list of priority rules for valid combined items, list of supplementary rules for reduced items, list of rules for excluded items, current executable weight of each rule, and upper limit of combined layer score. For items that do not belong to any combined display semantic unit, their supplementary rule channel is directly retained. This ensures that there is no duplicate scoring within the combined unit and that items outside the combined unit are not missed.

[0110] In project implementation, the final executable rule expression results are stored in the database in the form of structured rule fragments. Each rule fragment contains a rule version number, a combination identifier, an execution sequence number, and a current valid status. This allows for simpler calculation of the score under combination constraints in subsequent step S4, where the rules can be called in the execution order without requiring further rule semantic adjudication. The rule version number is managed in a two-tiered system of major and sub-versions. The execution sequence number prioritizes combination-priority rules before individual supplementary rules. The current valid status is expressed using three values: enabled, reduced, and excluded.

[0111] To quantify the total executable score potential of each combination under the final two-level rule structure, the total executable score potential value is calculated. : ; in, This is the set of valid combination priority rules for combination u. This is the theoretical score contribution value of the combination under the r-th effective combination priority rule; This is a set of supplementary rules for individual items related to the products within the combination u; The theoretical maximum score for the supplementary rule for item s before any reductions. This is the condition coverage rate of the supplementary rule for the s-th individual item, which is covered by the corresponding combination priority rule. This formula first includes the scoreable value of the combination layer, and then includes the remaining score capability of the individual item layer that has not been covered, thus forming a unified upper limit boundary for subsequent scoring calculations.

[0112] Through this step, the final two-layer rule structure set, the final combination-level rule mapping result, and the executable rule expression result will be directly used as input for the subsequent step S4 to perform scoring calculation under combination constraints. Thus, the previously flat rule system that would cause rule conflicts has been reconstructed into an executable, adjudicable, and duplicate scoring-avoiding two-layer rule structure, truly solving the problem of multiple rules repeatedly hitting the same spatial area in a combined display scenario.

[0113] In the technical solution of this disclosure, the original scoring rules are semantically reconstructed using a dynamic rule engine, transforming the flat, single-item rule system into a two-layer structure consisting of combination priority rules and single-item supplementary rules. This step fundamentally solves the conflict problem of multiple rules repeatedly hitting within the same spatial area in a combined display scenario, while also decoupling the rules from the image recognition results. Business personnel do not need to modify the program code when adjusting the scoring rules; rule changes can be dynamically and hot-updated, significantly improving the flexibility and responsiveness of marketing strategy iteration.

[0114] S4: Execute the overall combined score calculation and individual item supplementary score under the constraints of the two-level rules, and merge them to generate a unified shelf score result.

[0115] Based on the two-layer rule structure and mapping results output by S3, a score calculation is performed on each combined display semantic unit. First, the overall score of the combination is calculated, then supplementary individual scores are performed on product objects not covered by the combination rules, and finally, a unified shelf score result is generated. This step achieves refined scoring under combination constraints, taking into account both the overall value of the combination and the differentiated contribution of individual products. Specifically, it is implemented through the following sub-steps.

[0116] S4.1: Verify the type, number of layers, area and proportion conditions of the combination priority rule for each combination semantic unit, and generate a valid execution set containing the comprehensive hit rate.

[0117] Read the final two-layer rule structure set, the final combination-level rule mapping result, and the set of combined display semantic units output in step S3. For each combined display semantic unit, obtain its combination identifier, combination type identifier, and combination coverage layer number. Combined coverage area The system identifies the product composition structure within each combination and the list of combination priority rules associated with it. Then, it verifies the execution conditions of each combination priority rule, determining whether the combination meets the type, level, area, and proportion conditions specified in the rule.

[0118] During the specific verification process, the actual number of coverage layers in the combination will be compared with the target number of coverage layers required by the rules. The actual coverage area of ​​the combination is compared with the target coverage area required by the rules. The comparison involves comparing the proportion of each type of product within the combination with the proportion threshold required by the rules. If a semantic unit of a combination hits multiple combination priority rules simultaneously, all of them are retained as candidate hit results. However, only rules with the enabled status are included in the effective execution set. Rules with the reduced or excluded status are only recorded in this sub-step for their hit status and do not directly participate in the score calculation.

[0119] At the execution level, a rule-by-rule verification and combination-by-combination caching approach is adopted. That is, a set of combination rule hit results is established with the combination semantic unit as the main index and the hit rule number as the secondary index. This set stores at least the number of hit conditions. Total number of conditions The target threshold, actual value, and current mapping strength are specified. The hit determination threshold for combined rules is set to 0.60 to 0.95. For area-based rules, a floating tolerance can be set during rule configuration, with a recommended range of 5% to 15%. For layer-based rules, an offset tolerance can be set, with a recommended range of 0 to 1 layer, to enhance the robustness of rule determination.

[0120] To quantify the overall hit rate of each combination for each combination priority rule, the overall hit rate is calculated. : ; in, Let be the number of conditions that were actually hit for the u-th combination. Let r be the total number of conditions for the r-th combination priority rule. Let u be the actual coverage area of ​​the u-th combination. Let r be the target coverage area required by the r-th combination priority rule. Let u be the actual number of coverage layers for the u-th combination. Let be the target coverage layer required by the r-th combination priority rule. This formula simultaneously characterizes the hit condition ratio and key physical attribute deviation, avoiding the neglect of layer and area deviations simply because the number of hit conditions is high. The calculated... The overall hit rate of the rule is stored in the combined rule hit result set, which is used in subsequent steps to quantify the contribution of the hit rate to the score when calculating the theoretical score of a single rule.

[0121] This step transforms the executable rule expression result output in step S3 into a rule hit status that includes a comprehensive hit index, and includes rules with the status of "enabled" into the effective execution set, providing a clear quantitative basis for the target objects and their hit degree for subsequent calculation of the overall combined score.

[0122] S4.2: Calculate the theoretical score of a single rule by combining the rule's full score, overall hit rate, and execution weight. Compress and merge the scores of multiple rules to obtain the combined overall score.

[0123] For each combined object to be scored, the overall combined score is calculated one by one according to its effective combined priority rules. This overall combined score is not simply the sum of the full scores of the hit rules, but rather a weighted calculation combining rule hit rate, rule full score, and the execution weight of the rule in the current combined object. If a combined object hits multiple effective combined priority rules simultaneously, the theoretical score for each rule is first calculated, and then normalization is performed based on the coexistence relationship between rules to prevent excessive stacking of multiple combined rules on the same combined object.

[0124] In practice, a two-level scoring model is adopted: the first level calculates the theoretical score contribution of a single combination rule to the current combination; the second level compresses and merges all theoretical score contributions for the combination to obtain the final overall score. The compression and merging coefficient is set to 0.75 to 0.95 to suppress the repeated amplification of the same combination by multiple highly overlapping combination rules. If a combination only hits one combination priority rule, its overall score can be directly equal to the theoretical score contribution value of that rule. After obtaining the overall score of the combination, it is also necessary to calculate how much of the total scoring capacity the combination has already occupied and how much score space can be reserved for individual supplementary rules. It is particularly important to note that once the overall score of the combination is formed, the product objects within the combination cannot be scored again under the same conditions within the scope of conditions already covered by the combination rules.

[0125] Single Rule Theory Score Calculated by the following formula: ; in, Let r be the maximum score for the r-th combination priority rule. A value between 1 and 100 is recommended. Let be the overall hit rate of the u-th combination with the r-th combination priority rule. Let be the current execution weight of the r-th combined priority rule. It is dimensionless, ranging from 0.10 to 1.00, and is given by the rule decision result in step S3. This formula maps the hit rate to a scoreable value and adds the rule's own importance to form the basic contribution value of the combined score.

[0126] The total scores of multiple rules within the combination are then compressed and merged to obtain the overall combination score. : ; in, This is the sum of the single-rule theoretical scores for this combination under all valid combination priority rules. This represents the number of valid combination priority rules that this combination hits, ranging from 1 to 20. This is the rule overlap compression factor, and a value between 0.20 and 0.80 is recommended. This formula performs controlled compression on the total score when multiple rules are matched in a combination, preventing the score from being abnormally inflated due to the superposition of multiple rules.

[0127] This step directly yields the overall score set of the combination, which is the first main scoring channel in the entire step S4. At the same time, the output set of remaining assignable weights for the combination will be used by subsequent sub-steps to determine which individual items can still be scored.

[0128] S4.3: Based on the coverage and exclusivity relationships, classify the items within the combination, trim the individual item rule conditions that have been absorbed by the combination rules, and obtain the set of residual individual item conditions.

[0129] After the overall score of the combination is calculated, a unified attribution process is performed on product objects belonging to the same combined display semantic unit. This is because the overall score of the combination calculated in step S4.2... The combined rules have already reflected all the business value covered by the rules. To avoid double scoring, subsequent supplementary scoring for individual products will only be conducted on semantics that are not covered or not fully covered by the combined rules.

[0130] Based on this principle, the unified attribution process involves first locking the product member list by combination object, and then checking each product object to see if the corresponding individual product supplementary rules have been covered by the combination rules, excluded, or only partially reduced. If an individual product rule is completely covered by the combination rules, the rule is directly removed from the individual product scoring channel; if it is only partially covered, only the uncovered condition fragments are retained to form the individual product residual condition set; if it is completely uncovered, it is retained as is and added to the individual product executable rule set.

[0131] In terms of implementation, a reverse lookup approach within a product group is adopted: using the product identifier as an index, the group identifier to which it belongs is retrieved in reverse, and then the corresponding set of coverage relationships and exclusive relationships for that group are read. If a product does not belong to any group display semantic unit, all individual product rules are directly retained without reduction. For partially reduced individual product rules, their remaining execution weights need to be recalculated. The lower the remaining execution weight, the more the group rule has absorbed its scoring semantics. The remaining execution weight is limited to between 0.05 and 1.00; a weight below 0.05 can be directly considered invalid. The focus of this sub-step is not to immediately calculate the score, but to clarify which individual products can still be scored, under what conditions, and with how much weight they can still be scored, in preparation for the next sub-step to supplement individual product scoring.

[0132] To quantify the upper limit of the residual executable score for each individual supplementary rule under the current combination constraints, calculate... : ; in, The theoretical maximum score for the supplementary rule for item s before any reductions. This is a dimensionless formula representing the condition coverage rate of the supplementary rules for this item, which are covered by the corresponding combined priority rules. This formula directly converts the amount of semantics absorbed by the combined rules into how much scoreable space the item still retains. The calculated... The remaining executable score upper limit is stored in the individual item residual condition set. Simultaneously, the remaining execution weight of each individual item supplementary rule under the current combination constraints is calculated. : ; in, For the s-th item, supplement the retention weight value formed in step S3. This formula calculates the condition coverage rate of the supplementary rule for this individual item, which is covered by the corresponding combination priority rule. This formula further establishes the coverage relationship at the rule execution level; the higher the coverage, the lower the weight of the individual item's rule remaining in effect. The calculated... As the remaining execution weight, it is also stored in the individual product residual condition set.

[0133] This step transforms the items within the combination from a state where they may be scored repeatedly to a state where they are scored based on remaining semantics. This is a key control step to avoid repeated scoring. The output, which includes the upper limit of the remaining executable score and the remaining execution weight, is the set of executable rules and the set of residual conditions for each item. These will be the direct input to sub-step S4.4.

[0134] S4.4: Perform supplementary scoring on individual items for items within the combination based on the residual condition set, and directly score items outside the combination according to the complete individual item rules to form a supplementary score set for individual items.

[0135] Supplementary scoring is performed on all candidate items for individual item scoring. For items within a combination, scoring is based solely on their residual condition set; for items outside a combination, scoring is performed directly according to the complete individual item rules. The scoring method still uses the rule's maximum score multiplied by the hit rate multiplied by the remaining execution weight to form the individual rule score. Multiple individual item rules hitting the same item are then partially merged to prevent duplicate amplification at the individual item level.

[0136] The hit rate calculation here follows the conditional hit concept from step S4.1, but the object is changed from combined semantic units to product objects. For example, for a single-item rule that the brand must be located in a specified layer, the hit rate is determined by the deviation between the actual layer number and the target layer number. If the actual layer number equals the target layer number, the hit rate is 1; the larger the deviation, the lower the hit rate. For a rule that the area percentage of a single item reaches a threshold, the hit rate is determined by the degree of fit between the actual percentage and the target percentage. To ensure consistency in scoring criteria between products outside and inside the combination, the single-item score merging and compression coefficient is set to 0.10 to 0.50.

[0137] After execution, three types of results are generated: the first type is the set of supplementary scores for individual items within the combination, the second type is the set of scores for individual items outside the combination, and the third type is the set of supplementary scores for all individual items. This preserves the value of individual items that are not covered within the combination and avoids the products within the combination being scored repeatedly and thus receiving full marks.

[0138] Single rule score for single item replenishment rules Calculated by the following formula: ; in, The upper limit of the residual executable score for supplementary rules for the s-th item. Let be the conditional hit rate of the i-th product object for the s-th supplementary rule, dimensionless, with a value ranging from 0 to 1. The remaining execution weight for the s-th item supplementary rule is dimensionless. This formula unifies the remaining score, the actual degree to which the conditions are met, and the remaining effective weight of the current rule to obtain the actual executable item supplementary score.

[0139] This step completes the supplementary scoring at the individual product level, providing detailed individual product scoring for the final integration.

[0140] S4.5: Summarize the overall combined score and the individual supplementary score, perform upper limit constraints and outlier checks, and output a unified shelf score result and detailed scores for each dimension.

[0141] The combined scoring results and individual item supplementary scoring results are merged to generate a unified shelf scoring result. In practice, the overall scores of all combinations are first summarized, then the supplementary scores of individual items within a combination and the scores of individual items outside the combination are summarized to form the original total shelf score. Subsequently, upper limit constraints, outlier checks, and scoring detail encapsulation are applied to the total score, outputting result data that can be directly displayed on the front end and used for historical recalculation. This fusion is not a simple summation; it also requires boundary control of the total shelf scoring capacity to prevent the total shelf score from exceeding the business-defined upper limit due to a large number of combined and individual item rules simultaneously.

[0142] In terms of business operations, the maximum total score for a shelf is usually set to a fixed value such as 100, 120, or 200 points. If the maximum score is reached, the incremental scores at the end of the combination layer and the individual product layer are compressed proportionally. The output results should include at least the shelf identifier, total score, summary of combination scores, summary of individual product supplementary scores, detailed scores for each combination semantic unit, detailed supplementary scores for each product object, and the rule version number on which it is based.

[0143] Original total score of the shelf and final total score Calculated using the following formulas respectively: ; ; in, The sum of the overall scores for all combined semantic units in the display. Supplement the total score of all individual items according to the rules. The recommended upper limit for shelf ratings set for the business is between 80 and 200 points. The `max` function is used to initiate compression when the original total score exceeds the upper limit. The first formula aggregates the combined rating channel and the single-item supplementary rating channel. The second formula truncates the results that exceed the rating upper limit, ensuring that the final total score does not exceed the upper limit and that the score remains within the boundaries set by the business.

[0144] This step outputs a unified shelf rating result, a detailed score for combinations, and a detailed score for individual item replenishment. This output will be directly used as input for the subsequent step S5, which performs structured storage and rule reuse processing of the rating results.

[0145] In the technical solution of this disclosure embodiment, scoring calculation is performed under the constraints of a two-layer rule structure. First, the overall score of the combination is calculated according to the combination rules. Then, supplementary scoring is performed on individual items not covered by the combination rules, and double scoring is avoided through the adjudication of coverage and exclusivity relationships. This step achieves a balanced quantification of the overall value of the combined display and the differentiated contribution of individual items. The scoring results reflect both the marketing synergy of the combined display and retain a detailed evaluation at the individual item level. At the same time, the rationality and business interpretability of the scoring results are ensured through compression and merging and upper limit constraints.

[0146] S5: Archive the scoring results together with the combined semantic units, and quickly re-scorp based on the stored structured data and build a multi-version scoring chain when the rules change.

[0147] The scoring results output from S4 are jointly stored with the combined semantic units. When the scoring rules change, steps S3 and S4 are directly re-executed based on the stored structured data, without the need for re-image recognition, thus achieving rapid re-scoring of historical data. This step completely decouples the rules from the recognition results, significantly reducing the computational cost of rule iteration. Specifically, this is achieved through the following sub-steps.

[0148] S5.1: Jointly archive shelf rating results, combined display semantic units, and product structured data to construct a historical rating master record that includes a master record table, a combination detail table, and an individual product detail table.

[0149] The unified shelf scoring results, combination score details, and individual item replenishment score details output in step S4 are jointly archived with the combination display semantic unit set output in step S2 and the product structured result set output in step S1 to construct a historical scoring master record. This master record includes at least the shelf identifier, image identifier, scoring time, total score, combination score, individual item replenishment score, list of combination display semantic unit identifiers, list of product structured object identifiers, and the rule version number on which this scoring was based.

[0150] When storing, not only the numerical results are saved, but also the binding relationship between the results and the structure is saved. That is, each rating detail must be able to be traced back to the corresponding combined semantic unit and the corresponding product structured object. In this way, when the rules change in the future, the existing combined structure and product structure can be read directly without having to re-identify the image.

[0151] In terms of database design, a hierarchical archiving approach is adopted, consisting of a master record table, a combination detail table, an individual item detail table, and a rule version table. The master record table stores shelf-level scoring overviews, including fields such as shelf identifier, image identifier, scoring time, total score, summary of combination scores, summary of individual item supplementary scores, and rule version number. The combination detail table stores the score and boundary information of each combination display semantic unit, including fields such as combination identifier, combination type, overall combination score, combination coverage area, number of combination coverage layers, and the proportion of goods within the combination. The individual item detail table stores the individual item supplementary scoring results, including fields such as product identifier, associated combination identifier, matched rule identifier, single rule score, and residual condition information. The rule version table stores fields such as rule release time, version number, activation status, and rule summary. The tables are linked through scoring task identifiers, which use a combination of timestamp, shelf identifier, and rule version number to ensure that multiple independent historical records can be generated for the same shelf under different rule versions.

[0152] To verify the integrity and consistency of archived data, an archive consistency coefficient is calculated. : ; in, The total archived score for the q-th scoring task is calculated by summing the combined scores. Summary of scores for individual items The sum is obtained, that is , This is the final total score for the q-th scoring task output in step S4. This represents the upper limit of the score for the q-th scoring task. This formula converts the difference between the archived total score and the final total score into a consistency coefficient; the coefficient is 1 when they are completely consistent, and the larger the difference, the closer the coefficient is to 0. This coefficient is used to subsequently determine whether the archived data is complete. The calculated... As an archival quality indicator, it is stored in the scoring archive record set.

[0153] This step transforms the computable process into a reusable historical record containing archived consistency coefficients, establishing a stable data foundation for subsequent rule reuse and rapid re-evaluation.

[0154] S5.2: Create reusable indexes by shelf dimension, version dimension, and time dimension, generate structure reuse pointers, and calculate the validity and reusability of historical indexes.

[0155] After completing the basic archiving, a reusable index is built for the historical records. This reusable index doesn't simply store a single rating result; it stores the retrieval entry points that support direct re-execution of steps S3 and S4. To this end, shelf identifiers, image identifiers, combination identifiers, rule version numbers, rating times, and combination type identifiers are extracted from the rating archive records to form a multi-dimensional index. Then, combination boundaries, combination coverage layers, combination coverage area, and the proportion of goods within a combination are extracted from the combination structure archive set to form a structure reuse pointer. Subsequently, when rules change, the system does not re-invoke the image recognition module. Instead, it directly locates the corresponding historical shelf's structured product result set and combination display semantic unit set based on the re-rating retrieval key, and then injects the new rule version into the execution chain of steps S3 and S4, thus completing rapid re-rating.

[0156] When building the index, it is organized into three levels: shelf dimension, version dimension, and time dimension. The shelf dimension is used to locate the historical records of the same physical shelf. By using the shelf identifier, you can query all the rating records of the shelf at different points in time. The version dimension is used to distinguish the rating results under different rule versions. By using the rule version number, you can compare the rating differences of the same shelf under different rules. The time dimension is used to support trend analysis and interval recalculation. By using the rating time, you can analyze the changing trend of display quality within a certain period of time.

[0157] For pointers that reuse composite structures, the structural integrity state is additionally saved. If the composite structure of a certain historical record is incomplete, it is prohibited from participating in the fast re-evaluation to avoid incorrect recalculation of damaged historical data.

[0158] To quantify the ability of each historical record to support subsequent re-scoring, the effectiveness of the historical index is calculated. : ; in, This represents the number of combined semantic units archived in this scoring task, ranging from 1 to 100. This represents the number of archived product structured objects in this scoring task, ranging from 1 to 1000. This represents the number of missing key reusable fields in the scoring task, ranging from 0 to 20. This formula measures whether a single historical record is sufficient to support subsequent re-scoring; the more complete the combined and product structures and the fewer missing fields, the higher the index effectiveness. The calculated... As a measure of the validity of the historical index, it is stored in the historical scoring index set.

[0159] Simultaneously calculate the reusability of each historical record to the new rule version. : ; in, To determine the validity of the historical index for this scoring task, This is the version sequence value corresponding to the new rule version number. The version sequence value is an integer that increments sequentially according to the release time of the rule version. The earliest version has a sequence value of 1, and the sequence value increases by 1 for each subsequent new version. This is the version sequence value corresponding to the old rule version number of the original historical record. In this formula, the more complete the historical structure and the smaller the difference with the new rule version, the more suitable the record is to directly enter the fast re-evaluation process. The calculated result... As a measure of reusability readiness, it is stored in the historical rating index set.

[0160] This step establishes a retrieval channel from historical records to re-scoring that includes index validity and reusability readiness, providing efficient index support for subsequent batch re-scoring.

[0161] S5.3: When a new rule version is released, directly read the historically stored product structured results and combined semantic units, and re-execute steps S3 and S4 to achieve rapid re-scoring without re-identifying the image.

[0162] When the business side releases a new rule version, the system reads the historical rating index set within the shelf area to be updated, and finds the corresponding product structured result set and combined display semantic unit set based on the re-rating retrieval key. Subsequently, instead of performing image recognition, these historical structures are directly input into step S3 to re-execute rule semantic reconstruction and combined-level rule mapping processing. The new output results from step S3 are then input into step S4 to re-execute the rating calculation processing under combined constraints, ultimately obtaining the recalculated rating result set under the new rule version. The key to this process lies in reusing which old structures and replacing which new rules. Therefore, during recalculation, the product structure and combined structure must be frozen, allowing only changes to the rule version and rule-driven rating paths. Historical recognition results and historical combination boundaries cannot be written back or modified. This ensures that the rating differences between different rule versions truly stem from rule changes, rather than recognition structure drift.

[0163] In terms of execution scheduling, a batch recalculation mode is adopted, which can generate historical re-scoring tasks in batches according to store, shelf type, or time interval. The batch size of the task is set to 100 to 5000 historical records to avoid excessive database read pressure due to excessively large single batch tasks. If the rule change is limited to the adjustment of numerical parameters such as scoring factor weight, score threshold, and floating tolerance, and the rule type, condition structure, coverage relationship, and exclusivity relationship remain unchanged, it is considered that the rule difference is small, and only step S4 needs to be re-executed; if the rule type, condition structure, rule object ownership, coverage relationship, or exclusivity relationship changes, steps S3 and S4 must be re-executed in their entirety. The above judgment is based on the rule type identifier and condition structure hash value recorded in the rule version management system.

[0164] To quantify the structure reuse efficiency of batch re-scoring tasks, the structure reuse efficiency is calculated. : ; in, This represents the number of structured product objects directly reused in this batch, i.e., the number of individual items, with a value ranging from 1 to 100,000. This represents the total number of structured objects for all products in the historical records of this batch, ranging from 1 to 100,000. This formula quantifies the proportion of old structures directly reused in this recalculation. The calculated... As a measure of structural reuse efficiency, it is stored in the historical recalculation task set. Simultaneously, the overall recalculation effectiveness is calculated. : ; in, This formula determines the reusability of the corresponding historical records. It couples structural reuse efficiency with rule version adaptability to assess whether the current recalculation has sufficient engineering value. The calculated... As part of the overall recalculation validity, it is also stored in the historical recalculation task set.

[0165] This step truly achieves the core goal of eliminating the need for image re-identification after rule updates, and evaluates the reuse efficiency and effectiveness with quantitative indicators. It is the most critical reuse step in this scheme compared to traditional image scoring schemes.

[0166] S5.4: Re-archive the recalculation results according to the new rule version, store them in parallel with the original historical records to form a multi-version scoring chain, and output the total score difference between the old and new versions and the difference in the hit rules for comparative analysis.

[0167] The recalculated scoring results are re-archived according to the new rule version and stored in parallel with the original historical scoring records, forming a multi-version historical scoring chain for the same shelf under different rule versions. Each new version historical scoring record must save the new rule version number, the reused combination structure version identifier, the recalculation time, the total score, the combination score, the single item supplement score, and the difference value corresponding to the old version. At the same time, the system compares the results of the new and old versions using rule reuse, and outputs at least the difference in total score, the difference in combination score, the difference in single item supplement score, and the difference in rule hit. Through these comparison results, the business side can quickly determine whether a rule version adjustment has improved the ability to depict the target display scenario.

[0168] Ultimately, the multi-version scoring chains, combined structure archives, and product structure archives are all written into the final traceable scoring history database, providing a foundation for subsequent trend analysis, rule backtesting, and marketing effectiveness verification. In terms of engineering implementation, version sequence management is implemented for multiple versions of results from the same shelf on the same day, preserving the complete scoring chain without overwriting old records. This allows for rapid rescoring and review of exactly which scoring components were altered by each rule change.

[0169] To quantify the degree of score change between the old and new rule versions, the relative rate of change in total score was calculated. : ; in, For the q-th historical record, the final total score under the new rule version is... This represents the final total score for the q-th historical record under the old rule version. This formula normalizes the score changes caused by rule modifications and measures the impact of rule reuse on historical results. The calculated... As a key indicator of the rule reuse comparison results, it is stored in the rule reuse comparison result set.

[0170] Through step S5.4, the construction of a multi-version historical scoring chain containing the relative change rate of total score and the comparison of rule reuse were completed, providing data support for business decisions and realizing in-depth mining of the value of historical data.

[0171] In the technical solution of this disclosure, scoring results, combined semantic units, and product structured data are jointly archived to construct a historical record system containing multiple versions of scoring chains. When rules change, the system can directly reuse the stored product structure and combination structure for rapid re-scoring without having to re-call the image recognition model. This step enables low-cost backtracking analysis of historical data and rule iteration verification, significantly reducing computational costs, while providing quantitative evidence for marketing strategy optimization through multi-version comparative analysis.

[0172] This disclosure also provides an intelligent scoring system for product display images based on dynamic rule configuration. Figure 3 This is a schematic diagram of a product display image intelligent scoring system based on dynamic rule configuration, provided according to an embodiment of this disclosure. This system is used to run the product display image intelligent scoring method based on dynamic rule configuration described in the above embodiments. (Refer to...) Figure 3 The system may include: The image recognition and modeling module is used to recognize and perform basic structured modeling on the collected shelf images, generating a structured dataset. The combined semantic unit construction module is used to identify a collection of products that are cross-category and spatially continuous based on the structured data set, and to construct a combined display semantic unit with business semantics. The rule dynamic reconstruction module is used to semantically reconstruct the original scoring rules through a dynamic rule engine, transforming the flat single-item rule system into a two-layer rule structure consisting of combination priority rules and single-item supplementary rules, and establishing a mapping relationship between each of the combined display semantic units and the combination priority rules. The scoring calculation module is used to perform overall score calculation on the combined display semantic unit according to the mapping relationship under the constraints of the two-layer rule structure, and to perform supplementary individual scores on products not covered by the combined priority rule, and to generate shelf score results by merging them. The archiving and re-scoring module is used to jointly archive the shelf scoring results with the combined display semantic units and the structured data set, and in response to changes in scoring rules, to re-trigger the rule dynamic reconstruction module and the scoring calculation module based on the archived combined display semantic units and the structured data set to achieve rapid re-scoring.

[0173] According to embodiments of this disclosure, an electronic device is also provided, which may include a processor, a communications interface, a memory, and a communication bus, wherein the processor, the communications interface, and the memory communicate with each other via the communication bus. The processor can invoke logical instructions stored in the memory to execute the methods provided in the above embodiments.

[0174] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and sold or used as independent products, and can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this disclosure, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this disclosure. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0175] On the other hand, this disclosure also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the methods provided in the above embodiments.

[0176] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0177] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0178] It should be understood that the above embodiments are only used to illustrate the technical solutions of this disclosure, and not to limit them; although this disclosure has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features; and 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 this disclosure.

Claims

1. A method for intelligent scoring of merchandise display images based on dynamic rule configuration, the method comprising: The method includes: S1. Recognize and perform basic structured modeling on the collected shelf images to generate a structured data set; S2. Based on the structured data set, identify a collection of products that are cross-category and spatially continuous, and construct a combined display semantic unit with business semantics; S3. The original scoring rules are semantically reconstructed through a dynamic rule engine, transforming the flat single-item rule system into a two-layer rule structure consisting of combination priority rules and single-item supplementary rules, and establishing a mapping relationship between each of the combined display semantic units and the combination priority rules. S4. Under the constraints of the two-layer rule structure, the combined display semantic unit is calculated according to the mapping relationship, and the individual product supplementary score is performed on the products not covered by the combined priority rule, and the shelf score result is generated by fusion. S5. The shelf scoring results are jointly archived with the combined display semantic unit and the structured data set. In response to changes in the scoring rules, S3 and S4 are re-executed based on the archived combined display semantic unit and the structured data set to achieve rapid re-scoring.

2. The method of claim 1, wherein, S1 includes: The shelf images are preprocessed to establish a unified spatial reference, and a deep learning object detection network is used to identify the goods and calculate the comprehensive credibility and actual projected area. Based on the product bounding box, determine the layer number and column area, and generate a basic structured data set. Each product record includes product identifier, product category, product brand, center coordinates, layer number, column area number, and projected area. Calculate the spatial adjacency strength between any two products, merge products that meet the adjacency conditions into an initial spatial clustering unit, and record the cluster number, cover layer number, starting column area, ending column area, product quantity and brand composition for each clustering unit; The structured dataset includes the basic structured dataset and the initial spatial clustering unit.

3. The method of claim 1, wherein, S2 includes: Based on the initial spatial clustering units, a region growing merging algorithm is used to identify cross-category and spatially continuous candidate combinations; By combining fill rate, category diversity and inter-layer span, the cross-category effectiveness strength is calculated, and effective combinations are selected from the candidate combinations. A unique identifier is assigned to each valid combination, the member product list and spatial boundaries are recorded, the semantic strength and the number of coverage layers are calculated, and the basic information of the combination is formed. Based on the basic information of the combination, according to the category and brand composition of the effective combination, a predefined template library is matched to generate a combination type identifier. The coverage area is calculated according to the spatial boundary, and the internal product proportion structure is calculated according to the member product list. The combination display semantic unit containing the combination identifier, member product list, spatial boundary, number of coverage layers, coverage area, internal product proportion structure and combination type identifier is constructed.

4. The method of claim 1, wherein, S3 includes: The original scoring rules are broken down into rule objects, conditions, scoring factors, score ranges, and applicable scope; Based on whether the rule conditions involve multiple categories co-occurring or combined attributes, the rules are divided into combination priority rules and single-item supplementary rules, thus constructing a two-layer rule framework; Scan each of the combination priority rules one by one, establish a mapping relationship between each combination display semantic unit and the combination priority rule, and generate a scoring expression unit that includes a combination identifier, a hit rule identifier, and a score limit. Based on the scoring expression unit, the coverage and exclusion relationships of the combination priority rule on the single item supplement rule are determined by the difference between the condition coverage rate and the scoring contribution, and the single item supplement rule is marked as exclusion, reduction or retention state according to the determination result. Based on the mapping relationship, the coverage relationship, and the exclusive relationship, an executable rule expression result containing the rule version number, execution sequence number, and current valid state is generated, forming a stable two-layer rule structure.

5. The method of claim 4, wherein, The determination of the coverage and exclusion relationship between the combination priority rule and the single-item supplementary rule includes: Calculate the condition coverage of the single-item replenishment rule, and the exclusivity strength between the combination priority rule and the single-item replenishment rule; When the condition coverage rate is greater than or equal to the condition coverage rate threshold and the exclusion strength is greater than or equal to the score contribution difference threshold, the single item supplement rule is marked as exclusionary. When the condition coverage rate is greater than or equal to the condition coverage rate threshold and the exclusivity intensity is less than the score contribution difference threshold, the single item supplement rule is marked as a reduction state. When the condition coverage rate is less than the condition coverage rate threshold, the single-item supplement rule is retained as a valid rule.

6. The method of claim 4, wherein, S4 includes: For each of the combined display semantic units, verify the type, number of layers, area and proportion conditions in the mapped combination priority rules, and generate a valid execution set containing comprehensive hit rate; Based on the effective execution set, the theoretical score of a single rule is calculated by combining the full score of each rule, the comprehensive hit rate, and the execution weight. The scores of multiple rules in the same combination are then compressed and merged to obtain the overall combined score. Based on the determination results of the coverage relationship and the exclusivity relationship, the products in the combination are assigned, and the individual product rule conditions that have been absorbed by the combination priority rule are trimmed to obtain the individual product residual condition set. Based on the set of residual conditions for individual items, supplementary scoring is performed on the products within the combined display semantic unit, and products that do not belong to any combined display semantic unit are directly scored according to the complete individual item rules to obtain a set of supplementary individual item scores. The overall score of the combination and the supplementary scores of the individual items are summed, and the shelf score result is output after applying the upper limit constraint.

7. The method of claim 1, wherein, S5 includes: The shelf rating results, the combined display semantic units, and the structured data set are jointly archived to construct a historical rating master record that includes a master record table, a combination detail table, and an individual item detail table. Based on the aforementioned historical rating master record, a reusable index is established according to the shelf dimension, version dimension, and time dimension; When a new rule version is released, the stored combined display semantic units and the structured data set are located based on the reusable index, and the rule reconstruction and score calculation are directly read and re-executed. The recalculated scoring results are archived according to the new rule version and stored in parallel with the original historical records to form a multi-version scoring record.

8. A product display image intelligent scoring system based on dynamic rule configuration, applied to the method of any one of claims 1-7, characterized in that, The system includes: The image recognition and modeling module is used to recognize and perform basic structured modeling on the collected shelf images, generating a structured dataset. The combined semantic unit construction module is used to identify a collection of products that are cross-category and spatially continuous based on the structured data set, and to construct a combined display semantic unit with business semantics. The rule dynamic reconstruction module is used to semantically reconstruct the original scoring rules through a dynamic rule engine, transforming the flat single-item rule system into a two-layer rule structure consisting of combination priority rules and single-item supplementary rules, and establishing a mapping relationship between each of the combined display semantic units and the combination priority rules. The scoring calculation module is used to perform overall score calculation on the combined display semantic unit according to the mapping relationship under the constraints of the two-layer rule structure, and to perform supplementary individual scores on products not covered by the combined priority rule, and to generate shelf score results by merging them. The archiving and re-scoring module is used to jointly archive the shelf scoring results with the combined display semantic units and the structured data set, and in response to changes in scoring rules, to re-trigger the rule dynamic reconstruction module and the scoring calculation module based on the archived combined display semantic units and the structured data set to achieve rapid re-scoring.

9. An electronic device, characterized in that, The electronic device includes a memory and at least one processor. The memory stores a computer program, and the processor executes the computer program to implement the intelligent scoring method for merchandise display images based on dynamic rule configuration according to any one of claims 1-7.

10. A computer storage medium, characterized in that, It stores a computer program, which, when executed, implements a method for intelligent scoring of merchandise display images based on dynamic rule configuration according to any one of claims 1-7.

Images