Premix packaging label automatic verification method and system based on image recognition

Abstract

The invention belongs to the technical field of image recognition, and discloses a premix packaging label automatic verification method and system. The method comprises the following steps: firstly, collecting a label image, and constructing a label layout hypergraph according to a key field set after multi-scale semantic segmentation and optical character recognition; secondly, field topology and semantic features in the hypergraph are analyzed, differential semantic normalization is conducted on different fields, and field variation concept lattices are constructed through formal concept analysis in combination with field variation rules and historical legal change samples; and then generating a verification candidate set according to production batch standard information, establishing a partial order relationship of candidate interpretations, searching maximum candidate interpretations to complete global consistency judgment, and outputting a verification result containing a processing strategy and an abnormal classification report. According to the method, on the basis of the hypergraph and the concept lattice, the field relation, the organization standard and the legal variant equivalence relation are expressed in a unified mode, reasonable variation and substantive errors are distinguished, and automatic label verification with strict consistency of key fields and elastic fault tolerance of non-key fields is achieved.

Description

An Automatic Verification Method and System for Premix Packaging Labels Based on Image Recognition Technical Field

[0001] This invention relates to the field of image recognition and data processing technology, and discloses an automatic verification method and system for premixed product packaging labels based on image recognition. Background Technology

[0002] In the production of premixed agents, the accuracy of packaging labels directly affects product quality and compliance. Currently, industrial sites commonly employ automated verification systems based on image recognition. These systems capture label images using cameras, extract text content using optical character recognition (OCR) technology, and compare it with standard information in a production database to determine label correctness. Such systems typically rely on pre-defined templates or fixed area rules to match character content, position, and format item by item. The technology primarily achieves pixel-level or character-level consistency verification and is suitable for scenarios with stable label formats and infrequent version changes.

[0003] In the automated verification process of premixed product packaging labels, existing image recognition systems struggle to distinguish between reasonable differences arising from compliance evolution or variations in visual expression and substantial errors caused by printing or labeling mistakes. Specifically, on the one hand, label content may undergo minor but legal changes due to regulatory updates or formula adjustments, such as updating the ingredient content label from 1.0% to 1.00%. On the other hand, the same version of the label may exhibit visual variations in different batches due to differences in printing equipment, suppliers, or typesetting habits, such as changes in line break positions, inconsistent unit capitalization, different date formats, or slight shifts in text areas. The root of these problems lies in the current verification methods' reliance on fixed templates or precise character-level matching, focusing only on formal consistency without considering the semantic equivalence of label content or the reasonableness of version evolution. Because the system cannot identify the essential differences between these two types of variations, it often misjudges compliance changes or legal typesetting variations as errors, leading to numerous invalid alarms and requiring repeated manual verification by operators. Relaxing the overall matching threshold to reduce the false alarm rate may result in missing genuine information mismatches. If we can achieve automatic identification and tolerance of reasonable differences, we can significantly reduce unnecessary interventions without sacrificing detection capabilities, making automatic verification truly adapt to the objective needs of dynamic changes and diverse expressions of label content in actual production. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of existing technologies, this invention provides an automatic verification method for premixed product packaging labels based on image recognition, comprising: S1, acquiring original images of premixed product packaging labels, performing multi-scale semantic segmentation and optical character recognition on the original image of the label to obtain a set of key fields, and constructing a label layout hypergraph based on the set of key fields; S2, analyzing the topological and semantic features between fields based on the label layout hypergraph, obtaining a set of field semantic features through semantic normalization, and constructing a field variation concept lattice based on the field semantic feature set; S3, constructing a label verification candidate set based on the field variation concept lattice in conjunction with production batch standard information, performing global consistency determination and anomaly classification decision based on the label verification candidate set, and generating a label verification result.

[0005] Further, the step of performing multi-scale semantic segmentation and optical character recognition on the original label image to obtain a key field set includes: inputting the original label image into a multi-scale semantic segmentation network to obtain an initial segmentation probability map containing probability values ​​of text area, graphic area, QR code area, and background area, and generating a pixel-level region mask based on pixel-level category determination and morphological closing operation; extracting a set of text area image blocks based on the pixel-level region mask, performing optical character recognition on the set of text area image blocks to obtain a character recognition result set containing character content and global bounding rectangle coordinates; performing row detection and column alignment detection on the character recognition result set according to the global bounding rectangle coordinates, aggregating discrete characters into fields, and performing association operations on cross-row fields to obtain the key field set.

[0006] Further, the step of constructing a tag layout hypergraph based on the key field set includes: generating a hypergraph node set with each field in the key field set as a node; calculating the relative orientation and spacing ratio between node pairs in the hypergraph node set, identifying clustered field combinations that are spatially closely arranged and semantically related, and generating a set of hyperedges that simultaneously connect all nodes within the clustered field combinations; supplementing each hyperedge in the hyperedge set with semantic relationship attributes, including modification relationships, parallel relationships, and subordinate relationships; assembling the hypergraph node set and the hyperedge set into the tag layout hypergraph, which is used to represent the spatial structure and content distribution of tags.

[0007] Furthermore, based on the topological and semantic features between fields in the label layout hypergraph, the step of obtaining the field semantic feature set through semantic normalization processing includes: reading the node attributes and hyperedge information in the label layout hypergraph; identifying the field type through lexical analysis and hyperedge adjacency relationships, wherein the field type includes numeric, unit, date, and text types; performing differentiated semantic normalization on the field text content according to the field type: unifying the floating-point format and precision for numeric fields, mapping standard unit notation for unit fields, reorganizing date fields into a unified date format, and performing character normalization and synonym replacement for text fields; and integrating the normalization results, field types, and the adjacency field list and binding field list provided by the label layout hypergraph to construct the field semantic feature set.

[0008] Further, the step of constructing a field variation concept lattice based on the field semantic feature set includes: introducing premixed label verification domain rules and historical legal change samples to construct a variation pattern attribute set, wherein the variation pattern attributes include precision variation, format variation, synonym variation, and layout variation; constructing a formal background matrix based on the field semantic feature set and the variation pattern attribute set; generating a concept node set and a concept partial order relation set using a formal concept analysis method; supplementing each concept node with a semantic label, wherein the semantic label includes standard concept, equivalent variant concept, extended variant concept, and anomalous concept; and assembling to generate the field variation concept lattice.

[0009] Furthermore, the step of constructing a candidate set for tag verification based on the field variation concept lattice and the standard information of the production batch includes: reading the standard information of the current production batch and constructing a standard field set, wherein the fields in the standard field set are marked as critical or non-critical; traversing the standard field set, searching for semantically equivalent concept nodes in the field variation concept lattice, and generating a resolved field matching result through ambiguity resolution; generating multiple candidate interpretations based on the resolved field matching results, assembling them into the candidate set for tag verification, wherein each candidate interpretation contains the mapping relationship from standard fields to concept nodes and the variation path.

[0010] Further, the step of performing a global consistency determination based on the tag verification candidate set includes: statistically analyzing the key-level field matching rate and non-key-level field variation for each candidate explanation in the tag verification candidate set, and establishing a partial order relationship between candidate explanations; the partial order relationship is defined as follows: if candidate explanation A has a higher key-level matching rate and fewer non-key anomalies and fewer unmatched standard fields, then candidate explanation A is determined to be superior to candidate explanation B; searching for the maximal candidate explanation that is not superior to other candidate explanations in the partial order relationship, performing uniqueness and compliance checks on the maximal candidate explanation set, and outputting a global consistency determination result, the result including complete consistency, existence of legal variation, existence of ambiguity, and existence of anomalies.

[0011] Furthermore, the step of performing anomaly classification decision based on the tag verification candidate set includes: determining the processing strategy type according to the global consistency judgment result, wherein the processing strategy type includes allowable, manual review required, and immediate alarm shutdown required; if the processing strategy type is manual review required or immediate alarm shutdown required, then extract the detailed information of the difference field that caused the judgment, generate an anomaly classification report, and clarify the difference type as precision difference, format difference, synonym difference, layout difference, or content error; if the processing strategy type is allowable and there is a legitimate variation, then record the variation path information to the statistical database.

[0012] Furthermore, the step of generating label verification results also includes: generating visual annotation evidence on the original label image based on the node coordinate information in the label layout hypergraph; drawing bounding boxes using a first annotation color for critical level differences and using a second annotation color for non-critical level differences; associating the visual annotation evidence with the anomaly classification report and outputting it as part of the label verification results.

[0013] An automatic verification system for premixed product packaging labels based on image recognition is provided to implement the aforementioned automatic verification method for premixed product packaging labels based on image recognition. The system includes: an image acquisition and layout hypergraph construction module for acquiring original images of premixed product packaging labels, extracting key fields through multi-scale semantic segmentation, and constructing a label layout hypergraph based on the spatial and logical relationships between fields; a field semantic parsing and variant concept lattice construction module for parsing the semantic features in the label layout hypergraph, performing field type identification and semantic normalization, and constructing a field variant concept lattice based on variant pattern attributes; and a global consistency determination and anomaly decision module for aligning production batch standard information with the field variant concept lattice to generate a candidate set, selecting the optimal interpretation through partial order relations, performing a global consistency determination, outputting the verification result, and triggering the corresponding processing flow.

[0014] This invention effectively solves the problem of distinguishing legitimate variations from substantive errors caused by compliance evolution or visual expression diversity in premixed packaging label verification by constructing a label layout hypergraph and a field variation concept lattice. First, the method utilizes the label layout hypergraph to unify discrete OCR characters and layout information into a clustered semantic structure, overcoming the representational limitations of traditional simple graph models when handling line breaks or offsets in multi-field combinations (such as numerical values ​​and units), thus ensuring the robustness of structured input. Second, based on the field variation concept lattice, this invention transforms the originally scattered variation rules into a hierarchical knowledge representation, achieving structured management and reasoning of legitimate variations such as precision, format, and synonymy. Furthermore, by establishing a global consistency judgment mechanism based on partial order relations, this invention breaks through the decision bottleneck of a single similarity threshold, enabling the search for the largest candidate explanation that satisfies "strict matching at the critical level and flexible tolerance at the non-critical level" within the candidate explanation set. This not only reduces the false alarm rate caused by printing and layout differences but also ensures zero omission of key compliance information, achieving automatic verification that adapts to actual production needs. This invention promotes the application of graph theory-based structured analysis and high-level semantic reasoning techniques in the field of industrial image data processing, improving the accuracy and reliability of image semantic understanding and intelligent decision-making under complex and variable conditions. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 is a flowchart of an automatic verification method for premixed product packaging labels based on image recognition according to the present invention; Figure 2 is a schematic diagram of the original image acquisition of the label in an embodiment of the present invention; Figure 3 is a schematic diagram of a multi-scale semantic segmentation network in an embodiment of the present invention; Figure 4 is a schematic diagram of the field aggregation process in an embodiment of the present invention; Figure 5 is a schematic diagram of a semantic normalization example in an embodiment of the present invention; Figure 6 is a schematic diagram of a field variation concept lattice in an embodiment of the present invention; Figure 7 is a schematic diagram of candidate interpretation and global consistency determination in an embodiment of the present invention; Figure 8 is a schematic diagram of visualized labeled evidence in an embodiment of the present invention; Figure 9 is a functional block diagram of an automatic verification system for premixed product packaging labels based on image recognition according to the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Example 1: Please refer to Figure 1. This example provides an automatic verification method for premixed product packaging labels based on image recognition, including: S1: acquiring the original image of the premixed product packaging label, performing multi-scale semantic segmentation and optical character recognition on the original image of the label to obtain a set of key fields, and constructing a label layout hypergraph based on the set of key fields.

[0019] This step focuses on the structured representation of premixed product packaging label images. It extracts a set of key fields from the label using multi-scale semantic segmentation and optical character recognition (OCR) techniques, and constructs a label layout hypergraph based on this set of key fields. This label layout hypergraph, as the final output of S1, will be used in S2 to parse the topological and semantic features between fields.

[0020] Specifically, the process of acquiring the original image of the premixed product packaging label, performing multi-scale semantic segmentation and optical character recognition on the original image of the label to obtain a set of key fields, and constructing a label layout hypergraph based on the set of key fields includes: S11: Acquiring the original image of the premixed product packaging label to obtain the original image of the label.

[0021] An industrial camera is deployed at the packaging station on the production line, facing the label affixing area of ​​the premixed packaging container. A data acquisition signal is triggered when the packaging container passes the inspection station. The original label image refers to the unprocessed digital image output by the industrial camera in a single trigger acquisition. Its data structure is a three-channel pixel matrix, with the number of rows and columns determined by the camera sensor resolution. Each pixel location stores the grayscale values ​​of the red, green, and blue channels. During acquisition, camera parameters are adaptively set according to the production line speed and label size. These parameters include resolution, frame rate, and exposure time. The resolution is determined based on the ratio of the minimum physical size of the label character to the desired character pixel width. The frame rate is determined based on the ratio of the production line conveyor speed to the distance between adjacent packages. The exposure time is determined through pre-calibration experiments based on the ambient lighting intensity and the camera's photosensitivity to ensure that the character edges in the original label image are clearly visible and without motion blur. After acquisition, the original label image is written to an image buffer and assigned a unique image identifier. This image identifier is generated by concatenating the acquisition timestamp and the station number, used for tracing the image source in subsequent processes. See Figure 2, which is a schematic diagram of original label image acquisition provided in this embodiment. As shown in the figure, this diagram illustrates the physical layout and data flow of the image acquisition stage in an automated verification system for premixed product packaging labels. The left side depicts a production line scene, with premixed product packaging containers placed on a conveyor belt. Packaging labels are affixed to the surface of the containers. An industrial camera is fixed above the conveyor belt via a bracket, with its lens facing the label area. The dashed line indicates the camera's field of view. When a packaging container moves to the inspection station, the system triggers an acquisition signal, and the image data enters the image buffer on the right side via a signal transmission link. Below the image buffer is the data structure of the original label image: a three-channel pixel matrix, where the red, green, and blue channels store the grayscale values ​​of their respective colors. In the premixed product packaging label verification scenario, the label surface often exhibits subtle visual differences due to factors such as printing batches, lighting conditions, and container curvature. Traditional fixed-parameter acquisition methods struggle to maintain image quality across different production environments. This application adaptively sets the camera resolution, frame rate, and exposure time based on the production line speed and label size, ensuring that character edges in the original label image are clearly visible and free of motion blur, providing a high-quality input data source for subsequent multi-scale semantic segmentation and optical character recognition.

[0022] S12: Perform multi-scale semantic segmentation on the original labeled image to obtain a pixel-level region mask.

[0023] Furthermore, the process of performing multi-scale semantic segmentation on the original label image includes the following steps: S121: Input the original label image into the multi-scale semantic segmentation network to obtain an initial segmentation probability map. The multi-scale semantic segmentation network refers to a lightweight convolutional neural network with an encoder-decoder architecture. The encoder part extracts feature representations of the original label image at different scales through multi-layer convolution and downsampling operations, and the decoder part restores the multi-scale features to the original resolution through upsampling and feature fusion operations. The data structure of the initial segmentation probability map is a four-channel pixel matrix, with the number of rows and columns consistent with the original label image. Each pixel position stores the probability value of the pixel belonging to four categories: text area, graphic area, QR code area, and background area, and the sum of the four probability values ​​is one. The multi-scale semantic segmentation network is trained offline on a pre-labeled premixed label sample set. During the training process, the cross-entropy loss function is used to measure the difference between the prediction result and the labeling result, and the network parameters are iteratively updated through the gradient descent algorithm until the loss converges. See Figure 3, which is a schematic diagram of the multi-scale semantic segmentation network provided in this application embodiment. As shown in the figure, this figure illustrates the network structure and processing flow for converting the original label image into a pixel-level region mask. The input on the left is the original image of the label. Different colored dashed boxes mark the typical distribution locations of the text, graphic, and QR code areas. The image first enters the encoder module, where multi-scale features are extracted step-by-step through multi-layer convolution and downsampling operations. The feature map size decreases from large to small, corresponding to the gradual abstraction process from shallow texture to deep semantics. It then enters the decoder module, where upsampling and feature fusion operations gradually restore the compressed feature map to its original resolution, with the feature map size increasing from small to large. The output on the far right is the initial segmentation probability map, containing four channels: text, graphic, QR code, and background. Darker areas indicate high-probability regions where the pixel belongs to the corresponding category. In premixed product packaging label verification, the label surface typically contains text content requiring field-by-field comparison, company logo graphics, and a traceability QR code. If optical character recognition is performed directly on the entire image, decorative lines in the graphic area and the dot matrix structure of the QR code will be misidentified as meaningless characters. This application achieves pixel-level region division through a multi-scale semantic segmentation network, enabling subsequent processing to adopt matching processing strategies for different regions and eliminate cross-regional interference from the source.

[0024] Specifically, premixed product packaging labels typically include product name, ingredient list, content values, production date, company logo, and traceability QR code. The processing logic for different content areas differs fundamentally in subsequent verification processes. The text area carries semantic content that needs to be compared field-by-field with standard information, while the graphic and QR code areas require processing using a dedicated decoder, and the background area contains no valid information. If optical character recognition (OCR) is directly applied to the entire original label image, decorative lines in the graphic area and the dot matrix structure of the QR code will be misidentified as meaningless characters, interfering with subsequent field extraction. By introducing a multi-scale semantic segmentation network, the original label image is divided into four regions based on semantic attributes, enabling subsequent processing to employ matching strategies for different regions, eliminating cross-regional interference at the source.

[0025] S122: Perform pixel-level category determination on the initial segmentation probability map to obtain a pixel-level region mask. For each pixel position in the initial segmentation probability map, the category corresponding to the maximum value of the four category probability values ​​is taken as the region label of that pixel. The data structure of the pixel-level region mask is a single-channel integer matrix, with the number of rows and columns consistent with the original label image. Each pixel position stores an integer label, with the integer label value ranging from zero to three, corresponding to the four categories of text area, graphic area, QR code area, and background area, respectively. For boundary pixels with similar probability values ​​during the category determination process, morphological closing operation is used to smooth the pixel-level region mask to eliminate isolated noise and jagged artifacts at the region edges. The structuring element used in the morphological closing operation is a square with a side length determined according to the average stroke width of the label characters.

[0026] S13: Extract the set of image blocks for text areas based on pixel-level region masks.

[0027] Furthermore, the process of extracting the text region image block set based on the pixel-level region mask includes the following steps: S131: Traverse all pixels with a label value of zero in the pixel-level region mask, and aggregate spatially connected pixels into independent text region connected components to obtain a set of text region connected components. The connectivity determination adopts the eight-neighbor connectivity criterion, that is, if two pixels are adjacent in the horizontal, vertical, or diagonal direction and both have a label value of zero, they are determined to be in the same connected component. The set of text region connected components is an ordered list containing all independent text region connected components, and each connected component records the set of pixel coordinates it contains.

[0028] S132: For each connected component in the text region connected component set, calculate its minimum bounding rectangle to obtain the text region bounding rectangle set. The minimum bounding rectangle is the axis-aligned rectangle with the smallest area that can completely enclose all pixels of the connected component, described by three attributes: the coordinates of the top-left vertex, the rectangle width, and the rectangle height. The data structure of the text region bounding rectangle set is a list, with each element storing the attributes of a minimum bounding rectangle.

[0029] S133: Based on the set of bounding rectangles for the text area, extract the corresponding region from the original label image to obtain a set of text area image blocks. For each smallest bounding rectangle in the set of bounding rectangles for the text area, starting from the coordinates of its top-left vertex and within the range of the rectangle's width and height, copy the corresponding pixel region from the original label image to generate an independent text area image block. The set of text area image blocks is an ordered list containing all text area image blocks. Each text area image block retains its absolute coordinate information in the original label image for subsequent field positioning.

[0030] S14: Perform optical character recognition on the set of image blocks in the text area to obtain a set of character recognition results.

[0031] Furthermore, the process of performing optical character recognition on the set of text area image blocks includes the following steps: S141: Perform image preprocessing on each text area image block in the set of text area image blocks to obtain preprocessed text area image blocks. The image preprocessing includes three operations: grayscale conversion, contrast enhancement, and binarization. Grayscale conversion converts the three-channel text area image blocks into a single-channel grayscale image, and the conversion weight is determined according to the difference in sensitivity of the human eye to different color channels. Contrast enhancement uses a histogram equalization method to stretch the grayscale value distribution to the full dynamic range, improving the grayscale difference between characters and the background. Binarization uses an adaptive thresholding method, dynamically determining the binarization threshold based on the grayscale mean value within the pixel neighborhood, setting character pixels as foreground values ​​and background pixels as background values, and the neighborhood size is determined according to the average stroke width of the label characters.

[0032] S142: Perform character recognition on the preprocessed text area image block to obtain single-block character recognition results. The character recognition employs a sequence recognition model based on a recurrent neural network. The preprocessed text area image block is divided into several vertical strips. The feature vector of each strip is extracted and input into the recurrent neural network in horizontal order, outputting the character sequence and its confidence score. The data structure of the single-block character recognition result is a record list. Each record contains the content of the recognized single character, the coordinates of the bounding rectangle of the character in the preprocessed text area image block, and the recognition confidence score. The recognition confidence score is a value between zero and one, reflecting the model's degree of certainty regarding the character recognition result, and is calculated by the normalized exponential function of the recurrent neural network output layer.

[0033] S143: Summarize the single-block character recognition results corresponding to all text area image blocks, and transform the coordinates of the character's bounding rectangle from the local coordinate system of the text area image block to the global coordinate system of the original label image, obtaining a character recognition result set. The coordinate transformation is achieved by adding the local coordinates to the absolute coordinates of the corresponding text area image block's top-left corner in the original label image. The data structure of the character recognition result set is a record list, with each record containing three attributes: character content, global bounding rectangle coordinates, and recognition confidence.

[0034] S15: Field aggregation is performed based on the character recognition result set to obtain a key field set. See Figure 4, which is a schematic diagram of the field aggregation process provided in this embodiment. As shown in the figure, this figure illustrates the process of converting discrete character recognition results into field units with semantic integrity. The upper left part shows the original pixel representation of the text area image block, where the characters "vitamin C" and "25.0%" are not neatly arranged; the right part shows the character recognition result set obtained after optical character recognition, with each character marked by an independent red bounding box, and its circumscribed rectangle coordinates and recognition confidence score recorded. The lower part shows the result of the field aggregation process. The system performs row detection based on the vertical center coordinates of the character's circumscribed rectangle, and then performs column alignment detection and field segmentation based on the horizontal spacing between adjacent characters, aggregating spatially closely arranged characters into field units marked with blue boxes. The final output key field set contains two records, corresponding to the field identifier, field text content, and its circumscribed rectangle coordinates, respectively. In premixed product packaging labels, ingredient names and corresponding content values ​​often appear adjacent to each other. If only the character-by-character output of optical character recognition is used as input for subsequent comparison, the system will be unable to distinguish between adjacent characters that belong to different semantic units. This application uses a field aggregation method based on spatial location relationships to enable subsequent semantic normalization and variant recognition to be processed at the field level, avoiding the semantic fragmentation problem introduced by character-level operations.

[0035] Furthermore, the process of field aggregation based on the character recognition result set includes the following steps: S151: Perform row detection on the character recognition result set according to the coordinates of the global bounding rectangle to obtain character row grouping results. The row detection is based on the vertical center coordinates of the character bounding rectangles, and characters whose vertical center coordinate difference is less than the row height threshold are grouped into the same row. The row height threshold is determined by multiplying the median height of the bounding rectangles of all characters in the character recognition result set by a preset ratio coefficient, which is determined by statistical analysis based on the ratio of the typical line spacing to the character height of the premixed label. The data structure of the character row grouping results is a nested list, where each element of the outer list corresponds to a character row, and the inner list contains records of all characters in that row.

[0036] S152: For each character row in the character row grouping result, sort them in ascending order according to the horizontal left boundary coordinates of the character's bounding rectangle, and perform column alignment detection and field segmentation based on the horizontal spacing between adjacent characters to obtain the in-row field grouping result. The column alignment detection is achieved by calculating the ratio of the horizontal gap between the bounding rectangles of adjacent characters to the average width of the characters. When this ratio exceeds the field spacing threshold, adjacent characters are determined to belong to different fields. The field spacing threshold is determined through pre-calibration based on the typical difference between the spacing between characters within a field and the spacing between spaces between fields in the premix label. The data structure of the in-row field grouping result is a nested list, where each element of the outer list corresponds to a field, and the inner list contains records of all characters within that field.

[0037] S153: Perform cross-line field association on the intra-line field grouping results of all character lines to obtain a key field set. This cross-line field association addresses multi-line fields commonly found in premixed labels. When fields in adjacent lines significantly overlap horizontally and their vertical spacing is less than a line height threshold, they are merged into a single field. Specifically, the overlap rate of the bounding rectangles of each field in adjacent lines is calculated horizontally. This overlap rate is defined as the ratio of the width of the overlapping area to the smaller of the widths of the two rectangles. When the overlap rate exceeds the cross-line merging threshold, a merging operation is performed. This cross-line merging threshold is determined through statistical analysis based on typical layout characteristics of cross-line fields in premixed labels. The bounding rectangle of the merged field is the smallest bounding rectangle among all the bounding rectangles of the characters involved in the merging. The field content is concatenated from the character content of each line in line order. The data structure of the key field set is a record list. Each record contains five attributes: field identifier, field text content, field bounding rectangle coordinates, number of characters within the field, and average field recognition confidence. The field identifier is automatically generated according to the extraction order, and the average field recognition confidence is the arithmetic mean of the recognition confidence of all characters within the field.

[0038] Specifically, the field distribution on premixed product packaging labels exhibits a hierarchical characteristic. Product names are typically displayed in a large font on a separate line, while ingredient lists are arranged in a compact multi-column format, with content values ​​and units often appearing adjacently. If only the character-by-character output from optical character recognition is used as input for subsequent comparisons, the system will be unable to distinguish between adjacent characters belonging to different semantic units, nor will it be able to recognize the phenomenon of the same field being distributed across lines due to line breaks. By introducing a field aggregation method based on row detection and column alignment detection, discrete characters are organized into semantically complete field units according to their spatial positional relationships. This allows subsequent semantic normalization and variant recognition to be processed at the field level, avoiding the semantic fragmentation problem introduced by character-level operations.

[0039] S16: Construct a tag layout hypergraph based on a set of key fields.

[0040] Furthermore, the process of constructing a label layout hypergraph based on the key field set includes the following steps: S161: Generate a hypergraph node set using each field in the key field set as a node. The data structure of the hypergraph node set is a record list, with each record corresponding to one node, containing two parts: a node identifier and node attributes. The node identifier directly inherits the field identifiers from the key field set. The node attributes include the field text content, the coordinates of the field's bounding rectangle, the number of characters within the field, and the average recognition confidence score of the field, all of which are copied from the key field set.

[0041] S162: Traverse all node pairs in the hypergraph node set, calculate the spatial relationship features between node pairs, and obtain the node pair spatial relationship set. The spatial relationship features include two attributes: relative orientation and spacing ratio. Relative orientation is determined based on the relative positions of the center points of the bounding rectangles of the two node fields, with values ​​including above, below, left, right, and overlapping. The determination is based on the relationship between the coordinate difference between the two center points in the vertical and horizontal directions and the rectangle dimensions. The spacing ratio is defined as the ratio of the shortest boundary distance between the bounding rectangles of the two fields to the average length of the diagonals of the two rectangles. It is used to quantify the spatial closeness of the two fields; a smaller spacing ratio indicates a closer spatial distance between the two fields. The data structure of the node pair spatial relationship set is a record list, with each record containing a node pair identifier, relative orientation, and spacing ratio.

[0042] S163: Identify clustered field combinations based on the set of spatial relationships between node pairs and generate a set of hyperedges. The clustered field combination refers to a group of fields in the premix label that are semantically related and spatially closely arranged. Typical examples include fields appearing in groups such as ingredient names and corresponding content values, indicator names and corresponding units, and production dates and expiration dates. The method for identifying clustered field combinations is as follows: First, select node pairs from the set of spatial relationships whose spacing ratio is less than the clustering threshold, and consider them as candidate related node pairs. The clustering threshold is determined through statistical analysis based on the typical spacing distribution of related fields in the premix label. Then, perform connected clustering on the candidate related node pairs, merging node pairs indirectly related through shared nodes into the same clustered field combination. Finally, generate a hyperedge for each clustered field combination, which connects all nodes within the clustered field combination. The data structure of the hyperedge set is a record list, where each record contains three parts: a hyperedge identifier, a list of hyperedge-connected nodes, and hyperedge attributes. Hyperedge identifiers are automatically numbered in the order of generation. The list of nodes connected by the hyperedge records all node identifiers associated with the hyperedge. Hyperedge attributes include cluster type label and maximum spacing ratio within the cluster. The cluster type label is automatically inferred based on the text content characteristics of the nodes connected by the hyperedge. The value range includes three types: numerical unit clusters, name value clusters, and date description clusters. The inference is based on whether the node text content contains numeric characters, unit of measurement keywords, or date format patterns.

[0043] S164: Supplement each hyperedge in the hyperedge set with semantic relationship attributes, and assemble the hypergraph node set and hyperedge set into a label layout hypergraph. The semantic relationship attributes describe the semantic dependency type between nodes connected by hyperedges, including three types: modifier relationship, parallel relationship, and subordinate relationship. A modifier relationship refers to one node supplementing the meaning of another node; a typical example is a unit node modifying a numerical node. A parallel relationship refers to nodes being at the same semantic level; a typical example is the relationship between component names in a component table. A subordinate relationship refers to one node semantically depending on another node; a typical example is small-font annotation text subordinate to its main content. The determination of semantic relationship attributes is based on the field text content and the size of the field's bounding rectangle of the hyperedge-connected nodes. Specific determination rules are pre-configured according to the premix label layout specifications. The data structure of the label layout hypergraph is a binary tuple, with the first element being the hypergraph node set and the second element being the hyperedge set. The label layout hypergraph fully represents the spatial structure and content distribution of the current premix packaging label, where nodes carry field-level information and hyperedges carry clustered associations and semantic dependencies between fields.

[0044] Specifically, traditional label layout methods typically employ a simple directed graph structure, using fields as nodes and binary relationships between fields as edges. This approach has limitations when dealing with the multi-field combinations common in premixed product labels. For instance, the description of ingredient content often consists of three fields: ingredient name, numerical value, and unit. A simple directed graph would require multiple edges to depict the relationships between name and numerical value, and numerical value and unit, failing to directly express the semantic coherence of these three elements. When label layout changes lead to line breaks between numerical values ​​and units, the edge structure of the simple directed graph alters, affecting the stability of subsequent comparisons. By introducing a hypergraph structure, a single hyperedge connects multiple semantically related field nodes, allowing the ingredient name, numerical value, and unit to be expressed and processed as a whole. Even with line breaks or slight positional shifts in the label, the integrity of the clustered field combinations remains stable as long as the set of nodes connected by the hyperedge remains unchanged. This representation method provides a more robust structured input basis for semantic normalization and variant recognition in S2, enabling the system to perform semantic analysis at the field combination level rather than the single field level, thereby more accurately determining the nature of label changes.

[0045] S2: Based on the label layout hypergraph, analyze the topological and semantic features between fields to obtain the field semantic feature set, and construct the field variation concept lattice based on the field semantic feature set.

[0046] This step, based on the label layout hypergraph output by S1, analyzes the topological structure and semantic features between fields. Semantic normalization is used to obtain the field semantic feature set, and a formal concept analysis method is introduced to construct a field variation concept lattice. This field variation concept lattice, as the final output of S2, will be called in S3 to perform global consistency determination and anomaly classification decisions in conjunction with production batch standard information.

[0047] Specifically, the process of parsing the topological and semantic features between fields based on the label layout hypergraph to obtain a set of field semantic features, and constructing a field variation concept lattice based on the set of field semantic features includes: S21: reading the node attributes and hyperedge information in the label layout hypergraph, performing field type identification, and obtaining field type labeling results.

[0048] Furthermore, the process of reading node attributes and hyperedge information in the label layout hypergraph and performing field type recognition includes the following steps: S211: Extract the field text content attributes of each node from the hypergraph node set of the label layout hypergraph, perform lexical analysis on the field text content, and obtain lexical analysis results. The lexical analysis refers to scanning and marking the field text content according to character categories, identifying the distribution of numeric character sequences, alphabetic character sequences, Chinese character sequences, punctuation marks, and special symbols. The data structure of the lexical analysis results is a record, containing attributes such as the proportion of numeric characters, the proportion of alphabetic characters, the proportion of Chinese characters, decimal point occurrence markers, percent sign occurrence markers, date separator occurrence markers, and unit of measurement keyword matching markers. The proportion of numeric characters is defined as the ratio of the number of numeric characters in the field text content to the total number of characters; the definitions of other proportion attributes are similar. The decimal point occurrence marker is a Boolean attribute, taking a true value when a decimal point character exists in the field text content. The date separator appearance flag is a Boolean attribute, taking a true value when the field text contains commonly used year-month-day separator characters, including hyphens, slashes, and Chinese characters for year, month, and day. The unit of measurement keyword matching flag is also a Boolean attribute, taking a true value when the field text matches any entry in a pre-defined unit of measurement thesaurus. This thesaurus is pre-configured based on commonly used units of measurement in the premix industry and includes categories such as mass units, volume units, concentration units, and proportion units.

[0049] S212: Based on the lexical analysis results, field type determination is performed to obtain field type labeling results. The field type determination adopts a rule-driven classification method, executing the following rules in priority order: when the date separator appears as true in the lexical analysis results and the proportion of numeric characters exceeds the threshold for the proportion of numeric characters in the date type, the field is determined to be a date field; when the unit of measurement keyword matches as true in the lexical analysis results and the proportion of Chinese characters or alphabetic characters exceeds the threshold for the proportion of text in the unit type, the field is determined to be a unit field; when the proportion of numeric characters exceeds the threshold for the proportion of numeric characters in the numeric type and the appearance of a decimal point is marked as true or the appearance of a percentage sign is marked as true, the field is determined to be a numeric field; when none of the above rules are met, the field is determined to be a text field. The thresholds for the proportion of numeric characters in the date type, the proportion of text in the unit type, and the proportion of numeric characters in the numeric type are all determined based on statistical analysis of the premixed label field samples. The data structure of the field type labeling results is a mapping table, where the key is the field identifier and the value is the field type label. The field type label can take four values: date, unit, numeric, and text.

[0050] S213: Identify the tight binding relationships between fields in the hyperedge set based on the label layout hypergraph, and update the field type labeling results. Traverse each hyperedge in the hyperedge set. For hyperedges whose cluster type label is "numerical unit cluster," extract the list of connected nodes, identify the numeric and unit fields, and mark them as tightly bound. For hyperedges whose cluster type label is "name value cluster," identify the text and numeric fields, and mark them as tightly bound. The tight binding relationships are recorded as additional attributes in the field type labeling results. Each field's additional attributes contain a list of bound field identifiers, pointing to other fields that have tight binding relationships with that field.

[0051] Specifically, the field types in premixed product packaging labels directly determine the subsequent semantic normalization processing method. Numerical fields require standardized decimal places and separator formats, unit fields need to be mapped to standard notation, date fields need to be converted to a unified date format, while text fields are primarily normalized through character standardization. If the same normalization strategy is applied to all fields without field type identification, the semantic equivalence relationships between different field types cannot be correctly handled. Through a field type identification method based on lexical analysis and hyperedge information, the system can automatically distinguish the semantic attributes of fields and identify semantically dependent field combinations such as numerical values ​​and units, names and values, providing a classification basis for subsequent differentiated semantic normalization processing.

[0052] S22: Based on the field type annotation results, perform semantic normalization on the field text content to obtain the field semantic feature set. See Figure 5, which is a schematic diagram of a semantic normalization example provided in this application embodiment. As shown in the figure, the figure displays the comparison results of different field types before and after semantic normalization in tabular form. The table contains three columns: field type, original field content, and normalized value. The first row is an example of a numeric field; the original content "1.00%" is converted to the standard floating-point number "0.01" after numeric normalization, achieving uniform precision. The second row is an example of a unit field; the original content "G / KG" is mapped to the standard notation "g / kg" after unit normalization. The third row is an example of a date field; the original content "2023-5-1" is reorganized into the unified format "2023-05-01" after date normalization. The fourth row is an example of a text field; the original content "Vitamin C" is compressed with whitespace after text normalization to obtain "Vitamin C". In the scenario of verifying premixed product packaging labels, labels of the same batch of products may exhibit visual variations such as inconsistent decimal places, mixed use of uppercase and lowercase units, and changes in date separators due to differences in printing equipment, suppliers, or typesetting habits. This application uses a differentiated semantic normalization strategy to convert the surface-formed but semantically equivalent field content into a unified normalized representation, enabling subsequent field variation concept grids to accurately identify the essential difference between legitimate variations and substantive errors.

[0053] Furthermore, the process of performing semantic normalization on the field text content based on the field type labeling results to obtain the field semantic feature set includes the following steps: S221: Perform numerical normalization on fields with numeric type labels in the field type labeling results to obtain the numerical normalization result. The numerical normalization process is as follows: First, identify the numerical part and the modifier part in the field text content. The numerical part includes the integer part, the decimal point, and the fractional part. The modifier part includes the plus / minus sign, the percent sign, and the thousandth sign. Then, convert the numerical part to standard floating-point representation, remove leading and trailing zeros, and retain significant digits. Next, adjust the numerical meaning according to the modifier part. If there is a percent sign, divide the value by one hundred to convert it to decimal form. If there is a thousandth sign, divide the value by one thousand to convert it to decimal form. Finally, round the normalized value according to a preset precision. The preset precision is configured according to the precision requirements of the premix industry for content labeling. The data structure of the numerical normalization result is a mapping table, where the key is the field identifier and the value is the normalized standard floating-point number.

[0054] S222: Perform unit normalization on fields with the type label "unit" in the field type annotation results to obtain the unit normalization result. The unit normalization process is as follows: First, convert the field text content to lowercase and remove whitespace characters; then, search for matching items in a preset unit synonym mapping table. The unit synonym mapping table records the correspondence between different writing forms of the same unit of measurement and its standard form, including uppercase and lowercase variants, Chinese and English variants, and abbreviation full name variants; if a matching item is found, replace the field text content with the corresponding standard form; if no matching item is found, retain the original text content and mark it as an unknown unit. The data structure of the unit normalization result is a mapping table, where the key is the field identifier and the value is the standard unit writing string or the original string with the unknown mark.

[0055] S223: Perform date normalization on fields with the type label "Date" in the field type annotation results to obtain the date normalization result. The date normalization process is as follows: First, attempt to parse the field text content using multiple date format templates. The date format templates include common formats such as year-month-day order, month-day-year order, and day-month-year order. Separators include hyphens, forward slashes, periods, and Chinese characters for "year, month, day". Then, extract the year, month, and day components from the successfully parsed results, and perform century completion for two-digit years based on the current year. Finally, reassemble the three components in year-month-day order using hyphens as separators into a unified date format string. The data structure of the date normalization result is a mapping table, where the key is the field identifier and the value is the unified format date string.

[0056] S224: Perform text normalization on fields with the type label "text" in the field type annotation results to obtain the text normalization result. The text normalization process is as follows: First, convert full-width characters to half-width characters and traditional Chinese characters to simplified Chinese characters; then remove leading and trailing whitespace characters from the text and compress consecutive whitespace characters within the text into single spaces; next, perform a lookup and replacement based on a preset text synonym mapping table, which records the correspondence between different expressions of the same concept in the premix industry and its standard expression. The data structure of the text normalization result is a mapping table, where the key is the field identifier and the value is the normalized text string.

[0057] S225: Integrate the numerical normalization results, unit normalization results, date normalization results, and text normalization results, and combine them with the field type annotation results and the adjacency relationships in the label layout hypergraph to construct a field semantic feature set. The data structure of the field semantic feature set is a record list, where each record corresponds to a field and includes six attributes: field identifier, original value, normalized value, field type, adjacent field list, and bound field list. The original value is directly taken from the field text content attribute of the corresponding node in the label layout hypergraph. The normalized value is obtained from the numerical normalization results, unit normalization results, date normalization results, or text normalization results, depending on the field type. The field type is taken from the field type annotation results. The adjacent field list is extracted from the hyperedge set of the label layout hypergraph, recording the identifiers of all other fields associated with this field through the same hyperedge. The bound field list is taken from the bound field identifier list additional attribute of this field in the field type annotation results.

[0058] S23: Introduce domain rules and historical legal change samples to construct a set of mutation pattern attributes.

[0059] Furthermore, the process of introducing domain rules and historical legal change samples and constructing a set of variation pattern attributes includes the following steps: S231: Read the variation rule configuration from the premixed product label verification domain knowledge base and parse it into a set of variation rules. The variation rule configuration records the legal variation types and their judgment conditions that may occur in the compliance evolution and production practice of premixed product packaging labels. The data structure of the variation rule set is a record list, and each record contains four attributes: rule identifier, applicable field type, variation type label, and variation judgment condition. The value range of the variation type label includes four types: precision variation, format variation, synonym variation, and layout variation. Precision variation refers to the situation where the decimal places of a numeric field change but the value itself remains equal within the allowable error range; format variation refers to the situation where the separator or arrangement order of a date field changes but the date represented is the same; synonym variation refers to the situation where a unit field or text field uses a synonym to replace the standard writing; layout variation refers to the situation where the line break position or alignment of a field changes but the field content remains unchanged. The variation judgment condition is recorded in the form of a field attribute comparison expression, used to determine whether a specific field belongs to the variation type.

[0060] S232: Retrieve historical verification samples marked as valid from the historical label verification record database, and extract the change pair instance set. The historical verification samples refer to label verification records from past production batches that have been manually reviewed and confirmed as valid changes. Each record contains three attributes: the field value before the change, the field value after the change, and the manual judgment conclusion. The data structure of the change pair instance set is a record list. Each record contains five attributes: change pair identifier, normalized value before the change, normalized value after the change, field type, and confirmed variant type label.

[0061] S233: Merge the set of mutation rules and the set of change pair instances to construct a set of mutation pattern attributes. For each rule in the set of mutation rules, convert its mutation type label and mutation judgment condition into a mutation pattern attribute; for each change pair instance in the set of change pair instances, abstract the relationship between its normalized value before and after the change into a mutation pattern attribute. The data structure of the mutation pattern attribute set is a set, where each element is a mutation pattern attribute descriptor, containing three parts: attribute identifier, attribute type, and attribute constraint. The attribute type corresponds to the mutation type label, and the attribute constraint describes the applicable conditions and equivalence judgment method of the mutation pattern.

[0062] S24: Based on the field semantic feature set and the mutation pattern attribute set, a field mutation concept lattice is constructed using a formal concept analysis method. See Figure 6, which is a schematic diagram of the field mutation concept lattice provided in this embodiment. As shown in the figure, the diagram uses a hierarchical Hasse diagram structure to illustrate the organization of the field mutation concept lattice. The top layer is the standard concept node, whose content is an empty set, indicating that the corresponding field uses the standard syntax and has no mutations; the right side is labeled with the example value "1.0%". The second layer contains two concept nodes: the left side is the equivalent variant concept node, whose content includes precision mutation and format mutation attributes, indicating that the corresponding field is semantically completely equivalent to the standard syntax; the right side is labeled with the example value "1.00%"; the right side is the extended variant concept node, whose content includes synonym mutation and layout mutation attributes, indicating that the corresponding field is semantically equivalent to the standard syntax but the expression form is extended; the right side is labeled with the example value "one percent". The bottom layer is the abnormal concept node, indicating that the normalized value of the corresponding field does not match any standard value and may contain a substantial error; the right side is labeled with the example value "2.0%". The connections between concept nodes represent partial order relationships, and the attribute set of a lower-level concept contains the attribute set of the upper-level concept. The bottom diagram illustrates the meaning of the three color codes: blue indicates complete semantic consistency, green indicates allowed legal variations, and red indicates substantial differences. In the verification of premixed product packaging labels, traditional methods based on character similarity struggle to systematically manage the various legal variation patterns arising from regulatory updates, formula adjustments, or supplier changes. This application uses a formal concept analysis method to model the normalized semantics and variation rules of fields into a unified concept lattice structure. This allows the system to manage the relationships between standard syntax, equivalent variants, and extended variants within an ordered hierarchical structure, providing structured variation knowledge support for subsequent global consistency determination.

[0063] Furthermore, the process of constructing a field variation concept lattice based on the field semantic feature set and the variation pattern attribute set includes the following steps: S241: Constructing a formal background matrix, which is the input data structure of the formal concept analysis method. The formal background matrix is ​​a two-dimensional Boolean matrix, with rows corresponding to each field in the field semantic feature set and columns corresponding to each variation pattern attribute in the variation pattern attribute set. The element at a given position in the matrix is ​​either true or false, indicating whether the corresponding field satisfies the applicable conditions of the corresponding variation pattern attribute. The method for determining whether a field satisfies a variation pattern attribute is as follows: condition matching is performed based on the attribute constraints of the variation pattern attribute and the normalized value, field type, and other features of the field. If the field features meet the applicable conditions of the attribute constraints, the matrix element is true; otherwise, it is false.

[0064] S242: Execute a concept lattice construction algorithm on the formal background matrix to generate a set of concept nodes and a set of concept partial order relations. The concept lattice construction algorithm adopts a bottom-up closure calculation method. The specific process is as follows: First, calculate the object closure corresponding to each attribute subset in the formal background matrix. The object closure refers to the set of fields that satisfy all attributes in the attribute subset. Then, calculate the attribute closure for each object closure. The attribute closure refers to the set of all attributes that the field set commonly satisfies. When there is a one-to-one correspondence between the object closure and the attribute closure, the pair constitutes a formal concept. Finally, establish partial order relations between formal concepts according to the inclusion relationship of the attribute sets. The data structure of the concept node set is a record list. Each record corresponds to a formal concept and contains three attributes: concept identifier, concept extension, and concept connotation. The concept extension is the set of field identifiers contained in the formal concept, and the concept connotation is the set of attribute identifiers of the variation pattern corresponding to the formal concept. The data structure of the concept partial order relation set is a directed edge list. Each edge records the parent concept identifier and the child concept identifier, indicating that the attribute set of the child concept contains the attribute set of the parent concept.

[0065] S243: Supplement each concept node in the concept node set with a semantic tag, and assemble the concept node set and the concept partial order relation set into a field variation concept lattice. The semantic tag is automatically generated based on the variation pattern attribute type contained in the concept connotation, and is used to identify the semantic position of the concept node in the variation hierarchy. The value range of the semantic tag includes four types: standard concept, equivalent variant concept, extended variant concept, and anomalous concept. A standard concept is a concept node whose concept connotation is an empty set, indicating that the corresponding field adopts the standard form and has no variation; an equivalent variant concept is a concept node whose concept connotation only contains precision variation or format variation attributes, indicating that the corresponding field is completely equivalent to the standard form in semantics; an extended variant concept is a concept node whose concept connotation contains synonym variation or layout variation attributes, indicating that the corresponding field is semantically equivalent to the standard form but the expression form is extended; an anomalous concept is a concept node whose concept extension contains fields whose normalized value does not match any standard value, indicating that the corresponding field may have a substantial error. The data structure of the field variation concept lattice is a triple. The first element is the set of concept nodes, the second element is the set of partial order relations, and the third element is the identifier of the top element and the identifier of the bottom element. The field variation concept lattice organizes the normalized semantics and variation patterns of fields in a hierarchical structure, so that fields with the same normalized semantics and variation characteristics are aggregated into the same concept node, and the inclusion and expansion relationships between different variation patterns are characterized by partial order relations.

[0066] Specifically, compliance verification of premixed product packaging labels faces the challenge of complex and exhaustive variation patterns. Label content may undergo subtle changes due to regulatory updates, formula tweaks, or supplier changes. These changes manifest as different text at the character level, but may be completely equivalent at the semantic level. Traditional verification methods based on character similarity or single-point rules struggle to systematically manage these variation patterns, often requiring separate judgment rules for each variation type. The lack of a unified organizational structure among these rules easily leads to conflicts or omissions. By introducing formal concept analysis to construct a field variation concept lattice, the normalized semantics of fields and variation rules are modeled as nodes and partial order relations within the concept lattice. This allows the system to manage the relationships between standard syntax, equivalent variants, and extended variants within an ordered hierarchical structure. When encountering a new variation pattern, only the new attribute needs to be added to the variation pattern attribute set, and the concept lattice recalculated; no modification to the verification logic is required. Furthermore, the partial order relations of the concept lattice naturally support transitive reasoning between variation patterns. If a field belongs to an equivalent variant concept, it automatically inherits the semantic equivalence of the standard concept, eliminating the need for repeated judgment. This representation method provides structured mutation knowledge support for global consistency determination in S3, enabling the system to distinguish between legitimate mutations and substantive errors.

[0067] S3: Construct a candidate set for label verification based on the concept of field variation and the standard information of production batch, and perform global consistency judgment and anomaly classification decision based on the candidate set for label verification.

[0068] This step constructs a candidate set for label verification based on the field variation concept lattice and production batch standard information output from S2, and outputs the final label verification result through global consistency determination and anomaly classification decision. The label verification result serves as the final output of S3, guiding subsequent processing strategies on the production line.

[0069] Specifically, the process of constructing a candidate set for label verification based on the field variation concept grid and the production batch standard information, and performing global consistency judgment and anomaly classification decision based on the candidate set for label verification includes: S31: Read the standard information of the current production batch and construct a standard field set.

[0070] Furthermore, the process of reading the standard information of the current production batch and constructing a set of standard fields includes the following steps: S311: Query the product standard record corresponding to the current production batch from the production management database to obtain the original standard information data. The original standard information data includes fields such as product name, batch number, ingredient list, content list, production date, expiration date, company name, and implementation standard. The ingredient list and content list are array types, storing all ingredient names and their corresponding content standard values ​​respectively. The arrays have equal lengths and their positions correspond one-to-one.

[0071] S312: Perform field splitting and type labeling on the original standard information data to obtain a standard field set. The field splitting expands the ingredient list and content list into independent ingredient and content fields, and establishes an association label for each pair of ingredients and contents. The type labeling classifies each split field according to the field type determination rules defined in S21, obtaining field type labels. The data structure of the standard field set is a record list, with each record containing six attributes: standard field identifier, standard field name, standard field value, standard field type, field level label, and associated field identifier. The field level label has two value ranges: critical and non-critical. Critical fields are those that must be completely consistent during verification, including batch number, ingredient content, and production date. Non-critical fields are those that are allowed to differ within the legal variation range, including formatted explanatory text and layout-related content. The field level labels are pre-configured according to the compliance requirements of the premix industry.

[0072] S32: Align the field variation concept lattice with the standard field set to generate a candidate set of label verifications.

[0073] Furthermore, the process of aligning the field variation concept grid with the standard field set and generating a candidate set for label verification includes the following steps: S321: Traverse each standard field in the standard field set, search for matching concept nodes in the concept node set of the field variation concept grid, and obtain a set of field matching results. The search and matching method is as follows: First, filter concept nodes with the same field type in the field variation concept grid according to the standard field type of the standard field; then, for the filtered concept nodes, check whether there are fields in their concept extension whose normalized values ​​are equal to the standard field values ​​under the semantic equivalence judgment condition; the semantic equivalence judgment condition is defined according to the field type. The semantic equivalence condition for numerical fields is that the absolute value of the difference between the normalized values ​​is less than the numerical equivalence threshold; the semantic equivalence condition for date fields is that the three components of year, month, and day are completely identical; the semantic equivalence condition for unit fields and text fields is that the normalized value strings are completely identical. The numerical equivalence threshold is configured according to the accuracy tolerance requirements for content labeling in the premix industry. The data structure of the field matching result set is a mapping table, where the key is the standard field identifier and the value is a list of matched concept node identifiers. When multiple concept nodes match the same standard field, the list contains all the matching concept node identifiers.

[0074] S322: Disambiguation is performed on standard fields in the field matching result set that have multiple matching concept nodes to obtain the disambiguated field matching result. The disambiguation method is as follows: for multiple concept nodes matching the same standard field, nodes with semantic labels of standard concepts are selected first; if no standard concept node exists, nodes with semantic labels of equivalent variant concepts are selected first; if multiple candidates still exist, spatial position consistency is compared between the adjacent field list of the field in the concept extension and the associated field identifier of the standard field, and the concept node with the most consistent spatial position is selected. The data structure of the disambiguated field matching result is the same as that of the field matching result set, but the length of the concept node identifier list corresponding to each standard field does not exceed one.

[0075] S323: Generate candidate explanations based on the resolved field matching results and assemble them into a tag verification candidate set. The candidate explanation refers to a complete description of the matching relationship between the current tag and standard information, including the correspondence between each standard field and a concept node, as well as the corresponding mutation path information. The method for generating candidate explanations is as follows: For the case where all standard fields in the resolved field matching results have matching concept nodes, a single candidate explanation is generated. This candidate explanation records the mapping from each standard field to its matching concept node, as well as the semantic label and conceptual connotation of that concept node. For the case where some standard fields have no matching concept nodes, multiple candidate explanations are generated, corresponding to different assumptions such as the unmatched field being judged as missing, replaced by other fields, or having content errors. The data structure of the tag verification candidate set is a record list, with each record corresponding to a candidate explanation, containing four attributes: candidate explanation identifier, field mapping table, list of unmatched standard fields, and list of unmatched tag fields. The field mapping table records the mapping relationship from standard field identifiers to concept node identifiers and the corresponding mutation paths. The list of unmatched standard fields records the standard field identifiers that exist in the standard field set but have not found matching concept nodes. The list of unmatched label fields records field identifiers that exist in the field variation concept cell but are not matched by any standard field.

[0076] S33: Perform a compliance assessment on each candidate interpretation in the label verification candidate set and establish a partial order relationship between candidate interpretations. See Figure 7, which is a schematic diagram of candidate interpretations and global consistency determination provided in this application embodiment. As shown in the figure, this diagram illustrates the evaluation and screening process of the label verification candidate set. The upper part displays three candidate interpretation cards side-by-side, each card divided into two areas: key-level field matching status and non-key-level variation statistics. Candidate interpretation A has all three key-level fields (batch number, ingredient content, production date) successfully matched, and among the non-key-level fields, the standard number is 2, the equivalent number is 1, and the anomaly number is 0, marked at the bottom as a maximum candidate interpretation; Candidate interpretation B also has all key-level fields matched, but the non-key-level anomaly number is 1; Candidate interpretation C failed to match the ingredient content field. The middle area shows the partial order relationship of candidate interpretations, with candidate interpretation A being superior to candidate interpretation B, which is superior to candidate interpretation C. The parentheses explain the basis for the partial order determination: those with higher key-level matching rates and fewer non-key anomalies are given priority. The bottom is the global consistency determination result box, displaying "Legitimate variation exists → Can be released". In the scenario of premixed product packaging label verification, traditional methods typically use a uniform character similarity threshold for binary judgment. When the threshold is set too strictly, legitimate variations are misjudged as errors; when the threshold is set too leniently, genuine errors may be missed. This application achieves an automated decision-making logic of "strict on critical information, flexible on non-critical information" by establishing a partial order relationship among candidate interpretations and searching for the maximum candidate interpretation. While ensuring the correctness of key information such as batch number and ingredient content, it automatically tolerates legitimate variations in non-critical fields.

[0077] Furthermore, the process of performing compliance assessment and establishing partial order relations on candidate interpretations in the tag verification candidate set includes the following steps: S331: For each candidate interpretation in the tag verification candidate set, the key-level field matching and non-key-level field variation are statistically analyzed to obtain the candidate interpretation compliance assessment result. The statistical method for the key-level field matching is as follows: traverse the field mapping table of the candidate interpretation, filter standard fields with key-level labels, check whether their corresponding concept node semantic labels are standard concepts or equivalent variant concepts, and calculate the ratio of the number of key-level fully matching fields to the total number of key-level fields as the key-level matching rate. The statistical method for the non-key-level field variation is as follows: traverse the field mapping table of the candidate interpretation, filter standard fields with non-key-level labels, check their corresponding concept node semantic labels, and count the number of fields with semantic labels of standard concepts, equivalent variant concepts, extended variant concepts, and anomalous concepts, which are respectively recorded as the number of non-key standard fields, the number of non-key equivalent fields, the number of non-key extended fields, and the number of non-key anomalous fields. The data structure of the candidate explanation compliance assessment results is a record, which includes eight attributes: candidate explanation identifier, critical level matching rate, number of non-critical standards, number of non-critical equivalences, number of non-critical extensions, number of non-critical anomalies, number of unmatched standard fields, and number of unmatched tag fields.

[0078] S332: Based on the compliance assessment results of candidate explanations, establish a partial order relationship between candidate explanations to obtain a set of candidate explanation partial order relationships. The partial order relationship is defined as follows: Candidate explanation A is superior to candidate explanation B if and only if the critical-level matching rate of candidate explanation A is greater than or equal to the critical-level matching rate of candidate explanation B, the number of non-critical anomalies in candidate explanation A is less than or equal to the number of non-critical anomalies in candidate explanation B, and the number of unmatched standard fields in candidate explanation A is less than or equal to the number of unmatched standard fields in candidate explanation B, and at least one of the above three conditions is strictly unequal. The data structure of the candidate explanation partial order relationship set is a directed edge list, with each edge recording the identifier of the superior candidate explanation and the identifier of the inferior candidate explanation.

[0079] S34: Search for the largest candidate explanation in the set of partial order relations of candidate explanations and perform a global consistency determination.

[0080] Furthermore, the process of searching for the maximum candidate explanation in the partial order relation set of candidate explanations and performing a global consistency determination includes the following steps: S341: Traverse the partial order relation set of candidate explanations, identify candidate explanations that are not superior to any other candidate explanation, mark them as maximal candidate explanations, and obtain the maximal candidate explanation set. The maximal candidate explanation refers to a candidate explanation for which there is no better candidate explanation under the partial order relation. The data structure of the maximal candidate explanation set is an identifier list containing the candidate explanation identifiers of all maximal candidate explanations.

[0081] S342: Perform uniqueness and compliance checks on the set of maximum candidate explanations, and execute a global consistency determination to obtain a global consistency determination result. The uniqueness check method is as follows: check the number of elements in the set of maximum candidate explanations. If the number of elements is one, the uniqueness check passes; if the number of elements is greater than one, the uniqueness check fails. The compliance check method is as follows: for maximum candidate explanations that pass the uniqueness check, check whether their critical-level matching rate is equal to one, whether their number of non-critical anomalies is equal to zero, and whether their number of unmatched standard fields is equal to zero. If all three conditions are met, the compliance check passes; otherwise, the compliance check fails. The data structure of the global consistency determination result is a record, containing two attributes: a consistency determination flag and a determination basis. The consistency determination flag has four values: completely consistent, contains a legal variation, contains ambiguity, and contains anomalies. When the uniqueness test and compliance test pass, and the number of non-critical equivalences and non-critical extensions of the maximum candidate interpretation are both zero, the consistency determination flag is "completely consistent." When the uniqueness test and compliance test pass, and the number of non-critical equivalences or non-critical extensions of the maximum candidate interpretation is greater than zero, the consistency determination flag is "there is a legitimate variation." When the uniqueness test fails, the consistency determination flag is "there is ambiguity." When the compliance test fails, the consistency determination flag is "there is anomaly." The determination criteria record the specific field information and concept node information that led to this determination result.

[0082] S35: Execute anomaly classification decisions based on the global consistency judgment results and generate label verification results.

[0083] Furthermore, the process of performing anomaly classification decisions and generating label verification results based on the global consistency determination results includes the following steps: S351: Determine the processing strategy type based on the consistency determination flag of the global consistency determination results to obtain the processing strategy type flag. The processing strategy type flag has three values: allowable, requires manual review, and requires immediate alarm shutdown. When the consistency determination flag is completely consistent or has a legitimate variation, the processing strategy type flag is allowed; when the consistency determination flag is ambiguous, the processing strategy type flag is required to be manually reviewed; when the consistency determination flag is abnormal, further check the field-level tags involved in the abnormality. If there is a critical field abnormality, the processing strategy type flag is required to immediately alarm shutdown; if there is only a non-critical field abnormality, the processing strategy type flag is required to be manually reviewed.

[0084] S352: For cases where the processing strategy type is marked as requiring manual review or immediate alarm shutdown, extract detailed information of the difference fields and generate an anomaly classification report. The method for extracting the detailed information of the difference fields is as follows: read the field information causing the anomaly or ambiguity from the judgment criteria of the global consistency judgment result, including the standard field identifier, standard field value, matching concept node identifier, actual field value in the concept node's extension, concept node semantic label, and variation pattern attribute in the concept node's conceptual connotation. The data structure of the anomaly classification report is a record list, with each record corresponding to a difference field, containing five attributes: field name, standard value, actual value, difference type, and difference level. The difference type is automatically generated based on the concept node's semantic label and conceptual connotation, and its value range includes five types: precision difference, format difference, synonym difference, layout difference, and content error. The difference level is determined based on the field level label, and its value range includes two types: critical difference and non-critical difference.

[0085] S353: For cases where the processing strategy type is marked as requiring manual review or immediate alarm shutdown, visual annotation evidence is generated based on the label layout hypergraph in S1. The method for generating the visual annotation evidence is as follows: read the coordinates of the bounding rectangle of the node corresponding to the difference field in the label layout hypergraph, draw rectangular borders with different colors on the original label image to mark the position of the difference field, use the first annotation color for critical differences and the second annotation color for non-critical differences, wherein the first annotation color and the second annotation color are pre-configured as a color combination with high visual contrast. At the same time, add text descriptions to the edge area of ​​the image, including the difference field name and difference type. See Figure 8, which is a schematic diagram of the visual annotation evidence provided in the embodiment of this application. As shown in the figure, this figure shows the visual annotation effect generated by the system when the label verification result requires manual review or alarm shutdown. The main body of the figure is a simulated premixed packaging label, which includes the product name "compound premixed agent", product specifications, ingredient content table (vitamin A, vitamin D3, vitamin E, iron, zinc, manganese and their corresponding content values), production information (batch number, production date, expiration date, implementation standard) and manufacturer name, etc. The system overlays location markers for the difference fields onto the label image: red rectangles circle two critical difference fields, "80000 IU / kg" (vitamin D3 content) and "2024 / 1 / 15" (production date format); a yellow rectangle circles a non-critical difference field, "GB / T 9840" (implementation standard). The legend in the lower left corner illustrates the color coding rules: red boxes represent critical differences, and yellow boxes represent non-critical differences. In a real production environment, when the system detects an anomaly on the label, operators need to quickly locate the problematic field and determine its severity. This application, by directly overlaying location markers and color coding for the difference fields onto the original label image, allows reviewers to intuitively identify the problem without having to compare each field individually, significantly improving the efficiency of manual review. It also provides visual evidence for post-production quality audits and traceability.

[0086] S354: Assemble the processing strategy type marker, anomaly classification report, and visual annotation evidence to generate the label verification result. The data structure of the label verification result is a record, containing seven attributes: verification timestamp, image identifier, batch identifier, processing strategy type marker, consistency judgment marker, anomaly classification report, and visual annotation evidence path. The verification timestamp records the time of verification, the image identifier is inherited from the image identifier of the original label image in S1, the batch identifier is taken from the current production batch, and the visual annotation evidence path records the storage location of the generated annotation image. The label verification result is written to the verification result database and pushed to the production line control system. The production line control system performs operations such as release, pause for manual review, or emergency shutdown according to the processing strategy type marker.

[0087] S355: For cases where the processing strategy type is marked as allowable and the consistency judgment is marked as having legitimate variations, the variation path information is recorded for subsequent statistical analysis. The method for recording the variation path information is as follows: extract field mapping records with semantic labels of equivalent variant concepts or extended variant concepts from the field mapping table of maximum candidate interpretations; read the variation pattern attribute identifier from the conceptual connotation of the corresponding concept node; assemble the standard field identifier, the normalized value of the actual field, and the variation pattern attribute identifier into a variation path record; and write this record into the variation path statistical database. The records in the variation path statistical database can be used to analyze the label variation patterns caused by different production batches, different printing equipment, or different suppliers, providing data support for subsequent optimization of variation rule configuration.

[0088] Specifically, the core challenge of automated verification of premixed product packaging labels lies in how to reasonably tolerate non-critical variations arising from compliance evolution and production practices while strictly ensuring the accuracy of key information. Traditional verification methods typically use a uniform character similarity threshold for judgment. When the threshold is set too strictly, legitimate format variations and synonym substitutions are misjudged as errors, leading to numerous invalid alarms and increasing the burden on operators. When the threshold is set too leniently, genuine content errors may be missed, posing compliance risks. By introducing a candidate interpretation generation and partial order relation filtering mechanism, the system can search for the optimal interpretation among multiple possible interpretations, rather than simply making a binary judgment based on the similarity threshold. The definition of the partial order relation incorporates three dimensions: the matching rate of key-level fields, the number of non-key-level anomalies, and the number of unmatched fields, ensuring that the complete consistency of key information takes precedence over the tolerance for variations in non-key information. When multiple extremely large candidate interpretations exist, the system marks them as ambiguous rather than directly judging them as pass or fail, triggering a manual review process to avoid the risk of misjudgment that may arise from automated decision-making. This globally consistent judgment mechanism implements a strict but flexible automatic decision-making logic for critical aspects, significantly reducing invalid alarms and manual intervention while ensuring compliance. Simultaneously, through explicit annotation of variation patterns in the field variation concept grid, the system can finely categorize differing fields into precision variations, format variations, synonym variations, layout variations, and content errors, outputting structured anomaly classification reports. This allows operators to quickly pinpoint the nature and severity of problems during review. The generation of visually labeled evidence further enhances the traceability and interpretability of verification results, facilitating post-audit and quality traceability. Continuous recording and statistical analysis of variation path information provides a data foundation for iterative optimization of verification rules, enabling the system to continuously improve its ability to identify legitimate variations as production practice accumulates.

[0089] For example, a premixed agent manufacturer deploys this method on its packaging line to implement automatic label verification. When the packaging label of a batch of products passes through the inspection station, an industrial camera captures the original image of the label. A multi-scale semantic segmentation network identifies the pixel-level distribution of the text area, graphic area, and QR code area. The optical character recognition module performs character recognition on the text area and obtains a set of key fields containing product name, ingredient content table, production date, and batch number through field aggregation. The label layout hypergraph constructed based on the key field set associates the ingredient name, content value, and unit of measurement into clustered field combinations through hyperedges. In S2, the system identifies that the original value of a certain content value field in the batch label is 1.00%, which differs from the 1.0% of historical batches in terms of decimal places. After value normalization, the normalized value is equal, and this difference is classified as a precision variation pattern by the field variation concept grid. In S3, the system reads the standard content value of the ingredient from the production batch standard information as 1%. It searches for a concept node with a normalized value of 0.01 in the field variation concept grid. The semantic label of this node is equivalent variant concept, and the concept connotation includes precision variation attribute. The constructed candidate explanations showed a critical-level matching rate of 1, zero non-critical anomalies, and a global consistency assessment indicating the existence of a legitimate variation. The processing strategy type was marked as allowable. The system recorded this precision variation path and allowed the packaging to continue circulating without triggering manual review or alarm shutdown. While ensuring the accuracy of the content information, it automatically identified and tolerated legitimate precision expression changes.

[0090] Example 2: Based on Example 1, this example provides an automatic verification system for premixed product packaging labels based on image recognition, as shown in Figure 9. It includes: an image acquisition and layout hypergraph construction module: used to acquire original images of premixed product packaging labels, extract key fields through multi-scale semantic segmentation, and construct a label layout hypergraph based on the spatial and logical relationships between fields; a field semantic parsing and variant concept lattice construction module: used to parse the semantic features in the label layout hypergraph, perform field type identification and semantic normalization, and construct a field variant concept lattice based on variant pattern attributes; and a global consistency determination and anomaly decision module: used to align production batch standard information with the field variant concept lattice to generate a candidate set, filter the optimal interpretation through partial order relations, perform global consistency determination, output verification results, and trigger corresponding processing flows.

Claims

1. A method for automatic verification of premix packaging labels based on image recognition, characterized in that, Includes the following steps: S1: Acquire the original image of the premix packaging label, perform multi-scale semantic segmentation and optical character recognition on the original label image to obtain a set of key fields, and construct a label layout hypergraph based on the set of key fields; S2: Based on the topological and semantic features between fields in the label layout hypergraph, obtain the field semantic feature set through semantic normalization, and construct the field variation concept lattice based on the field semantic feature set; S3: Construct a label verification candidate set based on the field variation concept lattice and production batch standard information, perform global consistency judgment and anomaly classification decision based on the label verification candidate set, and generate label verification results.

2. The method for automatic verification of premixed product packaging labels based on image recognition according to claim 1, characterized in that, The steps of performing multi-scale semantic segmentation and optical character recognition on the original label image to obtain a key field set include: inputting the original label image into a multi-scale semantic segmentation network to obtain an initial segmentation probability map containing probability values ​​of text area, graphic area, QR code area, and background area, and generating a pixel-level region mask based on pixel-level category determination and morphological closing operation; extracting a set of text area image blocks based on the pixel-level region mask, performing optical character recognition on the set of text area image blocks to obtain a character recognition result set containing character content and global bounding rectangle coordinates; performing row detection and column alignment detection on the character recognition result set according to the global bounding rectangle coordinates, aggregating discrete characters into fields, and performing association operations on cross-row fields to obtain the key field set.

3. The method for automatic verification of premix packaging labels based on image recognition according to claim 1, characterized in that, The steps of constructing a tag layout hypergraph based on the set of key fields include: generating a hypergraph node set with each field in the set of key fields as a node; calculating the relative orientation and spacing ratio between node pairs in the hypergraph node set, identifying clustered field combinations that are spatially closely arranged and semantically related, and generating a set of hyperedges that simultaneously connect all nodes within the clustered field combinations; supplementing each hyperedge in the set of hyperedges with semantic relationship attributes, including modification relationships, parallel relationships, and subordinate relationships; and assembling the hypergraph node set and the hyperedge set into the tag layout hypergraph, which is used to represent the spatial structure and content distribution of tags.

4. The method for automatic verification of premixed product packaging labels based on image recognition according to claim 1, characterized in that, The step of obtaining a set of semantic features of fields by parsing the topological and semantic features between fields based on the label layout hypergraph and performing semantic normalization processing includes: reading the node attributes and hyperedge information in the label layout hypergraph; identifying the field type through lexical analysis and hyperedge adjacency relationships, wherein the field type includes numeric, unit, date, and text types; performing differentiated semantic normalization on the text content of the fields according to the field type: unifying the floating-point format and precision for numeric fields, mapping standard unit notation for unit fields, reorganizing date fields into a unified date format, and performing character normalization and synonym replacement for text fields; and integrating the normalization results, field types, and the list of adjacent fields and the list of bound fields provided by the label layout hypergraph to construct the set of semantic features of the fields.

5. The method for automatic verification of premixed product packaging labels based on image recognition according to claim 1, characterized in that, The step of constructing a field variation concept lattice based on the field semantic feature set includes: introducing premixed label verification domain rules and historical legal change samples to construct a variation pattern attribute set, which includes precision variation, format variation, synonym variation, and layout variation; constructing a formal background matrix based on the field semantic feature set and the variation pattern attribute set; generating a concept node set and a concept partial order relation set using formal concept analysis methods; supplementing each concept node with a semantic label, which includes standard concept, equivalent variant concept, extended variant concept, and anomalous concept; and assembling to generate the field variation concept lattice.

6. The method for automatic verification of premixed product packaging labels based on image recognition according to claim 1, characterized in that, The step of constructing a candidate set for tag verification based on the field variation concept lattice and the standard information of the production batch includes: reading the standard information of the current production batch and constructing a standard field set, wherein the fields in the standard field set are marked as critical or non-critical; traversing the standard field set, searching for semantically equivalent concept nodes in the field variation concept lattice, and generating a resolved field matching result through ambiguity resolution; generating multiple candidate interpretations based on the resolved field matching results, assembling them into the candidate set for tag verification, wherein each candidate interpretation contains the mapping relationship from standard fields to concept nodes and the variation path.

7. The method for automatic verification of premix packaging labels based on image recognition according to claim 6, characterized in that, The step of performing a global consistency determination based on the tag verification candidate set includes: statistically analyzing the key-level field matching rate and non-key-level field variation for each candidate explanation in the tag verification candidate set, and establishing a partial order relationship between candidate explanations; the partial order relationship is defined as follows: if candidate explanation A has a higher key-level matching rate and fewer non-key anomalies and fewer unmatched standard fields, then candidate explanation A is determined to be superior to candidate explanation B; searching for the maximal candidate explanation that is not superior to other candidate explanations in the partial order relationship, performing uniqueness and compliance checks on the maximal candidate explanation set, and outputting a global consistency determination result, the result including complete consistency, existence of legal variation, existence of ambiguity, and existence of anomalies.

8. The method for automatic verification of premixed product packaging labels based on image recognition according to claim 7, characterized in that, The step of performing anomaly classification decision based on the tag verification candidate set includes: determining the processing strategy type according to the global consistency judgment result, wherein the processing strategy type includes allowable, manual review required, and immediate alarm shutdown required; if the processing strategy type is manual review required or immediate alarm shutdown required, then extract the detailed information of the difference field that caused the judgment, generate an anomaly classification report, and clarify the difference type as precision difference, format difference, synonym difference, layout difference, or content error; if the processing strategy type is allowable and there is a legitimate variation, then record the variation path information to the statistical database.

9. The method for automatic verification of premix packaging labels based on image recognition according to claim 1, characterized in that, The step of generating the label verification result further includes: generating visual annotation evidence on the original label image based on the node coordinate information in the label layout hypergraph; drawing bounding boxes using a first annotation color for critical level differences and using a second annotation color for non-critical level differences; and associating the visual annotation evidence with the anomaly classification report and outputting it as part of the label verification result.

10. An automatic verification system for premixed product packaging labels based on image recognition, used to implement the automatic verification method for premixed product packaging labels based on image recognition as described in any one of claims 1-9, characterized in that, The system includes: an image acquisition and layout hypergraph construction module, used to acquire original images of premixed product packaging labels, extract key fields through multi-scale semantic segmentation, and construct a label layout hypergraph based on the spatial and logical relationships between fields; a field semantic parsing and variant concept lattice construction module, used to parse the semantic features in the label layout hypergraph, perform field type identification and semantic normalization, and construct a field variant concept lattice based on variant pattern attributes; and a global consistency determination and anomaly decision module, used to align production batch standard information with the field variant concept lattice to generate a candidate set, select the optimal interpretation through partial order relations, perform global consistency determination, output verification results, and trigger corresponding processing flows.

Images