A canning product sorting and identifying method and system based on image processing

Abstract

The application discloses a kind of based on image processing's canned product sorting identification method and system, it is related to industrial vision technology field.This kind of based on image processing's canned product sorting identification method and system, including S1, canned body sorting observation data are collected and are gathered with same jar, time sequence calibration and standardization expression;S2, label visibility is received analysis, according to the analysis result, it is generated and label visibility is insufficiently marked;S3, label identity confirmation analysis is carried out, product coding is locked and review sorting rule import;S4, according to sorting execution synchronous analysis result, sorting mechanism action trigger and review channel import are carried out.The existing sorting system is solved, in non-directional conveying and canned body random rotation scene, it is difficult to distinguish label real missing and temporary invisible caused by posture shielding only by fixed station single frame image, leading to inconsistent identification result and sorting path mis-triggering problem.

Description

Technical Field

[0001] This invention relates to the field of industrial vision technology, specifically to a sorting and identification method and system for canned products based on image processing. Background Technology

[0002] With the development of automated production lines and intelligent manufacturing technologies, the demand for automatic identification and sorting of industrial products in production, packaging, warehousing, and outbound processes is constantly increasing. For batch products in continuous conveying conditions, manufacturers typically need to use automated inspection equipment to identify product appearance, labeling information, packaging status, or category attributes, and perform processes such as sorting, classification, verification, or rejection based on the identification results to improve production line efficiency and product flow management. Machine vision technology, due to its non-contact, continuous acquisition, fast response speed, and ease of integration with automated control systems, has been widely used in product inspection and sorting scenarios in fields such as food and beverage, daily chemical products, pharmaceutical packaging, electronics manufacturing, and logistics warehousing. Existing machine vision sorting systems typically complete online inspection tasks through image acquisition, image preprocessing, target localization, feature extraction, category recognition, and execution control. They can also be combined with batch information, order information, or basic product information in the production management system to achieve product flow management and quality traceability. As production pace increases and product variety grows, visual recognition systems are gradually evolving from simple appearance inspection to multi-category recognition, continuous target tracking, process data recording, and sorting result traceability. Sorting control systems are also gradually linking with conveyor control, actuator control, and production data management systems to meet the requirements of recognition accuracy, processing continuity, and management traceability in automated production scenarios.

[0003] For example, the invention patent with announcement number CN108176608B discloses a machine vision-based method and device for detecting nut defects, which relates to the field of parts defect detection. The purpose is to improve the efficiency of nut defect detection. The method first acquires top view, side view, and oblique view images of the nut and performs binarization processing on the above images. Then, it extracts the feature parameters of the nut to be detected from the binarized images and compares the extracted feature parameters with the pre-set feature parameters to determine whether the nut to be detected has problems such as excessively large or small inner diameter, excessive cutting of the upper groove, no upper groove, reversed upper groove, incorrect number of outer grooves, inverted top and bottom, no internal thread, or overlap with other nuts. Then, the nuts to be detected are classified to distinguish between scrap, defective, abnormal and qualified products, which improves the efficiency of nut inspection and increases the pass rate of nuts leaving the factory.

[0004] For example, invention patent CN105327864B discloses an automatic sorting device, including an integrated feeding device, a transition wheel receiving device, a bamboo strip identification device, a flipping device, a conveying device, and a sorting ejection device. The integrated feeding device, the transition wheel receiving device, and the conveying device are arranged in sequence. The bamboo strip identification device is installed in the middle of the conveying device, and a flipping device is installed inside the bamboo strip identification device. Multiple sorting ejection devices are respectively installed on the side ends of the conveying device. This device realizes automatic material feeding through the integrated feeding device, completes material flipping and identification through the flipping device in conjunction with the bamboo strip identification device, and automatically sorts the identified materials through the sorting ejection devices. It realizes automatic material feeding, automatic flipping, automatic identification, and automatic sorting, which improves product quality and product stability, increases product production efficiency, and reduces labor costs.

[0005] Existing machine vision-based sorting and recognition methods largely rely on single-frame images from fixed inspection stations for category determination and sorting control. When products are in a non-directional transport state, or when their surfaces undergo random rotation, partial occlusion, glare interference, or posture changes during transport, single-frame images typically only reflect a momentary visible state, making it difficult to accurately distinguish between genuine label loss and temporary invisibility caused by posture occlusion. For canned products, the label area may enter or leave the camera's field of view as the can rotates. Directly relying on single-frame images for label loss judgment or product category confirmation can easily lead to inconsistent recognition results for the same can, further causing erroneous sorting path triggers, incorrect rejection of abnormal items, or misjudgment of items under review. Therefore, existing sorting systems still face the problem of difficulty in distinguishing between genuine label loss and temporary invisibility due to posture occlusion in non-directional transport and random can rotation scenarios.

[0006] Therefore, in order to address the above problems, there is an urgent need for a sorting and identification method and system for canned products based on image processing. Summary of the Invention

[0007] Technical problems to be solved To address the shortcomings of existing technologies, this invention provides a sorting and identification method and system for canned products based on image processing. This solves the problem that existing sorting systems, in scenarios involving non-directional transport and random can rotation, struggle to distinguish between genuine label loss and temporary invisibility caused by posture occlusion using only a single frame image from a fixed workstation, leading to inconsistent identification results and erroneous sorting path triggering.

[0008] Technical solution To achieve the above objectives, the present invention provides the following technical solution: a sorting and identification method for canned products based on image processing, comprising: S1, collecting can sorting observation data and performing same-can aggregation, time-series calibration, and standardized expression; S2, performing label visibility continuity analysis on the can sorting observation data, and generating identifiable markers and insufficient label visibility markers based on the label visibility continuity analysis results; S3, performing label identity verification analysis based on identifiable markers and insufficient label visibility markers, and locking product codes and importing review sorting rules based on the label identity verification analysis results; S4, performing sorting execution synchronization analysis on the can sorting observation data, and triggering sorting mechanism actions and importing review channels based on the sorting execution synchronization analysis results.

[0009] Furthermore, the specific process of collecting tank sorting observation data is as follows: collecting tank sorting observation data, which includes: the current batch label template image set data, the current batch product code data, and the current batch sorting channel number data read by the production management system; the sorting mechanism action receipt timestamp data and sorting mechanism action status feedback data read by the sorting execution mechanism controller; tank multi-frame image data; image acquisition timestamp data; camera exposure time feedback data; light source current feedback data; conveyor encoder pulse data; encoder pulse timestamp data; and photoelectric trigger timestamp data. The current tank is bound to the current batch product code data through the photoelectric trigger timestamp data.

[0010] Further, the specific process of same-tank image data aggregation, timing calibration, and standardized representation is as follows: Multi-frame image data of the tank is sorted according to the image acquisition timestamp data, removing duplicate timestamp frames, empty frames, and failed acquisition frames; multi-frame image data of the tank is grouped according to photoelectric trigger timestamp data to form same-tank image groups for a single tank; image size unification, boundary cropping, background grayscale compression, and grayscale range correction are performed on the multi-frame image data of the tank; time reference alignment is performed on the conveyor encoder pulse data, encoder pulse timestamp data, photoelectric trigger timestamp data, and image acquisition timestamp data; field format unification processing is performed on the sorting mechanism action receipt timestamp data and sorting mechanism action status feedback data; batch field consistency processing is performed on the current batch label template image set data, current batch product code data, and current batch sorting channel number data; tank sorting observation data is standardized using a zero-mean unit variance standardization algorithm; and multi-frame image data of the tank, camera exposure time feedback data, light source current feedback data, and conveyor encoder pulse data are normalized using a maximum-minimum value normalization algorithm.

[0011] Further, the specific process of performing label visibility analysis on the tank sorting observation data is as follows: The outer contour of the tank body in multi-frame image data is fitted to obtain the tank body region through contour closure screening; the texture coherence of the tank body region is screened through color change density statistics and texture gradient extraction to obtain candidate label texture regions; the label candidate coverage is obtained based on the ratio of the area of ​​the candidate label texture region to the area of ​​the tank body region; the position of the candidate label texture regions in adjacent frames is matched through feature point registration, and the cross-frame label display amount is obtained by combining image acquisition timestamp data; the reflective band is extracted from the brightness saturated region in multi-frame image data of the tank body through brightness saturation connected component screening, and abnormal frames are corrected by combining camera exposure time feedback data and light source current feedback data to obtain the reflective occlusion interference amount; the distance between the candidate label texture region and the edge of the tank body is statistically analyzed through tank body boundary positioning to obtain the curved edge compression amount; the attitude analysis of the outer contour of the tank body is performed through tank body centerline fitting and boundary tilt angle calculation to obtain the tank body attitude deflection amount; the distance between the candidate label texture region and the edge of the tank body is statistically analyzed through adjacent frames. The rate of change of the tank's attitude deflection is calculated by differential conversion between the tank's attitude deflection and the image acquisition timestamp data. The white space area is extracted from the tank body region through texture connectivity filtering and color variance statistics to obtain the white space occupancy. The sampling window for the same tank is determined by the trigger time corresponding to the photoelectric trigger timestamp data and the image acquisition timestamp data after that trigger time. The candidate label coverage is calculated and incremented by one to obtain the candidate coverage enhancement term. The cross-frame label display amount within the same tank sampling window is integrated and incremented by one to obtain the cross-frame display accumulation term. The calculation... The sum of reflective occlusion interference and curved edge compression is taken as an inverse hyperbolic sine and then one is added to obtain the imaging interference compression term. The candidate coverage enhancement term is multiplied by the cross-frame display accumulation term, divided by the imaging interference compression term, one is added, and the natural logarithm is taken to obtain the visibility inheritance base term. The negative exponent of the tank attitude deflection change rate is taken to obtain the attitude change suppression term. The white space occupancy is added to one and then the reciprocal is taken to obtain the white space occupancy suppression term. The product of the visibility inheritance base term, attitude change suppression term, and white space occupancy suppression term is calculated to obtain the label visibility inheritance value.

[0012] Furthermore, the specific process for generating identifiable markers and insufficient label visibility markers based on the label visibility acceptance analysis results is as follows: Real-time comparison of label visibility acceptance value and label visibility acceptance threshold: When the label visibility acceptance value is greater than or equal to the label visibility acceptance threshold, identification suppression is performed on the brightness saturated connected domain corresponding to the reflective occlusion interference and the adjacent part of the tank edge corresponding to the arc edge compression. Only the label candidate texture region with the superimposed label candidate coverage greater than or equal to the label candidate coverage threshold is retained to participate in label identity confirmation and generate identifiable markers; When the label visibility acceptance value is less than the label visibility acceptance threshold, no label missing judgment is output before cross-frame supplementary confirmation is completed; Cross-frame supplementary confirmation of the same tank and the same tank image group is initiated, and the area superposition and texture coherence verification of the label candidate texture regions with matching positions in adjacent frames are performed; If the superimposed label candidate coverage is still less than the label candidate coverage threshold, an insufficient label visibility marker is generated, and the current tank is marked as a review object.

[0013] Furthermore, the specific process of label identity verification analysis based on identifiable markers and insufficient label visibility markers is as follows: The label template matching quantity is obtained based on the scale-invariant feature transform matching results between multi-frame image data of the tank and the current batch of label template image set data; the trademark outline, barcode stripe direction, text connected components, and main pattern edge structure are extracted from the candidate texture region of the label, and the key identifier completeness quantity is obtained based on the coverage ratio of the corresponding region in the current batch of label template image set data; the candidate category similarity is ranked using the current batch of label template image set data, the current batch of product code data, and multi-frame image data of the tank, and the product category separation quantity is obtained based on the similarity interval between the first and second candidate categories; the cross-frame label conflict quantity is obtained based on the number of times the candidate product codes are inconsistent in different image frames of the same tank image group obtained by ranking the candidate category similarity; the extracted trademark outline, barcode stripe direction, text connected components, and main pattern edge structure are further analyzed. The results are compared with the corresponding regions in the current batch of label template image set data. The label evidence gap is obtained based on the area ratio of the corresponding regions that do not form a valid match. The label template matching quantity is calculated and one is added to obtain the template matching enhancement term. The product of the key identifier completeness quantity and the product category separation quantity is calculated, and the square root of the product is added to obtain the key separation coordination term. The label evidence gap quantity within the same tank sampling window is time-weighted integraled according to the sampling time interval of adjacent frames corresponding to the image acquisition timestamp data to obtain the cumulative label evidence gap quantity. The absolute value of the cross-frame label conflict quantity, the sum of the cumulative label evidence gap quantity and the value of one are calculated to obtain the evidence conflict damping term. The template matching enhancement term and the key separation coordination term are multiplied and divided by the evidence conflict damping term, one is added and the natural logarithm is taken to obtain the identity confirmation basic term. The hyperbolic tangent of the label visibility acceptance value is taken and one is added to obtain the visibility enhancement term. The product of the identity confirmation basic term and the visibility enhancement term is calculated to obtain the label identity confirmation value.

[0014] Furthermore, the specific process of locking product codes and importing review sorting rules based on the label identity verification analysis results is as follows: Real-time comparison of label identity verification value and label identity verification threshold: When the label identity verification value is greater than or equal to the label identity verification threshold, the product code of the current tank is locked, and the regular sorting category is determined based on the correspondence between the current batch product code data and the current batch sorting channel number data. The locked product code is written into the identification status record of the current tank, and the product code is not updated based on the single frame identification result before the current tank completes sorting; When the label identity verification value is less than the label identity verification threshold, the review channel category is determined based on the review sorting rules, the current tank is marked with a label missing risk mark or an identity conflict mark, and the current tank is imported into the review sorting rules; The regular sorting category and the review channel category are used together as the sorting target category.

[0015] Furthermore, the specific process of performing synchronous analysis on the tank sorting observation data is as follows: The tank's center point is located using a target tracking algorithm across multiple frames of image data. A temporal mapping between the tank's center trajectory and conveyor displacement is completed by combining image acquisition timestamp data, conveyor line encoder pulse data, and encoder pulse timestamp data. The tank's position binding quantity is obtained based on the mapping consistency. The sorting channel matching quantity is obtained based on the mapping consistency between the current batch sorting channel number data and the sorting target category. The conveying speed disturbance is obtained based on the pulse interval changes in the conveyor line encoder pulse data and encoder pulse timestamp data. The conveying speed disturbance change rate is obtained by differentiating the conveying speed disturbances corresponding to adjacent encoder pulse timestamp data. The boundary relationship between the target tank and adjacent tanks is analyzed using a target detection algorithm. The interference quantity between adjacent tanks is obtained based on the boundary distance, boundary overlap ratio, and center trajectory interval. The estimated arrival time of the target tank at the sorting execution position is calculated using image acquisition timestamp data, conveyor line encoder pulse data, and encoder pulse timestamp data. The difference between the estimated arrival time and the sorting mechanism action feedback timestamp data is then calculated, and the results are combined with sorting mechanism action status feedback data to correct for inaction, delayed action, and action failure, yielding the execution feedback deviation. The tank position binding amount is calculated and incremented by one to obtain the position binding enhancement term. The sorting channel matching amount is calculated and incremented by one to obtain the channel matching enhancement term. The label identity verification value is taken as the hyperbolic tangent and incremented by one to obtain the identity verification enhancement term. The absolute value of the conveyor speed disturbance change rate, the interference amount of adjacent tanks, the inverse hyperbolic sine value of the execution feedback deviation, and the sum of these factors are calculated to obtain the execution synchronization damping term. The position binding enhancement term, channel matching enhancement term, and identity verification enhancement term are multiplied, divided by the execution synchronization damping term, and then the arctangent is taken to obtain the sorting execution synchronization value.

[0016] Furthermore, the specific process of triggering sorting mechanism actions and importing verification channels based on the sorting execution synchronization analysis results is as follows: Real-time comparison of sorting execution synchronization value and sorting execution synchronization threshold: When the sorting execution synchronization value is greater than or equal to the sorting execution synchronization threshold, the corresponding sorting mechanism action is triggered according to the locked product code and sorting category, controlling the sorting mechanism to act when the target can arrives at the execution window, and recording the action completion status; When the sorting execution synchronization value is less than the sorting execution synchronization threshold, the triggering of direct rejection action is stopped, the target can is imported into the verification channel, and adjacent cans are allowed to pass at the same time; Finally, sorting control instructions, verification channel instructions, and sorting execution results are output.

[0017] The second aspect of this invention provides an image processing-based sorting and identification system for canned products, comprising: a data acquisition and preprocessing module for acquiring can sorting observation data and performing same-can aggregation, time-series calibration, and standardized representation; a can rotation attitude and label visibility discrimination module for performing label visibility continuity analysis on the can sorting observation data, and generating identifiable markers and insufficient label visibility markers based on the label visibility continuity analysis results; a multi-frame label identity verification and missing risk discrimination module for performing label identity verification analysis based on identifiable markers and insufficient label visibility markers, and locking product codes and importing review sorting rules based on the label identity verification analysis results; and a sorting decision and execution verification module for performing sorting execution synchronization analysis on the can sorting observation data, and triggering sorting mechanism actions and importing review channels based on the sorting execution synchronization analysis results.

[0018] Beneficial effects The present invention has the following beneficial effects: (1) This invention, by jointly discriminating the rotation posture of the can, the candidate texture area of ​​the label, the reflection occlusion and the compression of the curved edge, performs visibility analysis first when the label image does not have the conditions for direct recognition, instead of directly outputting the label missing result, thereby achieving the effect of delayed confirmation of label missing risk and suppression of misjudgment, effectively solving the problem in the prior art that the label is temporarily invisible due to random rotation of the can and is misjudged as a label missing.

[0019] (2) This invention incorporates label template matching, key identification integrity, product category separation, cross-frame label conflict and label evidence gap into the identity verification process, so that product code verification is based on the consistency of multiple types of label evidence, thereby achieving the effect of multi-evidence verification and confirmation of product identity, and effectively solving the problem of inconsistent identification results of the same tank at different collection times in the prior art.

[0020] (3) This invention generates a label missing risk mark or identity conflict mark when the identity is not confirmed, and imports the corresponding tank into the review and sorting rules to avoid unconfirmed targets from directly entering the regular sorting path, thereby achieving the effect of abnormal target diversion and review, and effectively solving the problem that the sorting path is easily triggered when the identity is not confirmed in the prior art.

[0021] (4) This invention synchronizes the tank center trajectory, conveying displacement, sorting channel mapping and actuator feedback to ensure that the sorting action corresponds to the target tank's position, thereby achieving the effect that the identification result is consistent with the sorting execution action sequence, effectively solving the problem of incorrect rejection caused by conveying speed fluctuations or execution feedback deviations in the prior art.

[0022] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0023] Figure 1 This is a flowchart of a canned product sorting and identification method based on image processing according to the present invention; Figure 2 This is a structural diagram of a canned product sorting and identification system based on image processing according to the present invention; Figure 3 This is a bar chart showing the label visibility acceptance values ​​of the present invention. Figure 4 This is a dual-axis comparison chart of the candidate label coverage and the label visibility acceptance value of the present invention. Figure 5 This is a flowchart illustrating the label visibility acceptance determination process of the present invention. Detailed Implementation

[0024] 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.

[0025] Please see Figures 1-5 This invention provides a technical solution: a sorting and identification method for canned products based on image processing, comprising the following steps: S1, collecting can sorting observation data and performing same-can aggregation, time-series calibration, and standardized expression; S2, performing label visibility continuity analysis on the can sorting observation data, and generating identifiable markers and insufficient label visibility markers based on the label visibility continuity analysis results; S3, performing label identity verification analysis based on identifiable markers and insufficient label visibility markers, and locking product codes and importing review sorting rules based on the label identity verification analysis results; S4, performing sorting execution synchronization analysis on the can sorting observation data, and triggering sorting mechanism actions and importing review channels based on the sorting execution synchronization analysis results.

[0026] Specifically, the process of collecting tank sorting observation data is as follows: The tank sorting observation data includes: the current batch label template image set data, the current batch product code data, and the current batch sorting channel number data read from the production management system. The current batch label template image set data refers to the set of standard label images corresponding to the current batch product code data read from the production management system, including at least the trademark area template, barcode area template, text area template, and main pattern area template corresponding to each product code within the current batch; the sorting mechanism action receipt timestamp data and sorting mechanism action status feedback data read from the sorting execution mechanism controller; multi-frame image data of the tank; and images. The system collects timestamp data, camera exposure time feedback data, light source current feedback data, conveyor encoder pulse data, encoder pulse timestamp data, and photoelectric trigger timestamp data. The photoelectric trigger timestamp data is used to bind the current tank to the current batch product code data. The current batch label template image set data, the current batch product code data, and the current batch sorting channel number data are issued by the production management system and cached at the inspection station during batch switching. The sorting mechanism action receipt timestamp data and the sorting mechanism action status feedback data are transmitted back in real time by the sorting execution mechanism controller after the sorting action is completed. The current tank establishes a correspondence with the current batch data cached at the inspection station through the photoelectric trigger timestamp data.

[0027] This implementation plan establishes a complete, unified, and traceable data foundation for the can sorting and identification process. This ensures that cans entering the inspection station are identified and judged under the same batch standard, product coding system, and sorting channel system, avoiding discrepancies between image objects, batch templates, product categories, and sorting targets in subsequent analysis stages. By forming a unified observation basis for the current can, current batch data, and sorting execution feedback information, it supports the continuous calculation of subsequent label visibility acceptance values, label identity confirmation values, and sorting execution synchronization values. This ensures that label identification, product coding locking, sorting category determination, and action feedback traceability have a consistent data source, improving the data integrity, temporal consistency, and result verifiability of the online identification and sorting control process for canned products.

[0028] Specifically, the process of same-tank image aggregation, timing calibration, and standardized representation is as follows: Multi-frame image data of the tank are organized according to the image acquisition timestamp data, removing duplicate timestamp frames, empty frames, and failed acquisition frames, retaining multi-frame images corresponding to the same photoelectric trigger timestamp data to form a same-tank image group for a single tank; The multi-frame image data of the tank undergoes image size unification, boundary cropping, background grayscale compression, and grayscale range correction to ensure consistent data format and grayscale representation for images acquired at different times; The encoder pulse data, encoder pulse timestamp data, photoelectric trigger timestamp data, and image acquisition timestamp data of the conveyor line are aligned with the time reference to ensure that image frames, trigger signals, and conveyor pulses are in the same timing chain; The sorting mechanism action receipt timestamp data and sorting mechanism action status feedback data undergo field format unification processing, converting the action status feedback data into a unified status code, and establishing a correlation between the action receipt timestamp data and the corresponding photoelectric trigger timestamp data; The current batch label template image set data, the current batch product code data, and... The current batch sorting channel number data undergoes batch field consistency processing to establish a correspondence between label templates, product codes, and sorting channel numbers. A zero-mean unit variance standardization algorithm is used to standardize tank sorting observation data, mitigating the impact of differences in measurement dimensions between different acquisition times, conveyor cycles, and equipment feedback on subsequent calculations. A maximum-minimum value normalization algorithm is used to normalize multi-frame tank image data, camera exposure time feedback data, light source current feedback data, and conveyor encoder pulse data, transforming image intensity, time intervals, conveyor pulse variations, and execution feedback variations involved in subsequent formulas into dimensionless data. Both the zero-mean unit variance standardization algorithm and the maximum-minimum value normalization algorithm use historical sliding window data corresponding to the current batch as the calculation benchmark, ensuring that image intensity, time intervals, conveyor pulse variations, and execution feedback variations at different acquisition times within the same batch maintain a uniform numerical scale. This guarantees that label visibility acceptance values, label identity confirmation values, and sorting execution synchronization values ​​can be compared using the same threshold system.

[0029] In this implementation plan, the original tank sorting observation data is organized into a unified input basis for subsequent discrimination calculations. This ensures that the image sequence, conveying sequence, triggering sequence, execution feedback, and batch fields of the same tank form a stable and consistent data representation, reducing interference from invalid image frames, format differences, timing offsets, inconsistent status calibers, and differences in numerical units on subsequent analysis. The processed data can support continuous recognition of image groups of the same tank, correspondence verification between label templates and product codes, and process traceability of sorting action feedback. It also ensures that label visibility acceptance values, label identity confirmation values, and sorting execution synchronization values ​​have a unified calculation scale and a unified comparison basis, improving the data reliability, computational stability, and threshold determination consistency of the tank sorting and identification process.

[0030] Specifically, the process of performing label visibility continuation analysis on can sorting observation data is as follows: Contour closure screening is used to fit the outer contour of the can body in multi-frame image data to obtain the can body region. Contour closure screening uses edge connectivity ratio, the distance between the beginning and end of the contour, and the length and width constraints of the circumscribed rectangle as screening conditions. Local connections are performed on broken edges before boundary fitting is executed. Texture coherence screening is performed on the can body region using color change density statistics and texture gradient extraction to obtain candidate label texture regions. Color change density statistics are used to extract areas with concentrated color jumps within the can body region, and texture gradient extraction is used to retain pattern, text, or barcode texture regions with directional continuity. Based on the candidate label textures... The ratio of the area of ​​the texture region to the area of ​​the tank body is used to obtain the label candidate coverage. This ratio is limited to calculations within the same tank body area to avoid coverage shifts caused by changes in tank scale across different image frames. Feature point registration is used to match the position of label candidate texture regions in adjacent frames. The feature point registration uses stable corner points, texture intersections, and edge junctions within adjacent frames of the same tank as the matching basis, and outlier point pairs whose matching distance exceeds the registration tolerance are removed. The sampling time interval between adjacent frames is determined based on image acquisition timestamp data. The cross-frame label display amount is obtained according to the degree of change in the overlapping area of ​​the label candidate texture regions within adjacent sampling time intervals. The degree of change in overlapping area is determined according to the label candidate area in adjacent frames. The correlation between the overlap area difference of texture regions and the sampling time interval is calculated. Reflective bands are extracted from the brightness saturation regions in multi-frame image data of the tank using brightness saturation connected component filtering. This filtering uses grayscale saturation value, connected component extension direction, and connected component width as constraints to eliminate isolated noise bright spots, forming reflective band regions. Abnormal frame correction is performed using camera exposure time feedback data and light source current feedback data to obtain the amount of reflective occlusion interference. Abnormal frame correction is based on the contribution of sudden changes in exposure time and light source current to the brightness saturation region, reducing the impact of fluctuations in acquisition conditions on the amount of reflective occlusion interference. Finally, the candidate texture region of the label and the edge of the tank are located using tank boundary positioning. The distance between the two sides is statistically analyzed to obtain the compression amount of the curved edge. The distance statistics are based on the shortest distance from the boundary of the candidate texture region of the label to the left and right curved boundaries of the tank body and the degree of concentration of the distance distribution. The attitude analysis of the outer contour of the tank body is performed by fitting the center line of the tank body and calculating the boundary tilt angle to obtain the attitude deflection of the tank body. The center line of the tank body is obtained by fitting the sequence of the central axis points of the outer contour of the tank body, and the boundary tilt angle is obtained by the degree of deviation of the two sides of the tank body relative to the vertical direction of the image. The change rate of the attitude deflection of the tank body and the image acquisition timestamp data is converted by the difference between adjacent frames to obtain the attitude deflection change rate of the tank body. The difference between adjacent frames is obtained by dividing the difference of the attitude deflection of consecutive frames of the same tank body by the corresponding sampling time interval.The white space area of ​​the can body is extracted by texture connectivity component filtering and color variance statistics to obtain the white space occupancy. Texture connectivity component filtering is used to exclude text, barcode, and pattern areas, while color variance statistics are used to identify blank areas with limited grayscale or color variation. The sampling window for the same can is determined by the trigger time corresponding to the photoelectric trigger timestamp data and the image acquisition timestamp data after that trigger time. Image frames within the same sampling window are all bound to the same photoelectric trigger timestamp data and form an ordered frame sequence according to the image acquisition timestamp data.

[0031] The candidate tag coverage is incremented by one to obtain a candidate coverage enhancement term. This increment prevents the subsequent product link from being interrupted when the candidate tag coverage is zero, and preserves the positive support effect of coverage changes on tag visibility carryover values. The cross-frame tag display quantity within the same tank sampling window is integrated and incremented by one to obtain a cross-frame display accumulation term. Integration accumulates continuous display evidence for the same tank within the sampling window, while incrementing maintains a stable baseline in the calculation process when cross-frame display evidence is insufficient. The sum of reflective occlusion interference and curved surface edge compression is calculated, and the sum is... Taking the inverse hyperbolic sine and adding one yields the imaging interference compression term. The inverse hyperbolic sine processing is used to compress the sudden increase in reflective occlusion interference and curved edge compression, ensuring a controlled increase in the interference term's value. Adding one helps form a stable denominator. Multiplying the candidate coverage enhancement term by the cross-frame display accumulation term, dividing by the imaging interference compression term, adding one, and taking the natural logarithm yields the visible continuity term. The product structure simultaneously reflects the synergistic support of candidate tag coverage and cross-frame tag display, while the division structure introduces the suppressive effect of imaging interference on the visible continuity term. The natural logarithm... Numerical processing is used to compress numerical expansion when the acceptance evidence is too strong; a negative exponent is taken for the tank attitude deflection change rate to obtain an attitude change suppression term. The negative exponent processing is used to quickly weaken the contribution of attitude-unstable frames to the tag visibility acceptance value when the tank attitude deflection change rate increases, and to maintain an acceptance weight close to one when the attitude change tends to be stable; the white space occupancy is increased by one and then the reciprocal is taken to obtain a white space occupancy suppression term. The reciprocal structure is used to reduce the support of blank areas for tag visibility acceptance value when the white space occupancy increases, and to avoid mistaking unlabeled texture areas as recognizable tag areas; calculation As can be seen, the tag visibility acceptance value is obtained by multiplying the basic term, the attitude change suppression term, and the whitespace occupancy suppression term. The product fusion method is used to ensure that the coverage display evidence, attitude stability, and whitespace suppression relationship jointly constrain the final judgment result. When any suppression factor is enhanced, it can be directly reflected in the tag visibility acceptance value. The tag candidate coverage, cross-frame tag display, reflective occupancy interference, curved edge compression, tank attitude deflection, and whitespace occupancy are all obtained from tank sorting observation data and are standardized or normalized to form dimensionless parameters. The specific calculation formula is as follows: ; In the formula, The tag visibility acceptance value is used to characterize whether the current image group of the same tank has the acceptance conditions for tag identity verification under the combined effects of attitude change, tag display and imaging interference. This indicates the coverage of candidate labels, which characterizes the degree to which candidate label texture areas cover the can body area; This indicates the amount of label visibility across frames, used to characterize the degree of continuous visibility of the label candidate texture region of the same tank between adjacent image frames; This indicates the amount of reflective occlusion interference, used to characterize the degree to which the brightness saturated reflective band occludes the candidate texture area of ​​the label; This indicates the amount of compression at the curved edge, used to characterize the degree of visual compression that occurs when the texture area of ​​the label candidate approaches the curved edge of the can body; This represents the tank's attitude deflection, used to characterize the degree of attitude deviation of the tank's outer contour relative to the image reference direction; This represents the rate of change of the tank's attitude deflection, used to characterize how quickly the attitude of the same tank changes within the same sampling window; Indicates the amount of white space occupied, used to characterize the degree of blank area occupied by the lack of effective label texture within the can body area; This represents the sampling window within the same tank, used to limit the cumulative range of label display across frames and ensure that all image frames involved in the calculation originate from the same tank triggering process.

[0032] In this embodiment, Table 1 is an example table for calculating the tag visibility acceptance value. It records in detail all dimensionless parameters and final results involved in the calculation of the tag visibility acceptance value under different working conditions, including the tag candidate coverage, the integral of the cross-frame tag display amount within the same tank sampling window, the amount of reflective occupancy interference, the amount of curved edge compression, the rate of change of tank attitude deflection, the amount of white space occupation, and the tag visibility acceptance value obtained by formula fusion. It is used to quantify whether multiple frames of images of the same tank have the visibility acceptance conditions to enter the tag identity confirmation stage under non-directional transportation and random tank rotation scenarios. Specifically: For condition 1, the tag candidate coverage is 0.82, the integral of cross-frame tag display within the same tank sampling window is 2.40, the glare interference is 0.08, the curved edge compression is 0.12, the tank attitude deflection change rate is 0.15, the white space occupancy is 0.10, and the tag visibility continuity value is 1.425; for condition 2, the tag candidate coverage is 0.76, the integral of cross-frame tag display within the same tank sampling window is 1.80, the glare interference is 0.65, the curved edge compression is 0.58, the tank attitude deflection change rate is 0.22, the white space occupancy is 0.18, and the tag visibility continuity value is 0.814; for condition 3, the tag candidate coverage is 0.80, the integral of cross-frame tag display within the same tank sampling window is 2.10, and the glare interference is 0.1. 0.09 for curved edge compression, 2.80 for tank attitude deflection, 0.11 for white space, and 0.096 for tag visibility. For condition 4, the tag candidate coverage is 0.55, the integral of cross-frame tag display within the same tank sampling window is 0.60, the reflection occupancy interference is 0.15, the curved edge compression is 0.14, the tank attitude deflection change rate is 0.30, the white space occupancy is 0.85, and the tag visibility. For condition 5, the tag candidate coverage is 0.12, the integral of cross-frame tag display within the same tank sampling window is 0.45, the reflection occupancy interference is 0.20, the curved edge compression is 0.18, the tank attitude deflection change rate is 0.25, the white space occupancy is 0.08, and the tag visibility.

[0033] Table 1. Example of Label Visibility Acceptance Value Calculation like Figure 3The bar chart shown is a histogram of label visibility acceptance values. Combined with Table 1, it can be seen that there are significant differences in label visibility acceptance values ​​under different working conditions, intuitively reflecting the visibility discrimination results under the combined constraints of multi-evidence collaboration and multiple suppression factors. Among them, working condition 1 has the highest label visibility acceptance value, reaching 1.425, indicating that the label candidate texture coverage in the image group corresponding to this tank is sufficient, the cross-frame label display is stable, the interference from reflection and curved surface compression is weak, and the posture change is gentle, possessing ideal recognition acceptance conditions. Working condition 3 has the lowest label visibility acceptance value, only 0.096, with a tank posture deflection change rate as high as 2.80. Although the integral of the label candidate coverage and cross-frame label display amount remains at a high level, it is still overwhelmed by the negative exponent of drastic posture changes. The suppression factor drops sharply, failing to meet the requirements for direct identification acceptance. While the acceptance value of 0.814 in Condition 2 is relatively high, it is compressed by the combined effects of reflective occlusion (0.65) and curved edge compression (0.58), falling short of Condition 1's level. Conditions 4 and 5 have acceptance values ​​of 0.376 and 0.281 respectively. The former suffers from significant suppression due to a large blank area caused by a high white space occupancy of 0.85, while the latter suffers from insufficient positive evidence due to a mere 0.12 tag candidate coverage. Both are deemed insufficiently accepted. Overall, this bar chart provides a clear quantitative baseline for comparing tag visibility acceptance thresholds, which can be used for tiered decision-making in triggering tag identity verification or review processes.

[0034] like Figure 4 The figure shows a biaxial comparison of candidate tag coverage and tag visibility acceptance value. Combined with Table 1, it can be seen that candidate tag coverage has a basic positive supporting effect on tag visibility acceptance value, but it is not a linear determining factor. The acceptance value is the result of multiplication and fusion of multiple parameters. The candidate tag coverage values ​​for conditions 1, 2, and 3 are 0.82, 0.76, and 0.80, respectively, which are close and all at a relatively high level. However, the tag visibility acceptance values ​​for the three conditions show a significant divergence of 1.425, 0.814, and 0.096. This indicates that under similar coverage conditions, the influence of suppression factors such as reflective occlusion interference, curved edge compression, and tank attitude deflection rate on the acceptance value is more prominent. In particular, in condition 3, the tank attitude deflection rate reaches 2.80, almost completely suppressing the basic acceptance factor through a negative exponential term. In scenario 4, the candidate label coverage decreased to 0.55, while the white space occupancy increased to 0.85, causing the acceptance value to further decline to 0.376. In scenario 5, the candidate label coverage was only 0.12, and the integral of the cross-frame label display within the same sampling window was as low as 0.45. The collapse of positive evidence caused the acceptance value to fall to a low of 0.281. This dual-axis comparison chart fully presents the non-monotonic constraint relationship between positive coverage evidence and final visibility acceptance, verifying the rationality and necessity of the product fusion, exponential suppression, and reciprocal penalty structure in the formula.

[0035] In this implementation scheme, effective visual evidence reflecting the label's visibility status is extracted from the same image group of the can. The can body area, label candidate texture area, cross-frame display status, imaging interference status, curved surface compression status, posture change status, and white space occupancy status are transformed into calculable label visibility carrying capacity values. This allows a unified numerical value to determine whether a label meets the conditions for continuing in label identification. This processing avoids direct interference from local occlusion, reflective bands, can body curved edge compression, or random rotation posture in a single image frame on the label recognition results. It ensures a stable expression of the label display trend and identifiability of the same can during continuous sampling, improving the ability to distinguish between truly missing labels and temporarily invisible labels. This provides a reliable preliminary judgment basis for subsequent product coding confirmation, missing risk assessment, and review channel control.

[0036] Specifically, the process of generating identifiable tags and tags with insufficient visibility based on the tag visibility continuation analysis results is as follows: (e.g.) Figure 5 The flowchart shown is for determining tag visibility acceptance, which compares the tag visibility acceptance value and the tag visibility acceptance threshold in real time. When the label visibility acceptance value is greater than or equal to the label visibility acceptance threshold, the label image of the current tank is determined to have the recognition acceptance condition. The image frame in the current image group of the same tank that meets the label candidate coverage constraint is determined as the acceptance frame, and the acceptance frame is bound to the corresponding image acquisition timestamp data to avoid recognition drift caused by the use of non-acceptance frames in the subsequent identity verification stage. Product identity recognition is allowed based on the label candidate texture region. The brightness saturated connected component corresponding to the reflective occlusion interference and the adjacent part of the tank edge corresponding to the arc edge compression are recognized and suppressed. The suppressed part is set as a low confidence participation region, and only its contour constraint effect is retained. It is not used as the main basis for product coding confirmation. Only the label candidate texture region with the superimposed label candidate coverage greater than or equal to the label candidate coverage threshold is retained to participate in label identity verification and generate an identifiable mark. The identifiable mark is bound and stored with the corresponding image group of the same tank to prevent the subsequent single frame abnormal results from overwriting the current identifiable state.

[0037] When the label visibility acceptance value is less than the label visibility acceptance threshold, the label image of the current tank is determined not to have the direct recognition acceptance condition. The current image frame is marked as an insufficient acceptance frame, and its label candidate coverage, reflection occlusion interference, curved edge compression, and white space occupation are retained as the basis for subsequent verification. Before the cross-frame supplementation confirmation is completed, the label missing judgment is not output, and the abnormal rejection logic is not triggered, so that the current tank is given priority to enter the supplementation discrimination process within the same tank image group. The cross-frame supplementation confirmation of the same tank image group is initiated, and the label candidate texture regions with matching positions in adjacent frames are surface-based. The overlapping and texture coherence verification uses the feature point registration results to constrain the overlapping position, eliminates the misaligned overlapping parts caused by the tank's posture deviation, and judges whether the overlapping label candidate texture regions have the same label source based on the consistency of the texture gradient direction. If the overlapping label candidate coverage is still less than the label candidate coverage threshold, a label visibility deficiency mark is generated, and the current tank is marked as a review object. The review object does not directly enter the regular sorting trigger logic, but retains the corresponding image group of the same tank and the visibility discrimination process record for subsequent label identity confirmation value calculation and review channel control.

[0038] In this implementation scheme, the label images of the current tank are identified and triaged based on the label visibility acceptance value. This allows image groups from the same tank that meet the recognition acceptance conditions to enter the product identification process, preventing image frames with insufficient acceptance from being directly judged as missing labels. By distinguishing between identifiable states, acceptance frames, low-confidence participation areas, and insufficient label visibility markers, interference from reflective occlusion, curved edge compression, posture deviation, and single-frame abnormal results can be avoided in subsequent label identification confirmation. Through cross-frame supplementary confirmation and verification object marking, space for continued discrimination can be reserved when the label is temporarily invisible, reducing the risk of abnormal rejection logic being prematurely triggered and improving the stability of label missing determination, product identification, and verification channel control.

[0039] Specifically, the process of label identity verification analysis based on identifiable markers and insufficient label visibility markers is as follows: The label template matching quantity is obtained based on the scale-invariant feature transformation matching results between multi-frame image data of the tank and the current batch of label template image set data. Scale-invariant feature matching includes scale-invariant feature transformation algorithms or fast feature point and rotation descriptor algorithms. The matching process uses stable corner points, edge intersection points, and local texture descriptors within the candidate label texture area as matching objects, and filters out mismatched point pairs using the nearest neighbor distance ratio. The label candidate texture area is then processed for trademark contour extraction, barcode stripe direction extraction, text connected component extraction, and main pattern edge structure extraction. The trademark contour is obtained through closed edge chains and shape similarity constraints; the barcode stripe direction is obtained through the orientation gradient histogram; the text connected components are obtained through stroke width consistency and connected component scale constraints; and the main pattern edge structure is obtained through multi-scale edge response. The key identifier completeness is obtained based on the coverage ratio of the extracted areas in the corresponding areas of the current batch of label template image set data. The coverage ratio is calculated according to the ratio between the effective overlapping area of ​​each key identifier area and the area of ​​the corresponding template area, and for positions with offsets exceeding the mapping... The tolerance area is not included in the effective coverage; the candidate category similarity ranking is formed by the current batch label template image set data, the current batch product code data, and the tank multi-frame image data. The ranking process takes the feature descriptor set of each image frame in the same tank image group as input, calculates the matching similarity of the label template corresponding to each product code, and then sorts and merges them. The candidate category similarity is obtained by the cosine similarity between the image feature descriptor corresponding to the tank multi-frame image data and the template feature descriptor corresponding to the current batch label template image set data. Then, the product category separation quantity is obtained according to the similarity interval between the first and second candidate categories. The similarity interval is the difference between the similarity of the first and second candidate categories. The product category separation quantity uses the similarity interval to characterize the distinguishability between candidate product codes, and forms a separation insufficiency expression for cases where the similarity is tied or nearly tied. The cross-frame label conflict quantity is obtained by the number of inconsistencies of candidate product codes obtained by ranking different image frames in the same tank image group of the same tank through candidate category similarity. The number of inconsistencies is accumulated according to the number of changes in candidate product codes within the same tank sampling window, and the cases where consecutive adjacent frames point to different product codes are recorded as conflict evidence.The extracted results of trademark outline, barcode stripe direction, text connectivity, and main pattern edge structure are compared with corresponding regions in the current batch of label template image set data. The location mapping comparison establishes a corresponding region mapping relationship based on template region coordinates, current image region coordinates, and feature point registration relationships. For regions still lacking effective structural responses after mapping, gap statistics are performed. The label evidence gap is obtained based on the proportion of the corresponding region area that does not form an effective match. The area proportion is calculated as the ratio between the area of ​​the missing region and the total area of ​​the corresponding key identifier region in the template, forming a quantitative expression of the degree of insufficient label evidence.

[0040] The tag template matching quantity is calculated and incremented by one to obtain the template matching enhancement term. This increment prevents subsequent product calculations from losing their numerical basis when the tag template matching quantity is zero, while preserving the positive contribution of template matching degree to the basic identity verification item. The product of the complete key identifier quantity and the product category separation quantity is calculated, and the square root of the product is incremented by one to obtain the key separation coordination term. The product structure reflects the coordination relationship where the complete key identifier quantity and the product category separation quantity must both be true simultaneously. The square root operation mitigates the numerical amplification caused by an overly strong single term, and the increment maintains a stable starting point for the coordination term. The tag evidence gap quantity within the same sampling window is sampled according to the adjacent frames corresponding to the image acquisition timestamp data. The time-weighted integral of the sampling time interval is used to obtain the cumulative amount of tag evidence gap. The sampling window in the same tank is determined by the trigger time corresponding to the photoelectric trigger timestamp data and the sampling window of each image acquisition timestamp data after that trigger time. The time-weighted integral can reflect the persistence of the tag evidence gap in continuous image frames, making a numerical distinction between transient occlusion and persistent gap. The absolute value of the cross-frame tag conflict amount, the sum of the cumulative amount of tag evidence gap and the value of one, are calculated to obtain the evidence conflict damping term. The absolute value processing can eliminate the difference in conflict direction and retain only the conflict intensity. After forming the denominator damping together with the cumulative amount of tag evidence gap, it can suppress the impact of conflict and gap on the basic terms of identity confirmation. Interference; Multiply the template matching enhancement term and the key separation coordination term, divide by the evidence conflict damping term, add one, and take the natural logarithm to obtain the identity verification base term. The product structure can strengthen the consistent support of template matching and key identifier separation, the division structure can introduce the reverse constraint of cross-frame conflict and evidence gap, and the natural logarithm processing can compress the numerical expansion range of the identity verification base term; Take the hyperbolic tangent of the tag visibility acceptance value and add one to obtain the visibility enhancement term. The hyperbolic tangent processing can map the tag visibility acceptance value to a finite range, so that the visibility evidence produces a stable gain on the identity verification result, and the addition of one can avoid the enhancement term from having a non-positive value; Calculate the identity verification base term and the... The product of visibility enhancement terms yields the tag identity confirmation value. Product fusion enables a progressive constraint between identity matching evidence and visibility acceptance evidence. Only when template matching, key identifiers, category separation, and visibility acceptance are all supported will the tag identity confirmation value show a stable increase. Since the tag template matching quantity, key identifier completeness quantity, product category separation quantity, cross-frame tag conflict quantity, tag evidence gap quantity, and tag visibility acceptance value are all standardized or normalized before participating in the calculation, the calculations in the formula are all based on dimensionless numerical values. The tag identity confirmation value is not affected by image area units, time units, similarity scales, or differences in device acquisition units. The specific calculation formula is as follows: ; In the formula, The label identity confirmation value is used to characterize the degree of confirmation of the product code corresponding to the current tank under the joint constraints of multi-frame label evidence, template matching evidence, and visibility acceptance evidence. This represents the label template matching quantity, used to characterize the degree of feature matching between multi-frame image data of the tank and the current batch of label template image set data; This indicates the completeness of key identifiers, used to characterize the degree of completeness of the trademark outline, barcode stripe direction, text connectivity, and main pattern edge structure within the corresponding area of ​​the template. This represents the product category separation ratio, used to characterize the degree of similarity difference between the first and second candidate categories; This indicates the amount of cross-frame label conflict, used to characterize the degree of inconsistency in the encoding of candidate products output from different image frames within the same image group; This indicates the amount of missing label evidence, used to characterize the percentage of key labeled areas where no valid match was formed; This indicates the cumulative amount of label evidence gaps, used to characterize the degree of continuous accumulation of label evidence gaps within the same sampling window; This represents the tag visibility support value, used to characterize the degree of visibility support provided by the current image group within the same tank for the tag identification process.

[0041] In this implementation plan, label appearance evidence within the same can image group is transformed into product identification criteria. This transforms label identification from relying solely on local recognition results of a single image frame to a comprehensive assessment based on template matching, key identifier integrity, candidate category differentiation, cross-frame recognition consistency, and label evidence gaps. This process allows for further confirmation of the product code corresponding to the current can based on label visibility, reducing identification biases caused by partial occlusion, glare interference, pose changes, or missing features in a single frame. Simultaneously, it identifies candidate product code conflicts and insufficient key identifier evidence within the same can image group, providing stable judgment criteria for subsequent product code locking, label missing risk marking, identity conflict marking, and verification channel import, thereby improving the consistency and reliability of canned product sorting and identification results.

[0042] Specifically, the process of locking product codes and importing review and sorting rules based on the label identity verification analysis results is as follows: Real-time comparison of label identity verification values ​​and label identity verification thresholds: When the label identity confirmation value is greater than or equal to the label identity confirmation threshold, the product code of the current tank is locked. The optimal matching result corresponding to the label template matching quantity is used as the basis for locking the product code, and the complete quantity of key identifiers and the separation quantity of product category are used as the basis for locking verification. The regular sorting category is determined based on the correspondence between the current batch product code data and the current batch sorting channel number data. The locked product code is written into the identification status record of the current tank. Before the current tank completes sorting, the product code will not be updated based on the single frame identification result. The identification results generated by the low confidence image frames in the subsequent image group of the same tank are only used as traceability references and will not participate in the product code rewriting. A product identity confirmation mark is generated.

[0043] When the tag identity verification value is less than the tag identity verification threshold, the judgment action of entering the regular sorting channel based on the current tank is suspended to avoid triggering the sorting path in the state of unverified identity. The review channel category is determined according to the review sorting rules. Among them, when the tag evidence gap is greater than or equal to the tag evidence gap threshold in the same tank sampling window, a tag missing risk mark is generated; when the cross-frame tag conflict reaches the cross-frame conflict threshold, an identity conflict mark is generated. The review sorting rule table includes the correspondence between the tag missing risk mark and the review channel category, as well as the correspondence between the identity conflict mark and the review channel category. The tag missing risk mark or identity conflict mark is applied to the current tank, and the current tank is imported into the review sorting rules. The review sorting rules prioritize the allocation of the current tank to the manual review channel or the buffer review channel, and retain the same tank image group, tag visibility acceptance value, and tag identity verification value as the review basis.

[0044] The regular sorting category and the verification channel category are used together as the sorting target category. The sorting target category is used to determine the target channel of the target tank in the synchronous analysis of sorting execution, and serves as the category basis for the sorting mechanism action trigger or the verification channel import.

[0045] In this implementation scheme, the product code of the current can is confirmed and routed based on the tag identity verification value. This ensures that cans with sufficient identity verification can form a stable product code locking result, while cans with insufficient identity verification will not directly enter the regular sorting path. By writing the locked product code into the identification status record and restricting the rewriting of the product code by subsequent low-confidence results in a single frame, the consistency of the identification results for the same can before sorting is completed can be maintained. Through the routed processing of tag missing risk markers, identity conflict markers, and verification channel categories, cans with insufficient tag evidence or cross-frame identification conflicts can be verified and accepted, reducing the risk of erroneous triggering of the sorting path in the state of unconfirmed identity. This processing further unifies the regular sorting category and the verification channel category into the sorting target category, providing a clear target channel basis for subsequent synchronous analysis of sorting execution, and improving the connection stability between product code confirmation, verification import, and sorting action triggering.

[0046] Specifically, the process of performing synchronous analysis on the tank sorting observation data is as follows: The same tank center point in multiple frames of tank image data is located frame by frame using a target tracking algorithm to obtain the tank center trajectory. The target tracking algorithm uses the outer contour center of the tank, the centroid of the tank area, and the continuity of motion between adjacent frames as tracking constraints, and eliminates abnormal positioning results where the center point jumps beyond the trajectory tolerance. The corresponding time of each frame is determined based on the image acquisition timestamp data, and the conveying displacement between frames is calculated based on the encoder pulse data and encoder pulse timestamp data. The conveying displacement is obtained according to the conversion relationship between the encoder pulse increment and the conveying distance corresponding to a unit pulse, and is bound to the inter-frame time period corresponding to the image acquisition timestamp data, thus establishing the tank center trajectory. A time-series mapping is performed between the tank trajectory and the conveyor displacement. This mapping establishes a correspondence between image frame times and encoder pulse times under the same time base, ensuring that the tank image displacement and the actual conveyor line displacement are on the same comparison link. The tank position binding amount is obtained based on the reciprocal relationship between the image displacement corresponding to the tank center trajectory and the conveyor displacement converted from the encoder pulse data of the conveyor line. The reciprocal relationship uses the deviation between the image displacement and the conveyor displacement as the denominator constraint; the smaller the deviation, the better the tank position binding amount characterizes the binding consistency between the target tank and the conveyor timing. The sorting channel matching amount is obtained based on the mapping consistency between the current batch sorting channel number data and the sorting target category. The mapping consistency is based on the current batch product code data and the sorting target... The correspondence between the category and the current batch sorting channel number data is verified. The sorting channel matching value is taken as a valid match when the target channel corresponding to the sorting target category matches the current batch sorting channel number data; otherwise, an invalid match value is taken. Both valid and invalid match values ​​are expressed using preset status codes to facilitate subsequent standardized calculations with continuous position binding quantities. The conveyor speed disturbance is obtained based on the pulse interval changes in the conveyor line encoder pulse data and encoder pulse timestamp data. The conveyor speed disturbance is calculated according to the instantaneous speed change corresponding to adjacent pulse intervals, and missing and duplicate pulses are removed before being included in the statistics. The difference in conveyor speed disturbance corresponding to adjacent encoder pulse timestamp data is used to obtain... The rate of change of conveying speed disturbance is calculated as the ratio of the difference in conveying speed disturbance between two consecutive pulse time periods to the corresponding time interval. The target tank and adjacent tanks are analyzed by the target detection algorithm. The target detection algorithm outputs the bounding box of the target tank, the bounding box of the adjacent tank and the corresponding center point. The front adjacent tank and the rear adjacent tank are determined according to the conveying direction. The interference amount of the adjacent tank is obtained by normalization and fusion based on the reciprocal of the boundary distance, the boundary overlap ratio and the reciprocal of the center trajectory interval. Before normalization and fusion, the boundary distance, the boundary overlap ratio and the center trajectory interval are mapped to a unified numerical range so that the three types of interference evidence, namely, contact, occlusion and trajectory proximity, can jointly express the interference amount of the adjacent tank.Based on image acquisition timestamp data, conveyor line encoder pulse data, and encoder pulse timestamp data, the estimated arrival time of the target can at the sorting execution position is calculated. The estimated arrival time is based on the correspondence between the current position of the target can, the sorting execution position, and the conveyor speed, and is updated using the conveyor speed within the most recent encoder pulse time window. The estimated arrival time is obtained by dividing the remaining displacement from the current position of the target can to the sorting execution position by the conveyor speed obtained based on the encoder pulse interval. The remaining displacement is converted according to the distance between the current center position of the target can and the sorting execution position. The difference between the estimated arrival time and the sorting mechanism action feedback timestamp data is then calculated. Combined with the sorting mechanism action status feedback data, corrections are made for inaction, delayed action, and action failure states to obtain the execution feedback deviation. The correction process converts inaction, delayed action, and action failure states into corresponding deviation penalty items, which, together with the time difference, form the execution feedback deviation.

[0047] The tank position binding amount is calculated and incremented by one to obtain the position binding enhancement term. This increment prevents subsequent product calculations from being interrupted when the tank position binding amount is zero. The sorting channel matching amount is calculated and incremented by one to obtain the channel matching enhancement term, allowing the channel matching result to participate in the calculation as a positive enhancement term. The label identification value is taken as a hyperbolic tangent and incremented by one to obtain the identification enhancement term. This hyperbolic tangent processing is used to compress the numerical range of the label identification value, preventing it from excessively amplifying the sorting execution synchronization value. The absolute value of the conveyor speed disturbance change rate, the adjacent tank interference amount, the inverse hyperbolic sine value of the execution feedback deviation, and the sum of one are calculated to obtain the execution synchronization damping term. The absolute value is used to retain the disturbance change intensity, the inverse hyperbolic sine is used to compress the sudden increase in the execution feedback deviation, and one is used to keep the denominator stable. The position binding enhancement term, the channel matching enhancement term, and the identification enhancement term are then combined. The sorting execution synchronization value is obtained by multiplying the confirmation enhancement terms, dividing by the execution synchronization damping term, and then taking the arctangent. The product structure is used to reflect the joint support of position binding, channel matching, and identity confirmation. The division structure is used to reflect the reverse constraints of conveying disturbance, adjacent interference, and execution deviation. The arctangent is used to limit the numerical expansion of the sorting execution synchronization value. Since the tank position binding amount, sorting channel matching amount, tag identity confirmation value, conveying speed disturbance amount, adjacent tank interference amount, and execution feedback deviation amount are all standardized or normalized before participating in the calculation, the product, hyperbolic tangent, absolute value, inverse hyperbolic sine, division, and arctangent operations in the formula are all based on dimensionless numerical values. The sorting execution synchronization value is not affected by differences in image displacement units, conveying distance units, time units, and execution receipt status codes. The specific calculation formula is as follows: ; In the formula, This represents the sorting execution synchronization value, used to characterize whether the current target tank meets the synchronization conditions to trigger the sorting action; This indicates the tank position binding amount, used to characterize the degree of consistency between the tank center trajectory and the conveying displacement; This represents the matching quantity of the sorting channel, which characterizes the degree of matching between the sorting target category and the sorting channel number data of the current batch; This represents the label identification value, used to characterize the degree of identification of the product code corresponding to the current tank; This represents the amount of disturbance in the conveying speed, used to characterize the degree of fluctuation in the conveying linear speed; It represents the rate of change of conveyor speed disturbance, used to characterize the magnitude of change in the amount of conveyor speed disturbance; This indicates the amount of interference between adjacent tanks, used to characterize the degree of contact, overlap, and trajectory proximity between the target tank and its adjacent tanks; This indicates the execution feedback deviation, which characterizes the degree of deviation between the expected arrival time and the timestamp data of the sorting mechanism's action feedback.

[0048] In this implementation plan, the image trajectory of the current target can, its conveying position, sorting channel, identity verification result, adjacent can status, and execution feedback status are all uniformly incorporated into the sorting action triggering judgment. This ensures that sorting execution no longer relies solely on identification results or fixed delay control, but rather on a comprehensive judgment based on the target can's arrival status, channel matching status, and execution synchronization status. This process improves the accuracy of the correspondence between the current target can and the sorting execution position, reduces interference from conveying speed fluctuations, adjacent can contact, trajectory proximity, and action feedback deviations on the sorting action, and avoids mismatches between confirmed product codes and actual execution objects. The sorting execution synchronization value provides a unified judgment basis for sorting mechanism action triggering, channel import verification, and anomaly rejection control, improving the consistency of action timing, target object accuracy, and traceability of execution results during online sorting of canned products.

[0049] Specifically, the process of triggering sorting mechanism actions and importing verification channels based on the sorting execution synchronization analysis results is as follows: Real-time comparison of sorting execution synchronization values ​​and sorting execution synchronization thresholds: When the sorting execution synchronization value is greater than or equal to the sorting execution synchronization threshold, the position relationship between the target tank and the execution window is determined based on the tank position binding amount. The target sorting channel is checked for consistency with the locked product code based on the sorting channel matching amount. The corresponding sorting mechanism action is triggered according to the locked product code and sorting category, so that the sorting action is triggered according to the target tank arrival time, rather than being triggered directly according to a fixed delay. The sorting mechanism is controlled to act when the target tank arrives at the execution window. The sorting mechanism includes at least one of the diversion baffle, pneumatic rejection mechanism and mechanical lever. The specific configuration can be configured according to the conveyor line structure and sorting channel layout. The action completion status is recorded. The sorting mechanism action acknowledgment timestamp data and sorting mechanism action status feedback data are bound to the target tank to form a traceable execution closed loop.

[0050] When the sorting execution synchronization value is less than the sorting execution synchronization threshold, the direct rejection action is stopped, the current target tank is determined to be an object with insufficient execution synchronization, the immediate actions related to abnormal rejection are suspended, the target tank is imported into the review channel, and adjacent tanks are allowed to pass. Adjacent tanks refer to tanks located in the same image frame in front of or behind the target tank in the conveying direction, and whose distance from the boundary of the target tank is less than the adjacent tank distance threshold. This avoids the target tank and adjacent tanks from being mismatched in rejection objects within the execution window, and reduces the risk of erroneous rejection caused by conveying speed fluctuations, tank contact, or execution feedback deviation.

[0051] Finally, the sorting control instructions, verification channel instructions, and sorting execution results are output. The sorting execution synchronization value, action completion status, and product code of the corresponding target tank are bound and archived for subsequent sorting cycle correction and traceability of erroneous rejections.

[0052] In this implementation scheme, sorting actions are triggered and access controlled based on the sorting execution synchronization value. This ensures that the corresponding sorting mechanism action is triggered only when the target can is in position, the channel is matched, the product code is locked, and the execution timing meets the requirements, avoiding reliance on fixed delays or single identification results for sorting. By stopping direct rejection and importing the data into the verification channel when synchronization is insufficient, the risk of erroneous rejection caused by conveyor speed fluctuations, adjacent cans being too close together, execution window offsets, and abnormal action feedback can be reduced, ensuring consistency in object and timing between the target can and the sorting action. By outputting sorting control instructions, verification channel instructions, and sorting execution results, and binding and archiving the execution process, a closed-loop record from product code confirmation to sorting action completion can be formed, improving the accuracy, stability, and traceability of sorting execution.

[0053] like Figure 2As shown, the second aspect of this invention provides a canned product sorting and identification system based on image processing, comprising: a data acquisition and preprocessing module, used to acquire can sorting observation data and perform same-can aggregation, time-series calibration, and standardized expression to form a unified data foundation under the same can, the same batch, and the same time-series chain; and a can rotation attitude and label visibility discrimination module, used to perform label visibility acceptance analysis on the can sorting observation data, determine whether the current same-can image group has label recognition acceptance conditions, and generate identifiable marks and label visibility deficiency marks based on the label visibility acceptance analysis results. The system includes: a tag generation module; a multi-frame tag identity verification and missing risk judgment module, used to perform tag identity verification analysis based on identifiable markers and insufficient tag visibility markers, to perform multi-frame verification of the product code corresponding to the current tank, and to lock the product code and import the verification sorting rules based on the tag identity verification analysis results; and a sorting decision and execution verification module, used to perform sorting execution synchronization analysis on the tank sorting observation data, to determine whether the target tank, sorting channel and execution window meet the synchronization triggering conditions, and to trigger the sorting mechanism action and import the verification channel based on the sorting execution synchronization analysis results.

[0054] In this implementation scheme, a progressive processing chain—including data acquisition and preprocessing, label visibility acceptance, label identification, and synchronized sorting execution—enables continuous identification and sorting control of canned products based on images from the same can, even during non-directional transport and random rotation. By first determining whether the label meets the recognition acceptance criteria and then performing multi-frame verification of the product code, recognition deviations caused by pose occlusion, glare interference, or partial invisibility in a single frame image can be reduced. By verifying the synchronization relationship between the target can, the sorting channel, and the execution window before sorting execution, the risk of accidental triggering of the sorting path and mismatched rejection objects can be reduced, thereby improving the consistency, accuracy, and traceability of canned product sorting and identification.

[0055] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0056] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for sorting and identifying canned products based on image processing, characterized in that, Includes the following steps: S1, collect tank sorting observation data and perform same-tank aggregation, time-series calibration and standardized expression; S2, perform label visibility acceptance analysis on the can sorting observation data, and generate identifiable labels and labels with insufficient label visibility based on the label visibility acceptance analysis results; S3, based on identifiable markers and insufficient label visibility markers, performs label identity verification analysis, and locks product codes and imports review and sorting rules based on the label identity verification analysis results; S4, perform synchronous analysis of sorting execution on the tank sorting observation data, and trigger sorting mechanism actions and import verification channels based on the results of the synchronous analysis of sorting execution; The specific process of performing label visibility analysis on the tank sorting observation data is as follows: The outer contour of the tank in multi-frame image data is fitted to obtain the tank body region through contour closure screening; the texture coherence of the tank body region is screened through color change density statistics and texture gradient extraction to obtain candidate label texture regions; the label candidate coverage is obtained based on the ratio of the area of ​​the candidate label texture region to the area of ​​the tank body region; the position of the candidate label texture regions in adjacent frames is matched through feature point registration, and the cross-frame label display amount is obtained by combining the image acquisition timestamp data; the reflective band is extracted from the brightness saturation region in multi-frame image data of the tank through brightness saturation connected component screening, and abnormal frames are corrected by combining camera exposure time feedback data and light source current feedback data to obtain the reflective occlusion interference amount; through… The tank boundary localization process involves statistically analyzing the distance between the candidate label texture area and the tank edge to obtain the arc edge compression amount; performing attitude analysis on the outer contour of the tank body by fitting the tank centerline and calculating the boundary tilt angle to obtain the tank attitude deflection amount; converting the tank attitude deflection amount and image acquisition timestamp data into a rate of change by using adjacent frame difference to obtain the tank attitude deflection change rate; extracting the white space area of ​​the tank body by texture connected component filtering and color variance statistics to obtain the white space occupancy amount; determining the sampling window for the same tank based on the trigger time corresponding to the photoelectric trigger timestamp data and the image acquisition timestamp data after that trigger time; and calculating the label visibility acceptance value based on the label candidate coverage amount, cross-frame label display amount, reflective occlusion interference amount, arc edge compression amount, tank attitude deflection change rate, and white space occupancy amount. The specific process of label identification analysis based on identifiable markers and insufficient label visibility markers is as follows: The label template matching quantity is obtained based on the scale-invariant feature transform matching results between multi-frame image data of the tank and the current batch of label template image set data; the trademark outline, barcode stripe direction, text connected components, and main pattern edge structure are extracted from the candidate texture regions of the label, and the key identifier completeness quantity is obtained based on the coverage ratio of the corresponding regions in the current batch of label template image set data; the candidate category similarity is ranked using the current batch of label template image set data, the current batch of product code data, and multi-frame image data of the tank, and the similarity is determined based on the first and second candidate categories. The product category separation quantity is obtained from the similarity interval of candidate categories; the cross-frame label conflict quantity is obtained from the number of inconsistent candidate product codes obtained by sorting different image frames in the same image group of the same can based on the similarity of candidate categories; the extraction results of trademark outline, barcode stripe direction, text connected components and main pattern edge structure are compared with the corresponding areas in the current batch of label template image set data by position mapping, and the label evidence gap quantity is obtained based on the area ratio of the corresponding areas that have not formed an effective match; the label identity confirmation value is calculated based on the label template matching quantity, key identifier integrity quantity, product category separation quantity, label evidence gap quantity, cross-frame label conflict quantity and label visibility acceptance value. The specific process of performing synchronous analysis on the tank sorting observation data is as follows: The tank's center point is located using a target tracking algorithm across multiple frames of image data. A temporal mapping between the tank's center trajectory and the conveyor displacement is completed by combining image acquisition timestamp data, conveyor line encoder pulse data, and encoder pulse timestamp data. The tank's position binding amount is obtained based on the mapping consistency. The sorting channel matching amount is obtained based on the mapping consistency between the current batch sorting channel number data and the sorting target category. The conveying speed disturbance is obtained based on the pulse interval changes in the conveyor line encoder pulse data and encoder pulse timestamp data. The conveying speed disturbance change rate is obtained by differentiating the conveying speed disturbance amounts corresponding to adjacent encoder pulse timestamp data. The target... The detection algorithm analyzes the boundary relationship between the target tank and adjacent tanks, and obtains the interference amount of adjacent tanks based on the boundary distance, boundary overlap ratio, and center trajectory interval. It calculates the estimated arrival time of the target tank at the sorting execution position based on image acquisition timestamp data, conveyor line encoder pulse data, and encoder pulse timestamp data. The difference between the estimated arrival time and the sorting mechanism action feedback timestamp data is then calculated, and the algorithm combines this with the sorting mechanism action status feedback data to correct for inactive, delayed, and failed actions, resulting in the execution feedback deviation. Finally, based on the tank position binding amount, sorting channel matching amount, tag identification value, conveyor speed disturbance change rate, adjacent tank interference amount, and execution feedback deviation, the sorting execution synchronization value is calculated.

2. The image processing-based sorting and identification method for canned products according to claim 1, characterized in that: The specific process for collecting tank sorting observation data is as follows: The system collects tank sorting observation data, which includes: the current batch label template image set data, the current batch product code data, and the current batch sorting channel number data read by the production management system; the sorting mechanism action receipt timestamp data and sorting mechanism action status feedback data read by the sorting execution mechanism controller; multi-frame image data of the tank; image acquisition timestamp data; camera exposure time feedback data; light source current feedback data; conveyor encoder pulse data; encoder pulse timestamp data; and photoelectric trigger timestamp data. The photoelectric trigger timestamp data is used to bind the current tank to the current batch product code data.

3. The image processing-based sorting and identification method for canned products according to claim 1, characterized in that: The specific process of performing same-tank aggregation, time-series calibration, and standardized representation is as follows: The multi-frame image data of the tank is sorted according to the image acquisition timestamp data, and frames with duplicate timestamps, empty frames and failed acquisition frames are removed. The multi-frame image data of the tank is grouped according to the photoelectric trigger timestamp data to form the same tank image group for a single tank. The image size is unified, the boundary is cropped, the background grayscale is compressed, and the grayscale range is corrected for the multi-frame image data of the tank; the time base is aligned for the encoder pulse data, encoder pulse timestamp data, photoelectric trigger timestamp data, and image acquisition timestamp data of the conveyor line. The field format of the sorting mechanism action receipt timestamp data and sorting mechanism action status feedback data is standardized; the batch field of the current batch label template image set data, the current batch product code data and the current batch sorting channel number data is made consistent; and the tank sorting observation data is standardized using the zero mean unit variance standardization algorithm. The maximum and minimum value normalization algorithm is used to normalize the multi-frame image data of the tank, the camera exposure time feedback data, the light source current feedback data, and the encoder pulse data of the conveyor line.

4. The image processing-based sorting and identification method for canned products according to claim 1, characterized in that: The specific process for generating identifiable tags and tags with insufficient visibility based on the tag visibility acceptance analysis results is as follows: Real-time comparison of tag visibility acceptance value and tag visibility acceptance threshold: When the label visibility acceptance value is greater than or equal to the label visibility acceptance threshold, the brightness saturated connected domain corresponding to the reflective occlusion interference and the adjacent part of the can edge corresponding to the arc edge compression are identified and suppressed. Only the label candidate texture area with the superimposed label candidate coverage greater than or equal to the label candidate coverage threshold is retained to participate in the label identity confirmation and generate an identifiable mark. When the label visibility acceptance value is less than the label visibility acceptance threshold, no label missing judgment is output before the cross-frame supplementary confirmation is completed; cross-frame supplementary confirmation of the same tank and the same tank image group is initiated, and the area of ​​the label candidate texture regions with matching positions in adjacent frames is superimposed and the texture coherence is checked; if the superimposed label candidate coverage is still less than the label candidate coverage threshold, a label insufficient visibility mark is generated, and the current tank is marked as the review object.

5. The image processing-based sorting and identification method for canned products according to claim 1, characterized in that: The specific process for locking product codes and importing review and sorting rules based on the label identity verification analysis results is as follows: Real-time comparison of tag identity verification value and tag identity verification threshold: When the tag identity verification value is greater than or equal to the tag identity verification threshold, the product code of the current tank is locked. The regular sorting category is determined based on the correspondence between the current batch product code data and the current batch sorting channel number data. The locked product code is written into the identification status record of the current tank. The product code will not be updated based on the single frame identification result before the current tank completes sorting. When the label identity verification value is less than the label identity verification threshold, the verification channel category is determined according to the verification and sorting rules, the current tank is marked with a label missing risk mark or an identity conflict mark, and the current tank is imported into the verification and sorting rules. Both the regular sorting category and the verification channel category are used as the sorting target category.

6. The image processing-based sorting and identification method for canned products according to claim 1, characterized in that: The specific process for triggering sorting mechanism actions and importing verification channels based on the sorting execution synchronization analysis results is as follows: Real-time comparison of sorting execution synchronization value and sorting execution synchronization threshold: When the sorting execution synchronization value is greater than or equal to the sorting execution synchronization threshold, the corresponding sorting mechanism action is triggered according to the locked product code and sorting category, the sorting mechanism is controlled to act when the target tank reaches the execution window, and the action completion status is recorded. When the sorting execution synchronization value is less than the sorting execution synchronization threshold, the direct rejection action is stopped, the target tank is introduced into the review channel, and the adjacent tanks are allowed to pass. The final output includes sorting control instructions, verification channel instructions, and sorting execution results.

7. A canned product sorting and identification system based on image processing, employing the canned product sorting and identification method based on image processing as described in any one of claims 1-6, comprising: The data acquisition and preprocessing module is used to collect tank sorting observation data and perform same-tank aggregation, time-series calibration, and standardized representation. The can rotation attitude and label visibility discrimination module is used to perform label visibility acceptance analysis on the can sorting observation data, and generate identifiable labels and labels with insufficient visibility based on the label visibility acceptance analysis results; The multi-frame tag identity verification and missing risk judgment module is used to perform tag identity verification analysis based on identifiable markers and insufficient tag visibility markers, and to lock product codes and import review and sorting rules based on the tag identity verification analysis results. The sorting decision and execution verification module is used to perform synchronous analysis of sorting execution on tank sorting observation data, and to trigger sorting mechanism actions and import verification channels based on the results of the synchronous analysis.

Images