A smart monitoring and control method and system for mass production of anti-counterfeiting labels

By using blurred image enhancement and multi-level recognition mechanisms, combined with the improved YOLOv8 model and the standard YOLOv8 model, the problem of insufficient recognition capability and high false detection rate caused by image blurring in the mass production of anti-counterfeiting labels is solved, and efficient and accurate quality monitoring is achieved.

CN121724960BActive Publication Date: 2026-11-17ZHEJIANG HESHU SOFTWARE CO LTD
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
CN202511936545.8
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2025-12-22
Publication Date
2026-11-17
Estimated Expiration
2045-12-22

AI Technical Summary

Technical Problem

Traditional visual inspection methods suffer from insufficient recognition capabilities and high false detection rates in the mass production of anti-counterfeiting labels due to motion blur of the original images. A single detection model cannot balance efficiency and accuracy.

Method used

A collaborative detection process is adopted, which includes fuzzy image enhancement preprocessing, rapid coarse screening, and precise fine screening. An improved YOLOv8 model is used for initial screening, and a standard YOLOv8 model is used for precise screening. The SIFT global search algorithm is combined to establish image correspondence, and fuzzy kernel processing is used to improve recognition capabilities.

Benefits of technology

It improves the robustness and detection efficiency of the system, reduces the false detection rate, ensures the accuracy of detection results and optimizes computing resources, achieves a balance between speed and accuracy, and provides real-time and reliable quality monitoring for the production line.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN121724960B_ABST
    Figure CN121724960B_ABST
Patent Text Reader

Abstract

This invention discloses an intelligent monitoring and control method and system for mass production of anti-counterfeiting labels, relating to the field of image processing technology. The method involves acquiring images of various anti-counterfeiting labels to obtain a fuzzy image set, then enhancing the fuzzy images in the fuzzy image set to obtain an enhanced fuzzy image set. A target coarse screening model is used to identify the images in the fuzzy image set to obtain identification results. Images with abnormal identification results are combined to obtain an abnormal image set, which is then identified using a target fine screening model to obtain detection results. If the detection results are unqualified, the corresponding anti-counterfeiting label is removed. The collaborative process of fuzzy enhancement, rapid coarse screening using the target coarse screening model, and precise fine screening using the standard model improves the ability to identify fuzzy images and the system's robustness. It increases efficiency while ensuring high detection accuracy, optimizes resource allocation, achieves a balance between speed and accuracy, and provides real-time and reliable quality monitoring support for the production line.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and more specifically to an intelligent monitoring and control method and system for mass production of anti-counterfeiting labels. Background Technology

[0002] With the continuous expansion of the consumer goods market and the increasing awareness of brand protection, the demand for anti-counterfeiting labels, as a key carrier for product identity authentication and authenticity verification, is experiencing explosive growth. Traditional manual visual inspection or semi-automatic inspection methods face bottlenecks such as low efficiency, inconsistent standards, and high labor costs in mass production scenarios, making it difficult to meet the high-speed, zero-defect quality control requirements of modern production lines.

[0003] Publication No. CN120068902A discloses a method for controlling the production quality of electronic tags. In the second image, a multi-scale denoising algorithm is used to reduce noise in the image to address vibration and dust interference, resulting in a third image. The online detection system judges the printing quality of the electronic tags based on the OCR recognition results. If an abnormality is detected, an abnormal signal is generated and transmitted to the process control system. The process control system receives the abnormal signal and adjusts the operating parameters of the printing, bonding, and cutting processes to ensure smooth process connections and avoid production rhythm imbalance. The adjusted process parameters are compared with the preset product consistency standard. If the parameter deviation exceeds the preset range, the automatic calibration module is activated to ensure product consistency. The system monitoring module monitors the operating status of the entire production process in real time. If a decrease in system stability is detected, an early warning mechanism is triggered, generating a maintenance signal and transmitting it to the maintenance system.

[0004] In existing technologies, in the scenario of batch automated detection of anti-counterfeiting labels, the motion blur of the original image leads to insufficient recognition ability and high false detection rate of traditional visual detection methods, as well as the difficulty of a single detection model to balance efficiency and accuracy. Summary of the Invention

[0005] The purpose of this invention is to solve the problem mentioned in the background art, which is that traditional visual detection methods have insufficient recognition ability and high false detection rate due to motion blur of the original image. Since a single detection model cannot balance efficiency and accuracy, an intelligent monitoring and control method and system for mass production of anti-counterfeiting labels is proposed.

[0006] A first aspect of this invention provides an intelligent monitoring and control method for mass production of anti-counterfeiting labels, the method comprising: A blurred image set is obtained by combining images of various anti-counterfeiting labels, and the blurred images in the blurred image set are enhanced to obtain an enhanced blurred image set. The images in the blurred image set are identified using a target coarse screening model to obtain identification results; the identification results include qualified and abnormal. Images identified as abnormal are combined to form an abnormal image set. This abnormal image set is then analyzed using a target fine-screening model to obtain detection results, which include qualified and unqualified results. If the test result is unqualified, the corresponding anti-counterfeiting label will be removed.

[0007] Optionally, enhancing the blurred images in the blurred image set to obtain an enhanced blurred image set includes: Step 1: Obtain a reference image and key parameters; segment the reference image to obtain a reference subset; segment the blurred image to obtain a blurred subset; the reference subset contains multiple reference sub-images; the blurred subset contains multiple blurred sub-images; the key parameters are the exposure time and sampling frequency of the blurred image. Step 2: Determine the position labels of each reference sub-image in the reference subset using the SIFT global search algorithm, add the position labels to the corresponding reference sub-images to obtain labeled reference sub-images, and combine the labeled reference sub-images to obtain the target reference subset; Step 3: Calculate the similarity between the label reference sub-image in the target reference subset and the fuzzy sub-image in the fuzzy subset. If the similarity is greater than the similarity threshold, add the position label of the label reference sub-image to the corresponding fuzzy sub-image to obtain preliminary displacement data. Step 4: Calculate the inter-frame time based on the exposure time and sampling frequency; calculate the blur parameters based on the preliminary displacement data and inter-frame time; generate a blur kernel based on the blur parameters; and perform blur processing on the target reference subset based on the blur kernel to obtain a blurred reference subset. Step 5: Calculate the similarity between the fuzzy reference subset and the fuzzy subset. If the similarity is not greater than the similarity threshold, repeat step 4 until the similarity is greater than the similarity threshold, then stop iterating. Perform inverse fuzzing on the fuzzy subset according to the fuzzy kernel to obtain an inverse fuzzy subset, and generate an enhanced fuzzy image set based on the inverse fuzzy subset.

[0008] Optionally, the target coarse screening model is based on an improvement of the YOLOv8 model, including: The target coarse screening model is obtained by replacing the backbone network in the YOLOv8 model with an improved backbone network and replacing the neck structure in the YOLOv8 model with an improved neck structure. The working principle of the improved backbone network includes: Obtain the original feature map, input the original feature map into the improved Conv module to obtain the first feature map, input the first feature map sequentially into the improved Conv module and the C2f module to obtain the second feature map, input the second feature map sequentially into the improved Conv module and the C2f module to obtain the third feature map, input the third feature map sequentially into the improved Conv module and the C2f module to obtain the fourth feature map, and input the fourth feature map sequentially into the improved Conv module, the C2f module and the SPPF module to obtain the fifth feature map; The first feature map, the second feature map, the third feature map, the fourth feature map, and the fifth feature map.

[0009] Optionally, the improved Conv module operates as follows: Obtain the original tensor, divide the original tensor into original sub-tensor sets, concatenate the original sub-tensors in the original sub-tensor sets according to the channel direction to obtain an intermediate tensor, input the intermediate tensor into the Conv module to obtain the final tensor, and use the final tensor as the output of the Conv module.

[0010] Optionally, the working principle of the improved neck structure includes: The 5th feature map is upsampled and then concatenated with the 4th feature map to obtain the 6th feature map. The 6th feature map is then sequentially input into the C2f module and the phantom convolution module to obtain the target 6th feature map. The 2nd feature map is upsampled to obtain the target 2nd feature map. The target 6th feature map, the target 2nd feature map, and the 3rd feature map are then concatenated to obtain the 7th feature map. The seventh feature map is sequentially input into the C2f module and the phantom convolution module to obtain the target seventh feature map. The first feature map is upsampled to obtain the target first feature map. The target seventh feature map, the target first feature map, and the second feature map are concatenated to obtain the eighth feature map. The 8th feature image is input into the C2f module to obtain the 9th feature image. The 9th feature image, the 7th feature image, and the 3rd feature image are concatenated to obtain the 10th feature image. The 10th feature image is input into the C2f module to obtain the 11th feature image. The 11th feature image is concatenated with the 5th feature image to obtain the 12th feature image. The 12th feature image is input into the C2f module to obtain the 13th feature image. The 9th feature map, the 11th feature map, and the 13th feature map are used as the output of the improved neck structure.

[0011] A second aspect of this invention provides an intelligent monitoring and control system for mass production of anti-counterfeiting labels, the system comprising: The image enhancement module is used to obtain a blurred image set by combining images of various anti-counterfeiting labels, and to enhance the blurred images in the blurred image set to obtain an enhanced blurred image set. The label coarse screening module is used to identify images in the fuzzy image set using a target coarse screening model to obtain identification results; the identification results include qualified and abnormal. The label screening module is used to combine images with abnormal identification results to obtain an abnormal image set, and to identify the abnormal image set through a target screening model to obtain a detection result; the detection result includes qualified and unqualified. The label rejection module is used to reject the corresponding anti-counterfeiting label if the test result is unqualified.

[0012] Optionally, the image enhancement module includes: The first operation module is used to acquire a reference image and key parameters, segment the reference image to obtain a reference subset, and segment the blurred image to obtain a blurred subset; the reference subset contains multiple reference sub-images; the blurred subset contains multiple blurred sub-images; the key parameters are the exposure time and sampling frequency of the blurred image; The second operation module is used to determine the position labels of each reference sub-image in the reference subset through the SIFT global search algorithm, add the position labels to the corresponding reference sub-images to obtain labeled reference sub-images, and combine the labeled reference sub-images to obtain the target reference subset. The third operation module is used to calculate the similarity between the label reference sub-image in the target reference subset and the fuzzy sub-image in the fuzzy subset. If the similarity is greater than the similarity threshold, the position label of the label reference sub-image is added to the corresponding fuzzy sub-image to obtain preliminary displacement data. The fourth operation module is used to calculate the inter-frame time based on the exposure time and sampling frequency, calculate the blur parameters based on the preliminary displacement data and inter-frame time, generate a blur kernel based on the blur parameters, and perform blur processing on the target reference subset based on the blur kernel to obtain a blurred reference subset; The fifth operation module is used to calculate the similarity between the fuzzy reference subset and the fuzzy subset. If the similarity is not greater than the similarity threshold, the fourth operation module is executed again until the similarity is greater than the similarity threshold, at which point the iteration stops. The fuzzy subset is then subjected to inverse fuzzing processing based on the fuzzy kernel to obtain an inverse fuzzy subset. An enhanced fuzzy image set is then generated based on the inverse fuzzy subset.

[0013] Optionally, the target coarse screening model is based on an improvement of the YOLOv8 model, including: The target coarse screening model is obtained by replacing the backbone network in the YOLOv8 model with an improved backbone network and replacing the neck structure in the YOLOv8 model with an improved neck structure. The working principle of the improved backbone network includes: Obtain the original feature map, input the original feature map into the improved Conv module to obtain the first feature map, input the first feature map sequentially into the improved Conv module and the C2f module to obtain the second feature map, input the second feature map sequentially into the improved Conv module and the C2f module to obtain the third feature map, input the third feature map sequentially into the improved Conv module and the C2f module to obtain the fourth feature map, and input the fourth feature map sequentially into the improved Conv module, the C2f module and the SPPF module to obtain the fifth feature map; The first feature map, the second feature map, the third feature map, the fourth feature map, and the fifth feature map.

[0014] Optionally, the improved Conv module operates as follows: Obtain the original tensor, divide the original tensor into original sub-tensor sets, concatenate the original sub-tensors in the original sub-tensor sets according to the channel direction to obtain an intermediate tensor, input the intermediate tensor into the Conv module to obtain the final tensor, and use the final tensor as the output of the Conv module.

[0015] Optionally, the working principle of the improved neck structure includes: The 5th feature map is upsampled and then concatenated with the 4th feature map to obtain the 6th feature map. The 6th feature map is then sequentially input into the C2f module and the phantom convolution module to obtain the target 6th feature map. The 2nd feature map is upsampled to obtain the target 2nd feature map. The target 6th feature map, the target 2nd feature map, and the 3rd feature map are then concatenated to obtain the 7th feature map. The seventh feature map is sequentially input into the C2f module and the phantom convolution module to obtain the target seventh feature map. The first feature map is upsampled to obtain the target first feature map. The target seventh feature map, the target first feature map, and the second feature map are concatenated to obtain the eighth feature map. The 8th feature image is input into the C2f module to obtain the 9th feature image. The 9th feature image, the 7th feature image, and the 3rd feature image are concatenated to obtain the 10th feature image. The 10th feature image is input into the C2f module to obtain the 11th feature image. The 11th feature image is concatenated with the 5th feature image to obtain the 12th feature image. The 12th feature image is input into the C2f module to obtain the 13th feature image. The 9th feature map, the 11th feature map, and the 13th feature map are used as the output of the improved neck structure.

[0016] The beneficial effects of this invention are: This invention proposes an intelligent monitoring and control method for the mass production of anti-counterfeiting labels. Through a collaborative detection process of fuzzy enhancement preprocessing, rapid coarse screening, and precise fine screening, it solves the problem of insufficient recognition capability for fuzzy images in traditional methods, thus improving system robustness. A high-speed initial screening is achieved using an improved YOLOv8 model, significantly improving overall detection efficiency and reducing the burden of subsequent processing. Accurate identification of suspicious targets using the standard YOLOv8 model ensures the accuracy of detection results. While guaranteeing high-precision screening, it optimizes the allocation of computational resources, achieving a balance between speed and accuracy, and providing efficient technical support for real-time and reliable quality monitoring on the production line. Attached Figure Description

[0017] Figure 1 A flowchart of an intelligent monitoring and control method for mass production of anti-counterfeiting labels provided in an embodiment of the present invention; Figure 2 A schematic diagram of a model for an intelligent monitoring and control method for mass production of anti-counterfeiting labels provided in an embodiment of the present invention; Figure 3 This is a framework diagram of an intelligent monitoring and control system for mass production of anti-counterfeiting labels provided in an embodiment of the present invention. Detailed Implementation

[0018] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0019] This invention provides an intelligent monitoring and control method for mass production of anti-counterfeiting labels. See also... Figure 1 , Figure 1 A flowchart illustrating an intelligent monitoring and control method for mass production of anti-counterfeiting labels, provided as an embodiment of the present invention. The method includes the following steps: S101, obtain a fuzzy image set by combining the images of each anti-counterfeiting label, and enhance the fuzzy images in the fuzzy image set to obtain an enhanced fuzzy image set; S102, the recognition result is obtained by recognizing the images in the fuzzy image set through the target coarse screening model; S103, combine the images with abnormal identification results to obtain an abnormal image set, and use the target fine screening model to identify the abnormal image set to obtain the detection result; S104. If the test result is unqualified, the corresponding anti-counterfeiting label will be removed.

[0020] Identification results include qualified and abnormal; detection results include qualified and unqualified. Based on the embodiments of the present invention, an intelligent monitoring and control method for mass production of anti-counterfeiting labels is provided. Through the collaborative process of fuzzy enhancement, rapid coarse screening using a target coarse screening model and precise fine screening using a standard model, the recognition capability and system robustness of fuzzy images are improved. While ensuring high detection accuracy, efficiency is improved, resource allocation is optimized, and a balance between speed and accuracy is achieved, providing real-time and reliable quality monitoring support for the production line.

[0021] In one implementation, the target coarse screening model is an improvement on the YOLOv8 model, used for rapid screening of anti-counterfeiting labels, while the target fine screening model is the standard YOLOv8 model (without modification), used for precise screening. The target coarse screening model has high detection efficiency but also a high false detection rate, and it is easy to identify qualified anti-counterfeiting labels as unqualified. Therefore, it is necessary to use the target fine screening model for fine screening to extract the falsely detected anti-counterfeiting labels.

[0022] One implementation employs fuzzy image enhancement and a two-level recognition mechanism, significantly improving detection efficiency and accuracy. First, a rapid coarse screening process eliminates most qualified products, followed by fine-grained detection of abnormal images. This effectively balances speed and accuracy requirements, reduces the risk of missed detections and false positives, and ensures the reliability of the screening results.

[0023] In one implementation, the combined use of the improved YOLOv8 model and the standard model optimizes resource allocation and task division. The coarse screening model achieves rapid initial screening of blurry images through improvements, reducing the overall processing time; the fine screening model focuses on the accurate identification of suspicious samples. The combination of the two ensures high accuracy while improving the overall response speed of the system.

[0024] In one implementation, when the monitoring system determines that an anti-counterfeiting label is substandard, the system immediately triggers an automatic rejection mechanism, physically separating the corresponding label from the production line and marking it as defective. Simultaneously, specific information about the substandard product (such as image data, defect type, and timestamp) is automatically recorded and uploaded to the quality management system for generating defect statistical reports and conducting traceability analysis of the production process. This achieves real-time, automated handling of substandard products, effectively preventing them from flowing into subsequent stages and providing data support for adjusting and optimizing the production process.

[0025] In one embodiment, enhancing the blurred images in a blurred image set to obtain an enhanced blurred image set includes: Step 1: Obtain the reference image and key parameters. Segment the reference image to obtain a reference subset, and segment the blurred image to obtain a blurred subset. The reference subset contains multiple reference sub-images, and the blurred subset contains multiple blurred sub-images. The key parameters are the exposure time and sampling frequency of the blurred image. Step 2: Determine the position labels of each reference sub-image in the reference subset using the SIFT global search algorithm, add the position labels to the corresponding reference sub-images to obtain labeled reference sub-images, and combine the labeled reference sub-images to obtain the target reference subset; Step 3: Calculate the similarity between the label reference sub-image in the target reference subset and the fuzzy sub-image in the fuzzy subset. If the similarity is greater than the similarity threshold, add the position label of the label reference sub-image to the corresponding fuzzy sub-image to obtain preliminary displacement data. Step 4: Calculate the inter-frame time based on the exposure time and sampling frequency; calculate the blur parameters based on the preliminary displacement data and inter-frame time; generate a blur kernel based on the blur parameters; and perform blur processing on the target reference subset based on the blur kernel to obtain the blurred reference subset. Step 5: Calculate the similarity between the fuzzy reference subset and the fuzzy subset. If the similarity is not greater than the similarity threshold, repeat step 4 until the similarity is greater than the similarity threshold, then stop iterating. Perform inverse fuzzing on the fuzzy subset according to the fuzzy kernel to obtain the inverse fuzzy subset, and generate an enhanced fuzzy image set based on the inverse fuzzy subset.

[0026] One implementation uses SIFT global search and similarity calculation, with the core being the establishment of a spatial correspondence between reference sub-images and blurred sub-images. The calculation process is as follows: First, the SIFT algorithm is used to extract feature points (such as texture edges and speckle pattern features) from the reference sub-images, assigning a unique location label to each reference sub-image to form a labeled reference sub-image and a target reference subset. Then, the similarity (commonly measured by normalized cross-correlation coefficient) between each labeled reference sub-image in the target reference subset and all blurred sub-images in the blurred subset is calculated. If the similarity exceeds a preset threshold (e.g., 0.8), it is determined that the two are images of the same physical point, and the location label of the reference sub-image is assigned to the corresponding blurred sub-image, directly obtaining preliminary displacement data (displacement = blurred sub-image coordinates - reference sub-image coordinates). This process avoids the inefficiency of blind searching in traditional DIC, quickly locking the correspondence through feature matching, providing a reliable initial displacement input for subsequent blurred parameter calculations, and ensuring the accuracy of the starting point for subsequent processes.

[0027] In one implementation, the inter-frame time Δt = 1 / f is calculated based on key parameters. s (where f) sThe sampling frequency is the parameter that reflects the time interval between two adjacent frames and is the basis for velocity calculation. The second step involves calculating the subset velocity based on the preliminary displacement data using the finite difference method; the first-order approximation formula is... (u) i u is the current frame offset. i-1 (where the displacement is from the previous frame), the second-order approximation formula is: (Introducing the displacement u of the first two frames) i-2 (To improve accuracy), two methods can be selected based on the smoothness of the motion; the third step is to use the motion fuzzy physical model. (where b is the fuzzy parameter, v is the subset velocity, and t) exp The exposure time, combined with the time ratio parameter, simplifies to obtain the final calculation formula for the blur parameters (first order) or (second order). Through this calculation process, the blur intensity and direction can be quantized subset by subset, accurately adapting to local blur differences caused by non-uniform structural deformation or camera motion in real-world scenes.

[0028] One implementation method uses reference subset blurring and iterative optimization to ensure that both matching elements are in the same blur dimension, avoiding artifact interference from traditional deblurring. The calculation process is as follows: First, based on the blur parameters obtained in step two, a 2D blur kernel with corresponding direction and intensity is generated (e.g., if the blur in the x-direction is 5px, a kernel matrix extending along the x-axis is generated); then, the blur kernel is applied to each label reference sub-image of the target reference subset through 2D convolution to obtain a blurred reference subset (the convolution process needs to expand the sub-image region to avoid edge failure; for example, if the kernel width is 5px, two columns of pixels are added to the left and right sides of the subset); next, the similarity between the blurred reference subset and the original blurred subset is calculated. If the similarity does not reach the threshold, the blur parameters are corrected by substituting the displacement deviation corresponding to the current similarity (e.g., if the similarity is low, the current frame displacement is adjusted according to the deviation and the blur parameters are recalculated), and the blur kernel generation and reference subset blurring steps are repeated until the similarity exceeds the threshold. In this iterative process, the blur parameter calculation depends on the displacement data, and the displacement correction depends on the blur matching result.

[0029] In one implementation, the simplified fuzzing logic avoids additional high computational overhead. The efficiency of the calculation process is reflected in two aspects: First, the fuzzing parameter calculation reuses the existing displacement data and subset structure of DIC, without the need for additional image segmentation or the introduction of complex deblurring algorithms (such as Wiener filtering). The fuzzing parameters can be obtained through basic algebraic operations (difference, multiplication), resulting in low computational complexity. Second, during the iteration process, only the reference subset is fuzzified, rather than the fuzzy image is deblurred (deblurring requires deconvolution operations, which consume a lot of computational power and are prone to introducing artifacts). Fuzzification is achieved through simple convolution.

[0030] In one embodiment, the target coarse screening model is based on an improvement of the YOLOv8 model, including: The target coarse screening model is obtained by replacing the backbone network in the YOLOv8 model with an improved backbone network and replacing the neck structure in the YOLOv8 model with an improved neck structure. Improvements to the working principle of the backbone network include: Obtain the original feature map, input the original feature map into the improved Conv module to obtain the first feature map, input the first feature map into the improved Conv module and the C2f module in sequence to obtain the second feature map, input the second feature map into the improved Conv module and the C2f module in sequence to obtain the third feature map, input the third feature map into the improved Conv module and the C2f module in sequence to obtain the fourth feature map, and input the fourth feature map into the improved Conv module, the C2f module and the SPPF module in sequence to obtain the fifth feature map; The first feature map, the second feature map, the third feature map, the fourth feature map, and the fifth feature map.

[0031] In one implementation, the backbone network performs initial downsampling and feature transformation at each stage using an improved Conv module. This operation actively reorganizes spatial information into the channel dimension each time the feature map spatial resolution is reduced, helping to alleviate the loss of fine-grained information about small targets caused by continuous downsampling. The improved Conv module forms a collaborative unit with the subsequent C2f module. The detailed intermediate features output by the improved Conv module are used by the C2f module for deep feature interaction and semantic refinement through its multi-branch structure, enabling the network to simultaneously maintain sensitivity to small target features and representation depth during the construction of a multi-level feature pyramid.

[0032] In one implementation, the first to fifth feature maps, representing multi-scale features retained in stages within the backbone network, provide a foundation for feature fusion in the subsequent neck network. The shallow, high-resolution features (such as the first and second feature maps) retain rich localization and detail information, while the deeper features (such as the fourth and fifth feature maps) carry stronger semantic information.

[0033] In one embodiment, the improved operation of the Conv module includes: Obtain the original tensor, divide the original tensor into original sub-tensor sets, concatenate the original sub-tensors in the original sub-tensor sets according to the channel direction to obtain the intermediate tensor, input the intermediate tensor into the Conv module to obtain the final tensor, and use the final tensor as the output of the Conv module.

[0034] In one implementation, the Conv module is improved by finely dividing the input feature map into multiple sub-regions (original sub-tensor sets) in the spatial dimension and concatenating these sub-regions along the channel direction. This efficiently transforms high-resolution spatial information into a deeper channel representation (intermediate tensor) without losing pixel information, avoiding the loss of fine-grained information of small targets caused by the direct discarding of features in traditional strided convolution. The Conv module processes this spatially detailed intermediate tensor pixel by pixel to further extract and enhance features. The transformation from space to channel compresses the feature map size and reduces computational complexity while preserving and organizing complete contextual information of small targets for subsequent convolutional layers. The subsequent fine convolution further deepens the feature expression, improves feature sensitivity and representation ability, and maintains the lightweight characteristics of the model, achieving a balance between detection accuracy and computational efficiency.

[0035] In one embodiment, improving the working principle of the neck structure includes: The 5th feature map is upsampled and then concatenated with the 4th feature map to obtain the 6th feature map. The 6th feature map is then input into the C2f module and the phantom convolution module to obtain the target 6th feature map. The 2nd feature map is upsampled to obtain the target 2nd feature map. The target 6th feature map, the target 2nd feature map, and the 3rd feature map are then concatenated to obtain the 7th feature map. The 7th feature map is sequentially input into the C2f module and the phantom convolution module to obtain the target 7th feature map. The 1st feature map is upsampled to obtain the target 1st feature map. The target 7th feature map, the target 1st feature map and the 2nd feature map are concatenated to obtain the 8th feature map. The 8th feature map is input into the C2f module to obtain the 9th feature map. The 9th, 7th and 3rd feature maps are concatenated to obtain the 10th feature map. The 10th feature map is input into the C2f module to obtain the 11th feature map. The 11th feature map is concatenated with the 5th feature map to obtain the 12th feature map. The 12th feature map is input into the C2f module to obtain the 13th feature map. The 9th, 11th, and 13th feature maps are used as the output for improving the neck structure.

[0036] In one implementation, see [link to implementation details]. Figure 2 , Figure 2This is a schematic diagram of a model for an intelligent monitoring and control method for mass production of anti-counterfeiting labels provided by an embodiment of the present invention. Multi-level cross-scale stitching sequentially pulls the feature maps from the 5th to the 1st layer back to the same resolution, and then merges them with adjacent layers pairwise, allowing high-level semantics and low-level details to coexist in the same feature map. The C2f module then performs channel rearrangement and grouping reuse on the stitching results, preserving gradient information from different layers while distributing computational load across multiple sub-paths. The phantom convolution further uses linear expansion plus lightweight linear aggregation, first expanding and then shrinking the channel dimension, controlling the number of parameters to about one-third of that of a regular convolution. After the two are superimposed, the network can backpropagate denser gradients under the same computational budget, maintaining the response values ​​of small target edges, while compressing redundant channels in the background region, providing a feature representation for subsequent detection heads that balances resolution and semantic richness.

[0037] In one implementation, the concatenation path of high-level, mid-level, and low-level layers simultaneously retains the 9th, 11th, and 13th feature maps as neck outputs, allowing the detection head to obtain descriptors from shallow, mid, and deep receptive fields in the same inference stage. Since these descriptors share the same backpropagation path, gradients can be superimposed across stages during backpropagation: the gradient obtained from the small target layer (9th layer) can directly enhance the detail preservation of the mid-target layer (11th layer), and the learned signal from the mid-target layer can be further fed back to the semantic enhancement of the large target layer (13th layer) through the shared C2f branch. This linkage mechanism allows the loss functions of targets of different sizes to reference each other during updates, alleviating gradient competition caused by inter-layer isolation. This results in a relatively stable convergence curve and detection performance when the input image resolution fluctuates or the target scale distribution changes.

[0038] Based on the same inventive concept, this invention also provides an intelligent monitoring and control system for mass production of anti-counterfeiting labels. See also... Figure 2 , Figure 2 A framework diagram of an intelligent monitoring and control system for mass production of anti-counterfeiting labels provided in this embodiment of the invention includes: The image enhancement module is used to obtain a blurred image set by combining images of various anti-counterfeiting labels, and to enhance the blurred images in the blurred image set to obtain an enhanced blurred image set. The label coarse screening module is used to identify images in a fuzzy image set using a target coarse screening model to obtain identification results; the identification results include qualified and abnormal. The label screening module is used to combine images with abnormal identification results to obtain an abnormal image set. The abnormal image set is then identified by the target screening model to obtain the detection result. The detection result includes qualified and unqualified. The label rejection module is used to reject the corresponding anti-counterfeiting label if the test result is unqualified.

[0039] Based on the intelligent monitoring and control system for mass production of anti-counterfeiting labels provided in this invention, the system improves the recognition capability and robustness of fuzzy images through a collaborative process of fuzz enhancement, rapid coarse screening using a target coarse screening model, and precise fine screening using a standard model. This ensures high detection accuracy while improving efficiency, optimizes resource allocation, and achieves a balance between speed and accuracy, providing real-time and reliable quality monitoring support for the production line.

[0040] In one embodiment, the image enhancement module includes: The first operation module is used to acquire a reference image and key parameters, segment the reference image to obtain a reference subset, and segment the blurred image to obtain a blurred subset; the reference subset contains multiple reference sub-images; the blurred subset contains multiple blurred sub-images; the key parameters are the exposure time and sampling frequency of the blurred image. The second operation module is used to determine the position labels of each reference sub-image in the reference subset through the SIFT global search algorithm, add the position labels to the corresponding reference sub-images to obtain labeled reference sub-images, and combine the labeled reference sub-images to obtain the target reference subset. The third operation module is used to calculate the similarity between the label reference sub-image in the target reference subset and the fuzzy sub-image in the fuzzy subset. If the similarity is greater than the similarity threshold, the position label of the label reference sub-image is added to the corresponding fuzzy sub-image to obtain preliminary displacement data. The fourth operation module is used to calculate the inter-frame time based on the exposure time and sampling frequency, calculate the blur parameters based on the preliminary displacement data and inter-frame time, generate a blur kernel based on the blur parameters, and perform blur processing on the target reference subset based on the blur kernel to obtain a blurred reference subset. The fifth operation module is used to calculate the similarity between the fuzzy reference subset and the fuzzy subset. If the similarity is not greater than the similarity threshold, the fourth operation module is executed again until the similarity is greater than the similarity threshold, at which point the iteration stops. The fuzzy subset is then defuzzified according to the fuzzy kernel to obtain the defuzzy subset, and the enhanced fuzzy image set is generated based on the defuzzy subset.

[0041] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention should still fall within the scope of the claims of the present invention.

Claims

1. A smart monitoring control method for the mass production of anti-counterfeit labels, characterized in that, The method includes: A blurred image set is obtained by combining images of various anti-counterfeiting labels, and the blurred images in the blurred image set are enhanced to obtain an enhanced blurred image set. The images in the enhanced blurred image set are identified using a target coarse screening model to obtain identification results; the identification results include qualified and abnormal. Images identified as abnormal are combined to form an abnormal image set. This abnormal image set is then analyzed using a target fine-screening model to obtain detection results, which include qualified and unqualified results. If the test result is unqualified, the corresponding anti-counterfeiting label will be removed. Enhanced blurred images in the blurred image set are obtained by enhancing the blurred images in the blurred image set, including: Step 1: Obtain a reference image and key parameters; segment the reference image to obtain a reference subset; segment the blurred image to obtain a blurred subset; the reference subset contains multiple reference sub-images; the blurred subset contains multiple blurred sub-images; the key parameters are the exposure time and sampling frequency of the blurred image. Step 2: Determine the position labels of each reference sub-image in the reference subset using the SIFT global search algorithm, add the position labels to the corresponding reference sub-images to obtain labeled reference sub-images, and combine the labeled reference sub-images to obtain the target reference subset; Step 3: Calculate the similarity between the label reference sub-image in the target reference subset and the fuzzy sub-image in the fuzzy subset. If the similarity is greater than the similarity threshold, add the position label of the label reference sub-image to the corresponding fuzzy sub-image to obtain preliminary displacement data. Step 4: Calculate the inter-frame time based on the exposure time and sampling frequency; calculate the blur parameters based on the preliminary displacement data and inter-frame time; generate a blur kernel based on the blur parameters; and perform blur processing on the target reference subset based on the blur kernel to obtain a blurred reference subset. Step 5: Calculate the similarity between the fuzzy reference subset and the fuzzy subset. If the similarity is not greater than the similarity threshold, repeat step 4 until the similarity is greater than the similarity threshold, then stop iterating. Perform inverse fuzzing on the fuzzy subset according to the fuzzy kernel to obtain an inverse fuzzy subset, and generate an enhanced fuzzy image set based on the inverse fuzzy subset.

2. The intelligent monitoring and control method for mass production of anti-counterfeiting labels according to claim 1, characterized in that, The target coarse screening model is an improvement based on the YOLOv8 model, including: The target coarse screening model is obtained by replacing the backbone network in the YOLOv8 model with an improved backbone network and replacing the neck structure in the YOLOv8 model with an improved neck structure. The working principle of the improved backbone network includes: Obtain the original feature map, input the original feature map into the improved Conv module to obtain the first feature map, input the first feature map sequentially into the improved Conv module and the C2f module to obtain the second feature map, input the second feature map sequentially into the improved Conv module and the C2f module to obtain the third feature map, input the third feature map sequentially into the improved Conv module and the C2f module to obtain the fourth feature map, and input the fourth feature map sequentially into the improved Conv module, the C2f module and the SPPF module to obtain the fifth feature map.

3. The intelligent monitoring and control method for mass production of anti-counterfeiting labels according to claim 2, characterized in that, The working principle of the improved Conv module includes: Obtain the original tensor, divide the original tensor into original sub-tensor sets, concatenate the original sub-tensors in the original sub-tensor sets according to the channel direction to obtain an intermediate tensor, input the intermediate tensor into the Conv module to obtain the final tensor, and use the final tensor as the output of the Conv module.

4. The intelligent monitoring and control method for mass production of anti-counterfeiting labels according to claim 2, characterized in that, The working principle of the improved neck structure includes: The 5th feature map is upsampled and then concatenated with the 4th feature map to obtain the 6th feature map. The 6th feature map is then sequentially input into the C2f module and the phantom convolution module to obtain the target 6th feature map. The 2nd feature map is upsampled to obtain the target 2nd feature map. The target 6th feature map, the target 2nd feature map, and the 3rd feature map are then concatenated to obtain the 7th feature map. The seventh feature map is sequentially input into the C2f module and the phantom convolution module to obtain the target seventh feature map. The first feature map is upsampled to obtain the target first feature map. The target seventh feature map, the target first feature map, and the second feature map are concatenated to obtain the eighth feature map. The 8th feature image is input into the C2f module to obtain the 9th feature image. The 9th feature image, the 7th feature image, and the 3rd feature image are concatenated to obtain the 10th feature image. The 10th feature image is input into the C2f module to obtain the 11th feature image. The 11th feature image is concatenated with the 5th feature image to obtain the 12th feature image. The 12th feature image is input into the C2f module to obtain the 13th feature image. The 9th feature map, the 11th feature map, and the 13th feature map are used as the output of the improved neck structure.

5. An intelligent monitoring and control system for mass production of anti-counterfeiting labels, characterized in that, The system includes: The image enhancement module is used to obtain a blurred image set by combining images of various anti-counterfeiting labels, and to enhance the blurred images in the blurred image set to obtain an enhanced blurred image set. The label coarse screening module is used to identify images in the enhanced blurred image set through a target coarse screening model to obtain identification results; the identification results include qualified and abnormal. The label screening module is used to combine images with abnormal identification results to obtain an abnormal image set, and to identify the abnormal image set through a target screening model to obtain a detection result; the detection result includes qualified and unqualified. The label rejection module is used to reject the corresponding anti-counterfeiting label if the test result is unqualified. The image enhancement module includes: The first operation module is used to acquire a reference image and key parameters, segment the reference image to obtain a reference subset, and segment the blurred image to obtain a blurred subset; the reference subset contains multiple reference sub-images; the blurred subset contains multiple blurred sub-images; the key parameters are the exposure time and sampling frequency of the blurred image; The second operation module is used to determine the position labels of each reference sub-image in the reference subset through the SIFT global search algorithm, add the position labels to the corresponding reference sub-images to obtain labeled reference sub-images, and combine the labeled reference sub-images to obtain the target reference subset. The third operation module is used to calculate the similarity between the label reference sub-image in the target reference subset and the fuzzy sub-image in the fuzzy subset. If the similarity is greater than the similarity threshold, the position label of the label reference sub-image is added to the corresponding fuzzy sub-image to obtain preliminary displacement data. The fourth operation module is used to calculate the inter-frame time based on the exposure time and sampling frequency, calculate the blur parameters based on the preliminary displacement data and inter-frame time, generate a blur kernel based on the blur parameters, and perform blur processing on the target reference subset based on the blur kernel to obtain a blurred reference subset; The fifth operation module is used to calculate the similarity between the fuzzy reference subset and the fuzzy subset. If the similarity is not greater than the similarity threshold, the fourth operation module is executed again until the similarity is greater than the similarity threshold, at which point the iteration stops. The fuzzy subset is then subjected to inverse fuzzing processing based on the fuzzy kernel to obtain an inverse fuzzy subset. An enhanced fuzzy image set is then generated based on the inverse fuzzy subset.

6. The intelligent monitoring and control system for mass production of anti-counterfeiting labels according to claim 5, characterized in that, The target coarse screening model is an improvement based on the YOLOv8 model, including: The target coarse screening model is obtained by replacing the backbone network in the YOLOv8 model with an improved backbone network and replacing the neck structure in the YOLOv8 model with an improved neck structure. The working principle of the improved backbone network includes: Obtain the original feature map, input the original feature map into the improved Conv module to obtain the first feature map, input the first feature map sequentially into the improved Conv module and the C2f module to obtain the second feature map, input the second feature map sequentially into the improved Conv module and the C2f module to obtain the third feature map, input the third feature map sequentially into the improved Conv module and the C2f module to obtain the fourth feature map, and input the fourth feature map sequentially into the improved Conv module, the C2f module and the SPPF module to obtain the fifth feature map.

7. The intelligent monitoring and control system for mass production of anti-counterfeiting labels according to claim 6, characterized in that, The working principle of the improved Conv module includes: Obtain the original tensor, divide the original tensor into original sub-tensor sets, concatenate the original sub-tensors in the original sub-tensor sets according to the channel direction to obtain an intermediate tensor, input the intermediate tensor into the Conv module to obtain the final tensor, and use the final tensor as the output of the Conv module.

8. The intelligent monitoring and control system for mass production of anti-counterfeiting labels according to claim 6, characterized in that, The working principle of the improved neck structure includes: The 5th feature map is upsampled and then concatenated with the 4th feature map to obtain the 6th feature map. The 6th feature map is then sequentially input into the C2f module and the phantom convolution module to obtain the target 6th feature map. The 2nd feature map is upsampled to obtain the target 2nd feature map. The target 6th feature map, the target 2nd feature map, and the 3rd feature map are then concatenated to obtain the 7th feature map. The seventh feature map is sequentially input into the C2f module and the phantom convolution module to obtain the target seventh feature map. The first feature map is upsampled to obtain the target first feature map. The target seventh feature map, the target first feature map, and the second feature map are concatenated to obtain the eighth feature map. The 8th feature image is input into the C2f module to obtain the 9th feature image. The 9th feature image, the 7th feature image, and the 3rd feature image are concatenated to obtain the 10th feature image. The 10th feature image is input into the C2f module to obtain the 11th feature image. The 11th feature image is concatenated with the 5th feature image to obtain the 12th feature image. The 12th feature image is input into the C2f module to obtain the 13th feature image. The 9th feature map, the 11th feature map, and the 13th feature map are used as the output of the improved neck structure.

Citation Information

Patent Citations

  • Electronic tag production quality control method

    CN120068902A

  • Printing label defect detection method and system

    CN118429353A

  • Line component defect detection method based on deep learning

    CN121073927A

  • Method for testing printed articles

    EP2787489A1