Beverage label defect detection method and system based on industrial vision

Abstract

The invention provides a beverage label defect detection method and system based on industrial vision, and relates to the technical field of industrial vision detection.Texture features of a local area are extracted according to a qualified novel recycled paper-based label image, a normal texture spectrum is established, brightness normalization processing is carried out on a to-be-detected label image, and a beverage label defect detection result is obtained. And extracting the texture features of the local area of the to-be-detected label image, comparing the texture features of the local area of the to-be-detected label image with a normal texture spectrum, calculating to obtain a texture difference degree, and performing defect mode matching based on a preset defect mode to obtain a defect mode matching result. According to the texture difference degree, the defect mode matching result and the texture difference degree threshold value, whether the local area of the to-be-detected label image has defects or not is judged, inherent features and real defects of materials can be effectively distinguished, the false detection rate and the omission ratio are remarkably reduced, and therefore the accuracy and efficiency of product quality control are improved.

Description

Technical Field

[0001] This application relates to the field of industrial vision inspection technology, and more specifically, to a method and system for detecting defects in beverage labels based on industrial vision. Background Technology

[0002] In modern beverage packaging production lines, a typical industrial vision inspection system is usually deployed at critical points, its core task being to ensure that the label affixing quality on beverage bottles meets standards. This system generally consists of a high-resolution industrial camera, a precisely calibrated LED light source array, and an image processing unit. In the initial stages of system operation, its performance usually meets the expected inspection requirements. However, during the long-term continuous operation of the industrial production line, the LED light source array upon which the industrial vision system relies may experience slight differences in light decay rates in different areas due to continuous material fatigue and accumulated thermal stress. This non-uniform light decay phenomenon causes the light source, originally designed to provide uniform illumination, to project unevenly bright and dark spots onto the label surface. This unevenness in light output directly results in the light distribution on the label surface no longer being ideally uniform when the industrial camera captures the label image. Specifically, some areas of the label appear darker due to insufficient light, while other areas may maintain relatively normal brightness. This local brightness difference reduces the overall contrast of the image and results in different contrast levels in different areas of the image.

[0003] Under the influence of uneven image illumination and decreased contrast, the image acquisition exposure time parameters originally optimized for uniform lighting environments can no longer simultaneously ensure image quality across all areas. If the system shortens the exposure time to avoid overexposure in bright areas, image details in poorly lit dark areas will be further lost, becoming blurry. Conversely, if the exposure time is extended to fully capture details in dark areas, bright areas may experience severe overexposure, leading to information saturation and irreversible loss of detail. In this situation, the feature extraction and decision-making logic of the defect classification system will be severely disrupted when processing images with missing details or information saturation. For example, a tiny bubble may be indistinguishable from background noise in poorly lit areas, while a light-colored printing blur may be completely masked in overexposed areas, resulting in a significant decrease in the accuracy of identifying defects such as tiny wrinkles, fine bubbles, and light-colored printing blur.

[0004] In response to market demand for environmentally friendly products, production lines may introduce beverage labels using a new type of recycled paper-based material. This new material is not traditional smooth coated paper; rather, due to its manufacturing process, its surface naturally possesses irregular, fine fibrous textures. This texture is an inherent physical property of the material, not a defect. Furthermore, the surface porosity of recycled paper-based materials is typically higher than that of traditional coated paper, resulting in a more complex diffuse reflection effect on light. When light shines on this surface, it doesn't form a sharp reflection like on a smooth surface, but rather scatters in all directions, giving the image a "soft" or "blurred" appearance at the microscopic level. More importantly, because this porous surface results in uneven absorption and curing of printing ink, it directly leads to reduced edge sharpness of the printed pattern and insufficient color saturation in some light-colored text or patterns, creating naturally low-contrast areas.

[0005] Ultimately, when the production line switched to inspecting this new beverage model, the uneven light decay of the LED light source array caused by long-term operation, combined with the inherent complex physical characteristics of the new label material (including its irregular fine fiber texture, complex diffuse reflection of light, and reduced edge sharpness and inherent low contrast of the printed pattern due to uneven ink absorption), resulted in a severe deterioration in the image quality acquired by the system. In the dark areas caused by uneven light source, the fiber texture of the new material became blurred and difficult to identify under low light and complex diffuse reflection, easily misjudged as minor wrinkles or surface scratches by the defect classification logic, leading to a sharp increase in the false detection rate. In the bright areas, the sacrifice of image details to avoid overexposure caused printing blurry defects, which already had low contrast due to uneven ink absorption, to be completely submerged in the background information or confused with the material's own texture features, resulting in a significant increase in the missed detection rate. Under this dual influence, the criteria for defect classification became chaotic, the system could hardly function properly, and it could not effectively distinguish between the inherent characteristics of the material and real defects, posing a serious challenge to product quality control. Summary of the Invention

[0006] This application provides a method and system for detecting defects in beverage labels based on industrial vision. It aims to solve the problem that during long-term operation of industrial production lines, the uneven light decay of LED light source arrays and the inherent complex characteristics of new recycled paper-based label materials lead to severe deterioration of image quality. As a result, the system cannot effectively distinguish between the inherent characteristics of the material and real defects, causing a sharp increase in false detection rate and false negative rate, and posing a serious challenge to product quality control.

[0007] On the one hand, this application provides a method for detecting defects in beverage labels based on industrial vision, including:

[0008] Based on qualified images of new recycled paper-based labels, local texture features are extracted, and a normal texture spectrum describing the inherent texture of new recycled paper-based labels is established based on the extracted texture features.

[0009] After performing brightness normalization on the real-time acquired image of the label to be detected, the texture features of its local area are extracted.

[0010] The texture features of a local region of the label image to be detected are compared with the normal texture spectrum to calculate the texture difference.

[0011] Based on the texture difference degree and the preset defect pattern, defect pattern matching is performed to obtain the defect pattern matching result;

[0012] Based on the texture difference degree, the defect pattern matching result, and the texture difference degree threshold adjusted according to production environment parameters, it is determined whether there are defects in a local area of ​​the label image to be detected.

[0013] Optionally, the step of establishing a normal texture spectrum describing the inherent texture performance of the novel recycled paper-based label based on the extracted texture features includes:

[0014] When the system is deployed for the first time or a new batch of materials is launched, an initial label image is acquired, and texture features of local areas are extracted from the initial label image;

[0015] The extracted texture features are clustered to obtain an initial normal texture reference;

[0016] During normal production line operation, images of the labels to be inspected are acquired in real time, and texture features of local areas are extracted from the images of the labels to be inspected.

[0017] Calculate the texture difference degree between the texture features of a local region of the label image to be detected and the initial normal texture reference, and determine the qualified region based on the texture difference degree;

[0018] Based on the areas deemed qualified, assess the confidence level of the qualified areas;

[0019] The initial normal texture reference is periodically updated based on qualified regions with confidence levels higher than a preset confidence threshold to obtain an updated normal texture spectrum;

[0020] Receive defect area information from manual feedback, and adjust the updated normal texture spectrum based on the defect area information.

[0021] Optionally, the step of comparing the texture features of a local region of the label image to be detected with the normal texture spectrum to calculate the texture difference includes:

[0022] The texture features of a local region of the image to be detected are compared with the normal texture spectrum to mark potential texture anomalies;

[0023] Track the potential texture anomalies in consecutive frames of an image sequence and establish anomaly trajectories;

[0024] Based on the abnormal trajectory, determine whether it is a persistent defect. If it is determined to be a persistent defect, calculate the texture difference degree between the texture features of the local region and the normal texture spectrum.

[0025] Optionally, the step of extracting texture features of local areas from a qualified new recycled paper-based label image and establishing a normal texture spectrum describing the inherent texture of the new recycled paper-based label based on the extracted texture features includes:

[0026] Configure a multispectral industrial camera to acquire qualified multispectral images of new recycled paper-based labels;

[0027] The multispectral image is divided into multiple local regions, and multispectral texture features are extracted from the local regions. The multispectral texture features characterize the inherent texture, light diffuse reflection characteristics, and contrast of the printed pattern in different spectral channels of the novel recycled paper-based label.

[0028] By analyzing the correlation between the multispectral texture features in different spectral channels, texture features characterizing the inherent texture, light diffuse reflection characteristics, and printing pattern contrast of the novel recycled paper-based label are obtained, and a normal texture spectrum describing the inherent texture performance of the novel recycled paper-based label is established.

[0029] Optionally, the step of analyzing the correlation between the multispectral texture features in different spectral channels to obtain texture features characterizing the inherent texture, light diffuse reflection characteristics, and printed pattern contrast of the novel recycled paper-based label, and establishing a normal texture spectrum describing the inherent texture performance of the novel recycled paper-based label, includes:

[0030] Statistical analysis was performed on the correlation between the multispectral texture features in different spectral channels to establish an initial multispectral texture correlation reference, which characterizes the normal spectral response characteristics of the current batch of materials and inks;

[0031] During normal production line operation, multispectral images of the labels to be inspected are acquired in real time, the multispectral images of the labels to be inspected are divided into multiple local regions, and multispectral texture features are extracted from each local region.

[0032] The correlation difference degree between the extracted multispectral texture features and the initial multispectral texture association reference is calculated to obtain the correlation difference degree;

[0033] Based on the correlation difference degree and the preset difference degree threshold range, determine whether it is a qualified region and assess the confidence level of the qualified region.

[0034] The initial multispectral texture association reference is periodically updated based on the multispectral texture features of qualified regions with confidence levels higher than a preset confidence threshold, to obtain the updated normal texture spectrum;

[0035] Receive defect area information from manual feedback, and adjust and update the normal texture spectrum based on the defect area information.

[0036] Optionally, the step of analyzing the correlation between the multispectral texture features across different spectral channels includes:

[0037] The spectral response curves of the multispectral texture features in the local area are fitted to obtain the spectral response characteristic curves;

[0038] Calculate the rate of change of curvature of the spectral response characteristic curve, and compare the rate of change of curvature with a preset material density non-uniformity characteristic threshold to obtain the comparison result of material density non-uniformity.

[0039] The cross-correlation of the multispectral texture features across different spectral channels is calculated to obtain the cross-correlation results.

[0040] By combining the fitting residuals of the spectral response characteristic curves, correlation anomalies that match the printing blur defect pattern are identified, and correlation anomaly identification results are obtained.

[0041] Based on the comparison results of the material density non-uniformity, the cross-correlation results, and the correlation anomaly identification results, it is determined whether the texture anomaly in the local area originates from the inherent non-uniformity of the material density or a printing blur defect.

[0042] Optionally, the step of calculating the rate of change of curvature of the spectral response characteristic curve and comparing the rate of change of curvature with a preset material density non-uniformity characteristic threshold to obtain the comparison result of material density non-uniformity includes:

[0043] Identify local extrema on the spectral response characteristic curve;

[0044] Calculate the peak intensity and peak width of the rate of curvature change of the spectral response characteristic curve at each local extremum point;

[0045] Statistical analysis was performed on all the peak intensities and peak widths to obtain the distribution characteristics of the rate of change of curvature;

[0046] Based on the distribution characteristics, the threshold for preset material density non-uniformity is dynamically adjusted;

[0047] The rate of curvature change of each local region is compared with the adjusted preset threshold for material density non-uniformity to obtain the comparison result of material density non-uniformity.

[0048] Optionally, the step of calculating the cross-correlation of the multispectral texture features across different spectral channels to obtain the cross-correlation results includes:

[0049] Before calculating the cross-correlation, the filtering parameters are adjusted according to the scattering intensity and spatial frequency distribution of the multispectral texture features of each local region under different spectral channels to obtain the multispectral texture features after scattering noise removal.

[0050] Cross-correlation calculations were performed on the multispectral texture features after scattering noise removal across different spectral channels to obtain the cross-correlation results.

[0051] Optionally, the texture features are obtained by calculating the gray-level co-occurrence matrix statistics or Gabor filter bank response values ​​of the local region.

[0052] On the other hand, this application provides a beverage label defect detection system based on industrial vision, the system comprising:

[0053] The texture feature extraction module is used to extract texture features of local areas from qualified new recycled paper-based label images, and to establish a normal texture spectrum describing the inherent texture of new recycled paper-based labels based on the extracted texture features.

[0054] The local texture feature extraction module is used to extract the texture features of local areas after performing brightness normalization processing on the real-time acquired image of the label to be detected.

[0055] The texture difference calculation module is used to compare the texture features of a local region of the label image to be detected with the normal texture spectrum to calculate the texture difference.

[0056] The defect pattern matching module is used to perform defect pattern matching based on the texture difference degree and the preset defect pattern to obtain the defect pattern matching result.

[0057] The defect determination module is used to determine whether there are defects in a local area of ​​the label image to be detected based on the texture difference degree, the defect pattern matching result, and the texture difference degree threshold adjusted according to the production environment parameters.

[0058] This application relates to a method and system for detecting defects in beverage labels based on industrial vision. By extracting texture features from local areas of qualified new recycled paper-based label images and establishing a normal texture spectrum describing the inherent texture of the new recycled paper-based label, it effectively solves the problem of confusion between the inherent characteristics of the new recycled paper-based material—such as irregular fine fiber textures, complex diffuse reflection, and low contrast of printed patterns—and defects. By performing brightness normalization processing on the real-time acquired label image and extracting texture features from its local areas, it effectively eliminates the problem of local brightness differences and contrast reduction caused by uneven light decay of the LED light source array, ensuring the accuracy of image feature extraction. By comparing the texture features of local areas of the label image to be detected with the normal texture spectrum, the texture difference degree is calculated, which quantifies the degree of texture deviation between the detected label and a normal label. Furthermore, by performing defect pattern matching based on the texture difference degree and a preset defect pattern, the specific defect type can be identified. Finally, based on texture difference, defect pattern matching results, and a texture difference threshold adjusted according to production environment parameters, it is determined whether a local area of ​​the label image to be detected contains defects. This dynamically adjusted threshold can adapt to changes in the production environment, avoiding misjudging inherent material features as defects, while improving the ability to identify real defects. This application can effectively distinguish between inherent material features and real defects, significantly reducing false detection and false negative rates. It overcomes the problem of decreased detection accuracy caused by uneven light source and complex material properties when processing new recycled paper-based labels in existing technologies, thereby improving the accuracy and efficiency of product quality control. Attached Figure Description

[0059] To illustrate this application more clearly, the accompanying drawings used in the embodiments will be briefly described below. Obviously, those skilled in the art can obtain other drawings based on these drawings without any creative effort.

[0060] Figure 1 The diagram above illustrates a flowchart of a beverage label defect detection method based on industrial vision.

[0061] Figure 2 The diagram above illustrates a schematic of a beverage label defect detection system based on industrial vision.

[0062] Figure reference numerals: 100, Beverage label defect detection system based on industrial vision; 10, Texture feature extraction module; 20, Local texture feature extraction module; 30, Texture difference calculation module; 40, Defect pattern matching module; 50, Defect judgment module. Detailed Implementation

[0063] The technical solutions of this application will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments. The components of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0064] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0065] During industrial production line operation, long-term operation leads to non-uniform light decay of LED light sources, resulting in uneven illumination on the label surface and the formation of bright and dark areas. This causes a decrease in overall image contrast, with varying contrast levels across different areas. Existing exposure parameters are insufficient to cover all areas: shortening the exposure results in loss of detail in dark areas, while extending the exposure leads to overexposure in bright areas. This disrupts the defect classification logic, reducing the accuracy of identifying defects such as minor wrinkles, bubbles, and blurry light-colored printing. Simultaneously, the surface of new recycled paper-based materials has fine fibrous textures, easily causing complex diffuse reflection, resulting in a soft and blurry image at a microscopic level. Its porous structure also leads to uneven ink absorption, reduced printing edge sharpness, and the formation of inherently low-contrast areas. When uneven light source is combined with the characteristics of the new material, in dark areas, fibrous textures are easily misjudged as wrinkles or scratches, increasing the false detection rate; in bright areas, already low-contrast printing blur defects are more easily masked, leading to a higher missed detection rate. The system struggles to distinguish material characteristics from actual defects, severely impacting quality control effectiveness.

[0066] like Figure 1 The diagram illustrates an exemplary flowchart of a beverage label defect detection method based on industrial vision. This application proposes a beverage label defect detection method based on industrial vision, comprising:

[0067] S10, Extract texture features of local areas from qualified new recycled paper-based label images, and establish a normal texture spectrum describing the inherent texture of new recycled paper-based labels based on the extracted texture features;

[0068] Among them, the new type of recycled paper-based label refers to a label made primarily from recycled pulp. Its surface typically possesses a unique fiber texture and high porosity, exhibiting significant differences in optical properties and printing performance compared to traditional smooth labels. Texture features refer to the spatially repeating patterns of pixel grayscale or color in an image, which can be used to describe visual attributes such as surface roughness, directionality, and contrast. Texture features are used to characterize the inherent texture, diffuse reflection characteristics, and contrast of the printed pattern of the new type of recycled paper-based label. The normal texture spectrum is a reference model established by analyzing the texture features of a large number of qualified new type of recycled paper-based label images. It describes the range and distribution patterns of texture representation that a qualified label should exhibit in different local areas.

[0069] S20: After performing brightness normalization processing on the real-time acquired image of the label to be detected, the texture features of its local area are extracted.

[0070] S30, compare the texture features of a local region of the label image to be detected with the normal texture spectrum, and calculate the texture difference degree;

[0071] Among them, texture difference refers to the degree of difference between the texture features of a local area of ​​the label image to be detected and the normal texture spectrum, which is used to quantify the degree of deviation between the area to be detected and the qualified area.

[0072] S40, perform defect pattern matching based on the texture difference degree and the preset defect pattern to obtain the defect pattern matching result;

[0073] Among them, the defect mode refers to various predefined label defect types, such as wrinkles, bubbles, blurry printing, scratches, etc. Each defect mode corresponds to a specific texture abnormality.

[0074] S50, based on the texture difference degree, the defect pattern matching result, and the texture difference degree threshold adjusted according to the production environment parameters, determine whether there is a defect in a local area of ​​the label image to be detected.

[0075] The texture difference threshold is a critical value used to determine whether a defect exists. When the texture difference exceeds this threshold, a defect is considered to exist. This threshold can be dynamically adjusted based on production environment parameters.

[0076] First, in the initial stage of the inspection process, it is necessary to extract the texture features of local areas from qualified new recycled paper-based label images, and then establish a normal texture spectrum describing the inherent texture of the new recycled paper-based labels based on the extracted texture features. This step is the foundation for subsequent defect judgment. For example, a batch of qualified new recycled paper-based label images produced under standard lighting conditions can be pre-collected. These images can be divided into multiple small local regions, and then the texture features of each local region can be calculated. After obtaining the local texture features of a large number of qualified labels, statistical analysis can be performed on these features, such as calculating their mean, variance, and distribution range, thereby constructing a normal texture spectrum that can comprehensively describe the inherent texture of the new recycled paper-based labels. This normal texture spectrum can be a region in a multi-dimensional feature space or a set of statistical models for subsequent comparison.

[0077] Secondly, during normal production line operation, the real-time acquired images of the labels to be inspected need to undergo brightness normalization processing before extracting texture features from their local areas. Since uneven lighting may exist in the actual production environment, direct extraction of texture features can be interfered with. Therefore, brightness normalization is necessary before extracting texture features. Brightness normalization can be achieved in various ways. For example, histogram equalization can be used to broaden the image's grayscale histogram, making its distribution more uniform, thereby enhancing image contrast and reducing the impact of brightness differences. Another method is to use a local adaptive brightness adjustment algorithm, such as the Retinex algorithm. This algorithm can adjust based on the local brightness information of the image, thereby eliminating global and local brightness unevenness while preserving image details. After completing brightness normalization, the images of the labels to be inspected are again divided into local regions, and the texture features of these local regions are extracted using the same method as when establishing the normal texture spectrum, ensuring feature consistency and comparability.

[0078] Next, the texture features of a local region of the image to be detected are compared with the normal texture spectrum to calculate the texture difference. This step aims to quantify the degree of deviation in texture between the region to be detected and the qualified region. The comparison method can vary depending on how the normal texture spectrum is constructed. For example, if the normal texture spectrum is a region in a feature space, the distance between the texture feature vector of the region to be detected and the center point of the region can be calculated, or it can be determined whether the feature vector falls inside the region. Commonly used distance metrics include Euclidean distance and Mahalanobis distance. If the normal texture spectrum is a set of statistical models, the texture features of the region to be detected can be matched with these models to calculate their probability density or similarity, thereby obtaining the texture difference. The greater the texture difference, the greater the deviation between the texture of the region to be detected and the normal texture, and the higher the probability of defects.

[0079] Then, based on the texture difference degree and the preset defect patterns, defect pattern matching is performed to obtain the defect pattern matching result. In actual production, label defects may manifest in various types, such as wrinkles, bubbles, printing blur, scratches, etc. Each defect pattern may exhibit specific anomalies in texture features. For example, wrinkles may cause drastic changes in texture directionality, bubbles may cause abnormal local texture smoothness, and printing blur may cause a decrease in texture contrast. Therefore, it is necessary to predefine these defect patterns and establish their corresponding texture anomaly features. Defect pattern matching can be implemented using various classifiers, such as support vector machines (SVM), neural networks, decision trees, etc. These classifiers learn the mapping relationship between texture difference degree and different defect patterns during the training phase. During the detection phase, the calculated texture difference degree is input into the classifier, which outputs a defect pattern matching result, indicating which defect pattern the area to be detected is most likely to belong to, or giving the probability of belonging to various defect patterns.

[0080] Finally, based on the texture difference degree, the defect pattern matching result, and the texture difference degree threshold adjusted according to production environment parameters, it is determined whether a local area of ​​the label image to be detected has a defect. This step is the final decision-making stage. Multiple factors need to be considered when determining defects. First, texture difference degree is a direct indicator of the degree of anomaly. Second, the defect pattern matching result provides information on the type of anomaly. Furthermore, considering that production environment parameters (such as temperature, humidity, production speed, etc.) may have a slight impact on the texture performance of the label, the texture difference degree threshold needs to be dynamically adjusted. For example, in a high-humidity environment, paper-based labels may swell slightly, causing subtle changes in texture features; in this case, the texture difference degree threshold can be appropriately relaxed. Threshold adjustment can be achieved through a machine learning model that learns the optimal threshold based on historical production data and environmental parameters. Ultimately, only when the texture difference degree exceeds the adjusted threshold and the defect pattern matching result indicates the presence of a certain defect is it determined that a local area of ​​the label image to be detected has a defect.

[0081] Specifically, this method first extracts texture features from qualified new recycled paper-based label images and establishes a normal texture spectrum, providing a reliable benchmark for subsequent defect determination. This step fully considers the inherent complex texture characteristics of new recycled paper-based labels, avoiding misjudging their inherent textures as defects. Subsequently, the brightness normalization process is performed on the real-time acquired label images to be tested, effectively eliminating the image brightness differences and contrast reduction problems caused by uneven light decay of the LED light source array, ensuring the accuracy of subsequent texture feature extraction. After brightness normalization, the texture features of local areas of the label images to be tested are extracted and compared with the normal texture spectrum to calculate the texture difference degree, thereby quantifying the degree of deviation between the tested area and the qualified area.

[0082] Furthermore, this application introduces a defect pattern matching mechanism to identify specific defect types based on the texture difference and preset defect patterns. This mechanism not only detects anomalies but also classifies them, providing more detailed information for subsequent quality control. For example, when a high texture difference is detected, defect pattern matching can distinguish whether the anomaly is caused by micro-wrinkles, micro-bubbles, or printing blur.

[0083] Finally, this application comprehensively considers the texture difference degree, the defect pattern matching result, and the texture difference degree threshold adjusted according to production environment parameters to make the final defect determination. This multi-factor, adaptive determination strategy significantly improves the accuracy and robustness of defect detection. For example, in dark areas caused by uneven light sources, even if the details of the fiber texture of the new material become blurred under low light and complex diffuse reflection, the inherent characteristics of the material and the establishment of a normal texture spectrum can still be accurately distinguished due to brightness normalization processing and the establishment of a normal texture spectrum, thereby effectively reducing the false detection rate. In bright areas, overexposure is avoided through brightness normalization processing, allowing printing blurry defects that are inherently low in contrast due to uneven ink absorption to be accurately identified, thereby significantly reducing the missed detection rate.

[0084] Compared with existing technologies, the advantages of this application are as follows: Existing technologies, when faced with the superposition of uneven light decay in LED light source arrays and the complex characteristics of novel recycled paper-based label materials, suffer from chaotic judgment criteria for defect classification, making them almost impossible to function properly. They fail to effectively distinguish between inherent material characteristics and actual defects, leading to a sharp increase in false detection and false negative rates. This application, however, effectively solves the image quality degradation problem caused by uneven light source by introducing brightness normalization processing; it solves the problem of confusion between inherent textures and defects in novel materials by establishing a normal texture spectrum describing the inherent texture of novel recycled paper-based labels; and it further improves the accuracy and robustness of defect judgment through defect pattern matching and adaptive threshold adjustment. Therefore, this application can significantly improve the accuracy and efficiency of defect detection for novel recycled paper-based labels in complex production environments, providing more reliable technical support for quality control in beverage packaging production lines.

[0085] In some embodiments, the step of establishing a normal texture spectrum describing the inherent texture performance of novel recycled paper-based labels based on the extracted texture features includes:

[0086] When the system is deployed for the first time or a new batch of materials is launched, an initial label image is acquired, and texture features of local areas are extracted from the initial label image;

[0087] The extracted texture features are clustered to obtain an initial normal texture reference;

[0088] During normal production line operation, images of the labels to be inspected are acquired in real time, and texture features of local areas are extracted from the images of the labels to be inspected.

[0089] Calculate the texture difference degree between the texture features of a local region of the label image to be detected and the initial normal texture reference, and determine the qualified region based on the texture difference degree;

[0090] Based on the areas deemed qualified, assess the confidence level of the qualified areas;

[0091] The initial normal texture reference is periodically updated based on qualified regions with confidence levels higher than a preset confidence threshold to obtain an updated normal texture spectrum;

[0092] Receive defect area information from manual feedback, and adjust the updated normal texture spectrum based on the defect area information.

[0093] Specifically, during initial deployment or when a new batch of materials is brought online, a batch of initial label images needs to be acquired. These images are typically considered qualified samples. Subsequently, texture features are extracted from local regions of these initial label images to capture their inherent texture information. To better characterize these initial texture features, clustering can be performed, for example using K-means or DBSCAN clustering algorithms, grouping similar texture features into one class to obtain an initial normal texture reference. This reference can be understood as a typical representative of the qualified label textures in the current batch or initial state.

[0094] During normal production line operation, images of the labels to be inspected are acquired in real time, and texture features are extracted from local areas. These real-time extracted texture features are compared with previously established initial normal texture references to calculate the texture difference. Based on this texture difference, it is possible to initially determine which areas are qualified. To improve the accuracy of the determination, the confidence level of areas determined to be qualified is further evaluated. The confidence level can be calculated based on factors such as the magnitude of the texture difference and the stability of the area, reflecting the probability that the area is truly qualified.

[0095] To enable the normal texture spectrum to adapt to dynamic changes during the production process, the technical solution of this application is designed to periodically update the normal texture spectrum. Specifically, the initial normal texture reference is updated based on the texture features of qualified regions with confidence levels higher than a preset confidence threshold, resulting in an updated normal texture spectrum. This update mechanism allows the normal texture spectrum to continuously learn and adapt to subtle changes on the production line. Furthermore, to further improve accuracy and robustness, this application also considers a human feedback mechanism. When a human verifies or corrects certain defective regions, the system receives this feedback on defective regions and adjusts the updated normal texture spectrum accordingly to avoid misclassifying known defect patterns as normal textures or vice versa.

[0096] The technical solution of this application addresses the insufficient robustness of traditional static normal texture spectra in the face of changes in the production environment by introducing dynamic updates and human feedback mechanisms. First, an initial normal texture reference is established upon initial deployment or the arrival of a new batch of materials, providing a baseline for subsequent real-time detection. Second, during normal production line operation, the texture features of the labels to be inspected are continuously acquired and compared with the initial normal texture reference, enabling real-time monitoring of label texture deviations. More importantly, by assessing the confidence level of qualified regions and periodically updating the normal texture spectrum using high-confidence qualified regions, the spectrum can adaptively learn and reflect subtle changes in materials or the environment during production, avoiding misjudgments caused by "normal" texture drift. Furthermore, the introduction of a human feedback mechanism allows for precise adjustments to the normal texture spectrum by incorporating the experience and knowledge of human experts, especially when dealing with complex or boundary-based defects, further improving the accuracy and reliability of detection.

[0097] Through the above technical solution, this application can establish a dynamically adaptive normal texture spectrum, significantly improving the accuracy and robustness of the industrial vision-based beverage label defect detection method. This technical solution can effectively handle subtle changes in material batches, ink characteristics, or environmental parameters during production, avoiding false alarms or missed detections caused by normal texture drift. Furthermore, combined with a human feedback mechanism, it enables continuous optimization of the understanding of "normal" textures, reducing reliance on manual intervention, improving the efficiency and reliability of automated detection, thereby reducing production costs and improving product quality.

[0098] In some embodiments, the step of comparing the texture features of a local region of the label image to be detected with the normal texture spectrum to calculate the texture difference includes:

[0099] The texture features of a local region of the image to be detected are compared with the normal texture spectrum to mark potential texture anomalies;

[0100] Track the potential texture anomalies in consecutive frames of an image sequence and establish anomaly trajectories;

[0101] Based on the abnormal trajectory, determine whether it is a persistent defect. If it is determined to be a persistent defect, calculate the texture difference degree between the texture features of the local region and the normal texture spectrum.

[0102] Specifically, when comparing the texture features of a local region of the image to be detected with the normal texture spectrum, the process begins by comparing the texture features of the local region of the image to be detected with the normal texture spectrum to identify and mark all regions that show significant differences from the normal texture spectrum. These regions are preliminarily identified as potential texture anomalies. Various statistical methods can be used for this comparison, such as Euclidean distance, Mahalanobis distance, or similarity measures, to quantify the degree of deviation in texture features.

[0103] Furthermore, to distinguish between transient anomalies and genuine defects, it is necessary to track the potential texture anomalies in consecutive frames of the image sequence, thereby establishing anomaly trajectories. This means recording the changes in position, size, shape, and other information of each potential texture anomaly over time, forming trajectory data that evolves over time. Tracking algorithms can employ Kalman filtering, particle filtering, or feature-matching-based tracking methods.

[0104] The process involves determining whether an abnormal trajectory constitutes a persistent defect. Persistent defects typically manifest as anomalies that are stable across multiple consecutive frames with minimal changes in position and shape, while transient anomalies may only appear in a single frame or a few frames. The determination logic can be based on parameters such as the length, duration, and spatial stability of the abnormal trajectory. For example, if a potential texture anomaly is tracked in a preset series of N consecutive frames, it can be determined to be a persistent defect.

[0105] If the defect is determined to be persistent, the texture difference between the texture features of the local region and the normal texture spectrum is calculated. This means that only regions confirmed as persistent defects will have their texture difference calculated and used in subsequent defect determination processes, thereby avoiding misjudgments of transient anomalies.

[0106] The technical solution of this application effectively solves the false alarm problem caused by transient interference in traditional methods by introducing a mechanism for tracking and determining the persistence of potential texture anomalies. Specifically, when a texture anomaly appears in a local area of ​​the image to be detected, its texture difference degree is not immediately calculated and a defect is determined. Instead, it is first marked as a potential texture anomaly. Subsequently, by analyzing consecutive frames in the image sequence, the dynamic changes of these potential texture anomalies are tracked, thereby establishing an anomaly trajectory. It is precisely because of the analysis of the anomaly trajectory that it is possible to distinguish between transient anomalies that appear briefly in time and are unstable in space (such as dust and reflections) and real defects that exist stably in consecutive frames and have a certain persistence (such as scratches and stains). Only when the anomaly is confirmed as a persistent defect is the texture difference degree accurately calculated, thereby ensuring the accuracy and reliability of subsequent defect determination.

[0107] Through the above technical solution, this application can significantly improve the accuracy and robustness of beverage label defect detection. By introducing the tracking and continuous determination of potential texture anomalies, it can effectively filter out common transient interferences in the production line, such as local texture fluctuations caused by brief dust, water droplets, or changes in lighting, avoiding misjudging these non-defect factors as defects. This not only reduces the false alarm rate, minimizes unnecessary downtime for inspection and manual re-inspection, and improves production efficiency, but also makes the defect determination results more stable and reliable, ensuring the consistency of product quality.

[0108] In some optional embodiments, it is assumed that on a beverage label production line, an industrial camera continuously acquires images of the labels to be inspected at a rate of 30 frames per second. When a texture anomalous point is detected in a local area of ​​a frame, the point is marked as a potential texture anomalous point. The spatial location of the anomalous point in subsequent frames is continuously monitored. If the position and shape of the anomalous point remain within a preset threshold range in the next 10 frames (approximately 0.33 seconds), the anomalous point is determined to be a persistent defect, such as a scratch or an ink stain. Only then is the texture feature of the area precisely compared with the normal texture spectrum, the texture difference is calculated, and further defect pattern matching is performed. Conversely, if the anomalous point appears only in 1-2 frames and then disappears or its position changes drastically, it is determined to be a transient anomalous point (e.g., a rapidly passing dust particle), and the anomalous point is ignored, and no texture difference calculation is performed, thus avoiding false alarms. This time-series-based analysis method enables intelligent differentiation between real defects and transient interference, greatly improving the reliability of detection.

[0109] In some embodiments, the step of extracting texture features of local regions from a qualified novel recycled paper-based label image and establishing a normal texture spectrum describing the inherent texture of the novel recycled paper-based label based on the extracted texture features includes:

[0110] Configure a multispectral industrial camera to acquire qualified multispectral images of new recycled paper-based labels;

[0111] The multispectral image is divided into multiple local regions, and multispectral texture features are extracted from the local regions. The multispectral texture features characterize the inherent texture, light diffuse reflection characteristics, and contrast of the printed pattern in different spectral channels of the novel recycled paper-based label.

[0112] By analyzing the correlation between the multispectral texture features in different spectral channels, texture features characterizing the inherent texture, light diffuse reflection characteristics, and printing pattern contrast of the novel recycled paper-based label are obtained, and a normal texture spectrum describing the inherent texture performance of the novel recycled paper-based label is established.

[0113] Specifically, an imaging device capable of simultaneously or sequentially capturing multiple narrowband spectral images is employed. For example, the camera can be configured with multiple filters or employ tunable filter technology, enabling it to acquire near-infrared or ultraviolet spectral images beyond the visible light range (such as red, green, and blue). The aim is to obtain richer spectral information than traditional RGB images, thereby more comprehensively reflecting the physical and chemical properties of the new recycled paper-based labels. Specifically, during the initial deployment or when new batches of materials are introduced, multispectral imaging is performed on labels known to be qualified to acquire multispectral images. Through multispectral imaging, a series of images captured at different wavelengths or bands can be obtained, which together constitute the multispectral data cube of the label.

[0114] In practical applications, the image of each spectral channel is segmented into several smaller, overlapping or non-overlapping regions, and its texture features are calculated within each local region. These multispectral texture features can be understood as a set of texture descriptors extracted from multiple spectral channels. For example, they may include gray-level co-occurrence matrix statistics (such as energy, contrast, homogeneity, entropy, etc.) or Gabor filter bank response values ​​calculated for each spectral channel. These features not only reflect texture information under a single spectrum, but more importantly, they collectively characterize the inherent texture, diffuse reflection characteristics, and contrast of the printed pattern under different spectral channels of the novel recycled paper-based label. Inherent texture refers to the fiber structure and roughness of the paper-based material itself; diffuse reflection characteristics refer to the scattering and absorption behavior of the label surface at different wavelengths of light, which is closely related to the material composition and surface treatment; the contrast of the printed pattern under different spectral channels reflects the absorption spectral characteristics of the ink and its interaction with the paper-based material.

[0115] Furthermore, mathematical or statistical methods are used to explore the interrelationships between texture features extracted from different spectral channels, enabling the analysis of the correlation between these multispectral texture features across different spectral channels. For example, cross-correlation coefficients, covariance matrices, or principal component analysis can be performed on texture features from different spectral channels. The aim is to identify key texture feature combinations that can stably and robustly characterize the inherent texture, diffuse reflection properties, and printed pattern contrast of novel recycled paper-based labels. Through this correlation analysis, noise or redundant information that may exist in a single spectral channel can be filtered out, resulting in more representative and discriminative texture features. Based on these features, a normal texture spectrum describing the inherent texture performance of novel recycled paper-based labels can be established. This normal texture spectrum will serve as a benchmark for subsequent defect detection.

[0116] The technical solution of this application, by introducing a multispectral industrial camera, can acquire richer and more comprehensive label information than traditional single-spectral images. The material composition, ink properties, and potential defects of novel recycled paper-based labels exhibit unique spectral response characteristics in different spectral bands. For example, some material defects or contaminants may not be obvious under visible light, but will show significant differences in absorption or reflection in the near-infrared or ultraviolet bands. By extracting multispectral texture features and further analyzing the correlation between these features in different spectral channels, a more refined and comprehensive normal texture spectrum can be constructed. This method can effectively distinguish between normal texture fluctuations caused by inherent material properties (such as uneven paper fiber density) and abnormal texture manifestations caused by actual defects (such as printing blur, uneven ink, or foreign matter contamination). It is precisely because multispectral data provides deeper physical and chemical information that the true state of the label can be captured more accurately, thereby improving the sensitivity and specificity of defect detection.

[0117] Through the above technical solution, this application overcomes the limitations of traditional single-spectral images in characterizing the complex characteristics of novel recycled paper-based labels. The introduction of multispectral texture features enables the capture of the label's inherent texture, diffuse reflection characteristics, and printed pattern contrast from a broader spectral dimension, thereby establishing a more accurate and robust normal texture spectrum. This method significantly improves the accuracy and reliability of defect detection, especially in identifying and distinguishing subtle defects that are difficult to detect under a single spectrum, and in differentiating inherent material texture inhomogeneities from actual production defects. Furthermore, by analyzing the correlation between different spectral channels, environmental noise and redundant information can be filtered out more effectively, making defect judgment more accurate, thereby reducing false alarm and false negative rates and improving performance.

[0118] In some optional embodiments, this application is implemented as follows: Assume that on a beverage label production line, it is necessary to detect defects such as blurry printing, uneven ink distribution, or uneven material fiber distribution on novel recycled paper-based labels. First, a multispectral industrial camera is configured, capable of capturing multiple narrowband spectral images, including visible light (e.g., 450nm, 550nm, 650nm) and near-infrared (e.g., 850nm, 950nm). In the early stages of production, a batch of known qualified labels is subjected to multispectral imaging to obtain their multispectral images. Subsequently, these multispectral images are divided into 10x10 pixel local regions. Within each local region, for each spectral channel, the energy, contrast, homogeneity, and other statistical quantities of its gray-level co-occurrence matrix (GLCM) are calculated to obtain the multispectral texture features of that local region.

[0119] Furthermore, these multispectral texture features are analyzed. For example, cross-correlation coefficients of GLCM features between different spectral channels (e.g., 550nm and 850nm) can be calculated. For qualified labels, these cross-correlation coefficients exhibit stable patterns, reflecting the normal correlation between the paper base material, ink, and diffuse light reflection. By statistically analyzing these correlation data from a large number of qualified labels, a normal texture spectrum describing the inherent texture performance of novel recycled paper base labels is established.

[0120] When the image of the label to be detected is acquired in real time, multispectral imaging and multispectral texture feature extraction of local areas are also performed. Then, the multispectral texture features of the local area of ​​the label to be detected are compared with the established normal texture spectrum. For example, if the texture feature of a certain local area in the 850nm channel deviates significantly from the normal texture spectrum, and its cross-correlation with the texture feature in the 550nm channel is significantly lower than the normal level, this may indicate the presence of ink penetration or abnormal internal material structure. Through this multispectral correlation analysis, defects that are difficult to detect by traditional single-spectral methods can be identified more accurately. For example, it can distinguish between normal texture fluctuations caused by slight unevenness in paper fiber density and printing blurring defects caused by ink diffusion during the printing process.

[0121] In some embodiments, the step of analyzing the correlation between the multispectral texture features in different spectral channels to obtain texture features characterizing the inherent texture, light diffuse reflection characteristics, and printed pattern contrast of the novel recycled paper-based label, and establishing a normal texture spectrum describing the inherent texture performance of the novel recycled paper-based label includes:

[0122] Statistical analysis was performed on the correlation between the multispectral texture features in different spectral channels to establish an initial multispectral texture correlation reference, which characterizes the normal spectral response characteristics of the current batch of materials and inks;

[0123] During normal production line operation, multispectral images of the labels to be inspected are acquired in real time, the multispectral images of the labels to be inspected are divided into multiple local regions, and multispectral texture features are extracted from each local region.

[0124] The correlation difference degree between the extracted multispectral texture features and the initial multispectral texture association reference is calculated to obtain the correlation difference degree;

[0125] Based on the correlation difference degree and the preset difference degree threshold range, determine whether it is a qualified region and assess the confidence level of the qualified region.

[0126] The initial multispectral texture association reference is periodically updated based on the multispectral texture features of qualified regions with confidence levels higher than a preset confidence threshold, to obtain the updated normal texture spectrum;

[0127] Receive defect area information from manual feedback, and adjust and update the normal texture spectrum based on the defect area information.

[0128] Specifically, upon initial deployment or the arrival of a new batch of materials, multispectral images of qualified new recycled paper-based labels are first acquired, and multispectral texture features are extracted from their local areas. Subsequently, statistical analysis is performed on the correlation between these multispectral texture features across different spectral channels. For example, the covariance matrix, correlation coefficient, or mutual information of texture features across different spectral channels can be calculated, thereby establishing an initial multispectral texture correlation reference. This reference is used to characterize the spectral response characteristics of the current batch of materials and inks under normal conditions, providing a benchmark for subsequent defect detection.

[0129] During normal production line operation, multispectral images of the labels to be inspected are acquired in real time. These images are divided into multiple local regions, and multispectral texture features are extracted from each local region. The extracted multispectral texture features are then compared with the initial multispectral texture association reference established above, and the association difference degree is calculated. This association difference measure quantifies the degree of deviation of the spectral response characteristics between the area to be inspected and the normal reference.

[0130] Furthermore, based on the calculated correlation difference degree and the preset difference degree threshold range, it is determined whether the current local area is a qualified area. For example, if the correlation difference degree is lower than the preset threshold, the area is considered qualified. At the same time, a confidence assessment is performed on the qualified areas, and the confidence degree can be calculated based on the magnitude of the difference degree, the stability of the area, or historical data.

[0131] As an optional implementation, the initial multispectral texture association reference is periodically updated based on the multispectral texture features of qualified regions with confidence levels higher than a preset confidence threshold, thereby obtaining an updated normal texture spectrum. This periodic update mechanism allows the normal texture spectrum to dynamically adapt to minor changes in materials or the environment during the production process, maintaining its accuracy and representativeness.

[0132] Furthermore, to further improve robustness and accuracy, this application also accepts defect region information provided by human feedback. When human inspection detects misjudgments or omissions, the defect region information can be fed back. Based on this defect region information, the updated normal texture spectrum is adjusted, for example through weight adjustment, outlier sample exclusion, or model retraining, to correct and optimize the normal texture spectrum so that it better reflects the normal state and defect patterns in actual production.

[0133] The technical solution of this application effectively solves the problem of insufficient adaptability of traditional static normal texture spectra to changes in the production environment by introducing dynamic updating and manual feedback mechanisms. Specifically, the establishment of an initial multispectral texture association reference provides a benchmark for detection, while real-time acquisition of the label image to be detected and extraction of multispectral texture features ensures the immediacy of detection. By calculating the correlation difference, the degree of deviation between the detected area and the normal reference can be quantified. Among them, the determination of qualified areas and the confidence assessment are the key to achieving dynamic updating. Only qualified areas with high confidence are used to update the normal texture spectrum, which avoids the erroneous inclusion of potentially abnormal areas in the normal reference, thereby ensuring the accuracy of the update. Periodically updating the initial multispectral texture association reference based on the multispectral texture features of these high-confidence qualified areas allows the normal texture spectrum to continuously learn and adapt to subtle changes in material, ink, or environmental parameters during the production process, such as slight color differences between material batches or fine adjustments to printing pressure. Furthermore, receiving defect area information from manual feedback and adjusting the normal texture spectrum accordingly is an important part of improving the robustness of this technical solution. Human feedback can correct misjudgments in complex or novel defect patterns, excluding actual defect patterns from the normal texture spectrum, thus avoiding misidentifying certain defects as normal variations. This human-machine collaborative approach enables the normal texture spectrum to more accurately represent "normal" states and effectively distinguish "defect" states, thereby significantly improving the accuracy and reliability of defect detection.

[0134] Through the above technical solutions, this application achieves significant optimization of the defect detection method for novel recycled paper-based labels. The dynamic update mechanism allows the normal texture spectrum to adaptively adjust to cope with gradual changes in materials, inks, or environmental factors during production, thereby effectively reducing the false alarm rate and false negative rate caused by reference drift. Simultaneously, the introduction of a human feedback mechanism allows the method to learn from actual production experience and correct its own judgments, further improving the ability to identify complex or novel defect patterns and the detection accuracy. Therefore, the defect detection method of this application exhibits higher robustness, adaptability, and reliability under long-term operation and multi-batch production scenarios, ensuring product quality stability and improved production efficiency.

[0135] In some optional embodiments, assuming this detection system is deployed for the first time on a beverage production line, multispectral images of a batch of qualified new recycled paper-based labels are initially acquired. Multispectral texture features are extracted and statistically analyzed from local areas of these images to establish an initial multispectral texture correlation reference. After several weeks of normal operation on the production line, a change in the production batch of new recycled paper-based labels by the supplier causes slight but widespread changes in certain spectral response characteristics of the label material. Relying solely on the initial static reference might misjudge these widespread, non-defective changes as defects, leading to numerous false alarms. However, in the technical solution of this application, multispectral images of the labels to be detected are acquired in real time, and the correlation difference between them and the initial multispectral texture correlation reference is calculated. Since this change is widespread, a large number of areas are identified as qualified areas with high confidence. Periodically, the initial multispectral texture correlation reference is updated based on the multispectral texture features of these high-confidence qualified areas, gradually adapting it to the spectral response characteristics of the new batch of materials. For example, every 24 hours, the normal texture spectrum is updated using a weighted average of the average multispectral texture features of all high-confidence qualified areas over the past 24 hours. Furthermore, suppose that at some point, a new, previously unseen blurry defect appears on the label due to printhead clogging. Initially, this new defect might not be accurately identified, or it might be misjudged as a normal texture variation. In this case, the production line operator manually inspects the defect and reports the defect area. Upon receiving this feedback, the updated normal texture spectrum is adjusted based on this defect area information. For example, the features of this defect pattern might be excluded from the normal texture spectrum distribution, or it might be marked as a new defect pattern. In this way, the system can quickly learn and adapt to new defect patterns, avoiding subsequent missed detections and continuously optimizing detection performance.

[0136] In some embodiments, the step of analyzing the correlation between the multispectral texture features across different spectral channels includes:

[0137] The spectral response curves of the multispectral texture features in the local area are fitted to obtain the spectral response characteristic curves;

[0138] Calculate the rate of change of curvature of the spectral response characteristic curve, and compare the rate of change of curvature with a preset material density non-uniformity characteristic threshold to obtain the comparison result of material density non-uniformity.

[0139] The cross-correlation of the multispectral texture features across different spectral channels is calculated to obtain the cross-correlation results.

[0140] By combining the fitting residuals of the spectral response characteristic curves, correlation anomalies that match the printing blur defect pattern are identified, and correlation anomaly identification results are obtained.

[0141] Based on the comparison results of the material density non-uniformity, the cross-correlation results, and the correlation anomaly identification results, it is determined whether the texture anomaly in the local area originates from the inherent non-uniformity of the material density or a printing blur defect.

[0142] This process involves using mathematical models (such as polynomial fitting and spline fitting) to describe the intensity or distribution trends of multispectral texture features in local areas under different spectral channels, thereby achieving spectral response curve fitting for the multispectral texture features of local areas. This yields a continuous spectral response characteristic curve, which intuitively reflects the reflection or absorption characteristics of the material or printed pattern under different wavelengths of light. Furthermore, the rate of curvature change of the spectral response characteristic curve can be understood as the rate at which the curvature of the curve changes with the spectral channel. Non-uniform material density often leads to local changes in the scattering or absorption characteristics of light at different wavelengths, which is reflected in the local shape of the spectral response characteristic curve as a significant change in curvature. By comparing the calculated rate of curvature change with a preset threshold for non-uniform material density, it is possible to preliminarily determine whether there are texture anomalies in local areas caused by non-uniform material density.

[0143] In practical applications, the cross-correlation of multispectral texture features across different spectral channels refers to an indicator measuring the similarity or synchronicity between texture features in different spectral channels. For example, this correlation can be quantified by calculating the Pearson correlation coefficient or mutual information. Normal label areas typically exhibit high cross-correlation between texture features in different spectral channels; however, this correlation may change when defects are present. Furthermore, the fitting residual of the spectral response curve refers to the difference between the actual measured multispectral texture feature values ​​and the predicted values ​​of the fitted curve. A large fitting residual may indicate that the spectral response characteristics of the area deviate from the normal pattern, which may be related to printing blurring defects in some cases. By analyzing these residuals and matching them with preset printing blurring defect patterns, correlation anomalies related to printing blurring can be identified. Finally, by comprehensively analyzing the comparison results of the above material density inhomogeneity, cross-correlation results, and correlation anomaly identification results, it is possible to more accurately determine whether the texture anomalies in local areas are due to the inherent density inhomogeneity of the new recycled paper-based label material itself or to blurring defects occurring during the printing process.

[0144] The technical solution of this application effectively captures the subtle differences in spectral response characteristics caused by the inherent density inhomogeneity of materials by introducing spectral response curve fitting and curvature change rate analysis of multispectral texture features. Simultaneously, by calculating the cross-correlation between different spectral channels, the overall consistency of texture features can be evaluated. Furthermore, the fitting residuals of the spectral response characteristic curves are used to identify printing blur defects because printing blur usually causes unexpected and irregular changes in the spectral reflectance or absorption characteristics of local areas, resulting in a large residual between the fitted curve and the actual data. By integrating these independent analytical results, this application can cross-validate and finely classify texture anomalies from multiple dimensions, thereby avoiding misjudgments that may arise from single feature analysis and achieving accurate differentiation of different defect sources.

[0145] Through the above technical solution, this application can significantly improve the accuracy and robustness of beverage label defect detection. Specifically, by conducting multi-dimensional and refined analysis of multispectral texture features, especially distinguishing between inherent material density inhomogeneity and printing blur defects, false alarms and missed alarms can be effectively reduced. This not only helps improve product quality control and reduce production costs and resource waste caused by defect misjudgment, but also provides a more accurate basis for tracing the source of problems in the production process, thereby achieving more efficient and intelligent defect management.

[0146] In some optional embodiments, it is assumed that a local area is identified as having a texture anomaly during the detection process. First, the multispectral image data of this local area is fitted with a spectral response curve to obtain its spectral response characteristic curve. If the curve exhibits an abnormal peak in the rate of curvature change within a specific wavelength range, and this peak exceeds a preset threshold for material density inhomogeneity, a preliminary judgment is made that a material density inhomogeneity problem may exist. Simultaneously, the cross-correlation between different spectral channels of this local area is calculated. If the cross-correlation decreases significantly, and the fitting residual of the spectral response characteristic curve is large, and these residual patterns highly match known printing blur defect patterns, it is likely to be judged as a printing blur defect. By combining this information, for example, if the rate of curvature change is abnormally significant but the cross-correlation is normal and the fitting residual is small, it is judged as inherent material density inhomogeneity; conversely, if the cross-correlation decreases significantly and the fitting residual is large and matches a blur pattern, it is judged as a printing blur defect. This multi-feature fusion judgment mechanism enables accurate classification of defect types, for example, distinguishing between uneven fiber distribution in the paper itself and blurring caused by ink diffusion during the printing process.

[0147] In some embodiments, the step of calculating the rate of change of curvature of the spectral response characteristic curve and comparing the rate of change of curvature with a preset material density non-uniformity characteristic threshold to obtain a comparison result of material density non-uniformity includes:

[0148] Identify local extrema on the spectral response characteristic curve;

[0149] Calculate the peak intensity and peak width of the rate of curvature change of the spectral response characteristic curve at each local extremum point;

[0150] Statistical analysis was performed on all the peak intensities and peak widths to obtain the distribution characteristics of the rate of change of curvature;

[0151] Based on the distribution characteristics, the threshold for preset material density non-uniformity is dynamically adjusted;

[0152] The rate of curvature change of each local region is compared with the adjusted preset threshold for material density non-uniformity to obtain the comparison result of material density non-uniformity.

[0153] Specifically, identifying local extrema on a spectral response curve involves mathematically analyzing the curve, such as calculating the first derivative and finding its zero point, or using local search algorithms, to determine the points where the slope of the curve changes direction. These points typically correspond to significant changes in the absorption, reflection, or transmission properties of the material within a specific spectral band, with the aim of accurately capturing the location of characteristic changes in the material's spectral response.

[0154] Calculating the peak intensity and peak width of the rate of curvature change of the spectral response curve at each local extremum point can be understood as a quantitative description of the spectral characteristics represented by these extremum points. The peak intensity reflects the severity of the spectral response change at the extremum point, while the peak width indicates the persistence or range of influence of this change within the spectral wavelength range. The purpose is to provide refined quantitative characteristic parameters for subsequent statistical analysis.

[0155] In practical applications, statistical analysis is performed on all the peak intensities and peak widths to obtain the distribution characteristics of the rate of change of curvature. For example, statistical measures such as the mean, variance, skewness, and kurtosis of these characteristics can be calculated, or histograms and probability density functions can be constructed to comprehensively describe the overall performance of the material density non-uniformity characteristics in the current batch of materials or under the current production state. The purpose is to establish a benchmark based on actual production data to assess the degree of anomaly of subsequent labels to be detected.

[0156] Furthermore, based on the aforementioned distribution characteristics, the threshold for preset material density non-uniformity is dynamically adjusted. For example, the threshold can be set based on a certain percentile of the statistical distribution (such as 95% or 99%), or an adaptive algorithm can be used to iteratively adjust it based on historical data and real-time statistical results. The purpose is to enable the threshold to better adapt to the inherent characteristics and fluctuations of materials in actual production, avoiding misjudgments caused by using a fixed threshold.

[0157] Therefore, by comparing the rate of curvature change of each local region with the adjusted preset material density non-uniformity characteristic threshold, the comparison result of material density non-uniformity is obtained, thereby more accurately determining whether there is a defect caused by material density non-uniformity in the local region.

[0158] Through the above technical solution, this application achieves more accurate and robust detection of defects related to uneven material density. Compared to the basic technical solution that uses a fixed threshold for comparison, this application significantly reduces the false alarm rate and false negative rate caused by batch differences in materials or fluctuations in the production environment by dynamically adjusting the threshold. This adaptive threshold adjustment mechanism better adapts to the complexity and diversity of actual production, improves the accuracy and reliability of defect judgment, and thus effectively enhances the quality control level of beverage labels.

[0159] In some optional embodiments, it is assumed that upon initial deployment or the arrival of a new batch of materials, a large number of qualified new recycled paper-based labels are first acquired using multispectral images. For each local region of these images, a spectral response curve is fitted, and local extrema are identified. Subsequently, the peak intensity and peak width of the rate of curvature change of the spectral response characteristic curve at each extrema are calculated. For example, Gaussian fitting can be used to quantify the peak intensity and peak width. Next, statistical analysis is performed on the peak intensity and peak width data of all acquired qualified samples to construct their probability distribution model; for example, the mean and standard deviation of these features can be calculated, assuming they follow a normal distribution. During production, when the rate of curvature change of the spectral response characteristic curve of a local region of the label to be tested is detected, its peak intensity and peak width are compared with the previously established statistical distribution model. If the feature value of the local region deviates from the mean of the statistical distribution by more than a preset multiple of a certain standard deviation (e.g., 2 or 3 standard deviations), the threshold for material density inhomogeneity is dynamically adjusted to more accurately reflect the actual fluctuation range of the current batch of materials. For example, if the average peak strength of the current batch of material is slightly higher than the historical average, the threshold will be adjusted accordingly to avoid misjudging normal fluctuations as defects. Finally, the rate of curvature change in this local area is compared with the adjusted threshold to determine whether there are defects related to material density inhomogeneity.

[0160] In some embodiments, the step of calculating the cross-correlation of the multispectral texture features across different spectral channels to obtain the cross-correlation result includes:

[0161] Before calculating the cross-correlation, the filtering parameters are adjusted according to the scattering intensity and spatial frequency distribution of the multispectral texture features of each local region under different spectral channels to obtain the multispectral texture features after scattering noise removal.

[0162] Cross-correlation calculations were performed on the multispectral texture features after scattering noise removal across different spectral channels to obtain the cross-correlation results.

[0163] Specifically, before performing cross-correlation calculations, the original multispectral texture features need to be preprocessed. Scattering intensity refers to the degree to which light is scattered on the label surface, and its level reflects the strength of scattering noise; spatial frequency distribution characterizes the fineness of the image texture and the frequency characteristics of noise. This information can be obtained through frequency domain analysis methods such as Fourier transform or wavelet transform on the multispectral texture features of local regions. For example, high-frequency components are usually related to details and noise, while low-frequency components are related to the overall structure of the image.

[0164] Based on the analysis of scattering intensity and spatial frequency distribution, filtering parameters can be adjusted. The adjustment of filtering parameters aims to specifically suppress scattering noise while preserving as much of the tag's inherent texture information as possible. For example, when high scattering intensity and specific high-frequency noise are detected, an adaptive low-pass filter or Wiener filter can be used, dynamically adjusting its cutoff frequency or filter coefficients according to the statistical characteristics of the noise. Alternatively, spatial domain filters such as median filtering and Gaussian filtering can be used, or specific frequency ranges can be suppressed directly in the frequency domain. In this way, scattering noise can be effectively removed from the original multispectral texture features, resulting in noise-reduced multispectral texture features.

[0165] Subsequently, cross-correlation calculations are performed on these de-scattered multispectral texture features across different spectral channels. Cross-correlation calculations can employ standard correlation coefficient methods, such as the Pearson correlation coefficient, to quantify the similarity or correlation of texture features across different spectral channels. Since the input data has already undergone noise reduction, the calculated cross-correlation results will be more accurate and reliable, more realistically reflecting the inherent correlation between label materials and printed patterns across different spectral channels.

[0166] The technical solution of this application solves the above problems by introducing adaptive filtering based on scattering intensity and spatial frequency distribution before performing cross-correlation calculation. During image acquisition, scattering noise often exists in multispectral texture features with specific spatial frequency and intensity distributions, obscuring the true correlation between textures in different spectral channels. By accurately analyzing the characteristics of this noise and dynamically adjusting the filtering parameters accordingly, these interferences can be specifically removed, allowing subsequent cross-correlation calculations to be based on cleaner, more realistic texture data. This preprocessing mechanism effectively isolates the impact of noise on texture correlation evaluation, ensuring the accuracy of the cross-correlation results.

[0167] Through the above technical solution, this application significantly improves the accuracy and robustness of multispectral texture feature cross-correlation calculation. By adjusting the filtering parameters and removing scattering noise based on scattering intensity and spatial frequency distribution before cross-correlation calculation, the interference of noise on texture correlation judgment can be effectively avoided. Therefore, the obtained cross-correlation results can more accurately reflect the inherent texture, diffuse reflection characteristics, and true correlation of printed patterns in different spectral channels of the novel recycled paper-based label, thereby improving the sensitivity and specificity of defect detection, reducing false alarm and false negative rates, and providing more reliable technical support for the quality control of beverage labels.

[0168] In some embodiments, the texture features are extracted by calculating the gray-level co-occurrence matrix statistics or Gabor filter bank response values ​​of the local region.

[0169] Specifically, gray-level co-occurrence matrix (GLCM) statistics are a commonly used texture analysis method. They describe texture by calculating the joint probability distribution of gray values ​​of two pixels in an image that have a specific spatial relationship (e.g., distance and orientation). Various statistics can be extracted from the GLCM, such as contrast, correlation, energy, entropy, and homogeneity. These statistics can quantify texture properties such as roughness, directionality, and regularity. For example, high contrast may indicate drastic texture changes, while high energy may represent texture uniformity.

[0170] As an alternative implementation, Gabor filter bank response values ​​can also be used to extract texture features. A Gabor filter is a linear filter that can effectively extract local features in both the spatial and frequency domains, and is particularly suitable for capturing directional and scale information in images. By configuring Gabor filters with different orientations and frequencies, the response intensity of the image at different orientations and scales can be obtained; these response values ​​constitute Gabor texture features. Gabor filter banks are robust to changes in illumination and rotation, and can better adapt to the complexities of real-world production environments.

[0171] By employing the aforementioned technical solutions and utilizing gray-level co-occurrence matrix statistics or Gabor filter bank response values ​​to extract texture features, the descriptive power and discriminative power of texture features can be significantly improved. This allows the established normal texture spectrum to more accurately reflect the inherent texture characteristics of qualified labels, and even subtle texture differences in local areas of the label image to be detected can be effectively captured. Therefore, the defect detection method of this application exhibits higher accuracy and robustness in identifying various types of label defects, especially those related to the texture of the material itself or printing details, thereby improving performance and reliability.

[0172] This application also proposes a beverage label defect detection system based on industrial vision, such as... Figure 2 As shown, a beverage label defect detection system 100 based on industrial vision is described. The system includes:

[0173] The texture feature extraction module 10 is used to extract texture features of local areas based on qualified new recycled paper-based label images, and to establish a normal texture spectrum describing the inherent texture performance of new recycled paper-based labels based on the extracted texture features.

[0174] The local texture feature extraction module 20 is used to extract the texture features of local areas after performing brightness normalization processing on the real-time acquired image of the label to be detected.

[0175] The texture difference calculation module 30 is used to compare the texture features of a local region of the label image to be detected with the normal texture spectrum to calculate the texture difference.

[0176] The defect pattern matching module 40 is used to perform defect pattern matching based on the texture difference degree and the preset defect pattern to obtain the defect pattern matching result.

[0177] The defect determination module 50 is used to determine whether there are defects in a local area of ​​the label image to be detected based on the texture difference degree, the defect pattern matching result, and the texture difference degree threshold adjusted according to the production environment parameters.

[0178] This application employs a local texture feature extraction module to perform brightness normalization processing before extracting texture features, effectively eliminating the image quality degradation caused by uneven light sources and ensuring the accuracy of subsequent analysis. The texture feature extraction module establishes a normal texture spectrum describing the inherent texture of the novel recycled paper-based label, enabling the system to accurately distinguish the inherent texture of the new material from actual defects, avoiding misjudgments. Furthermore, the introduction of a defect pattern matching module and a defect judgment module, combined with an adaptive texture difference threshold adjustment mechanism, further improves the accuracy and robustness of defect judgment. Compared with existing technologies, this application's industrial vision-based beverage label defect detection system represents a significant advancement.

[0179] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for detecting defects in beverage labels based on industrial vision, characterized in that, include: Based on qualified images of new recycled paper-based labels, local texture features are extracted, and a normal texture spectrum describing the inherent texture of new recycled paper-based labels is established based on the extracted texture features. After performing brightness normalization on the real-time acquired image of the label to be detected, the texture features of its local area are extracted. The texture features of a local region of the label image to be detected are compared with the normal texture spectrum to calculate the texture difference. Based on the texture difference degree and the preset defect pattern, defect pattern matching is performed to obtain the defect pattern matching result; Based on the texture difference degree, the defect pattern matching result, and the texture difference degree threshold adjusted according to production environment parameters, it is determined whether there are defects in a local area of ​​the label image to be detected.

2. The beverage label defect detection method based on industrial vision according to claim 1, characterized in that, The step of establishing a normal texture spectrum describing the inherent texture performance of the novel recycled paper-based label based on the extracted texture features includes: When the system is deployed for the first time or a new batch of materials is launched, an initial label image is acquired, and texture features of local areas are extracted from the initial label image; The extracted texture features are clustered to obtain an initial normal texture reference; During normal production line operation, images of the labels to be inspected are acquired in real time, and texture features of local areas are extracted from the images of the labels to be inspected. Calculate the texture difference degree between the texture features of a local region of the label image to be detected and the initial normal texture reference, and determine the qualified region based on the texture difference degree; Based on the areas deemed qualified, assess the confidence level of the qualified areas; The initial normal texture reference is periodically updated based on qualified regions with confidence levels higher than a preset confidence threshold to obtain an updated normal texture spectrum; Receive defect area information from manual feedback, and adjust the updated normal texture spectrum based on the defect area information.

3. The beverage label defect detection method based on industrial vision according to claim 1, characterized in that, The step of comparing the texture features of a local region of the label image to be detected with the normal texture spectrum to calculate the texture difference includes: The texture features of a local region of the image to be detected are compared with the normal texture spectrum to mark potential texture anomalies; Track the potential texture anomalies in consecutive frames of an image sequence and establish anomaly trajectories; Based on the abnormal trajectory, determine whether it is a persistent defect. If it is determined to be a persistent defect, calculate the texture difference degree between the texture features of the local region and the normal texture spectrum.

4. The beverage label defect detection method based on industrial vision according to claim 1, characterized in that, The steps of extracting texture features of local areas from qualified new recycled paper-based label images and establishing a normal texture spectrum describing the inherent texture of new recycled paper-based labels based on the extracted texture features include: Configure a multispectral industrial camera to acquire qualified multispectral images of new recycled paper-based labels; The multispectral image is divided into multiple local regions, and multispectral texture features are extracted from the local regions. The multispectral texture features characterize the inherent texture, light diffuse reflection characteristics, and contrast of the printed pattern in different spectral channels of the novel recycled paper-based label. By analyzing the correlation between the multispectral texture features in different spectral channels, texture features characterizing the inherent texture, light diffuse reflection characteristics, and printing pattern contrast of the novel recycled paper-based label are obtained, and a normal texture spectrum describing the inherent texture performance of the novel recycled paper-based label is established.

5. The beverage label defect detection method based on industrial vision according to claim 4, characterized in that, The steps of analyzing the correlation between the multispectral texture features in different spectral channels to obtain texture features characterizing the inherent texture, light diffuse reflection characteristics, and printed pattern contrast of the novel recycled paper-based label, and establishing a normal texture spectrum describing the inherent texture performance of the novel recycled paper-based label, include: Statistical analysis was performed on the correlation between the multispectral texture features in different spectral channels to establish an initial multispectral texture correlation reference, which characterizes the normal spectral response characteristics of the current batch of materials and inks; During normal production line operation, multispectral images of the labels to be inspected are acquired in real time, the multispectral images of the labels to be inspected are divided into multiple local regions, and multispectral texture features are extracted from each local region. The correlation difference degree between the extracted multispectral texture features and the initial multispectral texture association reference is calculated to obtain the correlation difference degree; Based on the correlation difference degree and the preset difference degree threshold range, determine whether it is a qualified region and assess the confidence level of the qualified region. The initial multispectral texture association reference is periodically updated based on the multispectral texture features of qualified regions with confidence levels higher than a preset confidence threshold, to obtain the updated normal texture spectrum; Receive defect area information from manual feedback, and adjust and update the normal texture spectrum based on the defect area information.

6. The beverage label defect detection method based on industrial vision according to claim 4, characterized in that, The step of analyzing the correlation between the multispectral texture features in different spectral channels includes: The spectral response curves of the multispectral texture features in the local area are fitted to obtain the spectral response characteristic curves; Calculate the rate of change of curvature of the spectral response characteristic curve, and compare the rate of change of curvature with a preset material density non-uniformity characteristic threshold to obtain the comparison result of material density non-uniformity. The cross-correlation of the multispectral texture features across different spectral channels is calculated to obtain the cross-correlation results. By combining the fitting residuals of the spectral response characteristic curves, correlation anomalies that match the printing blur defect pattern are identified, and correlation anomaly identification results are obtained. Based on the comparison results of the material density non-uniformity, the cross-correlation results, and the correlation anomaly identification results, it is determined whether the texture anomaly in the local area originates from the inherent non-uniformity of the material density or a printing blur defect.

7. The beverage label defect detection method based on industrial vision according to claim 6, characterized in that, The step of calculating the rate of change of curvature of the spectral response characteristic curve and comparing the rate of change of curvature with a preset material density non-uniformity characteristic threshold to obtain the comparison result of material density non-uniformity includes: Identify local extrema on the spectral response characteristic curve; Calculate the peak intensity and peak width of the rate of curvature change of the spectral response characteristic curve at each local extremum point; Statistical analysis was performed on all the peak intensities and peak widths to obtain the distribution characteristics of the rate of change of curvature; Based on the distribution characteristics, the threshold for preset material density non-uniformity is dynamically adjusted; The rate of curvature change of each local region is compared with the adjusted preset threshold for material density non-uniformity to obtain the comparison result of material density non-uniformity.

8. The beverage label defect detection method based on industrial vision according to claim 6, characterized in that, The step of calculating the cross-correlation of the multispectral texture features across different spectral channels to obtain the cross-correlation result includes: Before calculating the cross-correlation, the filtering parameters are adjusted according to the scattering intensity and spatial frequency distribution of the multispectral texture features of each local region under different spectral channels to obtain the multispectral texture features after scattering noise removal. Cross-correlation calculations were performed on the multispectral texture features after scattering noise removal across different spectral channels to obtain the cross-correlation results.

9. The beverage label defect detection method based on industrial vision according to claim 1, characterized in that, The texture features are extracted by calculating the gray-level co-occurrence matrix statistics or Gabor filter bank response values ​​of the local region.

10. A beverage label defect detection system based on industrial vision, characterized in that, The system includes: The texture feature extraction module is used to extract texture features of local areas from qualified new recycled paper-based label images, and to establish a normal texture spectrum describing the inherent texture of new recycled paper-based labels based on the extracted texture features. The local texture feature extraction module is used to extract the texture features of local areas after performing brightness normalization processing on the real-time acquired image of the label to be detected. The texture difference calculation module is used to compare the texture features of a local region of the label image to be detected with the normal texture spectrum to calculate the texture difference. The defect pattern matching module is used to perform defect pattern matching based on the texture difference degree and the preset defect pattern to obtain the defect pattern matching result. The defect determination module is used to determine whether there are defects in a local area of ​​the label image to be detected based on the texture difference degree, the defect pattern matching result, and the texture difference degree threshold adjusted according to the production environment parameters.

Images