Sealing self-adaptive soft package lithium battery packaging image detection method and system

Abstract

The invention provides a sealed self-adaptive soft package lithium battery packaging image detection method and system, relates to the technical field of image detection, and obtains feature information by analyzing and processing image information of a to-be-detected soft package lithium battery packaging area. Based on the feature information, whether material batch switching occurs or not is judged, when material batch switching is detected, an image sample of a current product is collected, local visual features are extracted for statistical analysis, and reference information is constructed. The method comprises the following steps: acquiring current image information of a to-be-detected product, extracting local visual features of the to-be-detected product, comparing the local visual features with reference information, identifying an abnormal region, quantifying the geometric shape, size and strength of the abnormal region, performing similarity matching and multiple verification in combination with a preset defect mode, and judging whether the abnormal region is a real defect or not. The problem that in the prior art, when facing different batches of raw materials, an automatic image detection system cannot adapt to material characteristic changes, so that qualified products are frequently misjudged as defects is effectively solved.

Description

Technical Field

[0001] This application relates to the field of image detection technology, and more specifically, to a method and system for image detection of sealed adaptive soft-pack lithium battery packaging. Background Technology

[0002] In the manufacturing of pouch lithium batteries, the packaging process directly affects battery performance and safety, and automated image inspection systems are often used to check the packaging area. However, in actual production, even if the raw materials meet quality standards, there may still be slight differences in surface optical properties between different batches. These differences appear as harmless visual artifacts in the images, but are misjudged as defects by existing systems, leading to the rejection of qualified products and severely reducing production efficiency and material utilization.

[0003] Traditional solutions rely on manual intervention or time-consuming system retraining, making them unsuitable for high-speed continuous production. For example, when switching batches of aluminum-plastic film on a production line, the differences in the microstructure and reflective properties of the new material are magnified by strong light from an industrial camera, resulting in overexposed bright spots or new textures. These changes do not affect battery sealing and safety, but because the detection program builds its knowledge base based on images of old materials, it is easy to misjudge bright spots as "bubbles" or "impurities," and textures as early forms of "wrinkles," leading to a large number of incorrect rejections and waste.

[0004] While reducing detection sensitivity can decrease false positives, it increases the risk of missing true defects. Re-establishing testing standards for new materials requires large-scale manual marking and program retraining, leading to production line shutdowns or ineffective testing, resulting in significant losses. Furthermore, it's impossible to predict whether similar problems will recur in subsequent batches; if retraining is required for each batch, the efficiency and stability of automated testing cannot be guaranteed. Summary of the Invention

[0005] This application provides a sealed adaptive soft-pack lithium battery packaging image detection method and system, which aims to solve the problem of misjudgment caused by material batch differences in soft-pack lithium battery packaging image detection, as well as the limitations of traditional detection systems that require manual intervention or time-consuming and labor-intensive retraining when faced with changes in material properties.

[0006] On one hand, this application provides a method for detecting images of sealed adaptive soft-pack lithium battery packaging, including:

[0007] Image information of the packaging area of ​​the soft-pack lithium battery to be detected is acquired, and dynamic background maintenance and fluctuation analysis processing are performed on the image information to obtain feature information reflecting the overall visual performance of the product. Based on the changes in the feature information, it is determined whether a material batch switch has occurred.

[0008] When a material batch switch is detected, image samples of the current product are collected, local visual features reflecting the optical performance of the material surface are extracted from the image samples, and statistical analysis is performed on the local visual features to construct a benchmark information representing the normal visual performance of the current product material.

[0009] The current image information of the product to be inspected is acquired, its local visual features are extracted, and the local visual features are compared with the reference information to identify abnormal areas that deviate from the reference information;

[0010] The geometry, size, and intensity of the abnormal region are quantified, and similarity matching and multiple verifications are performed in combination with a preset defect pattern to determine whether the abnormal region is a real defect.

[0011] Optionally, after the step of quantifying the geometry, size, and intensity of the abnormal region, performing similarity matching and multiple verifications in conjunction with a preset defect pattern, and determining whether the abnormal region is a real defect, the following steps are included:

[0012] During the testing process, image information of products deemed qualified is continuously acquired, and the reference information is adjusted progressively based on the changes in material properties reflected in the image information of products deemed qualified.

[0013] Optionally, the step of determining whether a material batch switch has occurred based on the changes in the feature information includes:

[0014] Maintain the short-term adaptive background and long-term stable background of the aforementioned feature information;

[0015] Based on the image information of the products that have been determined to be qualified, the short-term adaptive background is updated quickly, and the long-term stable background is updated slowly.

[0016] The feature information is compared with the short-term adaptive background. When the initial deviation exceeds a preset threshold, the overall deviation between the feature information and the long-term stable background is determined.

[0017] When the feature information deviates significantly from the overall long-term stable background, it is determined that a material batch switch has occurred; and after determining that a material batch switch has occurred, the short-term adaptive background is quickly adjusted.

[0018] Optionally, the step of acquiring image samples of the current product, extracting local visual features reflecting the optical performance of the material surface from the image samples, and performing statistical analysis on the local visual features to construct a benchmark information representing the normal visual performance of the current product material includes:

[0019] Continuously acquire image samples of the current product;

[0020] The image samples are initially screened for defects to obtain target image samples, so as to identify and exclude image samples containing serious defects;

[0021] Local visual features are extracted from the target image sample, statistical analysis is performed on the local visual features, and the benchmark information is constructed based on the statistical analysis results.

[0022] Optionally, the step of performing dynamic background maintenance and fluctuation analysis on the image information to obtain feature information reflecting the overall visual performance of the product, and determining whether a material batch switch has occurred based on the changes in the feature information, includes:

[0023] Dynamic background maintenance and fluctuation analysis are performed on the image information to obtain feature information that reflects the overall visual performance of the product, and the short-term fluctuation range and long-term stable range of the feature information are maintained.

[0024] Determine whether the short-term fluctuation range of the feature information in a continuous image sequence continuously exceeds the long-term stable range. When the short-term fluctuation range continuously exceeds the long-term stable range, determine whether the deviation between the feature information and the long-term stable range shows a convergence trend within a preset time.

[0025] When the short-term fluctuation range continues to exceed the long-term stable range, and the deviation does not show a convergence trend within a preset time, it is determined that a momentary or local change has occurred in the production environment.

[0026] When the short-term fluctuation range continues to exceed the long-term stable range, and the deviation shows a convergence trend within a preset time, it is determined to be a material batch switch.

[0027] Optionally, the step of extracting local visual features from the target image sample and performing statistical analysis on the local visual features includes:

[0028] Identify the distribution pattern of the local visual features and determine whether they exhibit multiple clustering trends;

[0029] For each central tendency, determine its internal normal fluctuation range;

[0030] Determine whether a single feature value in the local visual features falls outside the normal fluctuation range of any central tendency; and if the single feature value falls outside the normal fluctuation range of all central tendencies, mark it as an outlier and suppress it.

[0031] Optionally, the step of continuously acquiring image information of products judged to be qualified during the detection process, and progressively adjusting the reference information based on the changes in material properties reflected in the image information of products judged to be qualified, includes:

[0032] Continuously acquire image information of products that have been determined to be qualified;

[0033] Local visual features are extracted from the image information of products that have been determined to be qualified.

[0034] Perform statistical analysis on the local visual features to identify outliers in the statistical analysis;

[0035] The outliers are excluded from the local visual features, and the reference information is adjusted progressively based on the changes in material properties reflected in the local visual features after excluding the outliers.

[0036] Optionally, the step of quantifying the geometry, size, and intensity of the abnormal region, combining it with a preset defect pattern for similarity matching and multiple verifications, and determining whether the abnormal region is a real defect includes:

[0037] The visual representation of the abnormal region is quantified in multiple dimensions to obtain the first set of quantified information;

[0038] The first set of quantitative information is processed to obtain a second set of quantitative information that reflects the prominence of the abnormal region in different visual dimensions.

[0039] Based on the second set of quantification information, one or more candidate defect patterns that are similar to the quantification features of the abnormal region are identified in the preset defect pattern set.

[0040] Based on one or more identified candidate defect patterns, obtain their corresponding discrimination rules;

[0041] According to the discrimination rule, the abnormal region is subjected to multiple verifications to determine whether the abnormal region is a real defect; and according to the result of the multiple verifications, the defect type of the abnormal region is output.

[0042] Optionally, the step of performing multi-dimensional quantization of the visual representation of the abnormal region to obtain the first set of quantified information includes:

[0043] Statistical analysis of pixel grayscale values ​​is performed on the abnormal regions, and the average grayscale, standard deviation, skewness, and kurtosis of the abnormal regions are calculated.

[0044] Edge detection is performed on the abnormal region, and the perimeter, area, aspect ratio, circularity, and convexity of the abnormal region are calculated.

[0045] The first set of quantization information is obtained based on the average gray level, standard deviation, skewness, kurtosis, perimeter, area, aspect ratio, roundness, and convexity.

[0046] On the other hand, this application provides a sealed adaptive soft-pack lithium battery packaging image inspection system, the system comprising:

[0047] The image acquisition and feature judgment module is used to acquire image information of the packaging area of ​​the soft-pack lithium battery to be detected, perform dynamic background maintenance and fluctuation analysis on the image information to obtain feature information reflecting the overall visual performance of the product, and determine whether a material batch switch has occurred based on the changes in the feature information.

[0048] The benchmark construction module is used to collect image samples of the current product when it is determined that a material batch switch has occurred, extract local visual features reflecting the optical performance of the material surface from the image samples, and perform statistical analysis on the local visual features to construct a benchmark information representing the normal visual performance of the current product material.

[0049] An anomaly detection module is used to acquire the current image information of the product to be detected, extract its local visual features, and compare the local visual features with the reference information to identify abnormal areas that deviate from the reference information.

[0050] The defect judgment module is used to quantify the geometric shape, size and intensity of the abnormal area, and perform similarity matching and multiple verifications in combination with preset defect patterns to determine whether the abnormal area is a real defect.

[0051] This application relates to a sealed adaptive soft-pack lithium battery packaging image detection method and system. By acquiring image information of the packaging area of ​​the soft-pack lithium battery to be detected and performing dynamic background maintenance and fluctuation analysis on the image information, feature information reflecting the overall visual performance of the product can be obtained. Based on the changes in this feature information, it can intelligently determine whether a material batch switch has occurred, thereby effectively identifying minute changes in surface optical properties caused by batch differences in raw materials. When a material batch switch is detected, it can adaptively acquire image samples of the current product and extract local visual features reflecting the optical performance of the material surface for statistical analysis, thereby constructing a benchmark information representing the normal visual performance of the current product material. This benchmark information can accurately reflect the inherent visual characteristics of the new batch of materials, avoiding misjudging harmless visual artifacts as defects. Subsequently, the current image information of the product to be detected is acquired, its local visual features are extracted, and compared with the newly constructed benchmark information, thereby accurately identifying the abnormal areas that truly deviate from normal performance. Finally, by quantifying the geometric shape, size, and intensity of the abnormal area and combining it with a preset defect pattern for similarity matching and multiple verification, it can accurately determine whether the abnormal area is a real defect.

[0052] Through the above technical solution, this application effectively solves the problem in the prior art where automated image inspection systems frequently misclassify qualified products as defective due to their inability to adapt to changes in material properties when faced with different batches of raw materials. This method eliminates the need for manual intervention or time-consuming system retraining, significantly improving inspection accuracy and production efficiency, reducing material waste and capacity loss, overcoming the limitations of traditional solutions in high-speed continuous production, and achieving self-adaptation and intelligence in the inspection system, demonstrating significant and superior technical effects. Attached Figure Description

[0053] 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.

[0054] Figure 1 The diagram above illustrates a flowchart of a sealed adaptive soft-pack lithium battery packaging image detection method.

[0055] Figure 2 The diagram above illustrates a schematic of the structure of a sealed adaptive soft-pack lithium battery packaging image detection system.

[0056] Figure reference numerals: 100, Sealed adaptive soft-pack lithium battery packaging image inspection system; 10, Image acquisition and characteristic judgment module; 20, Benchmark construction module; 30, Anomaly recognition module; 40, Defect judgment module. Detailed Implementation

[0057] 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.

[0058] 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.

[0059] In the manufacturing process of pouch lithium batteries, quality control in the packaging stage is crucial. Traditional automated image inspection systems, when inspecting the packaging area, often misjudge harmless visual artifacts as defects due to minute differences in surface optical properties between different batches of raw materials. This leads to the incorrect rejection of qualified products, severely impacting production efficiency and material utilization. Traditional solutions often require manual intervention or time-consuming system retraining, which is unacceptable in the context of high-speed continuous production.

[0060] like Figure 1 The diagram illustrates an exemplary flowchart of a sealed adaptive soft-pack lithium battery package image detection method. This application proposes a sealed adaptive soft-pack lithium battery package image detection method, comprising:

[0061] S10, acquire image information of the packaging area of ​​the soft-pack lithium battery to be detected, perform dynamic background maintenance and fluctuation analysis on the image information to obtain feature information reflecting the overall visual performance of the product, and determine whether a material batch switch has occurred based on the changes in the feature information.

[0062] Image information typically refers to digital image data of the packaged area of ​​a soft-pack lithium battery, acquired through an industrial camera or other image acquisition equipment. It may include visual features such as grayscale values, color information, and texture.

[0063] Dynamic background maintenance and fluctuation analysis refers to a series of preprocessing and analysis steps performed on image information to distinguish between normal visual performance and abnormal changes in a product. This typically involves techniques such as image filtering, feature extraction, and statistical analysis.

[0064] Feature information is abstract data that, after processing, can reflect the overall visual performance of a product, such as the average brightness, contrast, and energy of specific frequency components of an image.

[0065] Material batch switching refers to the process of changing the raw materials (such as aluminum-plastic film) used in the production process from one batch to another, which may cause subtle changes in the optical properties of the product surface.

[0066] S20, when it is determined that a material batch switch has occurred, an image sample of the current product is collected, local visual features reflecting the optical performance of the material surface are extracted from the image sample, and statistical analysis is performed on the local visual features to construct a benchmark information representing the normal visual performance of the current product material.

[0067] Image samples refer to a set of representative images collected from qualified products under specific production conditions.

[0068] Local visual features refer to the visual attributes of specific areas in an image, such as texture, local brightness distribution, and edge information. These features can reflect the optical performance of a material surface.

[0069] Statistical analysis involves mathematically processing local visual features to obtain statistical quantities such as their distribution patterns, mean, and variance.

[0070] The baseline information is a reference model or dataset that represents the normal visual performance of the current product material, constructed based on the results of statistical analysis.

[0071] S30, acquire the current image information of the product to be detected, extract its local visual features, and compare the local visual features with the reference information to identify abnormal areas that deviate from the reference information;

[0072] An abnormal region refers to an area where the local visual features of the product to be tested differ significantly from the baseline information.

[0073] S40, quantify the geometry, size and intensity of the abnormal region, and perform similarity matching and multiple verifications in combination with a preset defect pattern to determine whether the abnormal region is a real defect.

[0074] Among them, geometric shape, size and intensity are indicators for quantifying abnormal areas, such as area, perimeter, aspect ratio, average gray value, maximum gray value, etc.

[0075] Defect patterns are predefined characteristic descriptions of various known defects (such as bubbles, scratches, foreign objects, etc.).

[0076] Similarity matching and multiple verification refers to comparing the quantified abnormal region features with the defect pattern and performing cross-verification through multiple rules or algorithms to improve the accuracy of defect judgment.

[0077] This method is typically implemented in an automated inspection system, which includes image acquisition equipment, an image processing unit (such as an industrial computer or an embedded vision system), and an actuator.

[0078] The core of the sealed adaptive soft-pack lithium battery packaging image detection method of this application lies in achieving accurate detection of the soft-pack lithium battery packaging area through a series of intelligent processing steps, and effectively dealing with the challenges brought about by material batch switching.

[0079] First, after acquiring image information of the packaged area of ​​the soft-pack lithium battery to be inspected, dynamic background maintenance and fluctuation analysis processing needs to be performed on the image information to obtain feature information reflecting the overall visual performance of the product. Then, based on the changes in the feature information, it is determined whether a material batch change has occurred. For example, images of the soft-pack lithium battery packaged area can be acquired using an industrial camera; these images can be high-resolution grayscale or color images. Various methods can be used to process this image information. One approach is to extract global brightness, contrast, and texture features from the image and monitor these features in real time. For example, the average grayscale value, the variance of the grayscale histogram, or the energy of specific frequency domain components can be calculated. This feature information is used to maintain a short-term adaptive background and a long-term stable background. The short-term adaptive background can be updated rapidly based on images of qualified products from the most recent period, while the long-term stable background is updated at a slower rate to reflect the long-term trend of the product's material properties. By comparing the deviations of the current feature information with the short-term adaptive background and the long-term stable background, it is possible to determine whether a material batch change has occurred. For example, when the initial deviation of the feature information from the short-term adaptive background exceeds a preset threshold, the overall deviation from the long-term stable background is further assessed. If the overall deviation is significant and consistent, it is determined that a material batch switch has occurred.

[0080] When a material batch change is detected, image samples of the current product need to be collected. Local visual features reflecting the optical properties of the material surface are extracted from these samples, and statistical analysis is performed on these features to construct a baseline representing the normal visual performance of the current product's material. For example, after confirming a material batch change, a sample collection phase is initiated. During this phase, image samples of qualified products from the current production batch are continuously acquired. These image samples first undergo preliminary defect screening to identify and exclude image samples containing serious defects, ensuring the accuracy of the subsequent baseline construction. From the screened target image samples, local visual features reflecting the optical properties of the material surface are extracted. These local visual features can be local texture features of the image (such as LBP, Gabor features), local brightness distribution features, or feature vectors extracted based on deep learning models. Statistical analysis is performed on these local visual features, such as calculating their mean, standard deviation, and distribution pattern. Based on the statistical analysis results, a baseline representing the normal visual performance of the current product's material is constructed. This baseline can be a multi-dimensional feature space model, a probability distribution model, or a set of statistical parameters.

[0081] Next, the current image information of the product to be inspected is acquired, its local visual features are extracted, and these local visual features are compared with benchmark information to identify abnormal regions that deviate from the benchmark information. For example, for each pouch lithium battery to be inspected, its current image information is acquired. From this image information, the same type of local visual features as those used when constructing the benchmark information are extracted. Then, the extracted local visual features are compared with the previously constructed benchmark information. Comparison methods may include calculating the distance between feature vectors, determining whether the feature values ​​fall outside the normal fluctuation range of the benchmark information, or using a classifier to determine whether it belongs to the "normal" category. Any region that differs significantly from the benchmark information is identified as an abnormal region.

[0082] Finally, the geometry, size, and intensity of the anomalous region are quantified, and similarity matching and multiple verifications are performed in conjunction with preset defect patterns to determine whether the anomalous region is a real defect. For example, once an anomalous region is identified, its visual performance is quantified in multiple dimensions. This includes statistical analysis of the pixel grayscale values ​​of the anomalous region, calculating its average grayscale, standard deviation, skewness, kurtosis, and other intensity features; and edge detection, calculating its perimeter, area, aspect ratio, roundness, convexity, and other geometric and dimensional features. This quantified information constitutes the first set of quantified information. Processing the first set of quantified information yields a second set of quantified information reflecting the prominence of the anomalous region in different visual dimensions. Based on the second set of quantified information, one or more candidate defect patterns with similar quantified features to the anomalous region are identified from the preset defect pattern set. For example, an anomalous region with a specific aspect ratio and low average grayscale may be matched with the defect patterns of "scratches" or "foreign objects." Based on the identified candidate defect patterns, their corresponding discrimination rules are obtained. These discrimination rules may include a series of threshold conditions, logical judgments, or machine learning models. Based on the discrimination rules, abnormal regions undergo multiple verifications to determine whether they are genuine defects. For example, if an abnormal region is initially identified as a "bubble," its internal grayscale uniformity, edge smoothness, and other characteristics may be further verified to eliminate interference from other similar visual representations. Based on the results of the multiple verifications, the defect type of the abnormal region is finally output.

[0083] The sealed adaptive soft-pack lithium battery packaging image detection method proposed in this application forms a complete and efficient detection process through a series of closely linked steps. Its working principle is as follows: firstly, through dynamic background maintenance and fluctuation analysis, the overall visual performance of the production line is monitored in real time. Once a significant and consistent deviation from the long-term stable background is detected, the occurrence of material batch switching can be accurately determined. This mechanism avoids misjudgments caused by material batch differences in traditional methods, laying the foundation for subsequent adaptive adjustments.

[0084] Once the material batch switch is confirmed, the process immediately enters the adaptive learning phase. This phase involves collecting image samples of qualified products from the current batch and extracting local visual features for statistical analysis. This constructs benchmark information representing the normal visual performance of the current material. This benchmark information is dynamically generated and accurately reflects the inherent characteristics of the new batch of materials, rather than relying on pre-set, potentially outdated standards.

[0085] During routine inspections, local visual features of the product under inspection are extracted in real time and compared with currently valid benchmark information. Any area deviating from the benchmark information is identified as an anomaly. This comparison mechanism ensures that only areas that truly deviate from the current normal product performance are marked, thus effectively distinguishing between harmless artifacts caused by changes in material properties and real defects.

[0086] Finally, for identified abnormal areas, their geometry, size, and intensity are quantified in multiple dimensions, and similarity matching and multiple verifications are performed in conjunction with preset defect patterns. This refined quantification and multiple verification mechanism greatly improves the accuracy and reliability of defect judgment, avoiding misjudgments and missed detections. By comprehensively analyzing the visual representation of abnormal areas and comparing them with known defect patterns, defect types can be identified more accurately, and visual interference caused by non-defect factors can be eliminated.

[0087] In summary, the sealed adaptive soft-pack lithium battery packaging image inspection method of this application forms a highly intelligent inspection closed loop through dynamic background maintenance, material batch switching judgment, adaptive benchmark construction, abnormal area identification, and a multi-verification defect judgment mechanism. The various technical features work together to solve the misjudgment problem caused by material batch differences in existing technologies, ensuring high accuracy and efficiency even in complex and ever-changing production environments.

[0088] This application represents a significant advancement and innovation compared to existing technologies. Traditionally, when dealing with different batches of raw materials, the pre-defined testing standards are based on specific batches. When batches change, even if the new batch meets quality standards, minute differences in surface optical properties often lead to harmless visual artifacts being misjudged as defects. This misjudgment not only results in the incorrect rejection of qualified products, severely impacting production efficiency and material utilization, but also requires time-consuming and labor-intensive manual intervention or retraining, which is unacceptable in the context of high-speed continuous production.

[0089] The core innovation of this application lies in its "sealing self-adaptation" characteristic. By introducing dynamic background maintenance and fluctuation analysis, it can monitor the characteristic information of the overall visual performance of the product in real time and intelligently determine whether a material batch switch has occurred based on its changes. This mechanism enables proactive perception of changes in the production environment, rather than passively waiting for manual adjustments. For example, when the production line switches to using different batches of aluminum-plastic film, the method of this application can accurately identify this change in material properties by maintaining and comparing the short-term adaptive background and the long-term stable background of the characteristic information, thereby avoiding misjudgments caused by the inability to identify such changes.

[0090] Furthermore, upon determining that a material batch change has occurred, the method of this application can adaptively acquire image samples of the current product, extract local visual features, and perform statistical analysis to construct a baseline information representing the normal visual performance of the current product material. This "baseline construction" module is key to the adaptive capability of this application. It enables the dynamic updating of the "qualification" standard based on the actual characteristics of the new batch of materials, rather than adhering to old, potentially inapplicable standards. Compared to traditional methods that require large-scale manual inspection and labeling and retraining of the analysis program, the method of this application can complete the updating of baseline information in an automated and efficient manner, greatly shortening the adaptation time to new materials and ensuring the continuity and stability of the production line.

[0091] Furthermore, after identifying anomalous regions that deviate from the baseline information, this application further improves the accuracy of defect judgment by quantifying the geometric shape, size, and intensity of the anomalous regions and combining them with preset defect patterns for similarity matching and multiple verifications. This multi-dimensional quantification and multiple verification mechanism enables a more precise distinction between genuine defects and visual interference caused by other non-defect factors, thereby effectively reducing the false judgment rate.

[0092] In summary, the sealed adaptive soft-pack lithium battery packaging image inspection method of this application effectively solves the problem of misjudgment caused by material batch differences in the prior art through its unique dynamic background maintenance, material batch switching judgment, adaptive benchmark construction, and multi-verification defect judgment mechanism. This significantly improves accuracy and robustness, thereby enhancing production efficiency and product quality. This adaptive capability enables it to maintain consistently high and stable performance in the face of complex and ever-changing production environments, demonstrating significant technological advancement and practical value.

[0093] In some embodiments, after the step of quantifying the geometry, size, and intensity of the abnormal region, performing similarity matching and multiple verifications in conjunction with a preset defect pattern, and determining whether the abnormal region is a real defect, the following steps are included:

[0094] During the testing process, image information of products deemed qualified is continuously acquired, and the reference information is adjusted progressively based on the changes in material properties reflected in the image information of products deemed qualified.

[0095] Specifically, during the testing process, image information of products that have been determined to be qualified is continuously acquired. This image information of qualified products is considered representative of "normal" products in the current production state. Subsequently, based on the changes in material properties reflected in these qualified product image information, the previously established benchmark information is adjusted in a gradual and smooth manner. The "gradual manner" means that the benchmark information is not updated in a one-time, large-scale modification, but rather through small, continuous iterations to gradually adapt to new material properties, thus avoiding instability of the benchmark information due to instantaneous fluctuations.

[0096] The above technical solution enables real-time and flexible adaptation to changes in product material properties, significantly improving the accuracy and robustness of defect detection. This solution effectively overcomes the limitations of static baseline information in traditional methods, reduces false alarm rates, and enhances the ability to identify real defects, thereby ensuring product quality stability and production efficiency.

[0097] For example, suppose that during the packaging process of pouch lithium batteries, slight fluctuations in the production environment (such as temperature, humidity, or minor differences in raw material suppliers) cause a slight but continuous drift in the surface gloss or texture of qualified products over a period of time. If the baseline information remains unchanged, these slight drifts may be misjudged as defects, or genuine defects may be difficult to identify due to confusion with the "normal" state after the drift. The technical solution of this application continuously collects images of batteries that have been determined to be qualified. For example, at regular intervals of a certain number of qualified products, their local visual features are extracted and compared with existing baseline information. If a slight statistical deviation is found between the feature distribution of qualified products and the existing baseline information, these new qualified product features are incorporated into the update of the baseline information in small steps with low weights. For example, methods such as exponentially weighted moving average (EWMA) can be used to ensure that the impact of new qualified product data on the baseline information is gradual, with newer data having a greater weight on the baseline information and older data having a gradually decreasing weight. In this way, the baseline information can smoothly track the long-term trend of material properties and always maintain its representativeness to the current normal product state, thereby ensuring the accuracy and stability of defect detection.

[0098] In some embodiments, the step of determining whether a material batch switch has occurred based on the changes in the feature information includes:

[0099] Maintain the short-term adaptive background and long-term stable background of the aforementioned feature information;

[0100] Based on the image information of the products that have been determined to be qualified, the short-term adaptive background is updated quickly, and the long-term stable background is updated slowly.

[0101] The feature information is compared with the short-term adaptive background. When the initial deviation exceeds a preset threshold, the overall deviation between the feature information and the long-term stable background is determined.

[0102] When the feature information deviates significantly from the overall long-term stable background, it is determined that a material batch switch has occurred; and after determining that a material batch switch has occurred, the short-term adaptive background is quickly adjusted.

[0103] The feature information refers to data reflecting the overall visual performance of the product, such as average brightness, texture features, and color distribution of an image. The short-term adaptive background can be understood as a rapid response model to recent changes in product feature information, capable of capturing instantaneous or short-term fluctuations during the production process. The long-term stable background represents the average or typical visual performance of the product under stable production conditions; its update speed is slower, aiming to reflect fundamental changes in material batches or production processes.

[0104] Specifically, maintaining the short-term adaptive background and long-term stable background of the aforementioned feature information aims to establish a multi-scale reference benchmark. The short-term adaptive background can quickly respond to minor changes on the production line, such as light and dust, while the long-term stable background provides a more stable reference point that is unaffected by short-term disturbances.

[0105] Furthermore, based on the image information of products deemed qualified, the short-term adaptive background is updated rapidly, while the long-term stable background is updated slowly. This means that only data from qualified products confirmed through quality inspection is used to update the background model, thereby ensuring the accuracy and reliability of the background. The rapid update of the short-term adaptive background allows it to quickly adapt to slight drifts in the production environment, while the slow update of the long-term stable background ensures its sensitivity to significant changes such as material batch switching, avoiding misjudgments due to short-term fluctuations.

[0106] Furthermore, the feature information is compared with the short-term adaptive background. When the initial deviation exceeds a preset threshold, the overall deviation between the feature information and the long-term stable background is determined. This two-stage judgment mechanism aims to improve the accuracy of the judgment. First, by comparing with the short-term adaptive background, most normal production fluctuations can be filtered out. Only when the deviation is large enough, indicating a possible abnormal change, is it further compared with the long-term stable background to confirm whether it is a deeper change in material properties. The preset threshold can be set according to actual production experience and requirements for false alarm rate and false negative rate.

[0107] Therefore, when the overall deviation of the characteristic information from the long-term stable background is significant and consistent, a material batch switchover is determined to have occurred. Here, "significant" means that the deviation is large enough to exceed the range of normal production fluctuations; "consistent" means that this deviation persists for a period of time, rather than being an occasional instantaneous fluctuation. This dual condition ensures that only genuine and stable changes in material properties will be identified as a batch switchover.

[0108] Finally, upon determining that a material batch change has occurred, the aforementioned short-term adaptation background is quickly adjusted. This is to enable rapid adaptation to the new material batch and avoid a large number of false alarms caused by the old background model in the early stages of new batch product deployment, thereby ensuring the continuity and accuracy of detection.

[0109] Through the above technical solution, this application can effectively distinguish between instantaneous fluctuations in the production process and actual material batch switching, significantly improving the accuracy and robustness of identifying changes in material properties. The introduction of a dual background allows for rapid adaptation to minor changes in the production environment while stably identifying fundamental changes in material batches, avoiding misjudgments or omissions that might occur with a single background model. In particular, the ability to quickly adjust the short-term adaptive background after determining that a material batch switching has occurred further enhances the adaptive capability, ensuring the continuity and efficiency of testing after the new batch of products is launched, thereby reducing production costs and improving the reliability of product quality control.

[0110] In some embodiments, the step of acquiring image samples of the current product, extracting local visual features reflecting the optical performance of the material surface from the image samples, and performing statistical analysis on the local visual features to construct a benchmark information representing the normal visual performance of the current product material includes:

[0111] Continuously acquire image samples of the current product;

[0112] The image samples are initially screened for defects to obtain target image samples, so as to identify and exclude image samples containing serious defects;

[0113] Local visual features are extracted from the target image sample, statistical analysis is performed on the local visual features, and the benchmark information is constructed based on the statistical analysis results.

[0114] Specifically, continuously acquiring image samples of the current product refers to continuously collecting image data of the current batch of products on the production line after a material batch change. These image samples are used for subsequent analysis and benchmark information construction.

[0115] The preliminary defect screening of the image samples to obtain target image samples, aimed at identifying and excluding image samples containing serious defects, can be understood as a preprocessing of all acquired image samples before constructing benchmark information. The purpose of this preprocessing is to identify and remove images with obvious serious defects, such as scratches, stains, and damage, to ensure that the samples used to construct benchmark information are high-quality images representing a normal state. Preliminary defect screening can be implemented using various methods, such as identifying abnormal regions based on simple image processing algorithms (e.g., threshold segmentation, edge detection), or using a pre-trained lightweight classification model to quickly classify images, marking and excluding suspected defective images.

[0116] In practical applications, after obtaining screened target image samples, local visual features reflecting the optical performance of the material surface are extracted from these images without serious defects. These features can include texture features, color distribution features, brightness uniformity features, etc. Subsequently, statistical analysis is performed on these local visual features, such as calculating their mean, variance, distribution range, and other statistical quantities. Based on these statistical analysis results, a benchmark information that can represent the normal visual performance of the current product material is constructed. This benchmark information will serve as the basis for comparing the local visual features of subsequent product images to be tested.

[0117] By employing the aforementioned technical solution, and by excluding image samples containing severe defects during the benchmark information construction process, the constructed benchmark information can more accurately and reliably reflect the normal visual performance of the current batch of materials. This significantly improves the accuracy of subsequent anomaly identification, effectively avoids misjudging defective features as normal features, thereby reducing the false negative and false positive rates, and enhancing the overall performance and robustness of the sealed adaptive soft-pack lithium battery packaging image detection method.

[0118] For example, suppose that after a material batch changeover, image samples of the packaging area of ​​the current batch of pouch lithium batteries are continuously acquired. To construct accurate benchmark information, these image samples are first subjected to preliminary defect screening. For example, a simple algorithm based on grayscale histogram analysis can be used: for each image, its grayscale histogram is calculated, and abnormal grayscale peaks or uneven distributions are detected, which may indicate severe scratches or stains. Simultaneously, a pre-defined simple classifier can be combined, which, by learning the features of a small number of known images with severe defects, can quickly identify whether there are large areas of color anomalies or structural damage in the image.

[0119] Through this initial screening, image samples deemed to contain severe defects are excluded, leaving only those considered normal or containing only minor, acceptable variations. These selected image samples are the target image samples. Subsequently, local visual features are extracted from these target image samples. For example, LBP (Local Binary Pattern) texture features or Gabor filter response features can be extracted to quantify the microscopic texture and optical reflectivity of the material surface. Statistical analysis is then performed on these extracted local visual features, such as calculating their mean vector and covariance matrix in the feature space, or constructing their probability density function model. Finally, based on these statistical analysis results, a baseline representing the normal visual performance of the current batch of materials is constructed, such as a multidimensional Gaussian distribution model or a set of feature cluster centers, for subsequent anomaly identification.

[0120] In some embodiments, the step of performing dynamic background maintenance and fluctuation analysis on the image information to obtain feature information reflecting the overall visual performance of the product, and determining whether a material batch switch has occurred based on the changes in the feature information, includes:

[0121] Dynamic background maintenance and fluctuation analysis are performed on the image information to obtain feature information that reflects the overall visual performance of the product, and the short-term fluctuation range and long-term stable range of the feature information are maintained.

[0122] Determine whether the short-term fluctuation range of the feature information in a continuous image sequence continuously exceeds the long-term stable range. When the short-term fluctuation range continuously exceeds the long-term stable range, determine whether the deviation between the feature information and the long-term stable range shows a convergence trend within a preset time.

[0123] When the short-term fluctuation range continues to exceed the long-term stable range, and the deviation does not show a convergence trend within a preset time, it is determined that a momentary or local change has occurred in the production environment.

[0124] When the short-term fluctuation range continues to exceed the long-term stable range, and the deviation shows a convergence trend within a preset time, it is determined to be a material batch switch.

[0125] Specifically, maintaining the short-term fluctuation range and the long-term stable range of the feature information means dynamically tracking and updating the variation range of the feature information at different time scales. The short-term fluctuation range refers to the variation range exhibited by the feature information within a short time window; its update speed is relatively fast, and it can quickly respond to small changes in the current production status. For example, it can be determined by calculating the standard deviation or range of the feature information of the most recent N image sequences. The long-term stable range refers to the stable range exhibited by the feature information over a longer time scale; its update speed is relatively slow, and it represents the inherent characteristics of the product material under normal production conditions. For example, it can be determined by calculating the average and standard deviation of the feature information over a relatively long period of time (such as several hours or days), or by using methods such as exponentially weighted moving averages for smoothing. The purpose of maintaining these two ranges is to provide reference benchmarks at different time scales to more accurately analyze the dynamic changes of the feature information.

[0126] Furthermore, it is determined whether the short-term fluctuation range of the feature information in a continuous image sequence consistently exceeds the long-term stable range. When the short-term fluctuation range consistently exceeds the long-term stable range, it is determined whether the deviation between the feature information and the long-term stable range shows a convergence trend within a preset time period. Here, "continuous image sequence" refers to a series of image data acquired continuously over time, used to observe the dynamic evolution of the feature information. "Consistently exceeding" means that the short-term fluctuation range of the feature information does not accidentally and briefly break through the long-term stable range, but rather exhibits a continuous deviation. This is usually achieved by setting a threshold for the number of consecutive exceedances or the duration of the deviation. "Convergence trend" means that after deviating from the long-term stable range, the numerical change of the feature information gradually tends to stabilize, forming a relatively stable interval at a new level. This can be determined by performing trend analysis, regression analysis, or statistical tests on the deviation sequence (e.g., determining whether the mean stabilizes near a new value within a preset time period). For example, the moving average of the deviation can be calculated, and its rate of change can be observed to approach zero. "Preset time" is a configurable parameter used to define the length of the time window required to observe the convergence trend. Its setting should take into account both the response speed of the production process and the real-time requirements of the detection.

[0127] Specifically, when the short-term fluctuation range continuously exceeds the long-term stable range, and the deviation does not show a convergence trend within a preset time, it is judged that a momentary or local change has occurred in the production environment. This condition indicates that although the feature information exhibits continuous fluctuations, these fluctuations have not formed a new stable state, but rather show randomness or transience. This typically corresponds to temporary disturbances on the production line, such as brief equipment vibrations, local contamination, or transient changes in lighting. When the short-term fluctuation range continuously exceeds the long-term stable range, and the deviation shows a convergence trend within a preset time, it is judged that a material batch has been switched. This condition indicates that the feature information not only continuously deviates from the original long-term stable range, but also tends towards a new stable state over a period of time. This strongly indicates that a substantial change has occurred in the product material itself, such as replacing different batches of raw materials, leading to a shift in the stability of its optical properties or surface texture.

[0128] Through the above technical solution, this application enables refined identification of changes in product material properties. Compared to judging solely based on simple changes in feature information, this application significantly improves the ability to distinguish between instantaneous or localized changes in the production environment and actual material batch switching by introducing a dynamic maintenance mechanism for short-term fluctuation ranges and long-term stable ranges, combined with the judgment of deviation convergence trends. This helps reduce false alarm rates and avoids frequent triggering of unnecessary baseline information updates or production adjustments due to temporary interference. Simultaneously, this technical solution can more timely and accurately identify genuine material batch switching, providing a reliable basis for subsequent baseline information updates and defect detection, ensuring that defect detection maintains high accuracy and adaptability even when material properties change.

[0129] For example, suppose that on a pouch lithium battery packaging production line, image information of the packaging area is continuously acquired, and feature information reflecting the overall visual performance of the product, such as average grayscale value and texture uniformity, is extracted from it. First, the short-term fluctuation range and long-term stable range of this feature information are maintained. The long-term stable range is established based on long-term historical data of the current batch of materials and represents the normal visual performance of that batch. The short-term fluctuation range is dynamically calculated based on feature information from a recent period (e.g., the most recent 50 image sequences).

[0130] Scenario 1: Dust suddenly drifts over the production line, causing a temporary change in localized lighting. This results in the short-term fluctuation range of feature information in 10 consecutive image sequences exceeding the long-term stable range. However, within a preset observation period (e.g., 20 image sequences), the deviation between the feature information and the long-term stable range does not show a convergence trend but quickly returns to within the long-term stable range. In this case, it would be judged as a momentary or localized change in the production environment, rather than a misjudgment of a material batch change.

[0131] Scenario 2: A new batch of packaging material was introduced on the production line, with slightly different surface reflectivity. In the initial period after the change, short-term fluctuations in feature information in consecutive image sequences consistently exceeded the original long-term stable range. Within a preset observation period (e.g., 20 image sequences), the deviation between the feature information and the long-term stable range was continuously monitored and gradually converged to a stable new value, exhibiting a convergence trend. At this point, it was accurately determined that a material batch switch had occurred, triggering a subsequent baseline information update process to adapt to the new material characteristics. Through the above mechanism, the detection method of this application can intelligently distinguish production fluctuations of different natures, ensuring the accuracy and robustness of the detection.

[0132] In some embodiments, the step of extracting local visual features from the target image sample and performing statistical analysis on the local visual features includes:

[0133] Identify the distribution pattern of the local visual features and determine whether they exhibit multiple clustering trends;

[0134] For each central tendency, determine its internal normal fluctuation range;

[0135] Determine whether a single feature value in the local visual features falls outside the normal fluctuation range of any central tendency; and if the single feature value falls outside the normal fluctuation range of all central tendencies, mark it as an outlier and suppress it.

[0136] Specifically, data analysis is performed on local visual features extracted from target image samples. This can be achieved through methods such as histogram analysis, density estimation, or clustering algorithms (e.g., K-means, Gaussian mixture models) to determine the numerical distribution pattern of these feature values. This process identifies the distribution morphology of the local visual features and determines whether they exhibit multiple clustering trends. The aim is to identify multiple independent or overlapping clusters of local visual features, which may correspond to different optical properties or texture types allowed in the normal production process of the material.

[0137] Specifically, for each central tendency, a normal fluctuation range is determined within it. This can be understood as, after identifying multiple central tendencies, performing independent statistical analysis on the data points within each tendency, such as calculating its mean and standard deviation, and defining a reasonable fluctuation range accordingly (e.g., mean ± 3 standard deviations). The aim is to establish a precise reference range for each specific normal material behavior pattern, in order to more precisely characterize the normal variability of materials.

[0138] In practical applications, when a new local visual feature value is acquired, it is compared with the normal fluctuation range of all established central tendency patterns. If the feature value does not belong to any of the defined normal fluctuation ranges, it is considered a potential outlier. The purpose is to initially screen features that do not conform to known normal patterns. Furthermore, if a single feature value falls outside the normal fluctuation range of all central tendency patterns, it is marked as an outlier and suppressed. This means identifying feature values ​​that do not conform to any normal fluctuation range as outliers and excluding them in subsequent statistical analysis for constructing benchmark information. Suppressing these outliers prevents them from negatively impacting the accuracy of the benchmark information, ensuring that the benchmark information accurately reflects the normal visual performance of the material.

[0139] Through the above technical solution, this application can more accurately capture the true distribution of local visual features of materials. Even when the material itself has multiple normal manifestations or a small number of atypical data points, it can construct more accurate and robust reference information. This significantly improves the accuracy of abnormal area identification and effectively reduces false alarms or missed alarms caused by inaccurate reference information, thereby enhancing the overall reliability and adaptability of the sealed adaptive soft-pack lithium battery packaging image detection method.

[0140] For example, suppose that when statistically analyzing the local visual features of a pouch lithium battery packaging material, two distinct peaks are found in the distribution of its texture feature values. This may correspond to two slightly different surface treatments that are acceptable in the production process. If a single statistical model is used, one of the normal treatment effects may be misjudged as an anomaly, or the average of the two effects may result in an excessively wide normal fluctuation range, thereby reducing the sensitivity of defect detection. The technical solution of this application first identifies these two central trends and establishes independent normal fluctuation ranges for them. For example, the texture feature values ​​of the first trend may be concentrated in the range of 50±5, while the second trend is concentrated in the range of 70±5. When a new local texture feature value of 60 is detected, it does not belong to the normal fluctuation range of either central trend, and is therefore marked as an outlier and suppressed to prevent it from interfering with the construction of the normal baseline. If feature values ​​of 52 or 73 are detected, they fall within their respective normal fluctuation ranges and are considered normal data. This processing method ensures that the baseline information can accurately reflect the various normal states of the material, while effectively excluding true outliers, thereby improving the accuracy and robustness of defect detection.

[0141] In some embodiments, the step of continuously acquiring image information of products determined to be qualified during the detection process, and progressively adjusting the reference information based on the material property changes reflected in the image information of products determined to be qualified, includes:

[0142] Continuously obtain the image information of the products determined to be qualified;

[0143] Extract local visual features from the image information of the products determined to be qualified;

[0144] Perform statistical analysis on the local visual features to identify outliers in the statistical analysis;

[0145] Exclude the outliers from the local visual features, and adjust the reference information in a progressive manner according to the material property changes reflected in the local visual features after excluding the outliers.

[0146] Specifically, continuously obtaining the image information of the products determined to be qualified means that during the normal operation of the production line, for the soft-pack lithium battery products that have passed the preliminary inspection and are determined to be defect-free, the image information of their packaging areas is continuously collected. These image information are regarded as samples representing the normal material properties of the current production batch. Among them, extracting local visual features from the image information of the products determined to be qualified can be understood as performing refined processing on these qualified product images to obtain characteristic data that can reflect the microscopic manifestations such as the surface texture, gloss, and color uniformity of the material. These local visual features can be grayscale histograms, texture features (such as local binary pattern LBP, gray-level co-occurrence matrix GLCM), color moments, etc.

[0147] Furthermore, perform data analysis on the set of extracted local visual features to discover the data points that significantly deviate from the overall distribution pattern. Outliers can be understood as the observed values that are statistically inconsistent with the behavior of most data points, and they may represent the occasional and unrepresentative minor abnormalities in the qualified products. The methods for identifying outliers can include but are not limited to distance-based methods (such as K-nearest neighbor), density-based methods (such as local outlier factor LOF), distribution-based methods (such as Z-score, interquartile range IQR), or clustering-based methods. In practical applications, excluding the outliers from the local visual features aims to purify the data set used for reference information adjustment. By removing these atypical data points, it can be ensured that the subsequent reference adjustment process is only based on the data that best represents the normal material properties, thereby improving the accuracy and robustness of the reference information. Thus, adjusting the reference information in a progressive manner according to the material property changes reflected in the local visual features after excluding the outliers means that after excluding the outliers, use the remaining and purer local visual feature data to correct or optimize the existing reference information in a small step and continuous update manner. This progressive adjustment ensures that the reference information can adapt to the minor and normal changes in material properties during the production process in real time, while avoiding incorrect adjustments caused by individual abnormal data points.

[0148] Through the above technical solution, this application can significantly improve the accuracy and stability of reference information. By eliminating outliers that may exist in images of qualified products, the reference information can more accurately capture the true normal characteristics of materials, thereby effectively reducing the false positive and false negative rates caused by reference deviations. In addition, this technical solution enhances robustness to minor and atypical changes in the production process, enabling it to maintain high-precision defect detection capabilities even in the face of complex and ever-changing production environments.

[0149] For example, suppose that after continuously acquiring image information of products deemed qualified, multiple local visual features are extracted from these images, such as the average grayscale value and texture contrast of a certain region. These feature values ​​are collected and form a dataset. To identify outliers, an interquartile range (IQR)-based method can be used. First, Q1 (first quartile) and Q3 (third quartile) are calculated for each feature, and IQR = Q3 - Q1. Then, any feature value that is less than Q1 - 1.5 * IQR or greater than Q3 + 1.5 * IQR is marked as an outlier. For example, if the texture contrast of an image of a qualified product is significantly higher or lower than that of most qualified products in a certain local region, it will be identified as an outlier and excluded from the dataset. After excluding these outliers, the remaining local visual feature data is used to update the baseline information in a progressive manner, such as exponentially weighted moving average (EWMA). For example, a new baseline value can be obtained by weighting the old baseline value with the current outlier-free data. The weighting factor can be adjusted according to actual needs to balance the response speed to new data and the degree of retention of historical data. In this way, the baseline information can smoothly and accurately adapt to normal changes in material properties while effectively resisting the interference of outlier data points.

[0150] In some embodiments, the step of quantifying the geometry, size, and intensity of the abnormal region, combining it with a preset defect pattern for similarity matching and multiple verifications, and determining whether the abnormal region is a real defect includes:

[0151] The visual representation of the abnormal region is quantified in multiple dimensions to obtain the first set of quantified information;

[0152] The first set of quantitative information is processed to obtain a second set of quantitative information that reflects the prominence of the abnormal region in different visual dimensions.

[0153] Based on the second set of quantification information, one or more candidate defect patterns that are similar to the quantification features of the abnormal region are identified in the preset defect pattern set.

[0154] Based on one or more identified candidate defect patterns, obtain their corresponding discrimination rules;

[0155] According to the discrimination rule, the abnormal region is subjected to multiple verifications to determine whether the abnormal region is a real defect; and according to the result of the multiple verifications, the defect type of the abnormal region is output.

[0156] Specifically, extracting a series of quantifiable visual features from identified anomalous regions to obtain the first set of quantifiable information refers to the following: These features can include geometric features, such as the area, perimeter, aspect ratio, roundness, and convexity of the anomalous region; they can also include grayscale or color features, such as the average grayscale value, grayscale standard deviation, skewness, and kurtosis of the anomalous region; and they can also include texture features, such as Local Binary Pattern (LBP) or Gray-Level Co-occurrence Matrix (GLCM). This multi-dimensional quantifiable information aims to comprehensively and objectively describe the visual characteristics of the anomalous region.

[0157] The first set of quantitative information is processed to obtain a second set of quantitative information reflecting the prominence of the abnormal region across different visual dimensions. This can be understood as optimizing and refining the original quantitative data. This processing may involve techniques such as feature normalization, feature selection, or feature dimensionality reduction to eliminate dimensional differences between different features, highlight features with higher discriminative power for defect identification, and reduce redundant information. Therefore, the resulting second set of quantitative information can more effectively characterize the key visual features of the abnormal region, providing more accurate input for subsequent defect pattern matching.

[0158] In practical applications, the processed features of the abnormal region are compared with a pre-established defect pattern library. This defect pattern set contains typical quantitative feature models or ranges of various known defect types (such as scratches, indentations, foreign objects, bubbles, etc.). By calculating similarity (e.g., Euclidean distance, cosine similarity, or matching through a classifier), one or more defect types that are closest to the current abnormal region can be initially screened as candidates.

[0159] Furthermore, based on the identified one or more candidate defect patterns, the corresponding discrimination rules are obtained. This means that for each initially matched candidate defect pattern, its own specific and more refined judgment criteria are invoked. These discrimination rules may be a set of logical conditions, threshold ranges, or more complex machine learning models, designed to make the final confirmation of a specific defect type.

[0160] Based on this, detailed quantitative features of the abnormal region are substituted into the acquired discrimination rules for multi-faceted cross-validation to confirm whether the abnormal region is a real defect. For example, an abnormal region initially identified as a "scratch" may need to simultaneously meet multiple conditions such as "length greater than X", "width less than Y", and "contrast greater than Z" before it can be finally confirmed as a real scratch. This multi-validation mechanism can effectively eliminate misjudgments caused by single feature similarity and significantly improve the accuracy of defect identification.

[0161] Finally, once the abnormal area passes multiple verifications of all discrimination rules, its defect type will be determined and output as the detection result. This includes not only the defect type but also detailed information such as its location, size, and severity, providing data support for subsequent quality control and process improvement.

[0162] The above technical solution enables accurate classification and judgment of abnormal areas in pouch lithium battery packaging images, significantly improving the accuracy and reliability of defect detection. This combination of multi-dimensional quantification, similarity matching, and multiple verifications effectively distinguishes between real defects and non-defect visual fluctuations, greatly reducing false alarm and false negative rates. This improves robustness and practicality, ensuring strict control over pouch lithium battery packaging quality and providing solid technical support for the automation and intelligentization of the production process.

[0163] In some embodiments, the step of performing multi-dimensional quantization of the visual representation of the abnormal region to obtain a first set of quantization information includes:

[0164] Statistical analysis of pixel grayscale values ​​is performed on the abnormal regions, and the average grayscale, standard deviation, skewness, and kurtosis of the abnormal regions are calculated.

[0165] Edge detection is performed on the abnormal region, and the perimeter, area, aspect ratio, circularity, and convexity of the abnormal region are calculated.

[0166] The first set of quantization information is obtained based on the average gray level, standard deviation, skewness, kurtosis, perimeter, area, aspect ratio, roundness, and convexity.

[0167] The process involves statistical analysis of pixel grayscale values ​​in abnormal regions to capture their visual characteristics from a brightness distribution perspective. Average grayscale reflects the overall brightness level of the region; standard deviation characterizes the dispersion of pixel grayscale values ​​within the region, which can be used to assess texture uniformity; skewness describes the asymmetry of the grayscale distribution, indicating the concentration trend of bright or dark pixels; and kurtosis reflects the sharpness of the grayscale distribution, which can be used to identify extreme brightness values. Furthermore, edge detection is performed on abnormal regions to capture their structural characteristics from a geometric perspective. Perimeter and area directly reflect the size of the abnormal region; aspect ratio describes the region's direction and shape; roundness measures how closely the region's shape resembles a circle; and convexity characterizes the degree of concavity or convexity of the region's boundaries. Thus, by combining the various indicators obtained from the above pixel grayscale value statistics and edge detection, a first set of quantitative information can be obtained. This first set of quantitative information comprehensively and multidimensionally describes the visual performance of abnormal regions, providing a solid data foundation for subsequent defect identification and classification.

[0168] The above technical solutions significantly improve the accuracy and robustness of identifying abnormal regions in pouch lithium battery packaging images. The multi-dimensional quantization method provides a more comprehensive feature description of abnormal regions, avoiding information loss or misjudgment that may result from a single feature description. This not only helps to more accurately distinguish between real defects and background noise or normal material fluctuations, but also provides more refined quantization fingerprints for different types of defects, thereby improving the success rate of defect pattern matching and ultimately achieving more efficient and reliable sealed adaptive pouch lithium battery packaging image detection.

[0169] This application also proposes a sealed adaptive soft-pack lithium battery packaging image detection system, such as... Figure 2 As shown, a sealed adaptive soft-pack lithium battery packaging image inspection system 100 is provided, the system comprising:

[0170] The image acquisition and feature judgment module 10 is used to acquire image information of the packaging area of ​​the soft-pack lithium battery to be detected, perform dynamic background maintenance and fluctuation analysis processing on the image information to obtain feature information reflecting the overall visual performance of the product, and determine whether a material batch switch has occurred based on the changes in the feature information.

[0171] The benchmark construction module 20 is used to collect image samples of the current product when it is determined that a material batch switch has occurred, extract local visual features reflecting the optical performance of the material surface from the image samples, and perform statistical analysis on the local visual features to construct a benchmark information representing the normal visual performance of the current product material.

[0172] The anomaly identification module 30 is used to acquire the current image information of the product to be detected, extract its local visual features, and compare the local visual features with the reference information to identify abnormal areas that deviate from the reference information.

[0173] The defect judgment module 40 is used to quantify the geometric shape, size and intensity of the abnormal area, and perform similarity matching and multiple verifications in combination with a preset defect pattern to determine whether the abnormal area is a real defect.

[0174] This application's system, through its modular design, organically combines dynamic background maintenance, material batch switching judgment, adaptive benchmark construction, and a multi-verification defect judgment mechanism, forming a highly intelligent detection closed loop. The image acquisition and characteristic judgment module can perceive changes in the production environment in real time and proactively identify material batch switching, avoiding misjudgments caused by the inability of traditional systems to recognize such changes. The benchmark construction module enables the system to dynamically update its "qualified" standards based on the actual characteristics of the new batch of materials, rather than adhering to old, potentially inapplicable standards. This adaptive capability significantly shortens the system's adaptation time to new materials, ensuring the continuity and stability of the production line. Furthermore, the collaborative work of the anomaly identification module and the defect judgment module, through multi-dimensional quantification and multiple verifications, further improves the accuracy of defect judgment, effectively distinguishing between genuine defects and visual interference caused by other non-defect factors.

[0175] In summary, the sealed adaptive soft-pack lithium battery packaging image inspection system of this application provides a robust and efficient solution through its unique structural and functional design. It effectively solves the misjudgment problem caused by batch differences in materials in the prior art, significantly improves the accuracy and automation level of the inspection system, and thus provides strong technical support for the production quality control of soft-pack lithium batteries.

[0176] 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 sealed adaptive soft-pack lithium battery packaging images, characterized in that, include: Image information of the packaging area of ​​the soft-pack lithium battery to be detected is acquired, and dynamic background maintenance and fluctuation analysis processing are performed on the image information to obtain feature information reflecting the overall visual performance of the product. Based on the changes in the feature information, it is determined whether a material batch switch has occurred. When a material batch switch is detected, image samples of the current product are collected, local visual features reflecting the optical performance of the material surface are extracted from the image samples, and statistical analysis is performed on the local visual features to construct a benchmark information representing the normal visual performance of the current product material. The current image information of the product to be inspected is acquired, its local visual features are extracted, and the local visual features are compared with the reference information to identify abnormal areas that deviate from the reference information; The geometry, size, and intensity of the abnormal region are quantified, and similarity matching and multiple verifications are performed in combination with a preset defect pattern to determine whether the abnormal region is a real defect.

2. The sealed adaptive soft-pack lithium battery packaging image detection method according to claim 1, characterized in that, After the step of quantifying the geometry, size, and intensity of the abnormal region, performing similarity matching and multiple verifications based on a preset defect pattern, and determining whether the abnormal region is a real defect, the following steps are included: During the testing process, image information of products deemed qualified is continuously acquired, and the reference information is adjusted progressively based on the changes in material properties reflected in the image information of products deemed qualified.

3. The sealed adaptive soft-pack lithium battery packaging image detection method according to claim 1, characterized in that, The step of determining whether a material batch switch has occurred based on the changes in the feature information includes: Maintain the short-term adaptive background and long-term stable background of the aforementioned feature information; Based on the image information of the products that have been determined to be qualified, the short-term adaptive background is updated quickly, and the long-term stable background is updated slowly. The feature information is compared with the short-term adaptive background. When the initial deviation exceeds a preset threshold, the overall deviation between the feature information and the long-term stable background is determined. When the feature information deviates significantly from the overall long-term stable background, it is determined that a material batch switch has occurred; and after determining that a material batch switch has occurred, the short-term adaptive background is quickly adjusted.

4. The sealed adaptive soft-pack lithium battery packaging image detection method according to claim 1, characterized in that, The steps of acquiring image samples of the current product, extracting local visual features reflecting the optical performance of the material surface from the image samples, and performing statistical analysis on the local visual features to construct a benchmark information representing the normal visual performance of the current product material include: Continuously acquire image samples of the current product; The image samples are initially screened for defects to obtain target image samples, so as to identify and exclude image samples containing serious defects; Local visual features are extracted from the target image sample, statistical analysis is performed on the local visual features, and the benchmark information is constructed based on the statistical analysis results.

5. The sealed adaptive soft-pack lithium battery packaging image detection method according to claim 1, characterized in that, The step of performing dynamic background maintenance and fluctuation analysis on the image information to obtain feature information reflecting the overall visual performance of the product, and determining whether a material batch change has occurred based on the changes in the feature information, includes: Dynamic background maintenance and fluctuation analysis are performed on the image information to obtain feature information that reflects the overall visual performance of the product, and the short-term fluctuation range and long-term stable range of the feature information are maintained. Determine whether the short-term fluctuation range of the feature information in a continuous image sequence continuously exceeds the long-term stable range. When the short-term fluctuation range continuously exceeds the long-term stable range, determine whether the deviation between the feature information and the long-term stable range shows a convergence trend within a preset time. When the short-term fluctuation range continues to exceed the long-term stable range, and the deviation does not show a convergence trend within a preset time, it is determined that a momentary or local change has occurred in the production environment. When the short-term fluctuation range continues to exceed the long-term stable range, and the deviation shows a convergence trend within a preset time, it is determined to be a material batch switch.

6. The sealed adaptive soft-pack lithium battery packaging image detection method according to claim 4, characterized in that, The steps of extracting local visual features from the target image sample and performing statistical analysis on the local visual features include: Identify the distribution pattern of the local visual features and determine whether they exhibit multiple clustering trends; For each central tendency, determine its internal normal fluctuation range; Determine whether a single feature value in the local visual features falls outside the normal fluctuation range of any central tendency; and if the single feature value falls outside the normal fluctuation range of all central tendencies, mark it as an outlier and suppress it.

7. The sealed adaptive soft-pack lithium battery packaging image detection method according to claim 2, characterized in that, The step of continuously acquiring image information of products deemed qualified during the detection process, and progressively adjusting the reference information based on changes in material properties reflected in the image information of qualified products, includes: Continuously acquire image information of products that have been determined to be qualified; Local visual features are extracted from the image information of products that have been determined to be qualified. Perform statistical analysis on the local visual features to identify outliers in the statistical analysis; The outliers are excluded from the local visual features, and the reference information is adjusted progressively based on the changes in material properties reflected in the local visual features after excluding the outliers.

8. The sealed adaptive soft-pack lithium battery packaging image detection method according to claim 1, characterized in that, The step of quantifying the geometry, size, and intensity of the abnormal region, combining it with a preset defect pattern for similarity matching and multiple verifications, and determining whether the abnormal region is a real defect includes: The visual representation of the abnormal region is quantified in multiple dimensions to obtain the first set of quantified information; The first set of quantitative information is processed to obtain a second set of quantitative information that reflects the prominence of the abnormal region in different visual dimensions. Based on the second set of quantification information, one or more candidate defect patterns that are similar to the quantification features of the abnormal region are identified in the preset defect pattern set. Based on one or more identified candidate defect patterns, obtain their corresponding discrimination rules; According to the discrimination rule, the abnormal region is subjected to multiple verifications to determine whether the abnormal region is a real defect; and according to the result of the multiple verifications, the defect type of the abnormal region is output.

9. The sealed adaptive soft-pack lithium battery packaging image detection method according to claim 8, characterized in that, The step of performing multi-dimensional quantization of the visual representation of the abnormal region to obtain the first set of quantified information includes: Statistical analysis of pixel grayscale values ​​is performed on the abnormal regions, and the average grayscale, standard deviation, skewness, and kurtosis of the abnormal regions are calculated. Edge detection is performed on the abnormal region, and the perimeter, area, aspect ratio, circularity, and convexity of the abnormal region are calculated. The first set of quantization information is obtained based on the average gray level, standard deviation, skewness, kurtosis, perimeter, area, aspect ratio, roundness, and convexity.

10. A sealed adaptive soft-pack lithium battery packaging image detection system, characterized in that, The system includes: The image acquisition and feature judgment module is used to acquire image information of the packaging area of ​​the soft-pack lithium battery to be detected, perform dynamic background maintenance and fluctuation analysis on the image information to obtain feature information reflecting the overall visual performance of the product, and determine whether a material batch switch has occurred based on the changes in the feature information. The benchmark construction module is used to collect image samples of the current product when it is determined that a material batch switch has occurred, extract local visual features reflecting the optical performance of the material surface from the image samples, and perform statistical analysis on the local visual features to construct a benchmark information representing the normal visual performance of the current product material. An anomaly detection module is used to acquire the current image information of the product to be detected, extract its local visual features, and compare the local visual features with the reference information to identify abnormal areas that deviate from the reference information. The defect judgment module is used to quantify the geometric shape, size and intensity of the abnormal area, and perform similarity matching and multiple verifications in combination with preset defect patterns to determine whether the abnormal area is a real defect.

Images