An intelligent packaging detection method and system

Abstract

The present invention relates to the field of computer vision technology, and specifically to an intelligent packaging inspection method and system, comprising the following steps: based on image preprocessing technology, feature extraction is employed, and multi-scale image information is processed through a filter set to generate a feature map set. In the present invention, the accuracy of feature maps is improved through image preprocessing and a multi-scale convolutional neural network. Structural feature extraction shifts packaging inspection from two-dimensional to three-dimensional. The application of a stereo vision algorithm enhances the accuracy of product packaging inspection. Autoencoders and generative adversarial networks identify more potential problems and improve the recognition of abnormal features. Time series analysis continuously monitors product quality, ensuring continuous quality control and enhancing sensitivity to potential problems. Transfer learning technology optimizes the model's update capability, increasing its flexibility and adaptability, thereby improving the efficiency and practicality of the inspection model.

Description

Technical Field

[0001] The present invention relates to the field of computer vision technology, and in particular to an intelligent packaging detection method and system. Background Art

[0002] Computer vision refers to the use of computers and corresponding algorithms to automatically acquire, analyze, and understand information in images or videos. This field involves fields such as image processing, pattern recognition, and machine learning, aiming to develop computer systems that can mimic and understand human visual capabilities.

[0003] Among them, the intelligent packaging inspection method and system utilizes computer vision technology to inspect packaging. Its purpose is to automatically inspect and evaluate the quality, integrity, and safety of packaging to improve production efficiency and reduce quality risks. To achieve this goal, the intelligent packaging inspection method and system uses image processing algorithms to pre-process and enhance packaging images. Next, feature extraction and pattern recognition techniques are used to analyze and identify the packaging, including its size, shape, logo, barcode, and so on. Machine learning and deep learning algorithms can also be used to classify, evaluate, and detect defects in packaging to determine whether it meets specified standards and requirements. Through these means, the intelligent packaging inspection method and system can achieve rapid and accurate inspection and evaluation of large quantities of packaging. This can improve the automation level of production lines and effectively reduce the incidence of human errors and quality issues. Furthermore, this technology can provide real-time monitoring and data analysis to assist companies in quality management and production optimization.

[0004] Traditional packaging quality inspection methods are mostly manual or semi-automated, resulting in low efficiency and persistent human error. Furthermore, traditional methods often fail to incorporate advanced image processing and deep learning technologies for feature extraction and anomaly identification, limiting their accuracy and inclusiveness. Furthermore, some traditional methods employ a phased approach to quality monitoring, failing to provide real-time quality tracking and early warning. Furthermore, traditional methods lack effective model optimization strategies, making them unable to adapt and optimize to new challenges and changes. Regarding three-dimensional analysis, due to technical and equipment limitations, traditional inspection methods are often unable to conduct in-depth and three-dimensional packaging inspections. Summary of the Invention

[0005] The purpose of the present invention is to solve the shortcomings of the prior art and to propose an intelligent packaging detection method and system.

[0006] In order to achieve the above object, the present invention adopts the following technical solution: an intelligent packaging detection method, comprising the following steps:

[0007] S1: Based on image preprocessing technology, feature extraction is used, and multi-scale image information is processed through a set of filters to generate a set of feature maps;

[0008] S2: Based on the feature graph set, a graph neural network algorithm is used to encode the product packaging structure, extract structural information, and generate a structural feature representation;

[0009] S3: Based on multi-view shooting technology and using stereo vision algorithms, it creates three-dimensional modeling of product packaging, performs depth analysis, and generates stereo vision analysis data;

[0010] S4: Based on the feature map set, an autoencoder or a generative adversarial network is used to learn normal features and identify abnormal features to generate an anomaly detection model;

[0011] S5: Based on the anomaly detection model and structural feature representation, a time series analysis method is used to continuously monitor product quality, perform trend judgment, and generate a quality monitoring report;

[0012] S6: Based on the stereoscopic vision analysis data and the quality monitoring report, the detection model is optimized and updated using transfer learning technology to generate a final detection model;

[0013] The feature map set specifically refers to product packaging image features extracted at multiple resolutions. The structural feature representation includes the overall composition of the product packaging and the association features between multiple components. The stereoscopic vision analysis data specifically refers to the three-dimensional model and depth information reconstructed based on packaging images captured from multiple angles. The anomaly detection model specifically refers to a deep learning model that identifies abnormal states based on the normal packaging feature distribution. The quality monitoring report specifically refers to quality change data and prediction trend chart based on time series.

[0014] As a further solution of the present invention, based on image preprocessing technology, feature extraction is adopted, and multi-scale image information is processed through a filter set to generate a feature map set. Specifically, the steps are as follows:

[0015] S101: Based on the image preprocessing technology, a z-score normalization algorithm is used to degenerate the illumination and scale influencing factors of the image to generate a preprocessed normalized image;

[0016] S102: Based on the preprocessed normalized image, a multi-scale convolutional neural network technology is used to perform a convolution operation through a convolution kernel to extract features and generate a convolution feature map;

[0017] S103: Based on the convolution feature map, using the ReLU activation function algorithm to enhance the nonlinear expression ability of the feature map and generate an activation feature map;

[0018] S104: Based on the activated feature map, a filter set is used to perform filtering optimization on the extracted multi-scale image information to enhance the difference and expressiveness of the characteristics, and generate a feature map set.

[0019] As a further solution of the present invention, based on the feature graph set, a graph neural network algorithm is used to encode the product packaging structure and extract structural information. The steps of generating a structural feature representation are specifically as follows:

[0020] S201: Based on the feature graph set, using a dimensionality reduction algorithm, convert it into a node matrix for describing product packaging structural features;

[0021] S202: Based on the node matrix, a graph neural network algorithm is used to calculate the association between nodes and generate an adjacency matrix;

[0022] S203: Based on the adjacency matrix and the node matrix, a convolutional neural network algorithm is used to aggregate and transform the feature vector information to generate encoded structural information;

[0023] S204: Based on the encoded structural information, a deconvolution algorithm is used to perform feature decoding, highlight the structural features of the product packaging, and establish a structural feature representation.

[0024] As a further solution of the present invention, based on multi-view shooting technology, a stereo vision algorithm is used to perform three-dimensional modeling of product packaging and perform depth analysis. The steps for generating stereo vision analysis data are as follows:

[0025] S301: Based on the multi-view shooting technology, using high-definition shooting equipment, shoot the product packaging from multiple angles to obtain multi-view image data;

[0026] S302: Calculating image depth based on the multi-view image data using a stereo vision algorithm to generate a depth map;

[0027] S303: Based on the depth map, a regression analysis algorithm is used to perform three-dimensional reconstruction, and a 3D model based on the product packaging is constructed to generate a 3D model;

[0028] S304: Based on the 3D model, a deep learning algorithm is used to optimize and analyze the model, present a three-dimensional visual effect of the product packaging, and generate three-dimensional visual analysis data.

[0029] As a further solution of the present invention, based on the feature graph set, an autoencoder or a generative adversarial network is used to learn normal features and identify abnormal features. The steps of generating an anomaly detection model are specifically as follows:

[0030] S401: Based on the feature map set, a dimensionality reduction algorithm is used to compress dense features, solve the curse of dimensionality problem, extract important features, and generate a mapping parameter set;

[0031] S402: Based on the mapping parameter set, using an autoencoder to compress and encode the input data and decode and restore it, by minimizing the difference between the input and the output, learning the main features in the data and generating a normal feature model;

[0032] S403: Based on the normal feature model, a generative adversarial network is used to train normal features. The generative adversarial network training of normal features includes a generator and a discriminator. The generator generates a simulated feature map, and the discriminator determines whether the feature map is real. After multiple rounds of training, a trained network model is generated.

[0033] S404: Based on the trained network model, a threshold method is used to identify abnormal features, screen features that deviate from the normal range, mark them as abnormal, and generate an abnormality detection model.

[0034] As a further solution of the present invention, based on the anomaly detection model and structural feature representation, a time series analysis method is used to continuously monitor product quality and perform trend judgment. The steps of generating a quality monitoring report are specifically as follows:

[0035] S501: Based on the anomaly detection model and the structural feature representation, evaluate the correlation between features and construct a matrix to generate a correlation coefficient matrix;

[0036] S502: Based on the correlation coefficient matrix, a K-means clustering algorithm is used to perform classification analysis on the data, classify data with similar characteristics into the same category, and generate a classification result;

[0037] S503: Based on the classification results, use time series analysis methods to track data changes over a long period of time, analyze data change trends, and generate a trend judgment report;

[0038] S504: Based on the trend judgment report, a comprehensive evaluation method is adopted to combine multiple evaluation indicators, continuously monitor product quality, and generate a quality monitoring report.

[0039] As a further solution of the present invention, based on the stereoscopic vision analysis data and quality monitoring report, the detection model is optimized and updated using transfer learning technology to generate the final detection model. Specifically, the steps are as follows:

[0040] S601: Based on the stereoscopic vision analysis data and the quality monitoring report, using a label propagation algorithm, propagating the labeled category information to the unlabeled data, completing the classification of the unlabeled data, and generating an optimized label set;

[0041] S602: Based on the optimized label set, a convolutional neural network is used to perform feature learning, capture image data features, and generate feature mapping data;

[0042] S603: Based on the feature map data, use the optical flow method to calculate the information of the object movement in the image, evaluate the accuracy of the current model, and generate a model evaluation result;

[0043] S604: Based on the model evaluation results, transfer learning technology is used to adapt the pre-trained deep model to the task environment, and the model is updated to generate a final detection model.

[0044] An intelligent packaging detection system is used to execute the above-mentioned intelligent packaging detection method. The system includes a preprocessing module, a feature extraction module, a three-dimensional model construction module, an anomaly detection module, a trend analysis module, a classification optimization module, and a model update module.

[0045] As a further solution of the present invention, the preprocessing module is based on the image and uses a z-score normalization method to perform normalization to eliminate interference factors including illumination and scale, and generate a preprocessed normalized image;

[0046] The feature extraction module uses a graph neural network algorithm to extract features based on the preprocessed normalized image, and performs dimension compression to generate a structural feature representation;

[0047] The 3D model building module uses a stereo vision algorithm to estimate depth based on multi-angle visual data and performs 3D reconstruction through regression analysis to generate a 3D model of the product packaging.

[0048] The anomaly detection module is based on a feature map set, uses a dimensionality reduction algorithm to compress dense features, performs feature learning through an autoencoder, and uses a generative adversarial network for network training to generate an anomaly detection model;

[0049] The trend analysis module classifies data based on anomaly detection models and structural feature representations through correlation analysis and clustering algorithms, and then uses time series analysis to analyze the trend of data changes and form a quality trend report;

[0050] The classification optimization module uses a label propagation algorithm to classify unlabeled data based on stereoscopic vision analysis data and quality monitoring reports, and performs feature learning through a convolutional neural network to obtain an optimized label set and feature mapping data;

[0051] The model updating module uses the optical flow method to evaluate the accuracy of the model based on the feature mapping data, and updates the model through the transfer learning method to form the final detection model.

[0052] As a further solution of the present invention, the preprocessing module includes an image normalization submodule, a convolution feature extraction submodule, a nonlinear enhancement submodule, and an optimization filtering submodule;

[0053] The feature extraction module includes a feature map generation submodule, a node matrix calculation submodule, a structure information encoding submodule, and a feature decoding submodule;

[0054] The three-dimensional model construction module includes a multi-view image acquisition submodule, a depth map generation submodule, a 3D model generation submodule, and a model optimization and analysis submodule;

[0055] The anomaly detection module includes a feature compression submodule, a network training submodule, and an abnormal feature recognition submodule;

[0056] The trend analysis module includes a correlation assessment submodule, a cluster classification submodule, a trend judgment submodule, and a quality monitoring submodule;

[0057] The classification optimization module includes a label propagation submodule, an unlabeled data classification submodule, and a feature learning submodule;

[0058] The model updating module includes a model evaluation submodule, a data preprocessing submodule, and a model updating submodule.

[0059] Compared with the prior art, the advantages and positive effects of the present invention are:

[0060] In the present invention, the generation of feature maps is made more accurate and vivid through image preprocessing and the use of multi-scale convolutional neural networks. The extraction and representation of structural features enable traditional packaging inspection work to shift from plane to three-dimensional. Three-dimensional modeling and depth analysis based on stereo vision algorithms improve the accuracy of product packaging inspection. Abnormal feature learning and recognition using autoencoders or generative adversarial networks enable packaging inspection to identify more potential problems. Continuous quality control is achieved through continuous monitoring of product quality through time series analysis, which improves sensitivity to potential problems and timeliness in problem solving. The use of transfer learning technology optimizes the model's update capability, making the detection model more flexible and adaptable, and improving the efficiency and practicality of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 It is a schematic diagram of the workflow of the present invention;

[0062] Figure 2 This is a detailed flow chart of S1 of the present invention;

[0063] Figure 3 This is a detailed flow chart of S2 of the present invention;

[0064] Figure 4 This is a detailed flow chart of S3 of the present invention;

[0065] Figure 5 This is a detailed flow chart of S4 of the present invention;

[0066] Figure 6 This is a detailed flow chart of S5 of the present invention;

[0067] Figure 7 This is a detailed flow chart of S6 of the present invention;

[0068] Figure 8 is a system flow chart of the present invention;

[0069] Figure 9 Schematic diagram of the system framework of the present invention. DETAILED DESCRIPTION

[0070] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0071] In the description of the present invention, it should be understood that the terms "length," "width," "up," "down," "front," "back," "left," "right," "vertical," "horizontal," "top," "bottom," "inside," "outside," and the like, indicating positions or relationships, are based on the positions or relationships shown in the accompanying drawings and are intended only to facilitate the description of the present invention and simplify the description. They do not indicate or imply that the devices or elements referred to must have a specific orientation, be constructed, or operate in a specific orientation. Therefore, they should not be construed as limiting the present invention. Furthermore, in the description of the present invention, "plurality" means two or more, unless otherwise expressly and specifically defined.

[0072] Example 1

[0073] See also Figure 1 The present invention provides a technical solution: an intelligent packaging detection method, comprising the following steps:

[0074] S1: Based on image preprocessing technology, feature extraction is used, and multi-scale image information is processed through a set of filters to generate a set of feature maps;

[0075] S2: Based on the feature graph set, a graph neural network algorithm is used to encode the product packaging structure, extract structural information, and generate a structural feature representation;

[0076] S3: Based on multi-view shooting technology and using stereo vision algorithms, it creates three-dimensional modeling of product packaging, performs depth analysis, and generates stereo vision analysis data;

[0077] S4: Based on the feature graph set, an autoencoder or generative adversarial network is used to learn normal features and identify abnormal features to generate an anomaly detection model;

[0078] S5: Based on the anomaly detection model and structural feature representation, time series analysis methods are used to continuously monitor product quality, determine trends, and generate quality monitoring reports;

[0079] S6: Based on the stereo vision analysis data and quality monitoring reports, transfer learning technology is used to optimize the detection model, update the model, and generate the final detection model;

[0080] The feature map set specifically refers to the product packaging image features extracted at multiple resolutions. The structural feature representation includes the overall composition of the product packaging and the correlation features between multiple components. The stereoscopic vision analysis data specifically refers to the three-dimensional model and depth information reconstructed based on the packaging images captured from multiple angles. The anomaly detection model specifically refers to a deep learning model that identifies abnormal states based on the normal packaging feature distribution. The quality monitoring report specifically refers to the quality change data and prediction trend chart based on time series.

[0081] First, multi-scale information processing is performed based on image preprocessing technology. Feature extraction and filter sets are used to generate a feature map set, providing high-quality input for subsequent analysis. This helps capture the detailed features of product packaging and lays the foundation for further processing.

[0082] Secondly, a graph neural network algorithm is used to encode the product packaging structure and, combined with structural information extraction, generate a structural feature representation. This step provides a deep understanding of the overall composition of the product packaging and the interrelated features of its components, helping to improve the accuracy of product structure identification and analysis. Multi-view photography technology and stereo vision algorithms are used for 3D modeling and in-depth analysis, generating stereoscopic visual analysis data to enhance understanding of product packaging. This helps capture the visual characteristics of the product packaging and enhances in-depth understanding of the packaging form. Using an autoencoder or generative adversarial network to learn normal features and identify abnormal features, generating an anomaly detection model, the system can effectively identify abnormal conditions, improving the accuracy and reliability of product quality inspection. This helps to proactively identify potential quality issues, thereby reducing defective product rates. Based on the anomaly detection model and structural feature representation, time series analysis methods are used to continuously monitor product quality, identify trends, and generate quality monitoring reports. This continuous monitoring and trend analysis provides the system with real-time insight into quality changes, helping companies quickly address potential issues and improve production efficiency.

[0083] Finally, the detection model is optimized and updated using transfer learning technology to generate the final detection model. This ensures that the system can continuously adapt to changing environments and new packaging characteristics, maintaining a high level of detection accuracy. Overall, the comprehensive performance of this method gives it a significant advantage in product packaging quality monitoring. From image processing to deep learning, to time series analysis and transfer learning, the interplay of these components creates a powerful quality monitoring system that provides comprehensive support and assurance for the production process.

[0084] See also Figure 2 Based on image preprocessing technology, feature extraction is adopted, and multi-scale image information is processed through a set of filters. The steps to generate a feature map set are as follows:

[0085] S101: Based on image preprocessing technology, the z-score normalization algorithm is used to simplify the image's illumination and scale factors and generate a preprocessed normalized image;

[0086] S102: Based on the preprocessed normalized image, a multi-scale convolutional neural network technology is used to perform convolution operation through the convolution kernel to extract features and generate a convolution feature map;

[0087] S103: Based on the convolution feature map, the ReLU activation function algorithm is used to enhance the nonlinear expression ability of the feature map and generate an activation feature map;

[0088] S104: Based on the activated feature map, a filter set is used to perform filtering optimization on the extracted multi-scale image information to enhance the difference and expressiveness of the features and generate a feature map set.

[0089] First, the input raw image is preprocessed using the z-score normalization algorithm. This step aims to simplify the effects of lighting and scale in the image and generate a preprocessed normalized image. This helps improve the stability and consistency of image features in subsequent processing steps.

[0090] Next, using multi-scale convolutional neural network technology, convolution kernels are used to perform convolution operations on the preprocessed normalized image to extract image features. This step generates a convolution feature map that includes key feature information at different scales.

[0091] Then, by applying the ReLU activation function algorithm, the nonlinear expression ability of the convolution feature map is enhanced to generate an activation feature map. This helps capture richer abstract features in the image and improves the expressive power of the model.

[0092] Finally, a filter set is introduced to perform filter optimization on the activated feature map. This step aims to process multi-scale image information and enhance the diversity and expressiveness of features. Through filter optimization, a feature map set is generated, which includes features that refine and enhance multi-scale image information.

[0093] During implementation, it's crucial to ensure the correct application of the z-score normalization algorithm during preprocessing, the proper architecture and parameter settings of the convolutional neural network, the appropriate use of the ReLU activation function, and the desired performance of the filter set design and application. This involves training and tuning the deep learning model to ensure optimal feature extraction and optimization throughout the image processing pipeline. At each step, careful experimentation and verification are required to ensure the reliability and effectiveness of the solution, including parameter selection and operational detail adjustments.

[0094] See also Figure 3 Based on the feature graph set, the graph neural network algorithm is used to encode the product packaging structure and extract structural information. The specific steps of generating structural feature representation are as follows:

[0095] S201: Based on the feature graph set, a dimensionality reduction algorithm is used to convert it into a node matrix for describing the structural features of the product packaging;

[0096] S202: Based on the node matrix, a graph neural network algorithm is used to calculate the correlation between nodes and generate an adjacency matrix;

[0097] S203: Based on the adjacency matrix and the node matrix, a convolutional neural network algorithm is used to aggregate and transform the feature vector information to generate encoded structural information;

[0098] S204: Based on the encoded structural information, a deconvolution algorithm is used to perform feature decoding, highlight the structural features of the product packaging, and establish a structural feature representation.

[0099] First, based on the acquired feature graph set, a dimensionality reduction algorithm is used to convert it into a node matrix to describe the characteristics of the product packaging structure. This node matrix can be regarded as an abstract representation of the packaging structure characteristics, providing a foundation for subsequent graph neural network processing.

[0100] Next, based on the resulting node matrix, a graph neural network algorithm is applied to calculate the degree of association between nodes, generating an adjacency matrix. This step helps characterize the connections and associations between different nodes, thereby constructing topological information about the product packaging structure.

[0101] Next, a convolutional neural network algorithm is used to aggregate and transform the feature vectors, combining the generated adjacency matrix and node matrix. This helps generate encoded structural information, thereby condensing and extracting the structural features of the product packaging. This step can be considered an abstraction of structural information and deep learning processing to obtain a higher-level, more meaningful feature representation.

[0102] Finally, based on the encoded structural information, a deconvolution algorithm is used to decode the features. This step helps restore and highlight the structural characteristics of the product packaging and establish a structural feature representation. The application of deconvolution allows interpretable and visual features to be regenerated from high-level, abstract structural information, making the structural features clearer and more understandable.

[0103] In practice, ensuring the correct selection and application of the dimensionality reduction algorithm, the graph neural network model design is consistent with the product packaging structure, the convolutional neural network parameters are appropriately set and trained, and the deconvolution algorithm accurately restores feature information is crucial. Parameter adjustment, network structure optimization, and model training are key steps to ensure successful implementation. Experimentation and verification will be crucial to confirm the effectiveness and accuracy of the method.

[0104] See also Figure 4 Based on multi-view shooting technology and stereo vision algorithm, the product packaging is 3D modeled and depth analyzed. The steps to generate stereo vision analysis data are as follows:

[0105] S301: Based on multi-view shooting technology, using high-definition shooting equipment, shoot the product packaging from multiple angles to obtain multi-view image data;

[0106] S302: Based on the multi-view image data, a stereo vision algorithm is used to calculate the image depth and generate a depth map;

[0107] S303: Based on the depth map, a regression analysis algorithm is used to perform three-dimensional reconstruction, a 3D model based on the product packaging is constructed, and a 3D model is generated;

[0108] S304: Based on the 3D model, a deep learning algorithm is used to optimize and analyze the model, present a three-dimensional visual effect of the product packaging, and generate three-dimensional visual analysis data.

[0109] First, using high-definition cameras, the product packaging is photographed from multiple angles to acquire multi-view image data. This step ensures sufficient perspective information so that the subsequent stereo vision algorithm can accurately calculate the image depth.

[0110] Next, based on the acquired multi-view image data, a stereo vision algorithm is used to calculate the image depth and generate a depth map. This depth map includes the distance information of each pixel from the camera, providing the basis for subsequent 3D reconstruction.

[0111] Then, using a regression analysis algorithm, a 3D reconstruction is performed based on the depth map to build a 3D model of the product packaging. This step involves converting the depth map information into 3D coordinates, thereby restoring the true shape and structure of the product packaging.

[0112] Finally, based on the resulting 3D model, deep learning algorithms are used for model optimization and analysis. This helps improve the model's accuracy and ability to express details, presenting a more realistic and three-dimensional visual effect for the product packaging. The generated stereoscopic visual analysis data can include model rotation and magnification operations to provide a comprehensive viewing experience.

[0113] In practice, it's crucial to ensure the high-definition performance of the camera equipment, select appropriate stereo vision and regression analysis algorithms, and properly design and train deep learning models. Parameter adjustment and model verification are crucial steps in the 3D reconstruction and model optimization process to ensure the resulting stereo vision analysis data is highly accurate and realistic. Experimentation and testing will be essential to confirm the effectiveness of the proposed solution.

[0114] See also Figure 5 Based on the feature graph set, an autoencoder or generative adversarial network is used to learn normal features and identify abnormal features. The specific steps for generating an anomaly detection model are as follows:

[0115] S401: Based on the feature map set, a dimensionality reduction algorithm is used to compress dense features, solve the curse of dimensionality problem, extract important features, and generate a mapping parameter set;

[0116] S402: Based on the mapping parameter set, the autoencoder is used to compress and encode the input data and decode and restore it, and the main features in the data are learned by minimizing the difference between the input and the output to generate a normal feature model;

[0117] S403: Based on the normal feature model, a generative adversarial network is used to train normal features. The generative adversarial network training of normal features includes a generator and a discriminator. The generator generates a simulated feature map, and the discriminator determines whether the feature map is real. After multiple rounds of training, a trained network model is generated.

[0118] S404: Based on the trained network model, a threshold method is used to identify abnormal features, screen features that deviate from the normal range, mark them as abnormal, and generate an anomaly detection model.

[0119] First, based on the feature map set, a dimensionality reduction algorithm is used to compress the dense features to address the curse of dimensionality, extract important features, and generate a set of mapping parameters. The goal of this step is to reduce the dimensionality of the data while retaining the key information to improve the efficiency of the model and reduce computational complexity.

[0120] Next, based on the resulting mapping parameter set, an autoencoder is used to learn normal features. An autoencoder is a neural network architecture that compresses and encodes input data and decodes it back to normal, learning the key features of the data by minimizing the difference between input and output. This step generates a normal feature model, which includes both the encoder and decoder components.

[0121] Then, based on the normal feature model, a generative adversarial network (GAN) is used to train normal features. A GAN consists of a generator and a discriminator. The generator is responsible for generating simulated feature maps, while the discriminator is responsible for determining whether the feature maps are real. Through multiple rounds of training, the generator gradually produces more realistic simulated feature maps, while the discriminator continuously improves the accuracy of its judgment. Ultimately, the trained network model is obtained, in which the generator portion includes the ability to generate the learned normal features.

[0122] Finally, based on the trained network model, a thresholding method is used to identify abnormal features. By comparing the generated feature map with the real data, features that deviate from the normal range are screened and marked as abnormal. This process allows the threshold to be adjusted according to the needs of the specific application scenario to balance precision and recall, thereby generating an anomaly detection model.

[0123] In practice, it is necessary to ensure the appropriate selection and parameter settings of the dimensionality reduction algorithm, as well as the architecture and hyperparameter settings of the autoencoder and generative adversarial network. Training the network model requires sufficient data and training time to obtain an accurate model of normal features. Finally, performance evaluation and adjustment of the anomaly detection model are also key steps to ensure the effectiveness of the solution. Experimentation and validation will be used to verify the accuracy and reliability of the anomaly detection model.

[0124] See also Figure 6 Based on the anomaly detection model and structural feature representation, the time series analysis method is used to continuously monitor product quality and make trend judgments. The specific steps for generating a quality monitoring report are as follows:

[0125] S501: Based on the anomaly detection model and structural feature representation, evaluate the correlation between features and construct a matrix to generate a correlation coefficient matrix;

[0126] S502: Based on the correlation coefficient matrix, a K-means clustering algorithm is used to perform classification analysis on the data, classify data with similar characteristics into the same category, and generate classification results;

[0127] S503: Based on the classification results, use time series analysis methods to track data changes over a long period of time, analyze data change trends, and generate a trend judgment report;

[0128] S504: Based on the trend judgment report, a comprehensive evaluation method is used to combine multiple evaluation indicators to continuously monitor product quality and generate a quality monitoring report.

[0129] First, based on the anomaly detection model and structural feature representation, we evaluate the correlation between features and construct a correlation coefficient matrix. This step aims to understand the degree of association between different features and provide a foundation for subsequent time series analysis. The resulting correlation coefficient matrix can reveal the underlying relationships between features.

[0130] Next, based on the resulting correlation matrix, we use the K-means clustering algorithm to classify the data. This helps group data with similar characteristics into the same category, generating a classification result. The goal of clustering is to discover patterns and structures within the data, making subsequent time series analysis more targeted.

[0131] Based on the classification results, time series analysis is then used to track data changes over time. By analyzing data trends, we can assess the long-term performance of product quality and detect potential anomalies or emerging trends. This step generates a trend analysis report that includes long-term data trends and anomalies.

[0132] Finally, based on the resulting trend assessment report, a comprehensive evaluation method is used to integrate multiple assessment indicators to continuously monitor product quality. This involves assigning weights to different trends and anomalies to more comprehensively assess the overall product quality performance. The resulting quality monitoring report provides a comprehensive assessment of long-term product quality trends.

[0133] In practice, the establishment of the correlation coefficient matrix and the application of the K-means clustering algorithm must be rational. The choice of time series analysis method must consider the characteristics of the data and the complexity of the problem. The weight distribution of the comprehensive evaluation method needs to be adjusted based on actual conditions to ensure the rationality of the comprehensive assessment of different factors. Experimentation and verification will be key steps in confirming the effectiveness of the implementation plan. This meticulous implementation ensures the full utilization of multiple technical means in product quality monitoring, thereby improving the overall performance of the system. From image processing to anomaly detection to time series analysis, each step is carefully designed to ensure effectiveness and reliability. In practice, timely data feedback and adjustments will optimize system performance. The meticulous consideration of the entire process makes this solution more feasible for practical application.

[0134] See also Figure 7 Based on the stereo vision analysis data and quality monitoring reports, the detection model is optimized and updated using transfer learning technology. The steps to generate the final detection model are as follows:

[0135] S601: Based on the stereoscopic vision analysis data and quality monitoring report, a label propagation algorithm is used to propagate the labeled category information to the unlabeled data, complete the classification of the unlabeled data, and generate an optimized label set;

[0136] S602: Based on the optimized label set, a convolutional neural network is used to perform feature learning, capture image data features, and generate feature mapping data;

[0137] S603: Based on the feature map data, use the optical flow method to calculate the information of the object movement in the image, evaluate the accuracy of the current model, and generate a model evaluation result;

[0138] S604: Based on the model evaluation results, transfer learning technology is used to adapt the pre-trained deep model to the task environment, and the model is updated to generate the final detection model.

[0139] First, based on stereo vision analysis data and quality control reports, a label propagation algorithm is used to propagate the labeled category information to the unlabeled data, completing the classification of the unlabeled data and generating an optimized label set. This step helps expand the influence of the labeled data, improves the model's generalization ability across the entire dataset, and ensures better adaptation to new task environments during transfer learning.

[0140] Next, based on the optimized label set, a convolutional neural network is used for feature learning. The goal of this stage is to capture the key features of the image data in order to more effectively describe and identify patterns in the data. The generated feature map data will become the key input for subsequent model updates.

[0141] Next, we use optical flow to calculate information about object motion in the image and evaluate the accuracy of the current model. This step helps us understand the model's performance when handling dynamic scenes and provides a benchmark for subsequent transfer learning. The resulting model evaluation results reflect the strengths and weaknesses of the current model.

[0142] Finally, based on the model evaluation results, transfer learning techniques are applied. Using the pre-trained deep learning model, the model is adapted to the task environment and updated. This can include fine-tuning model parameters, adjusting the network structure, or employing other transfer learning strategies to improve model performance. Ultimately, a final detection model optimized through transfer learning is generated.

[0143] In practice, we ensured the proper application of the label propagation algorithm, the optimization of the convolutional neural network structure and parameter settings, and the accurate calculation of the optical flow method. The choice of transfer learning required consideration of the similarity between the source and target tasks, as well as the choice of pre-trained models. Experimentation and verification were key steps in confirming the effectiveness of the proposed solution, involving cross-validation techniques to ensure the model's generalization capabilities. This meticulous consideration of the entire process made this solution highly feasible for practical application.

[0144] See also Figure 8 , an intelligent packaging detection system, the intelligent packaging detection system is used to execute the above-mentioned intelligent packaging detection method, the system includes a preprocessing module, a feature extraction module, a three-dimensional model construction module, an anomaly detection module, a trend analysis module, a classification optimization module, and a model update module.

[0145] The preprocessing module uses the z-score normalization method to standardize the image, eliminate interference factors including illumination and scale, and generate a preprocessed normalized image;

[0146] The feature extraction module uses a graph neural network algorithm to extract features based on the preprocessed normalized image, and performs dimension compression to generate a structural feature representation;

[0147] The 3D model building module uses stereo vision algorithms to estimate depth based on multi-angle visual data and performs 3D reconstruction through regression analysis to generate a 3D model of the product packaging.

[0148] The anomaly detection module is based on a feature map set, uses a dimensionality reduction algorithm to compress dense features, learns features through an autoencoder, and uses a generative adversarial network for network training to generate an anomaly detection model;

[0149] The trend analysis module classifies data based on anomaly detection models and structural feature representations through correlation analysis and clustering algorithms, and then uses time series analysis to analyze data change trends and form quality trend reports;

[0150] The classification optimization module uses the label propagation algorithm to classify unlabeled data based on stereo vision analysis data and quality monitoring reports, and performs feature learning through convolutional neural networks to obtain optimized label sets and feature mapping data;

[0151] The model update module uses the optical flow method to evaluate the accuracy of the model based on feature mapping data, and updates the model through the transfer learning method to form the final detection model.

[0152] First, the intelligent packaging inspection system utilizes a series of innovative modules, from preprocessing to model updating, to build a comprehensive and efficient quality control solution. Through the preprocessing module's z-score normalization, the system successfully eliminates illumination and scale interference in images, improving overall system reliability. The feature extraction module utilizes graph neural network algorithms and dimensionality compression to provide structural feature representations for subsequent analysis, enhancing the understanding and analysis of packaging characteristics. The 3D model construction module further utilizes stereo vision algorithms and 3D reconstruction to provide a new spatial dimension, enhancing comprehensive understanding of packaging form and structure.

[0153] Secondly, the anomaly detection module utilizes a dimensionality reduction algorithm and autoencoders, training the network through a generative adversarial network to establish a powerful anomaly detection model. This helps the system efficiently identify and model anomalies, ensuring a rapid response to packaging quality issues. The trend analysis module, based on the anomaly detection model and structural features, classifies data through correlation analysis and clustering algorithms, enabling timely monitoring and reporting of quality trends, providing crucial decision support for manufacturers.

[0154] Finally, the classification optimization module utilizes a label propagation algorithm and convolutional neural networks to process stereoscopic vision analysis data, improving the classification accuracy of unlabeled data and the effectiveness of feature mapping. The model update module continuously enhances model accuracy through optical flow and transfer learning, ensuring the system maintains high stability in a constantly changing environment. The integration of these key modules provides the packaging industry with a comprehensive, accurate, and efficient quality inspection solution, offering significant benefits and potential growth opportunities for production and manufacturing operations.

[0155] See also Figure 9 ,The preprocessing module includes image normalization submodule, convolution feature extraction submodule, nonlinear ,enhancement submodule, and optimization filtering submodule;

[0156] The feature extraction module includes a feature map generation submodule, a node matrix calculation submodule, a structure information encoding submodule, and a feature decoding submodule;

[0157] The 3D model construction module includes a multi-view image acquisition submodule, a depth map generation submodule, a 3D model generation submodule, and a model optimization and analysis submodule;

[0158] The anomaly detection module includes a feature compression submodule, a network training submodule, and an anomaly feature recognition submodule;

[0159] The trend analysis module includes a correlation assessment submodule, a cluster classification submodule, a trend judgment submodule, and a quality monitoring submodule;

[0160] The classification optimization module includes a label propagation submodule, an unlabeled data classification submodule, and a feature learning submodule;

[0161] The model update module includes a model evaluation submodule, a data preprocessing submodule, and a model update submodule.

[0162] Preprocessing Module: In intelligent packaging inspection systems, the preprocessing module is a critical initial step. Its submodules include image normalization, convolutional feature extraction, nonlinear enhancement, and optimized filtering. These steps work together to eliminate illumination and scale interference, extract key features, and improve image quality, laying a solid foundation for subsequent processing stages.

[0163] Feature Extraction Module: This module is responsible for deeply exploring abstract image features. It includes submodules for feature graph generation, node matrix calculation, structural information encoding, and feature decoding. Through graph neural networks, node matrix calculation, and structural encoding, this module generates high-dimensional feature representations, providing rich information for subsequent 3D model construction.

[0164] 3D Model Construction Module: Within the 3D Model Construction Module, submodules such as multi-view image acquisition, depth map generation, 3D model generation, and model optimization and analysis work together. This process acquires visual data from multiple angles, uses stereo vision algorithms to perform depth estimation, and ultimately generates an optimized 3D model, providing the system with accurate environmental information.

[0165] Anomaly Detection Module: This module is implemented through feature compression, feature learning, network training, and anomaly feature recognition submodules. Using autoencoders for feature learning and generative adversarial networks for model training, the system can accurately identify anomalies in images, providing a reliable basis for quality control.

[0166] Trend Analysis Module: This module encompasses submodules such as correlation assessment, cluster classification, trend determination, and quality monitoring. Through data correlation analysis, clustering algorithms, and time series analysis, the system can monitor dynamic changes in product packaging quality and determine trends, supporting timely quality adjustments.

[0167] Classification Optimization Module: The classification optimization module includes label propagation, unlabeled data classification, and feature learning submodules. Using stereoscopic visual analysis data and quality monitoring reports, the system uses a label propagation algorithm to classify unlabeled data. Finally, a convolutional neural network performs feature learning to optimize the classification results.

[0168] Model Update Module: This module is implemented through model evaluation, data preprocessing, and model update submodules. By using optical flow to evaluate model accuracy and preprocess input data, the system continuously updates and optimizes the detection model through transfer learning to adapt to changing environments and data.

[0169] The above are merely preferred embodiments of the present invention and do not limit the present invention in any other form. Any technician familiar with the profession may use the technical content disclosed above to change or modify it into an equivalent embodiment with equivalent changes and apply it to other fields. However, any modification, equivalent change and modification made to the above embodiment based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall still fall within the scope of protection of the technical solution of the present invention.

Claims

1. A smart packaging detection method, characterized in that: The following steps are involved: Based on image preprocessing technology, feature extraction is adopted, and multi-scale image information is processed through a set of filters to generate a set of feature maps; Based on the feature graph set, a graph neural network algorithm is used to encode the product packaging structure, extract structural information, and generate a structural feature representation; Based on multi-view shooting technology and using stereo vision algorithms, we can create three-dimensional models of product packaging, conduct depth analysis, and generate stereo vision analysis data. Based on the feature graph set, an autoencoder and a generative adversarial network are used to learn normal features to generate a trained network model. Based on the trained network model, a threshold method is used to identify abnormal features, screen features that deviate from the normal range, mark them as abnormal, and generate an anomaly detection model. Based on the anomaly detection model and structural feature representation, a time series analysis method is used to continuously monitor product quality, perform trend judgment, and generate a quality monitoring report; Based on the stereoscopic vision analysis data and quality monitoring report, the detection model is optimized and updated using transfer learning technology to generate a final detection model; The feature map set specifically refers to product packaging image features extracted at multiple resolutions. The structural feature representation includes the overall composition of the product packaging and the association features between multiple components. The stereoscopic vision analysis data specifically refers to the three-dimensional model and depth information reconstructed based on packaging images captured from multiple angles. The anomaly detection model specifically refers to a deep learning model that identifies abnormal states based on the normal packaging feature distribution. The quality monitoring report specifically refers to quality change data and prediction trend chart based on time series.

2. The intelligent packaging detection method according to claim 1, characterized in that: Based on image preprocessing technology, feature extraction is used, and multi-scale image information is processed through a set of filters. The steps to generate a feature map set are as follows: Based on the image preprocessing technology, a z-score normalization algorithm is used to simplify the illumination and scale influencing factors of the image and generate a preprocessed normalized image; Based on the preprocessed normalized image, a multi-scale convolutional neural network technology is used to perform a convolution operation through a convolution kernel to extract features and generate a convolution feature map; Based on the convolution feature map, a ReLU activation function algorithm is used to enhance the nonlinear expression ability of the feature map and generate an activation feature map; Based on the activated feature map, a filter set is used to perform filtering optimization on the extracted multi-scale image information to enhance the difference and expressiveness of the characteristics and generate a feature map set.

3. The intelligent packaging detection method according to claim 1, characterized in that: Based on the feature graph set, a graph neural network algorithm is used to encode the product packaging structure and extract structural information. The steps of generating a structural feature representation are as follows: Based on the feature graph set, using a dimensionality reduction algorithm, converting it into a node matrix for describing product packaging structural features; Based on the node matrix, a graph neural network algorithm is used to calculate the correlation between nodes and generate an adjacency matrix; Based on the adjacency matrix and the node matrix, a convolutional neural network algorithm is used to aggregate and transform the feature vector information to generate encoded structural information; Based on the encoded structural information, a deconvolution algorithm is used to perform feature decoding, highlight the structural features of the product packaging, and establish a structural feature representation.

4. The intelligent packaging detection method according to claim 1, characterized in that: Based on multi-view shooting technology and using stereo vision algorithms, we create three-dimensional modeling of product packaging and conduct depth analysis. The specific steps for generating stereo vision analysis data are as follows: Based on the multi-view shooting technology, high-definition shooting equipment is used to shoot product packaging from multiple angles to obtain multi-view image data; Based on the multi-view image data, a stereo vision algorithm is used to calculate the image depth and generate a depth map; Based on the depth map, a regression analysis algorithm is used to perform three-dimensional reconstruction, a 3D model based on the product packaging is constructed, and a 3D model is generated; Based on the 3D model, a deep learning algorithm is used to optimize and analyze the model, present a three-dimensional visual effect of the product packaging, and generate three-dimensional visual analysis data.

5. The intelligent packaging detection method according to claim 1, characterized in that: Based on the feature graph set, an autoencoder and a generative adversarial network are used to learn normal features and generate a trained network model. Based on the trained network model, a threshold method is used to identify abnormal features, filter out features that deviate from the normal range, and mark them as abnormal. The steps of generating the anomaly detection model are specifically as follows: Based on the feature map set, a dimensionality reduction algorithm is used to compress dense features, solve the dimensionality curse problem, extract important features, and generate a mapping parameter set; Based on the mapping parameter set, the autoencoder is used to compress and encode the input data and decode and restore it, and the main features in the data are learned by minimizing the difference between the input and the output to generate a normal feature model; Based on the normal feature model, a generative adversarial network is used to train normal features. The generative adversarial network training of normal features includes a generator and a discriminator. The generator generates a simulated feature map, and the discriminator determines whether the feature map is real. After multiple rounds of training, a trained network model is generated; Based on the trained network model, a threshold method is used to identify abnormal features, screen features that deviate from the normal range, mark them as abnormal, and generate an anomaly detection model.

6. The intelligent packaging detection method according to claim 1, characterized in that: Based on the anomaly detection model and structural feature representation, a time series analysis method is used to continuously monitor product quality and perform trend judgment. The steps for generating a quality monitoring report are as follows: Based on the anomaly detection model and the structural feature representation, evaluating the correlation between features and constructing a matrix to generate a correlation coefficient matrix; Based on the correlation coefficient matrix, the K-means clustering algorithm is used to perform classification analysis on the data, classify data with similar characteristics into the same category, and generate classification results; Based on the classification results, use time series analysis methods to track data changes over a long period of time, analyze data change trends, and generate trend judgment reports; Based on the trend judgment report, a comprehensive evaluation method is adopted to combine multiple evaluation indicators, continuously monitor product quality, and generate a quality monitoring report.

7. The intelligent packaging detection method according to claim 1, characterized in that: Based on the stereoscopic vision analysis data and quality monitoring report, the detection model is optimized and updated using transfer learning technology to generate the final detection model. The specific steps are as follows: Based on the stereoscopic vision analysis data and quality monitoring report, a label propagation algorithm is used to propagate the labeled category information to the unlabeled data, complete the classification of the unlabeled data, and generate an optimized label set; Based on the optimized label set, a convolutional neural network is used to perform feature learning, capture image data features, and generate feature mapping data; Based on the feature map data, use the optical flow method to calculate the information of the object movement in the image, evaluate the accuracy of the current model, and generate a model evaluation result; Based on the model evaluation results, transfer learning technology is used to adapt the pre-trained deep model to the task environment, and the model is updated to generate the final detection model.

8. An intelligent packaging detection system, characterized in that: The intelligent packaging detection method according to any one of claims 1 to 7, wherein the system includes a preprocessing module, a feature extraction module, a three-dimensional model construction module, an anomaly detection module, a trend analysis module, a classification optimization module, and a model update module.

9. The intelligent packaging detection system according to claim 8, characterized in that: The preprocessing module is based on the image and uses the z-score standardization method to perform standardization, eliminate interference factors including illumination and scale, and generate a preprocessed normalized image; The feature extraction module uses a graph neural network algorithm to extract features based on the preprocessed normalized image, and performs dimension compression to generate a structural feature representation; The 3D model building module uses a stereo vision algorithm to estimate depth based on multi-angle visual data and performs 3D reconstruction through regression analysis to generate a 3D model of the product packaging. The anomaly detection module is based on a feature map set, uses a dimensionality reduction algorithm to compress dense features, performs feature learning through an autoencoder, and uses a generative adversarial network for network training to generate an anomaly detection model; The trend analysis module classifies data based on anomaly detection models and structural feature representations through correlation analysis and clustering algorithms, and then uses time series analysis to analyze the trend of data changes and form a quality trend report; The classification optimization module uses a label propagation algorithm to classify unlabeled data based on stereoscopic vision analysis data and quality monitoring reports, and performs feature learning through a convolutional neural network to obtain an optimized label set and feature mapping data; The model updating module uses the optical flow method to evaluate the accuracy of the model based on the feature mapping data, and updates the model through the transfer learning method to form the final detection model.

10. The intelligent packaging detection system according to claim 8, characterized in that: The preprocessing module includes an image normalization submodule, a convolution feature extraction submodule, a nonlinear enhancement submodule, and an optimization filtering submodule; The feature extraction module includes a feature map generation submodule, a node matrix calculation submodule, a structure information encoding submodule, and a feature decoding submodule; The three-dimensional model construction module includes a multi-view image acquisition submodule, a depth map generation submodule, a 3D model generation submodule, and a model optimization and analysis submodule; The anomaly detection module includes a feature compression submodule, a network training submodule, and an abnormal feature recognition submodule; The trend analysis module includes a correlation assessment submodule, a cluster classification submodule, a trend judgment submodule, and a quality monitoring submodule; The classification optimization module includes a label propagation submodule, an unlabeled data classification submodule, and a feature learning submodule; The model updating module includes a model evaluation submodule, a data preprocessing submodule, and a model updating submodule.

Images