A computer vision processing method and system based on big data
By using big data computer vision processing methods, combined with adversarial examples and AI policy controllers, the model is dynamically updated to adapt to different workpiece states, solving the problem of inaccurate workpiece state judgment in existing technologies and achieving fast, flexible workpiece state recognition and efficient detection.
Patent Information
- Application Number
- CN202611407991.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-11
- Publication Date
- 2026-11-06
AI Technical Summary
Existing technologies cannot accurately and quickly determine the status of workpieces, resulting in poor flexibility in use, low detection efficiency and accuracy, and the identification and judgment criteria need to be redefined when the workpiece processing type changes, which also leads to poor operational flexibility.
A computer vision processing method based on big data is adopted. Image data is acquired, preprocessed, and features are extracted. The abnormal state of the workpiece is analyzed by combining big data, adversarial examples are generated, and the model is trained to achieve rapid and accurate judgment of the workpiece state. AI policy controller and adversarial network are used to generate abnormal workpiece data and dynamically update the main model to adapt to different workpiece states.
It enables rapid and accurate judgment of workpiece status, broadens the scope of abnormal data identification, improves the flexibility and accuracy of judgment, reduces the influence of miscellaneous factors, and improves processing efficiency and accuracy.
Smart Images
Figure CN123052291A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of computer vision processing technology, and in particular to a computer vision processing method and system based on big data. Background Technology
[0002] With the development of computer vision processing technology, more and more automated production equipment is beginning to use related technologies to identify various feature data of the equipment during the workpiece processing process, so as to achieve the effect of monitoring the status of the equipment and judging the workpiece processing, reducing the cost of manual supervision, and ensuring the overall processing efficiency.
[0003] However, most current processing methods rely on pre-setting workpiece processing parameters and combining them with the final processing effect to confirm the overall visual parameters. In some processing environments, issues such as light spots, irregular workpiece surfaces, or worn processing tools can lead to discrepancies between the final processing effect and the pre-set processing effect. This results in the visual processing detecting abnormalities and triggering alarms. However, in actual products, such problems do not necessarily lead to products failing to meet corresponding production standards. Existing technologies also lack accurate judgment of the actual workpiece condition, resulting in poor overall detection efficiency and accuracy, which affects the production efficiency of the entire production line. Furthermore, if the production line needs to change the workpiece processing type, corresponding identification and judgment standards need to be redefined, resulting in a narrow overall identification and judgment range and poor operational flexibility.
[0004] Therefore, there is an urgent need in the market for a processing system that can quickly and accurately determine the state of a workpiece and adapt to changes in workpiece processing requirements to solve the above problems. Summary of the Invention
[0005] The purpose of this invention is to provide a computer vision processing method and system based on big data, which solves the technical problems of existing technologies that cannot accurately and quickly determine the status of workpieces and have poor flexibility of use, and achieves the technical effect of quickly and accurately determining the status of workpieces and improving the overall flexibility of use.
[0006] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows:
[0007] A computer vision processing method based on big data includes acquiring image data content, performing data preprocessing and core feature information extraction and recognition, and the acquired workpiece data information includes the workpiece's external dimensions and specifications at a specific angle, the position to be processed, and the recognition data of the processed position.
[0008] The abnormal workpiece status data obtained by analysis and identification is combined with a large database for further identification, obtain corresponding solutions to the identification problems that have been generated, and automatically generate corresponding processing programs. If the corresponding solution strategy cannot be identified, the corresponding abnormal data information is extracted and put into the AI strategy controller, and adversarial examples are generated based on logical data and feature sample data.
[0009] The core feature data of adversarial examples are labeled and fed into the model along with normal data for processing and learning. By combining adversarial loss and misjudged sample data, a corresponding adversarial network is generated, resulting in a generative model that can generate actual abnormal data of workpieces. Based on the generative model, abnormal data of workpieces is generated, and abnormal situations of equipment and process links are inferred.
[0010] As an improvement, the image data acquisition content is obtained by acquiring image data from multiple processing equipment on the production line through the cloud and judging the current status of the workpiece according to the preset process content. The current status of the workpiece includes the workpiece processing coordinate position, the preset process positioning, and the workpiece appearance status. The current stage image is acquired and filtered and noise-reducing preprocessing is performed.
[0011] As an improvement, the current stage image content obtains the features of image pixels through a convolutional neural network, matches them with the database, automatically retrieves the meaning of the features and sends them to the processing end, and performs preliminary feature judgment in conjunction with preset process content.
[0012] As an improvement, the processing end also includes a diagnostic module. The diagnostic module will generate a workpiece quality report and a demand signal based on the feature information. The demand signal includes unfinished processing requirements and processing environment requirements. After the AI strategy controller retrieves the demand signal, it will generate samples in the model and simulate workpiece data with similar problems by combining logical causality and rules to provide high-quality simulation data.
[0013] As an improvement, the AI strategy controller can manually intervene to formulate corresponding labels and make detailed adjustments when generating samples. The AI strategy controller feeds the data into an adversarial model independent of the main model to generate data and verify it, which is used to derive the identification logic for similar workpiece problems.
[0014] As an improvement, the identification logic for similar workpiece problems derived from the adversarial model will be re-input into the data processing module for dynamic hybrid cleaning and adaptive repair correction, and feature extraction and fusion to generate a new model. The main model will then be updated and put into production.
[0015] As an improvement, after the main model is updated, it will be evaluated and judged by the evaluation module. It will analyze specific logical defects and shortcomings, and re-enter the adversarial model for multiple processing. During the main model processing and running phase, it will be uploaded to the management terminal via the Internet, and the management terminal will monitor and collect the running data.
[0016] A computer vision processing system based on big data includes a data collection layer, a data processing layer, a model training layer, and an application layer.
[0017] The data collection layer acquires the workpiece processing status through multiple sets of visual image collection devices and sensors;
[0018] The data processing layer filters and extracts features from the acquired data, and then feeds it into the processing end to identify and judge it in conjunction with the pre-stored processing information in the large database.
[0019] The model training layer collects abnormal feature information obtained by the data processing layer and generates corresponding adversarial samples after modification and debugging to simulate the problems that may arise from the corresponding problematic workpiece.
[0020] The application collects problem data from the model training layer, retrieves solutions using a large database, and interacts with the administrator for confirmation.
[0021] The beneficial effects of this invention are: the workpiece status can be quickly and accurately judged by multiple sets of sensors and acquired visual information; with the help of pre-stored data in a large database, specific workpiece status data information can be quickly determined, which facilitates remote processing supervision by operators.
[0022] By training adversarial models, higher quality simulated workpiece data can be obtained, thereby broadening the scope of abnormal data identification, improving the overall judgment flexibility, and more accurately identifying the corresponding problems of the workpiece, while reducing the impact of miscellaneous factors on the data.
[0023] It can be used in conjunction with big data to perform multiple simulations and can obtain preset information on different workpiece processing techniques, further improving processing efficiency and accuracy. It can also directly generate and export abnormal problems and processing results, making it convenient for managers to view and control. Attached Figure Description
[0024] Figure 1 This is a schematic diagram of a computer vision processing method based on big data according to the present invention;
[0025] Figure 2 This is a flowchart of a computer vision processing system based on big data according to the present invention. Detailed Implementation
[0026] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.
[0027] It should be noted that the terms "first" and "second" in this application are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or apparatuses.
[0028] like Figure 1 As shown, a computer vision processing method based on big data includes acquiring image data, performing data preprocessing and core feature information extraction and recognition. The acquired workpiece data includes the workpiece's external dimensions and specifications at specific angles, the position to be processed, and the already processed position recognition data. The method analyzes and identifies abnormal workpiece state data, combining it with a large database for further identification, obtaining corresponding solutions to existing identification problems, and automatically generating corresponding processing programs. If no corresponding solution strategy is identified, the corresponding abnormal data information is extracted and fed into an AI strategy controller. Adversarial examples are generated based on logical data and feature sample data. The core feature data of the adversarial examples are labeled and fed into the model along with normal data for processing and learning. By combining adversarial loss and misjudged sample data, a corresponding adversarial network is generated, resulting in a generative model capable of generating actual abnormal workpiece data. Based on the generative model, abnormal workpiece data is generated, and abnormal conditions in equipment and processes are inferred. The large database can access the internet to obtain model running data uploaded by other users, preset workpiece judgment labels, and process knowledge as constraints, reducing operation and maintenance pressure, while also acquiring more high-quality data for model training.
[0029] The image data acquisition process involves obtaining image data from multiple processing devices on the production line via the cloud and determining the current state of the workpiece based on preset process requirements. The current workpiece state includes the workpiece's processing coordinates, preset process positioning, and appearance. The image at the current stage is acquired and preprocessed with filtering and noise reduction. The current stage image content is processed using a convolutional neural network to obtain the features of image pixels. After matching with a database, the feature meanings are automatically retrieved and sent to the processing end, where preliminary feature judgment is performed in conjunction with the preset process requirements. The processing end also includes a diagnostic module. Based on the feature information, the diagnostic module generates a workpiece quality report and a demand signal. The demand signal includes unfinished processing requirements and processing environment requirements. After retrieving the demand signal, the AI strategy controller generates samples within the model and simulates similar problem workpiece data using logical causality and rules to provide high-quality simulation data. The AI strategy controller can be manually input with logical causal relationship data and supplemented with rule knowledge of different process requirements. This allows for the derivation of corresponding processing demand judgment information for different workpieces, thus addressing different workpiece states.
[0030] The AI strategy controller allows for manual intervention during sample generation, specifying corresponding labels and making detailed adjustments. It then feeds data into an adversarial model independent of the main model to generate and verify data, deriving problem identification logic for similar workpieces. This logic is then fed back into the data processing module for dynamic mixing, cleaning, and adaptive correction, followed by feature extraction and fusion to create a new model. This updates the main model, which is then deployed to production. After updating, the main model undergoes evaluation by an assessment module, analyzing specific logical flaws and shortcomings, and is repeatedly processed within the adversarial model. During the main model's processing and execution phase, the data is uploaded to a management terminal via the internet, which monitors and collects the operational data. The management terminal can also upload the summarized process data and workpiece judgment criteria to a cloud-based database via the internet, enabling remote access and use from different devices.
[0031] When in use, the operator needs to input the initial process data and build or retrieve the corresponding model. During the AI strategy controller's demand signal reception stage, a conditional generative adversarial network, physical constraint loss, and post-verification screening are added. The conditional generative adversarial network must ensure that the generation process has a clear and controllable objective. The physical constraint loss can add physical consistency constraints to the generator's training loss function to ensure that abnormal images conform to actual production and processing rules and avoid producing incomprehensible noise patterns. The post-verification screening is automatically screened by an independent simulation discriminator, or it can be further improved by manual confirmation to ensure the quality of the source data.
[0032] During the main model iteration phase, a dynamic hybrid data pool needs to be built to balance the proportions of historical data, new real data, and synthetic data, preventing the model from being biased towards one aspect; this is combined with a continuous learning algorithm to protect the important data parameters that have been learned and to continue learning corresponding new knowledge.
[0033] like Figure 2 As shown, a computer vision processing system based on big data includes a data collection layer, a data processing layer, a model training layer, and an application terminal. The data collection layer acquires the workpiece processing status through multiple sets of visual image collection devices and sensors. The data processing layer filters and extracts features from the acquired data, and then inputs it into the processing terminal to combine with pre-stored processing information in a large database for identification and judgment. The model training layer collects abnormal feature information acquired by the data processing layer, and generates corresponding adversarial samples after modification and debugging to simulate the problems that may occur with the corresponding problematic workpieces. The application terminal collects the problem data from the model training layer, retrieves solutions in conjunction with the large database, and interacts with the administrator for confirmation.
[0034] Example 1: When scratch identification processing is required on a workpiece surface, the AI strategy controller receives a diagnostic module "demand signal" containing anomaly category (e.g., "scratch"), suspected location, and process parameters (e.g., current tool ID, feed rate). This information serves as a conditional vector input. The generator then takes random noise and the aforementioned conditional vector as input and outputs a synthesized anomaly image. The discriminator determines whether the input image is a real anomaly image or a generated image, and also checks if the conditions are met. At this point, a pre-trained depth estimation network is used to ensure that the generated scratches conform to actual wear characteristics in the depth dimension; or pixel-level gradient constraints are added to make the scratch edge texture match the real material as a physical consistency constraint. After completion, the data needs to be filtered to select high-quality batches, and then quickly confirmed by a manual rule engine before finally saving and uploading for later use.
[0035] Example 2: To further clarify the specific scenarios to which the present invention is applicable and to solve the problems of adaptability and robustness under diverse production lines and extreme working conditions, this example further specifies the method and system based on Example 1.
[0036] This embodiment provides a computer vision processing method based on big data. The method is specifically applied to a workpiece processing visual inspection scenario in an industrial automated production line. The processed image data is real-time image data of the workpiece during the processing. The workpiece includes, but is not limited to, industrial objects with clearly defined processing procedures and quality standards, such as metal cutting parts, die-cast parts, injection molded parts, electronic component packages, and PCB boards.
[0037] The method automatically selects or switches the following processing modes based on the production line configuration and workpiece type:
[0038] Single-station independent processing mode: suitable for small-batch, multi-variety workpiece production scenarios. The image data acquisition content is obtained by acquiring multi-angle images of the workpiece in a single processing equipment through a local image acquisition terminal, and feature extraction and anomaly judgment are performed based on a preset single workpiece process file.
[0039] Multi-station serial production line mode: suitable for mass production of standardized workpieces. The acquired image data is aggregated from multiple consecutive processing stations on the production line through cloud or edge computing nodes, and time-series correlation analysis and cross-station anomaly tracing are performed based on the process sequence and intermediate states of each station.
[0040] Mixed-flow production mode: Suitable for flexible production lines that process multiple types of workpieces simultaneously. The method dynamically retrieves corresponding preset process data and recognition models by reading the carrier or barcode information on the workpiece, thereby achieving parallel recognition and judgment of different workpieces.
[0041] To address extreme working conditions that may occur in actual production, this embodiment proposes the following handling mechanism:
[0042] Operating Condition 1: Harsh environments such as strong light, dust, and oil stains cause a severe decline in image quality.
[0043] In the data preprocessing stage, an adaptive environment compensation module is added: based on the local image grayscale distribution and noise model, it automatically calls the multi-scale Retinex enhancement algorithm, dark channel dehazing algorithm, or real-time dust removal network based on lightweight CNN to restore the image.
[0044] If the recovered image still fails to meet the recognition confidence threshold, a multimodal fusion judgment is triggered: combining non-visual sensor data (such as laser displacement, ultrasonic ranging, and vibration signals) to assist in judging the workpiece status and avoid production line shutdown due to the lack of visual signals.
[0045] Condition 2: Unprecedented defect types caused by severe tool wear or abnormal workpiece material.
[0046] When the AI policy controller fails to identify the corresponding solution strategy, it not only generates adversarial examples but also introduces an additional physical simulation constraint module.
[0047] Based on current process parameters (such as cutting speed, feed rate, tool angle, material hardness, etc.), this module calls the built-in physical models of the machining process (such as cutting force model, thermal deformation model) to simulate and generate virtual abnormal image data that conforms to physical laws, as a supplement to adversarial examples.
[0048] The adversarial examples and physics simulation examples are used together in the model training to ensure that the generated model can produce anomalous data with real physical possibilities, rather than meaningless noise patterns.
[0049] Scenario 3: New workpiece type launched, historical anomaly data extremely scarce (cold start problem):
[0050] The large database contains a general abnormal feature library for multiple standard workpiece types and corresponding generation model parameters.
[0051] When a new workpiece type is launched, the AI policy controller retrieves the closest existing workpiece model from the big data database based on the geometric structure, material properties and process similarity of the new workpiece, and performs transfer learning initialization.
[0052] Meanwhile, the system automatically generates a set of initial adversarial samples based on rules and logical causality to quickly start model training and continuously fine-tune them after accumulating real production data.
[0053] In the corresponding system of this embodiment, the data collection layer further includes an environmental sensor group for collecting environmental parameters such as light intensity, dust concentration, temperature, and humidity as control inputs for image preprocessing; the data processing layer embeds a working condition recognition module for automatically switching preprocessing strategies and fusion judgment logic based on environmental parameters and image quality indicators; the model training layer further includes a physical simulation interface for receiving virtual abnormal images generated by external process simulation software or a built-in physical model; the application terminal provides an extreme working condition configuration interface for administrators to preset degradation strategies, alarm thresholds, and manual intervention rules under different working conditions.
[0054] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A computer vision processing method based on big data, characterized in that, This includes acquiring image data, performing data preprocessing and core feature extraction and recognition. The acquired workpiece data includes the workpiece's external dimensions and specifications at a specific angle, as well as the identification data of the position to be processed and the position already processed. The abnormal workpiece status data obtained by analysis and identification is combined with a large database for further identification, obtain corresponding solutions to the identification problems that have been generated, and automatically generate corresponding processing programs. If the corresponding solution strategy cannot be identified, the corresponding abnormal data information is extracted and put into the AI strategy controller, and adversarial examples are generated based on logical data and feature sample data. The core feature data of adversarial examples are labeled and fed into the model along with normal data for processing and learning. By combining adversarial loss and misjudged sample data, a corresponding adversarial network is generated, resulting in a generative model that can generate actual abnormal data of workpieces. Based on the generative model, abnormal data of workpieces is generated, and abnormal situations of equipment and process links are inferred.
2. The computer vision processing method based on big data according to claim 1, characterized in that, The image data is acquired by obtaining image data from multiple processing devices on the production line through the cloud and determining the current status of the workpiece according to the preset process content. The current status of the workpiece includes the workpiece processing coordinate position, the preset process positioning, and the workpiece appearance. The image of the current stage is acquired and preprocessed by filtering and noise reduction.
3. The computer vision processing method based on big data according to claim 2, characterized in that, The current stage image content obtains the features of image pixels through a convolutional neural network, matches them with the database, automatically retrieves the meaning of the features and sends them to the processing end, and performs preliminary feature judgment in combination with preset process content.
4. The computer vision processing method based on big data according to claim 3, characterized in that, The processing end also includes a diagnostic module, which generates a workpiece quality report and a demand signal based on the feature information. The demand signal includes unfinished processing requirements and processing environment requirements. After the AI strategy controller retrieves the demand signal, it generates samples within the model and simulates workpiece data with similar problems by combining logical causality and rules to provide high-quality simulation data.
5. The computer vision processing method based on big data according to claim 4, characterized in that, The AI strategy controller can manually intervene to formulate corresponding labels and make detailed adjustments when generating samples. The AI strategy controller feeds the data into an adversarial model independent of the main model to generate data and verify it, which is used to derive the identification logic for similar workpiece problems.
6. The computer vision processing method based on big data according to claim 5, characterized in that, The problem identification logic of the same type of workpiece obtained by the adversarial model will be re-entered into the data processing module for dynamic hybrid cleaning and adaptive repair correction, and feature extraction and fusion to generate a new model. The main model will then be updated and put into production.
7. The computer vision processing method based on big data according to claim 6, characterized in that, After the main model is updated, it will be evaluated and judged by the evaluation module. It will analyze specific logical defects and shortcomings, and re-enter the adversarial model for multiple processing. During the main model processing and running phase, it will be uploaded to the management terminal via the Internet, and the management terminal will monitor and collect the running data.
8. A computer vision processing system based on big data, characterized in that, It includes a data collection layer, a data processing layer, a model training layer, and an application layer. The data collection layer acquires the workpiece processing status through multiple sets of visual image collection devices and sensors; The data processing layer filters and extracts features from the acquired data, and then feeds it into the processing end to identify and judge it in conjunction with the pre-stored processing information in the large database. The model training layer collects abnormal feature information obtained by the data processing layer and generates corresponding adversarial samples after modification and debugging to simulate the problems that may arise from the corresponding problematic workpiece. The application collects problem data from the model training layer, retrieves solutions using a large database, and interacts with the administrator for confirmation.