A high-definition visual inspection and recognition system based on an image processing industrial control computer

Abstract

This invention belongs to the field of visual inspection and relates to a high-definition visual inspection and recognition system based on an image processing industrial control computer. It includes: a historical data acquisition module for acquiring image data of historical records of the same product; a feature point extraction module for constructing a convolutional neural network in deep learning to extract key feature points and other feature points from the image data; a 3D model construction module for calculating the spatial coordinates of other feature points in a 3D prediction model; an error threshold setting module for setting a first error threshold and a second error threshold; a sorting scheme selection module for selecting a suitable sorting scheme; a pass threshold setting module for setting a first pass threshold and a second pass threshold; and a production scheme optimization module for selecting a suitable optimization scheme. This invention solves the problem that visual inspection methods are easily affected by human factors, leading to difficulties in guaranteeing the accuracy and consistency of inspection results.

Description

Technical Field

[0001] This invention belongs to the technical field of visual inspection and relates to a high-definition visual inspection and recognition system based on an image processing industrial control computer. Background Technology

[0002] Traditional visual inspection methods rely primarily on manual annotation and experience-based judgment. This approach is not only inefficient and ill-suited for large-scale, high-intensity production tasks, but also susceptible to human factors such as operator fatigue and emotional fluctuations, which can compromise the accuracy and consistency of inspection results. In modern industrial production, such uncertain visual inspection methods are unacceptable, as even minor errors can significantly impact product quality and production efficiency.

[0003] With the rapid development of industrial automation and intelligent technologies, modern production places increasingly higher demands on product quality inspection. Uncertain visual inspection methods can no longer meet these needs for efficient, accurate, and stable inspection. There is an urgent need for a solution that can automatically, quickly, and accurately complete inspection tasks to ensure the stability and consistency of product quality.

[0004] To address the aforementioned issues, traditional visual inspection methods are susceptible to human factors, which can easily lead to difficulties in ensuring the accuracy and consistency of inspection results. Summary of the Invention

[0005] To address the problem that traditional visual inspection methods are susceptible to human error, making it difficult to guarantee the accuracy and consistency of inspection results, this invention provides a high-definition visual inspection and recognition system based on an image processing industrial control computer.

[0006] A high-definition visual inspection and recognition system based on an image processing industrial control computer includes the following modules:

[0007] The historical data acquisition module acquires image data of the same product from different angles, under different lighting conditions, and with different background environments based on historical records of the same product.

[0008] The feature point extraction module is used to build convolutional neural networks in deep learning and extract key feature points and other feature points from image data.

[0009] The 3D model building module is used to build a 3D prediction model of feature points and calculate the spatial coordinates of other feature points in the 3D prediction model.

[0010] The error threshold setting module obtains the error values ​​between the spatial coordinates of other feature points in the calculation of the 3D prediction model and the spatial coordinates of other feature points in different images, and sets the first error threshold and the second error threshold.

[0011] The sorting scheme selection module is used to calculate the error parameters between the spatial coordinates of other feature points of the actual product and the spatial coordinates of other feature points of the 3D prediction model, determine the relationship between the error parameters and the first error threshold and the second error threshold, and select a suitable sorting scheme.

[0012] The pass threshold setting module is used to monitor the pass rate achieved during the product manufacturing process and set the first pass threshold and the second pass threshold.

[0013] The production plan optimization module compares the product pass rate with the first pass threshold and the second pass threshold, and selects an appropriate optimization plan based on the judgment results.

[0014] Preferably, acquiring the image data includes the following steps:

[0015] Collect multiple image data from the historical records of the same product, including image data with different lighting conditions, angles, and background environments;

[0016] Image data is preprocessed using filtering techniques to reduce noise, contrast adjustment and sharpening, normalization, and correction.

[0017] Preferably, constructing a convolutional neural network in deep learning includes the following steps:

[0018] Label key feature points and other feature points of image data, construct a convolutional neural network in deep learning, and divide the labeled image data into training set, validation set and test set;

[0019] The training set trains a convolutional neural network to extract key feature points and other feature points from image data with different angles, lighting conditions, and background environments. The validation set verifies whether the convolutional neural network can extract key feature points and other feature points from the image data. The test set tests the convolutional neural network and ultimately evaluates the feature point extraction performance.

[0020] Preferably, constructing the 3D model construction module includes the following steps:

[0021] Extract key feature points from image data, including product corners, product edges, and specific markers; filter out key feature points with identifiability and stability from different image data.

[0022] A three-dimensional spatial coordinate system is constructed based on key feature points, the three-dimensional position of the camera, and the shooting direction of the lens. A three-dimensional prediction model of the feature points is obtained, and the spatial coordinates of other feature points in the three-dimensional spatial coordinate system are calculated.

[0023] Preferably, constructing the three-dimensional spatial coordinate system includes the following steps:

[0024] The camera's motion trajectory is tracked using a camera tracking algorithm to calculate the lens's shooting direction; the camera's three-dimensional position is measured using a laser rangefinder; and the camera's internal parameters, such as focal length, sensor size, and distortion coefficient, are obtained.

[0025] Select one or more key feature points as the origin of the coordinate system, and determine the direction of the coordinate axes using the camera's shooting direction and internal parameters; establish a three-dimensional spatial coordinate system using the origin and the direction of the coordinate axes; the X-axis and Y-axis are parallel to the two-dimensional image plane, and the Z-axis is consistent with the camera's shooting direction.

[0026] Preferably, constructing the three-dimensional prediction model includes the following steps:

[0027] By utilizing the camera's intrinsic and extrinsic parameters and combining the correspondence of key feature points, a projection matrix is ​​solved to describe the mapping relationship from three-dimensional space to a two-dimensional image plane. By combining the establishment of a three-dimensional spatial coordinate system and the solution of the projection matrix, the three-dimensional prediction model is constructed. The coordinates of other feature points of the two-dimensional image plane in three-dimensional space are calculated using the back projection process of the projection matrix.

[0028] Preferably, setting the first error threshold and the second error threshold includes the following steps:

[0029] Key feature points and other feature points of different images are obtained and input into the three-dimensional prediction model. The key feature points of different images are matched with the three-dimensional spatial coordinate system of the three-dimensional prediction model. The least squares method is used to approximate the real three-dimensional coordinate points. There are errors between the three-dimensional coordinate points of other feature points of the three-dimensional prediction model and the spatial coordinate points of other feature points actually measured. The calculated errors are combined to form a dataset and set a first error threshold and a second error threshold.

[0030] For each feature point in the 3D prediction model, calculate the spatial coordinate error between other feature points in the 3D prediction model and the corresponding feature points in actual measurement, and calculate the average error and maximum error of the spatial coordinate error of other feature points;

[0031] The spatial coordinate error is calculated to satisfy the following formula:

[0032] ;

[0033] in, , , It is the third in the three-dimensional prediction model The x, y, z coordinates of each feature point; , , It is the first actual measurement The x, y, z coordinates of each feature point; It is spatial coordinate error;

[0034] The average error is calculated to satisfy the following formula:

[0035] ;

[0036] in, It is the average error. Indicates the number of feature points. It is the sum of the spatial coordinate errors of all feature points;

[0037] The maximum error is calculated according to the following formula:

[0038] ;

[0039] in, This indicates finding the largest error among all the spatial coordinate errors of all feature points.

[0040] Preferably, selecting a suitable sorting solution includes the following steps:

[0041] Feature points of image data of small batches of products in the same batch are extracted and input into the three-dimensional prediction model. The key feature points of the product are matched with the three-dimensional spatial coordinate system of the three-dimensional prediction model. The error parameters between the spatial coordinates of other feature points of the three-dimensional prediction model and the spatial coordinates of other feature points of the product are calculated. The relationship between the error parameters and the first error threshold and the second error threshold is determined. A suitable sorting scheme is selected based on the judgment result.

[0042] If the error parameter is less than the first error threshold, it means that the accuracy of the 3D prediction model is acceptable.

[0043] If the error parameter is greater than or equal to the first error threshold but less than the second error threshold, it means that the accuracy of the 3D prediction model needs further fine-tuning.

[0044] If the error parameter is greater than or equal to the second error threshold, it is determined that the error parameter between the spatial coordinates of other feature points in the 3D prediction model and the spatial coordinates of other feature points in the actual product exceeds the prediction range, indicating that the 3D prediction model needs to be reconstructed.

[0045] Preferably, setting the first qualified threshold and the second qualified threshold includes the following steps:

[0046] Real-time monitoring of the pass rate achieved by products in the same batch during the production process. The product meets the pass standard when the error parameter between the spatial coordinates of other feature points of the trained 3D prediction model and the spatial coordinates of other feature points of the actual product is less than the first error threshold. The first pass threshold and the second pass threshold are set according to the product pass rate.

[0047] Preferably, selecting a suitable optimization scheme includes the following steps:

[0048] If the product's pass rate is greater than the second pass threshold, it means that the product can directly pass the quality inspection.

[0049] If the product pass rate is greater than the first pass threshold but less than or equal to the second pass threshold, it indicates that the company needs to further check whether there are any errors in the production process and procedures.

[0050] If the product pass rate is less than or equal to the first pass threshold, it means that the company needs to further check whether there are any errors in the shape and size of the batch of products.

[0051] In summary, the present invention has the following beneficial technical effects:

[0052] 1. By utilizing convolutional neural networks in deep learning to extract key feature points and other feature points from image data and constructing a 3D prediction model, the precise coordinates of these feature points in 3D space can be calculated. This high-precision feature point extraction and spatial coordinate calculation significantly improves the accuracy of detection and recognition. Simultaneously, because the system can automatically process and analyze image data, it also significantly improves the efficiency of detection and recognition, reducing manual intervention and errors.

[0053] 2. By setting a first and second pass threshold, the system can determine in real time whether products meet the pass standards and select appropriate optimization schemes based on changes in the pass rate. This dynamic, real-time quality monitoring method enables enterprises to promptly identify problems and take corresponding measures, thereby ensuring the stability and reliability of product quality. The system continuously optimizes and adjusts the detection and identification algorithm based on historical and real-time data to adapt to changes in different products and production environments.

[0054] 3. Through automated and intelligent image processing and data analysis, this system can provide enterprises with abundant production data and information, helping them better understand the production process and product quality. This data and information can provide strong support for enterprises to formulate production plans and optimize production processes. The system can also be integrated and linked with other production management systems to achieve data sharing and collaborative management, further improving the efficiency and intelligence of production management. Attached Figure Description

[0055] Figure 1This document discloses a schematic diagram of the framework of a high-definition visual inspection and recognition system based on an image processing industrial control computer. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0057] The following is in conjunction with the appendix Figure 1 A preferred description of the present invention is provided below.

[0058] See appendix Figure 1 As shown, this invention proposes a high-definition visual inspection and recognition system based on an image processing industrial control computer, including: a historical data acquisition module, a feature point extraction module, a three-dimensional model construction module, an error threshold setting module, a sorting scheme selection module, a qualified threshold setting module, and a production scheme optimization module.

[0059] The historical data acquisition module is used to acquire image data of the same product from different angles, under different lighting conditions, and with different background environments. The image data is then preprocessed.

[0060] Specifically, based on historical data and enterprise databases, we collect multiple image data of the same product's historical records; these multiple image data include images with different angles, lighting conditions, and background environments.

[0061] Image data is preprocessed using filtering techniques to reduce noise, contrast adjustment and sharpening, normalization, and correction.

[0062] The feature point extraction module is used to build a convolutional neural network (CNN) in deep learning. Preprocessed image data is input into the convolutional neural network (CNN) in deep learning to extract key feature points and other feature points from the image data.

[0063] Specifically, key feature points and other feature points in image data are labeled using manual annotation methods to construct a convolutional neural network (CNN) in deep learning. The labeled image data is divided into a training set, a validation set, and a test set. The training set trains the CNN to extract key feature points and other feature points from the image data. The validation set verifies whether the CNN can extract key feature points and other feature points from the image data. The test set tests the effectiveness of the CNN in extracting key feature points and other feature points from the image data.

[0064] For example, image data of a product's historical records is obtained, and the unique key feature point and multiple other feature points in the image are manually labeled. The training set, validation set, and test set are allocated in a 7:2:1 ratio. A convolutional neural network (CNN) in deep learning is constructed, which includes multiple convolutional layers, pooling layers, and fully connected layers. The multiple convolutional layers are used to extract the unique key feature point and multiple other feature points from the image data.

[0065] The training set trains a convolutional neural network (CNN) to extract key feature points and other feature points from image data with different angles, lighting conditions, and background environments; the validation set verifies whether the CNN can extract key feature points and other feature points from image data; and the test set ultimately evaluates the feature point extraction performance.

[0066] The 3D model building module constructs a 3D spatial coordinate system based on key feature points, camera position, and lens shooting direction, obtains the 3D prediction model of feature points, and calculates the spatial coordinates of other feature points in the 3D spatial coordinate system.

[0067] Specifically, key feature points are extracted from image data. These key feature points are categorized into types such as product corners, product edges, and specific markers. Key feature points with identifiability and stability are selected from different image data sets to form the basis for constructing a 3D coordinate system. Based on the key feature points, the camera's 3D position, and the lens's shooting direction, a 3D spatial coordinate system is constructed to obtain a 3D prediction model for the feature points. Based on the obtained 3D prediction model, the spatial coordinates of other feature points in the 3D spatial coordinate system are calculated.

[0068] In one embodiment of the present invention, establishing a three-dimensional spatial coordinate system includes the following steps:

[0069] The camera's motion trajectory is tracked using a camera tracking algorithm, and the camera's attitude is analyzed. The camera's shooting direction is consistent with the optical axis, and the lens's shooting direction is calculated. The three-dimensional position of the camera is measured using a laser rangefinder. Internal parameters such as the camera's focal length, sensor size, and distortion coefficient are obtained.

[0070] Select one or more key feature points as the origin of the coordinate system. The selection of the origin facilitates subsequent 3D reconstruction and data analysis. Use the camera's shooting direction and internal parameters to determine the direction of the coordinate axes. Establish a 3D spatial coordinate system using the origin and the direction of the coordinate axes. The X-axis and Y-axis are parallel to the 2D image plane, and the Z-axis is consistent with the camera's shooting direction.

[0071] In one embodiment of the present invention, constructing the three-dimensional prediction model includes the following steps:

[0072] By utilizing the camera's intrinsic and extrinsic parameters and the correspondence of key feature points, a projection matrix describing the mapping relationship from 3D space to the 2D image plane is solved. Combining the establishment of the 3D coordinate system and the solution of the projection matrix, the 3D prediction model is constructed. The coordinates of other feature points on the 2D image plane in 3D space are calculated using the back-projection process of the projection matrix.

[0073] For example, assuming a camera has already been calibrated, obtain The projection matrix is ​​M:

[0074] ;

[0075] in, and It's the camera's focal length. and These are the coordinates of the optical center. It was obtained through the camera calibration process.

[0076] The two-dimensional coordinate point P(100,200) is extracted from the captured two-dimensional image and converted into homogeneous coordinate form. If finding the inverse projection matrix (100, 200, 1) is not feasible, the 3D coordinates are continuously adjusted through iterative optimization. The least squares iterative optimization method is used to approximate the true 3D coordinates, and the 3D spatial coordinates corresponding to the 2D coordinates P(u, v) of other feature points are obtained. (x,y,z).

[0077] The error threshold setting module acquires key feature points and other feature points from different images, inputs a 3D prediction model of the feature points, matches the key feature points of different images with the 3D spatial coordinate system of the 3D prediction model, calculates the error values ​​between the spatial coordinates of other feature points in the 3D prediction model and the spatial coordinates of other feature points in different images, and sets a first error threshold and a second error threshold based on the dataset composed of the error values.

[0078] Specifically, the least squares method is used to approximate the real three-dimensional coordinates. There are errors between the three-dimensional coordinates of other feature points in the three-dimensional prediction model and the spatial coordinates of other feature points in actual measurement. The calculated errors are combined to form a dataset to set the first error threshold and the second error threshold.

[0079] For each feature point in the 3D prediction model, the spatial coordinate errors between other feature points in the 3D prediction model and the corresponding measured feature points are calculated. The average error and maximum error of the spatial coordinate errors of other feature points are also calculated, including the following steps:

[0080] The spatial coordinate error is calculated to satisfy the following formula:

[0081] ;

[0082] in, , , It is the third in the three-dimensional prediction model The x, y, z coordinates of each feature point; , , It is the first actual measurement The x, y, z coordinates of each feature point; It is spatial coordinate error;

[0083] The average error is calculated to satisfy the following formula:

[0084] ;

[0085] in, It is the average error. Indicates the number of feature points. It is the sum of the spatial coordinate errors of all feature points;

[0086] The maximum error is calculated according to the following formula:

[0087] ;

[0088] in, This indicates finding the largest error among all the spatial coordinate errors of all feature points.

[0089] For example, assuming other feature points in two image data sets have spatial coordinate errors of 3 mm and 5 mm, we know that the average error... millimeters, maximum error millimeters;

[0090] The first error threshold and the second error threshold are set based on the calculated average error and maximum error;

[0091] The first error threshold is set to 80% of the maximum error, and the first error threshold is 4 mm;

[0092] The second error threshold is set to twice the average error, and the second error threshold is 8 mm.

[0093] The sorting scheme selection module is used to calculate the error parameters between the spatial coordinates of other feature points of the actual product and the spatial coordinates of other feature points of the 3D prediction model, determine the relationship between the error parameters and the first error threshold and the second error threshold, and select a suitable sorting scheme.

[0094] Specifically, key feature points and other feature points are extracted from the image data of smaller batches of products within the same batch. These are then input into a 3D prediction model, where the key feature points of the product are matched with the 3D spatial coordinate system of the prediction model. Error parameters are calculated between the spatial coordinates of other feature points in the 3D prediction model and the spatial coordinates of other feature points in the actual product. The relationship between these error parameters and the first and second error thresholds is determined, and a suitable sorting scheme is selected based on the results.

[0095] If the error parameter is less than the first error threshold, it means that the accuracy of the 3D prediction model is acceptable.

[0096] If the error parameter is greater than or equal to the first error threshold but less than the second error threshold, it means that the accuracy of the 3D prediction model needs further fine-tuning.

[0097] If the error parameter is greater than or equal to the second error threshold, it can be determined that the error parameter between the spatial coordinates of other feature points of the 3D prediction model and the spatial coordinates of other feature points of the actual product exceeds the prediction range, indicating that the 3D prediction model needs to be reconstructed.

[0098] For example, setting error parameters This indicates that the 3D prediction model is acceptable and can perform high-definition visual inspection and recognition of products.

[0099] set up This indicates that the 3D prediction model is further fine-tuned by calculating the error parameters between the spatial coordinates of other feature points in the 3D prediction model and the spatial coordinates of other feature points in the actual product, and inputting these parameters into the 3D prediction model to fine-tune the accuracy of the 3D prediction model.

[0100] set up This indicates that the 3D prediction model needs to be rebuilt, and multiple historical image data should be selected for training.

[0101] The pass threshold setting module is used to monitor the pass rate achieved during the product manufacturing process and set the first pass threshold and the second pass threshold based on the product pass rate.

[0102] Specifically, the pass rate of products in the same batch is monitored in real time during the production process. The standard for a product to pass is that the error parameter between the spatial coordinates of other feature points of the trained 3D prediction model and the spatial coordinates of other feature points of the actual product is less than the first error threshold. The first pass threshold and the second pass threshold are set according to the product pass rate.

[0103] For example, a precision mechanical part is set, which needs to undergo rigorous quality testing to ensure that it meets design requirements. The first pass threshold is set to 90%, and the second pass threshold is set to 95%.

[0104] The production plan optimization module sets a first pass threshold and a second pass threshold, compares the product pass rate with the first pass threshold and the second pass threshold, and selects an appropriate optimization plan based on the judgment result.

[0105] Specifically, the pass rate of products in the same batch is monitored in real time, and the relationship between the pass rate and the first and second pass thresholds is compared. Based on the judgment results, an appropriate optimization scheme is selected, including:

[0106] If the product's pass rate is greater than the second pass threshold, it means that the product can directly pass the quality inspection.

[0107] If the product pass rate is greater than the first pass threshold but less than or equal to the second pass threshold, it indicates that the company needs to further check whether there are any errors in the production process and procedures.

[0108] If the product pass rate is less than or equal to the first pass threshold, it means that the company needs to further check whether there are any errors in the shape and size of the batch of products.

[0109] For example, if the product pass rate This indicates that the product passed quality inspection directly;

[0110] if Product pass rate To further examine and correct errors in the product's manufacturing process and procedures;

[0111] If the product pass rate Further inspection was conducted to check for errors in the shape and size of the batch of products.

[0112] Define the interfaces and communication protocols between each module, clarify the interface specifications between each module, including: interface name, parameter type, return value, etc., and formulate corresponding communication protocols to ensure that data and information can be transmitted correctly and efficiently between modules.

[0113] The above embodiments are only used to illustrate the technical solutions of the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A high-definition visual inspection and recognition system based on an image processing industrial control computer, characterized in that, Includes the following modules: The historical data acquisition module acquires image data of the same product from different angles, under different lighting conditions, and with different background environments based on historical records of the same product. The feature point extraction module is used to build convolutional neural networks in deep learning and extract key feature points and other feature points from image data. The 3D model construction module uses a camera tracking algorithm to track the camera's motion trajectory, measures the camera's 3D position using a laser rangefinder, and constructs a 3D spatial coordinate system by combining the lens shooting direction and corner or edge points among key feature points. The X and Y axes are parallel to the 2D image plane, and the Z axis is consistent with the lens shooting direction. Based on the camera's intrinsic and extrinsic parameters, the projection matrix is ​​solved, and the spatial coordinates of other feature points in the 3D prediction model are calculated through a back-projection process. The error threshold setting module obtains the error values ​​between the spatial coordinates of other feature points in the calculation of the 3D prediction model and the spatial coordinates of other feature points in different images, and dynamically sets the first error threshold and the second error threshold according to the distribution of the error dataset. In the sorting scheme selection module, if the error parameter is less than the first error threshold, the 3D prediction model is acceptable; if the error parameter is greater than or equal to the first error threshold but less than the second error threshold, the 3D prediction model needs to be fine-tuned; if the error parameter is greater than or equal to the second error threshold, the 3D prediction model needs to be reconstructed. The pass threshold setting module is used to monitor the pass rate achieved during the product manufacturing process and set the first pass threshold and the second pass threshold. Key feature points and other feature points of different images are obtained and input into the three-dimensional prediction model. The key feature points of different images are matched with the three-dimensional spatial coordinate system of the three-dimensional prediction model. The least squares method is used to approximate the real three-dimensional coordinate points. There are errors between the three-dimensional coordinate points of other feature points of the three-dimensional prediction model and the spatial coordinate points of other feature points actually measured. The calculated errors are combined to form a dataset and set a first error threshold and a second error threshold. Real-time monitoring of the pass rate achieved by products in the same batch during the production process. The product meets the pass standard when the error parameter between the spatial coordinates of other feature points of the trained 3D prediction model and the spatial coordinates of other feature points of the actual product is less than the first error threshold. The first pass threshold and the second pass threshold are set according to the product pass rate. In the production plan optimization module, if the product pass rate is greater than the second pass threshold, the product passes the test directly; if the product pass rate is greater than the first pass threshold but less than or equal to the second pass threshold, the production process is checked; if the product pass rate is less than or equal to the first pass threshold, the batch of products is checked for compliance.

2. The high-definition visual inspection and recognition system based on an image processing industrial control computer according to claim 1, characterized in that, There are errors between the 3D coordinates of other feature points in the 3D prediction model and the spatial coordinates of other feature points actually measured. A dataset is composed of these calculated errors to set a first error threshold and a second error threshold, including the following steps: For each feature point in the 3D prediction model, calculate the spatial coordinate error between other feature points in the 3D prediction model and the corresponding feature points in actual measurement, and calculate the average error and maximum error of the spatial coordinate error of other feature points; The spatial coordinate error is calculated and satisfies the following formula: ； in, , , It is the third in the three-dimensional prediction model The x, y, z coordinates of each feature point; , , It is the first actual measurement The x, y, z coordinates of each feature point; It is spatial coordinate error; The average error is calculated and satisfies the following formula: ； in, It is the average error. Indicates the number of feature points. It is the sum of the spatial coordinate errors of all feature points; The maximum error is calculated according to the following formula: ； in, This indicates finding the largest error among all the spatial coordinate errors of all feature points.

3. The high-definition visual inspection and recognition system based on an image processing industrial control computer according to claim 1, characterized in that, Acquiring the image data includes the following steps: Collect multiple image data from the historical records of the same product. These multiple image data include images with different lighting conditions, angles, and background environments. Image data is preprocessed using filtering techniques to reduce noise, contrast adjustment and sharpening, normalization, and correction.

4. The high-definition visual inspection and recognition system based on an image processing industrial control computer according to claim 1, characterized in that, Building a convolutional neural network in deep learning includes the following steps: Label key feature points and other feature points of image data, construct a convolutional neural network in deep learning, and divide the labeled image data into training set, validation set and test set; The training set trains a convolutional neural network to extract key feature points and other feature points from image data with different angles, lighting conditions, and background environments. The validation set verifies whether the convolutional neural network can extract key feature points and other feature points from the image data. The test set tests the convolutional neural network and ultimately evaluates the feature point extraction performance.

5. A high-definition visual inspection and recognition system based on an image processing industrial control computer according to claim 4, characterized in that, Constructing the three-dimensional prediction model includes the following steps: By utilizing the camera's intrinsic and extrinsic parameters and combining the correspondence of key feature points, a projection matrix is ​​solved to describe the mapping relationship from three-dimensional space to a two-dimensional image plane. By combining the establishment of a three-dimensional spatial coordinate system and the solution of the projection matrix, the three-dimensional prediction model is constructed. The coordinates of other feature points of the two-dimensional image plane in three-dimensional space are calculated using the back projection process of the projection matrix.

Images