Machine vision-assisted method for monitoring size accuracy of automobile injection molded part

Abstract

The invention discloses a machine vision-assisted method for monitoring the size accuracy of an automobile injection molded part, and relates to the technical field of machine vision, and the method comprises the steps: irradiating the injection molded part through the combination of a visible light source and a near-infrared light source, synchronously collecting a three-view image, segmenting an injection molded part region, selecting a key-sized ROI region, and carrying out the fusion to generate a multi-view fusion image; identifying the minimum key feature size from the ROI region, decomposing and fusing the image through multi-scale morphological iteration processing, extracting a real contour signal by adopting an ICA blind source separation algorithm, and generating a complete contour model through multi-view edge point fusion reconstruction; marking feature points based on a CAD standard model, positioning the feature points by adopting particle swarm optimization and a Bayesian iterative algorithm, and calculating critical dimension parameters; environment and injection molding part surface error factors are collected, a dynamic error calibration model is constructed to compensate dimensional deviation, a final size value after calibration is output, and the accuracy and reliability of automobile injection molding part size detection are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of machine vision technology, specifically to a machine vision-assisted method for monitoring the dimensional accuracy of automotive injection molded parts. Background Technology

[0002] In the production process of automotive injection molded parts, monitoring dimensional accuracy is a key step in ensuring product quality. Currently, traditional inspection methods mainly include manual measurement and contact measuring equipment, such as micrometers, calipers, and coordinate measuring machines (CMMs). With the development of machine vision technology, non-contact optical measurement methods have been gradually introduced, such as laser scanning and structured light projection technology. These technologies achieve dimensional measurement through image acquisition and processing, improving the degree of automation and avoiding surface damage problems that may be caused by contact measurement. Some advanced systems have begun to adopt multi-view image fusion and light source combination technology to improve measurement stability under complex surface conditions.

[0003] However, in existing technologies, machine vision-based size monitoring methods mostly use a single light source and single-view image acquisition, which can easily lead to image quality degradation due to specular reflection or uneven surface texture, thus affecting measurement accuracy. Traditional image processing algorithms have difficulty accurately segmenting injection molded parts under complex backgrounds or noise interference, and the contour extraction process is easily affected by interference signals such as reflection and deformation, resulting in incomplete or distorted contour reconstruction. These limitations make existing methods perform poorly in high-precision and high-stability industrial inspection scenarios.

[0004] Furthermore, existing technologies rely heavily on single optimization algorithms for feature point localization and size calculation, lacking dynamic compensation mechanisms for historical data and environmental errors. For example, particle swarm optimization may get stuck in local optima during initial localization, while Bayesian iterative algorithms may suffer from low iteration efficiency or positioning deviations if the prior distribution is not set properly. The influence of environmental factors such as temperature and humidity on the dimensions of injection molded parts is often overlooked, and the lack of an effective dynamic error calibration model further reduces the reliability and repeatability of the detection results. Summary of the Invention

[0005] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides a machine vision-assisted method for monitoring the dimensional accuracy of automotive injection molded parts. It employs a combination of visible and near-infrared light sources, simultaneous acquisition and fusion of multi-view images, multi-scale morphological processing combined with the ICA blind source separation algorithm to extract the true contour signal, and particle swarm optimization and Bayesian iterative algorithms to achieve high-precision feature point localization. This method solves the problems of low measurement accuracy and poor stability caused by specular reflection, noise interference, contour distortion, and environmental errors in traditional detection methods, significantly improving the accuracy and reliability of automotive injection molded part dimensional inspection.

[0006] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: a machine vision-assisted method for monitoring the dimensional accuracy of automotive injection molded parts, characterized by comprising: The injection molded part is illuminated by a combination of visible light and near-infrared light sources. Three-view images are acquired simultaneously, the injection molded part area is segmented and the ROI area of ​​key dimensions is selected, and the visible light and near-infrared images are fused to generate a multi-view fused image. The smallest key feature size is identified in the ROI region, the image is decomposed and fused through multi-scale morphological iterative processing, the real contour signal is extracted by the ICA blind source separation algorithm, and a complete contour model is generated by multi-view edge point fusion reconstruction. Feature points are marked based on the CAD standard model. Particle swarm optimization and Bayesian iterative algorithm are used to locate the feature points, calculate key size parameters and perform correlation verification. Collect environmental and injection molded part surface error factors, construct a dynamic error calibration model to compensate for dimensional deviations, output the final dimensional value after calibration, compare it with the qualified threshold for classification, and generate an inspection report.

[0007] Furthermore, the combined illumination method of visible light and near-infrared light sources involves the visible light source shining obliquely at a 45° angle onto the surface of the injection molded part, while the near-infrared light source illuminates vertically. Three high-definition industrial cameras are deployed at positions directly in front, 45° to the left front, and 45° to the right front, respectively, to simultaneously acquire visible light and near-infrared images from three perspectives. Image denoising employs an improved median filter, selecting the filter window size based on the grayscale standard deviation, and retaining pixels with grayscale deviation ≤ grayscale deviation threshold.

[0008] Furthermore, a checkerboard calibration plate is placed at the injection molding part detection position, and the camera intrinsic and extrinsic parameters are calculated using the Zhang calibration method to correct image distortion; with the front view as the reference image, reference holes are preset on the injection molding part in the front view, and corresponding feature points are matched in the left and right views using the SIFT feature matching algorithm.

[0009] Furthermore, the Otsu adaptive threshold segmentation algorithm is used to binarize the front view, segmenting the injection molded part area and the background area; the CAD design model of the automotive injection molded part is imported, the area corresponding to the key dimensions in the model is extracted, the coordinates of the CAD model are mapped to the pixel coordinates of the image, and the ROI covering the edge of the key dimension area is selected in the image; the visible light image and the near-infrared image from the same viewpoint are weighted and fused.

[0010] Furthermore, based on the pixel scale of the camera calibration, the size of the identified minimum key feature is converted into the pixel size, generating three sets of disk-shaped structural elements of different scales. The diameters of scales 1 to 3 are 1 / 3, 1 / 2, and 1 times the minimum aperture pixel scale, respectively. Morphological iterative processing is performed in ascending order of scale, sequentially completing erosion, dilation, and reconstruction operations. The number of erosion and dilation operations at each scale are 3, 2, and 1 times, respectively, and the contour similarity between the decomposed image and the original fused image is greater than or equal to the preset similarity threshold.

[0011] Furthermore, the three sets of decomposed images are used as input signal matrices, and the ICA blind source separation algorithm is used for optimization. Three independent signals to be separated are defined: true contour signal S1, reflection interference signal S2, and deformation interference signal S3. The separation matrix W is initialized as a 3×3 random matrix, and the kurtosis value of each signal is calculated. When the kurtosis value of the three signals deviates from 3 by a factor greater than or equal to the deviation threshold, the signals are determined to be statistically independent. Otherwise, the separation matrix W is iteratively updated using the gradient descent method. The iteration objective is to maximize the non-Gaussianity of the signal kurtosis. After each iteration, a new signal kurtosis value is calculated until the separation matrix W converges, and the true contour signal S1 is separated.

[0012] Furthermore, the Canny edge detection algorithm is used to extract key edge points of the real contour signal from multiple views. The edge points of the left and right views are transformed to the coordinates of the front view. The cubic spline interpolation algorithm is used to connect the edge points of the same feature from multiple views to fill the missing contour areas of a single view to generate a complete contour model. Gaussian filtering is then applied to the complete contour model for smoothing.

[0013] Furthermore, N key feature points are marked in the CAD standard model. These feature points include the center point of the hole, the midpoint of the edge of the wall thickness area, the center point of the boss, and the intersection of the edges. The three-dimensional standard coordinates of each feature point and the relative positional constraints between feature points are recorded to construct a standard feature point template. Initial positioning is performed using a particle swarm optimization algorithm. A particle swarm is generated around the theoretical pixel coordinate range of the standard feature points in the complete contour model. Particle inertia weights and learning factors are set. The fitness function is the reciprocal of the Euclidean distance between the candidate coordinates and the standard feature point template. After a preset number of iterations, the particle coordinates with the highest fitness are selected as the preliminary positioning result. Fine positioning is performed using a Bayesian iterative optimization algorithm. A prior distribution is constructed based on the feature point coordinate dataset of historical qualified injection molded parts. The posterior distribution is iteratively updated using the preliminary positioning result as the observation value until the feature point coordinate deviation is ≤ the coordinate deviation threshold, at which point the iteration stops.

[0014] Furthermore, based on the pixel scale of the camera calibration, the pixel coordinates of the feature points are converted into actual physical coordinates. Key dimension parameters are calculated according to the relative positional relationship between the feature points. Key dimension parameters include linear dimensions and geometric dimensions. Each dimension parameter is calculated multiple times and the average value is taken. Based on the inherent constraint relationship of each dimension of the injection molded part in the CAD model, the calculated original dimension values ​​are verified for correlation. If the verification deviation exceeds the preset range, the prior distribution parameters of the Bayesian iteration are adjusted, and the fine positioning and dimension calculation are re-executed until the correlation verification is passed.

[0015] Furthermore, sensors are deployed in the inspection area to collect ambient temperature, ambient humidity, surface temperature of the injection molded part, and intensity values ​​of deformation interference signals. These parameters are stored in a database as error factors. Based on the error factor database and historical inspection data, a dynamic error calibration model is constructed using a multiple linear regression algorithm to obtain the final calibrated dimensional value, which is then compared with the pass threshold for graded judgment. An inspection report containing basic information of the injection molded part, dimensional data, and judgment results is generated. If the failure rate reaches a preset proportion, a parameter adjustment prompt for the injection molding production equipment is triggered.

[0016] (III) Beneficial Effects This invention provides a machine vision-assisted method for monitoring the dimensional accuracy of automotive injection molded parts, which has the following beneficial effects: (1) By combining visible light and near-infrared light sources for illumination and simultaneously acquiring data from three perspectives, the specular reflection problem of a single light source is effectively avoided, enhancing the contrast of surface texture. Improved median filtering and weighted fusion techniques balance noise suppression and edge preservation, improving image quality. Through checkerboard calibration and SIFT feature matching, high-precision image correction and registration are achieved, laying a reliable foundation for subsequent dimensional measurements. This significantly improves the accuracy of injection molded part region segmentation and critical dimension ROI selection, ensuring the stability and repeatability of the inspection.

[0017] (2) By decomposing and fusing images through multi-scale morphological iterative processing, and combining the ICA blind source separation algorithm, the real contour signal is accurately extracted, effectively separating reflection and deformation interference, significantly improving the contour restoration accuracy. Multi-view edge point fusion and cubic spline interpolation are used to reconstruct the complete contour model, fill in the missing data in the occluded area, ensure the continuity and integrity of the three-dimensional contour, and Gaussian filtering smoothing optimization further eliminates jagged noise, making the contour closer to the actual physical characteristics, providing a high-fidelity basis for subsequent size calculation, and improving the overall measurement stability and repeatability.

[0018] (3) High-precision positioning of feature points is achieved through particle swarm optimization and Bayesian iterative algorithm. Combined with CAD standard model and historical data, the accuracy and robustness of feature point coordinates are significantly improved. Based on physical coordinate conversion and multiple measurements and averaging, the reliability of key dimension parameters is ensured. Through the correlation verification of inherent constraint relationship, the positioning parameters are dynamically adjusted, which effectively avoids cumulative error and makes the dimension measurement results closer to the actual value. Overall, the repeatability and consistency of detection are improved, meeting the needs of high-precision industrial detection.

[0019] (4) Through dynamic acquisition of multi-dimensional error factors and multivariate linear regression modeling, high-precision size deviation compensation is achieved. Combined with qualified threshold grading judgment and automatic report generation, the reliability and traceability of the test results are significantly improved. When the non-conformity rate exceeds the limit, the equipment parameter adjustment prompt is triggered, forming a closed-loop quality control, effectively reducing the risk of batch defects, and meeting the high-precision and high-stability production and testing needs of automotive injection molded parts. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the steps of the machine vision-assisted method for monitoring the dimensional accuracy of automotive injection molded parts according to the present invention; Figure 2 This is a schematic diagram of the machine vision-assisted method for monitoring the dimensional accuracy of automotive injection molded parts according to the present invention. Detailed Implementation

[0021] 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, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Please see Figures 1-2 This invention provides a machine vision-assisted method for monitoring the dimensional accuracy of automotive injection molded parts, comprising the following steps: Step 1: Irradiate the injection molded part with a combination of visible light and near-infrared light sources, simultaneously acquire three-view images, segment the injection molded part area and select the ROI area of ​​key dimensions, and fuse the visible light and near-infrared images to generate a multi-view fused image; Step one includes the following: Step 101: Construct a dual-light source, multi-view acquisition system, using a combination of visible light and near-infrared light sources. The visible light source is directed at a 45° angle to the surface of the injection molded part to avoid specular reflection caused by vertical illumination, while the near-infrared light source provides vertical illumination to enhance the contrast of surface texture. Simultaneously, deploy three high-definition industrial cameras, installed at the front, left front, and right front positions of the injection molded part, respectively, to simultaneously acquire visible light and near-infrared images from the front, left, and right views. Step 102: Denoise the visible light and near-infrared images from the three perspectives separately, and calculate the grayscale standard deviation σ of each image. When σ>20, a 9×9 pixel filtering window is used; when 10<σ≤20, a 5×5 pixel filtering window is used; when σ≤10, a 3×3 pixel filtering window is used. Based on the selected window, perform improved median filtering. During filtering, retain pixels whose grayscale value deviates from the center pixel's grayscale value by less than or equal to the grayscale deviation threshold (e.g., 15), and filter out noise pixels with deviations exceeding the threshold, thus achieving a balance between edge detail preservation and noise suppression. Step 103: Place the checkerboard calibration plate at the injection molded part inspection position. The checkerboard size is 2mm×2mm. Acquire images of the calibration plate. Calculate the intrinsic and extrinsic parameters of the three cameras using the Zhang calibration method. Based on the intrinsic and extrinsic parameters, perform radial and tangential distortion correction on the three-view images. Using the front view as the reference image, pre-set three reference holes on the injection molded part in the front view. Match the corresponding feature points in the left and right views using the SIFT feature matching algorithm. Step 104: The Otsu adaptive threshold segmentation algorithm is used to binarize the corrected and registered front view, segmenting the injection molded part area and the background area. The CAD design model of the automotive injection molded part is imported, and the areas corresponding to key dimensions in the model are extracted, such as wall thickness area, hole area, and boss spacing area. The coordinates of the CAD model are mapped to the pixel coordinates of the image, and the ROI (Region of Interest) is automatically selected in the image. The ROI range extends 5-10 pixels beyond the edge of the key dimension area to ensure that the key features are fully included. At the same time, irrelevant background areas are removed to reduce the amount of data processing. Step 105: Perform weighted fusion of the visible light image and the near-infrared image from the same viewpoint. Calculate the edge gradient magnitude G1 of the visible light image and the edge gradient magnitude G2 of the near-infrared image. When G1>G2, the weight coefficient α of the visible light channel is 0.6~0.7, and the weight coefficient β of the near-infrared channel is 0.3~0.4. When G1≤G2, α=0.3~0.4, β=0.6~0.7. Generate a fused image with both clear edges and low reflectivity interference, thus completing the generation of the three-view fused image.

[0023] When using this method, refer to steps 101 to 105: By combining visible and near-infrared light sources with simultaneous three-view acquisition, the specular reflection problem of a single light source is effectively avoided, enhancing surface texture contrast. Improved median filtering and weighted fusion techniques balance noise suppression and edge preservation, improving image quality. High-precision image correction and registration are achieved through checkerboard calibration and SIFT feature matching, laying a reliable foundation for subsequent dimensional measurements. This significantly improves the accuracy of injection molded part region segmentation and critical dimension ROI selection, ensuring the stability and repeatability of the inspection.

[0024] Step 2: Identify the smallest key feature size from the ROI region, decompose and fuse the image through multi-scale morphological iterative processing, extract the real contour signal using the ICA blind source separation algorithm, and generate a complete contour model through multi-view edge point fusion reconstruction. Step two includes the following: Step 201: Identify the smallest key feature size from the ROI, such as the smallest aperture and the thinnest wall thickness, and convert them to pixel size based on the pixel scale of the camera calibration; generate 3 sets of disk-shaped structural elements, scale 1 structural element diameter = round(smallest aperture × pixel scale / 3), corresponding to tiny details, scale 2 structural element diameter D2 = round(smallest aperture × pixel scale / 2), corresponding to medium contours, scale 3 structural element diameter D3 = round(smallest aperture × pixel scale), corresponding to the overall shape, to ensure that the structural element size matches the feature size and avoid feature loss or over-smoothing; Step 202: Perform morphological iterative processing on the three-view fused image. Use a scale 1 structuring element to perform erosion, dilation, and reconstruction operations. Erosion is performed 3 times, removing burrs from one pixel layer each time. Dilation is performed 3 times to restore the true edges. Reconstruction is performed once, constraining the dilation range based on the original image. Repeat erosion, dilation, and reconstruction using a scale 2 structuring element. Erosion is performed 2 times, dilation is performed 2 times, and reconstruction is performed once. Use a scale 3 structuring element to perform erosion, dilation, and reconstruction once, resulting in 3 sets of decomposed images. After each iteration, calculate the contour similarity J between the decomposed image and the original fused image. Decomposition is considered valid when J ≥ a preset similarity threshold (e.g., 0.95). Step 203: Use the three sets of decomposed images as the input signal matrix X. Each column represents the pixel grayscale signal of a set of decomposed images. Optimize using the ICA blind source separation algorithm. Set three independent signals to be separated: true contour signal S1, reflection interference signal S2, and deformation interference signal S3. Initialize the separation matrix W as a 3×3 random matrix with elements ranging from -0.5 to 0.5. Calculate the kurtosis value of each signal. When the kurtosis values ​​of the three signals deviate from 3 by ≥ the deviation threshold (e.g., 1.5), the signals are considered statistically independent. Otherwise, iteratively update the separation matrix W using gradient descent. The iteration objective is to maximize the non-Gaussianity of the signal kurtosis, i.e., to make the kurtosis of each signal deviate from 3 as much as possible. Calculate the new signal kurtosis value after each iteration until the separation matrix W converges. The change in W elements between two adjacent iterations should be ≤0.001. After separation, filter by signal features: retain signals with edge gradient magnitude ≥30 and adjacent pixel grayscale changes ≤20 as the pure true contour signal S1. Step 204: Perform multi-view contour fusion reconstruction on the real contour signal S1 after separation of the three views. Use the Canny edge detection algorithm to extract key edge points. Set the edge gradient threshold to 50~100 and adjust it according to the image grayscale. This includes edge points of hole positions, edge points on both sides of the wall thickness, and vertices of bosses. Transform the edge points of the left and right views to the coordinates of the front view. Use the cubic spline interpolation algorithm to connect the multi-view edge points with the same feature to fill the missing contour areas of a single view, such as the occluded parts of the injection molded part, and generate a complete contour model in three-dimensional space. Step 205: Perform contour smoothing optimization on the complete contour model, calculate the curvature value of each contour segment. When the curvature k>0.1, set the Gaussian filter standard deviation σ1=0.5~0.8; when k≤0.1, set σ1=0.8~1.2. Perform Gaussian filter smoothing on the contour, calculate the continuity error of the contour after filtering, and ensure that the deviation of the tangent angle between adjacent edge points is ≤0.5° to ensure that the contour is smooth and continuous without jagged deformation.

[0025] When using this method, refer to steps 201 to 205: By decomposing and fusing images through multi-scale morphological iterative processing, and combining the ICA blind source separation algorithm, the true contour signal is accurately extracted, effectively separating reflection and deformation interference, significantly improving the contour restoration accuracy. Multi-view edge point fusion and cubic spline interpolation are used to reconstruct the complete contour model, filling in missing data in occluded areas, ensuring the continuity and integrity of the three-dimensional contour. Gaussian filtering smoothing optimization further eliminates jagged noise, making the contour closer to the actual physical characteristics, providing a high-fidelity basis for subsequent size calculations, and improving the overall measurement stability and repeatability.

[0026] Step 3: Mark feature points based on the CAD standard model, use particle swarm optimization and Bayesian iterative algorithm to locate feature points, calculate key size parameters and perform correlation verification; Step three includes the following: Step 301: Import the CAD standard model of the automotive injection molded part, mark the feature points corresponding to the key dimensions in the CAD standard model, including the center point of each hole, the midpoint of the two sides of the wall thickness area, the center point of the boss, and the intersection of the edges, mark a total of N feature points, N≥8; record the three-dimensional standard coordinates of each feature point and the relative position constraints between feature points, such as the distance between the center of hole A and the center of hole B, the perpendicularity requirement between the center of hole C and the edge of the wall thickness, etc., and construct a standard feature point template; Step 302: Initial localization is performed using a particle swarm optimization algorithm. In the complete contour model, a particle swarm of 50 particles is generated around the theoretical pixel coordinate range of each standard feature point, ±5 pixels. The coordinates of each particle represent the coordinates of a candidate feature point. The particle inertia weight is set with an initial value of 0.9, which is linearly decreased to 0.4 during iteration. The individual learning factor C1=1.5 and the swarm learning factor C2==2.0. The fitness function is the reciprocal of the Euclidean distance between the candidate coordinates and the standard feature point template. The number of iterations is preset, such as 10 iterations. The particle position and velocity are updated in each iteration. After the iteration, the coordinates of the particle with the highest fitness are selected as the preliminary localization result of the feature point. Step 303: Use the Bayesian iterative optimization algorithm for fine localization. Take the preliminary localization results of feature points as observations, import the feature point coordinate dataset of historical qualified injection molded parts, with a sample size of ≥100 groups, calculate the prior distribution of feature point coordinates, which follows a normal distribution N(μ, σ2), where μ is the average of historical coordinates and σ2 is the prior standard deviation, construct a Bayesian posterior probability model, adjust the observation accuracy to ±0.1 pixels, re-collect observations, iteratively update the posterior distribution, and stop iterating when the feature point coordinate deviation between two adjacent iterations is ≤ the coordinate deviation threshold (e.g., 0.001 pixels), and output the three-dimensional coordinates of the feature points. Step 304: Based on the pixel scale of the camera calibration, convert the pixel coordinates of the feature points into actual physical coordinates. Based on the relative positional relationship between the feature points, calculate the key size parameters, including linear dimensions and geometric dimensions. Linear dimensions include length, width, wall thickness, hole diameter, and hole center distance. Geometric dimensions include roundness, parallelism, and perpendicularity. Each size parameter is calculated 3 times and the average value is taken. Step 305: Based on the inherent constraints in the CAD model, i.e., the geometrical logical relationships between the dimensions of the injection molded part, such as the sum of the center distances of holes along the same axis equaling the total length, perform a correlation verification on the calculated original dimensional values. For example, if the distance from hole center A to hole center B in the CAD model is... The distance from the center of hole B to the center of hole C is If the centers of holes A, B, and C are collinear, then the associated constraint relationship is as follows: + = , The total distance from hole center A to hole center C; calculate the verification deviation ΔL = |( + )- If the verification deviation is ≤0.003mm, the size calculation is deemed valid; if the verification deviation is >0.003mm, return to step 303, adjust the prior distribution parameters of the Bayesian iteration, expand the prior standard deviation to 1.2 times the original value, and re-execute the fine positioning and size calculation until the correlation verification passes.

[0027] When using this method, refer to steps 301 to 305: High-precision feature point positioning is achieved through particle swarm optimization and Bayesian iterative algorithms. Combined with CAD standard models and historical data, the accuracy and robustness of feature point coordinates are significantly improved. Based on physical coordinate conversion and averaging of multiple measurements, the reliability of key dimensional parameter calculations is ensured. By verifying the correlation of inherent constraint relationships, positioning parameters are dynamically adjusted to effectively avoid cumulative errors, making the dimensional measurement results closer to the actual values. Overall, the repeatability and consistency of the inspection are improved, meeting the needs of high-precision industrial inspection.

[0028] Step 4: Collect environmental and injection molded part surface error factors, construct a dynamic error calibration model to compensate for dimensional deviations, output the final dimensional value after calibration, compare it with the qualified threshold for classification, and generate an inspection report.

[0029] Step four includes the following: Step 401: Deploy sensing devices in the detection area to extract multi-dimensional error factors, deploy temperature and humidity sensors and infrared temperature measurement modules to collect ambient temperature, ambient humidity, and injection molded part surface temperature in real time; extract the intensity value of the separated deformation interference signal S3, the normalized result of the signal amplitude, ranging from 0 to 1; store the ambient temperature, ambient humidity, injection molded part surface temperature, and S3 intensity value as four error factors in the error factor database; Step 402: Based on the error factor database and historical test data, a dynamic error calibration model is constructed using a multiple linear regression algorithm. The inputs are ambient temperature, ambient humidity, injection molded part surface temperature, and the S3 strength value. The output is the corresponding dimensional correction value. The model parameters are fitted using the least squares method to achieve a goodness of fit R0. 2 ≥0.98, ensuring that the dimensional correction value can accurately compensate for dimensional deviations caused by various error factors; Step 403: Obtain the final calibrated dimensional value using key dimensional parameters and dimensional correction values; perform repeatability verification on the final dimensional value by performing 5 consecutive measurements and calculating the standard deviation of the 5 measurements. If the standard deviation is ≤0.002mm, the verification is considered successful; if it fails, readjust the error calibration model parameters and calibrate again. Step 404: Based on design requirements, preset the pass thresholds for each key dimension of the automotive injection molded part. Compare the calibrated final dimension values ​​with the pass thresholds and classify them into three levels: dimension values ​​within the pass thresholds with a deviation ≤ 0.002mm are considered excellent; dimension values ​​within the pass thresholds with a deviation of 0.002~0.005mm are considered average; and dimension values ​​exceeding the pass thresholds are considered unqualified. Step 405: Automatically generate an inspection report, including the injection molded part number, inspection time, original values, calibration values, deviation values ​​and judgment levels of each key dimension, and mark the specific location of the non-conforming dimensions; at the same time, feed the inspection data back to the injection molding production equipment. If the non-conforming rate reaches the preset ratio, such as a non-conforming rate ≥3%, a prompt to adjust the equipment parameters will be triggered.

[0030] When using this method, refer to steps 401 to 405: By dynamically acquiring multi-dimensional error factors and modeling multiple linear regression, high-precision dimensional deviation compensation is achieved. Combined with qualified threshold grading and automatic report generation, the reliability and traceability of the test results are significantly improved. When the non-conformance rate exceeds the limit, the equipment parameter adjustment prompt is triggered, forming a closed-loop quality control, effectively reducing the risk of batch defects, and meeting the high-precision and high-stability production and testing needs of automotive injection molded parts.

[0031] In the application, the various formulas mentioned are all calculated by removing dimensions and taking their numerical values. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The coefficients in the formulas are set by those skilled in the art according to the actual situation.

[0032] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, and combinations thereof. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0033] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0034] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A machine vision-assisted method for monitoring the dimensional accuracy of automotive injection molded parts, characterized in that: include: The injection molded part is illuminated by a combination of visible light and near-infrared light sources. Three-view images are acquired simultaneously, the injection molded part area is segmented and the ROI area of ​​key dimensions is selected, and the visible light and near-infrared images are fused to generate a multi-view fused image. The smallest key feature size is identified in the ROI region, the image is decomposed and fused through multi-scale morphological iterative processing, the real contour signal is extracted by the ICA blind source separation algorithm, and a complete contour model is generated by multi-view edge point fusion reconstruction. Feature points are marked based on the CAD standard model. Particle swarm optimization and Bayesian iterative algorithm are used to locate the feature points, calculate key size parameters and perform correlation verification. Collect environmental and injection molded part surface error factors, construct a dynamic error calibration model to compensate for dimensional deviations, output the final dimensional value after calibration, compare it with the qualified threshold for classification, and generate an inspection report.

2. The machine vision-assisted method for monitoring the dimensional accuracy of automotive injection molded parts according to claim 1, characterized in that: The combined illumination method of visible light and near-infrared light sources involves the visible light source shining obliquely at a 45° angle onto the surface of the injection molded part, while the near-infrared light source illuminates it vertically. Three high-definition industrial cameras are deployed at positions directly in front, 45° to the left front, and 45° to the right front, respectively, to simultaneously acquire visible light and near-infrared images from three perspectives. Image denoising employs an improved median filter, selecting the filter window size based on the grayscale standard deviation, and retaining pixels with grayscale deviation ≤ grayscale deviation threshold.

3. The machine vision-assisted method for monitoring the dimensional accuracy of automotive injection molded parts according to claim 2, characterized in that: A checkerboard calibration plate is placed at the injection molded part detection position, and the camera intrinsic and extrinsic parameters are calculated using the Zhang calibration method to correct image distortion. Using the front view as the reference image, reference holes are preset on the injection molded part in the front view, and corresponding feature points are matched in the left and right views using the SIFT feature matching algorithm.

4. The machine vision-assisted method for monitoring the dimensional accuracy of automotive injection molded parts according to claim 3, characterized in that: The Otsu adaptive threshold segmentation algorithm is used to binarize the front view, segmenting the injection molded part area and the background area; the CAD design model of the automotive injection molded part is imported, the area corresponding to the key dimensions in the model is extracted, the coordinates of the CAD model are mapped to the pixel coordinates of the image, and the ROI covering the edge of the key dimension area is selected in the image; the visible light image and the near-infrared image from the same viewpoint are weighted and fused.

5. The machine vision-assisted method for monitoring the dimensional accuracy of automotive injection molded parts according to claim 1, characterized in that: Based on the pixel scale of the camera calibration, the size of the identified minimum key feature is converted into the pixel size, generating three sets of disk-shaped structural elements of different scales. The diameters of scales 1 to 3 are 1 / 3, 1 / 2, and 1 times the minimum aperture pixel scale, respectively. Morphological iterative processing is performed in ascending order of scale, sequentially completing erosion, dilation, and reconstruction operations. The number of erosion and dilation operations at each scale are 3, 2, and 1 times, respectively, and the contour similarity between the decomposed image and the original fused image is greater than or equal to the preset similarity threshold.

6. The machine vision-assisted method for monitoring the dimensional accuracy of automotive injection molded parts according to claim 5, characterized in that: Three sets of decomposed images are used as input signal matrices. The ICA blind source separation algorithm is used for optimization. Three independent signals to be separated are defined as the true contour signal S1, the reflection interference signal S2, and the deformation interference signal S3. The separation matrix W is initialized as a 3×3 random matrix. The kurtosis value of each signal is calculated. When the kurtosis value of the three signals deviates from 3 by a factor greater than or equal to the deviation threshold, the signals are considered statistically independent. Otherwise, the separation matrix W is iteratively updated using the gradient descent method. The iteration objective is to maximize the non-Gaussianity of the signal kurtosis. After each iteration, a new signal kurtosis value is calculated until the separation matrix W converges and the true contour signal S1 is separated.

7. The machine vision-assisted method for monitoring the dimensional accuracy of automotive injection molded parts according to claim 6, characterized in that: The Canny edge detection algorithm is used to extract key edge points of the real contour signal from multiple views. The edge points of the left and right views are transformed to the coordinates of the front view. The cubic spline interpolation algorithm is used to connect the edge points of the same feature from multiple views to fill the missing contour areas of a single view to generate a complete contour model. Gaussian filtering is then applied to the complete contour model for smoothing.

8. The machine vision-assisted method for monitoring the dimensional accuracy of automotive injection molded parts according to claim 1, characterized in that: In the CAD standard model, N key feature points are marked, including the center point of the hole, the midpoint of the edge of the wall thickness area, the center point of the boss, and the intersection of the edges. The three-dimensional standard coordinates of each feature point and the relative positional constraints between the feature points are recorded to construct a standard feature point template. Initial positioning is performed using a particle swarm optimization algorithm. A particle swarm is generated around the theoretical pixel coordinate range of the standard feature points in the complete contour model. The particle inertia weight and learning factor are set. The fitness function is the reciprocal of the Euclidean distance between the candidate coordinates and the standard feature point template. After a preset number of iterations, the particle coordinates with the highest fitness are selected as the preliminary positioning result. Fine positioning is performed using a Bayesian iterative optimization algorithm. A prior distribution is constructed based on the feature point coordinate dataset of historical qualified injection molded parts. The posterior distribution is iteratively updated using the preliminary positioning result as the observation value until the feature point coordinate deviation is ≤ the coordinate deviation threshold, at which point the iteration stops.

9. The machine vision-assisted method for monitoring the dimensional accuracy of automotive injection molded parts according to claim 8, characterized in that: Based on the pixel scale of the camera calibration, the pixel coordinates of feature points are converted into actual physical coordinates. Key dimension parameters are calculated according to the relative positional relationship between feature points. Key dimension parameters include linear dimensions and geometric dimensions. Each dimension parameter is calculated multiple times and the average value is taken. Based on the inherent constraint relationship of each dimension of the injection molded part in the CAD model, the calculated original dimension values ​​are verified for correlation. If the verification deviation exceeds the preset range, the prior distribution parameters of Bayesian iteration are adjusted, and the fine positioning and dimension calculation are re-executed until the correlation verification passes.

10. The machine vision-assisted method for monitoring the dimensional accuracy of automotive injection molded parts according to claim 9, characterized in that: Sensors are deployed in the detection area to collect ambient temperature, ambient humidity, surface temperature of the injection molded part, and intensity values ​​of deformation interference signals. These parameters are stored in the database as error factors. Based on the error factor database and historical test data, a dynamic error calibration model is constructed using a multiple linear regression algorithm to obtain the final calibrated dimensional value, which is then compared with the qualified threshold for graded judgment. An inspection report containing basic information of the injection molded part, dimensional data, and judgment results is generated. If the failure rate reaches the preset ratio, a parameter adjustment prompt for the injection molding production equipment is triggered.

Images