Visual inspection system for hub baking varnish quality

Abstract

The invention belongs to the technical field of hub production, and particularly relates to a hub baking varnish quality visual detection system, which comprises S1, a transmission and positioning module for stably transmitting a hub to be detected to a detection station and clamping and rotating, and S2, a multi-modal data acquisition module for synchronously acquiring multi-dimensional information of the surface of the hub, the module comprises a high-resolution 2D visual unit, a high-precision 3D contour scanning unit and a dynamic structured illumination unit. According to the visual detection system for the hub baking varnish quality, the high-resolution 2D visual unit, the high-precision 3D contour scanning unit and the dynamic structured illumination unit are combined, multi-dimensional information of the surface of a hub is synchronously collected, illumination compensation and reflection elimination are performed on 2D images in different illumination modes by using a Phong illumination model and a pixel value compensation algorithm, and the hub baking varnish quality is detected. A standard image without reflection distortion is generated, clean and reliable input data is provided for subsequent defect recognition, and the misjudgment and omission ratio is remarkably reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of hub production, and in particular to a visual detection system for the quality of hub baking paint. BACKGROUND

[0002] As a key appearance part of a vehicle, the surface baking paint quality of a hub directly affects the visual aesthetics and brand image of the vehicle. At present, in the hub manufacturing industry, the quality final inspection link of the baking paint surface mainly has the following technical status and challenges:

[0003] For example, most domestic manufacturers still mainly rely on manual visual detection. This method has problems such as low efficiency, high labor intensity, and non-uniform detection standards. Moreover, the detection results are easily affected by the subjective experience, visual fatigue, and emotional state of the detection personnel, resulting in poor consistency in judgment, high rates of missed detection and false detection. In addition, manual detection cannot achieve digital recording and tracing of detection data, and cannot meet the requirements of modern intelligent manufacturing for process control and quality management.

[0004] To replace manual work, automatic detection technology based on machine vision has been introduced into this field, but when applied to the baking paint surface of a hub, it has encountered serious challenges. The main problem is that 2D vision technology is heavily dependent on stable lighting conditions. However, the baking paint surface of a hub has the characteristics of high reflectivity and strong mirror effect. Under fixed lighting, mirror reflection can cause local overexposure of the image, completely covering the real texture or color information in that area. At the same time, the resulting shadows are easily misjudged as recesses and other geometric defects. Therefore, 2D vision is extremely sensitive to lighting in this application, and its stability and reliability are insufficient. Meanwhile, 3D vision technology is not sensitive to color changes, subtle texture differences, and minor surface imperfections, making it difficult to effectively identify typical baking paint appearance defects such as orange peel, color difference, and shallow scratches. Therefore, the present application solves the above technical problems. SUMMARY

[0005] Based on the above technical problems, the present application proposes a visual detection system for the quality of hub baking paint.

[0006] The visual detection system for the quality of hub baking paint proposed by the present application comprises:

[0007] S1, a conveying and positioning module, for stably conveying the hub to be detected to a detection station and clamping and rotating it.

[0008] S2, a multi-modal data acquisition module, for synchronously acquiring multi-dimensional information of the hub surface. The module comprises:

[0009] a high-resolution 2D vision unit for acquiring color image information of the hub surface.

[0010] A high-precision 3D profile scanning unit is used to collect three-dimensional point cloud data of the hub surface.

[0011] A dynamic structured lighting unit comprises a plurality of independently controllable light sources, which are switched according to control instructions to switch a plurality of different lighting modes for sequential shooting.

[0012] S3, an intelligent cooperative motion control module, is used to drive the multi-modal data acquisition module to move along a predetermined path and adjust the observation posture thereof.

[0013] S4, a data processing and defect identification module, is electrically connected with the multi-modal data acquisition module, and the module comprises:

[0014] A data preprocessing and registration unit is used to accurately align the 2D images, 3D point cloud and corresponding lighting mode information collected at the same position in space and time.

[0015] A multi-modal data fusion unit is used to calculate the surface normal and curvature geometric information corresponding to each pixel point based on the 3D point cloud data, and based on the geometric information, the 2D images under different lighting modes are subjected to illumination compensation and reflection elimination to generate a standard image without reflection distortion.

[0016] A hybrid intelligent detection model unit is used to detect defects in the standard image, including: locating a suspected defect area through a primary algorithm, and then classifying and rating the suspected defect area through a deep learning convolutional neural network.

[0017] S5, a man-machine interaction and closed-loop learning module, is used to visually display the detection results and allow manual labeling of the detection results, and the labeled data is fed into a self-learning database to incrementally train and optimize the hybrid intelligent detection model.

[0018] Preferably, the 2D vision unit comprises an industrial area array camera, the 3D profile scanning unit comprises a high-speed line laser 3D scanner, and the intelligent cooperative motion control module comprises a six-axis robot, the end flange plate of which is integrated with the industrial area array camera and the laser 3D scanner, and the six-axis robot adaptively plans an optimal scanning path and posture according to the hub model.

[0019] Preferably, the spatial alignment of the data preprocessing and registration unit in step S4 is realized by the following method:

[0020] A rigid transformation matrix between the 2D vision unit and the 3D profile scanning unit is obtained through joint calibration , so that for a point in the 3D profile scanning unit coordinate system , the corresponding point in the 2D vision unit coordinate system is calculated by the following formula:

[0021] , and then project the transformed 3D point cloud to the 2D image plane by camera model to establish pixel-level correspondence, wherein and are the homogeneous coordinates of the point in the 3D profile scanning unit coordinate system and the 2D vision unit coordinate system, respectively, , , is the three-dimensional space coordinate of the point in the 3D profile scanning unit coordinate system, , , is the three-dimensional space coordinate of the point in the 2D vision unit coordinate system, and the transformation matrix is a 4x4 rigid transformation matrix containing rotation and translation transformation.

[0022] The time alignment is achieved by the following way:

[0023] When the six-axis robot moves to each scanning position, the control module sends a hardware trigger signal to synchronously control the dynamic structured illumination unit to switch mode, the 2D vision unit to collect images, and the 3D profile scanning unit to collect profile data, and all the data collected for the same position are given the same position identifier.

[0024] Preferably, the multi-modal data fusion unit in step S4 comprises the following steps:

[0025] The normal vector of each point on the surface is calculated using the registered 3D point cloud data .

[0026] Based on the unit normal vector of the surface point , the unit direction vector from the surface point to the light source , and the perspective direction vector from the surface point to the 2D vision unit optical center , the surface illumination intensity is predicted according to the reflection model.

[0027] The reflection model is the Phong illumination model combined with diffuse reflection component and specular reflection component, which is expressed as follows:

[0028] , wherein:

[0029] represents the total illumination intensity predicted by the model.

[0030] represents the ambient light intensity, represents the sum of the contributions of multiple light sources.

[0031] represents the intensity of a single light source, is the specular reflection coefficient of the surface material.

[0032] is the reflection light direction vector calculated according to the incident light direction and the normal vector .

[0033] is the highlight index representing the smoothness of the surface, is the diffuse reflection coefficient of the surface material.

[0034] The fusion image is generated by analyzing multiple 2D images collected under different illumination modes, identifying and separating the specular highlight and diffuse reflection components using the reflection model, and then eliminating the highlight interference through a pixel value compensation algorithm.

[0035] Preferably, the elimination of highlight interference through a pixel value compensation algorithm to generate a fusion image includes, for each pixel position , the following steps are performed:

[0036] For each illumination mode , the specular highlight intensity is calculated according to the reflection model.

[0037] Compare with a preset threshold , when > , mark the pixel in the image as a highlight pixel.

[0038] Define a set , which is the set of values of all pixels in the image that are not overexposed.

[0039] When is not empty, then , that is , a value can be selected from , including selecting the value with the highest brightness in , another strategy is to select the pixel value in the image with the largest diffuse reflection component in the set , the diffuse reflection component estimate value , where:

[0040] , is a set function of a function.

[0041] When is empty, then If a pixel is a highlight in all images, then we can estimate its value using spatial interpolation. It is a pixel-based Surrounding Neighborhood Interpolation function for non-highlight pixel values.

[0042] Preferably, the hybrid intelligent detection model unit in step S4 employs a two-level progressive detection architecture to identify defects in the fused non-reflective distortion image, including:

[0043] The primary filtering module is configured to extract edge features, texture features, and color consistency features of the fused image in parallel, and quickly locate all suspected defective regions of interest (ROIs) based on adaptive threshold segmentation and geometric feature filtering. Color consistency detection is achieved in the CIELAB color space by calculating local color differences.

[0044] And the final ROI set is composed of Confirmed, among which:

[0045] :Location The color difference value at that location.

[0046] , , :Location The brightness, red-green hue, and yellow-blue hue components in the CIELAB color space.

[0047] , , : The corresponding component of the reference standard area in the CIELAB color space.

[0048] The final set of suspected defective areas.

[0049] Candidate region :area The area.

[0050] Minimum area threshold :area The aspect ratio.

[0051] , The minimum and maximum thresholds for aspect ratio. : Logical AND operator.

[0052] The secondary fine inspection module is configured to use a deep convolutional neural network to refine the processing of each ROI. This network employs a multi-task learning framework and simultaneously outputs the defect type classification probability. and severity rating The final detection result is generated by combining the type confidence score and the severity score.

[0053] ,in:

[0054] : Predicted probability distribution of defect types : Normalized exponential function, used to convert network output into a probability distribution.

[0055] : Weight matrix of the defect type classification layer The feature vector output by the last layer of a deep convolutional neural network. : Bias vector of the defect type classification layer.

[0056] : Predicted defect severity rating Weight matrix of the severity level regression layer Bias vector of the severity level regression layer.

[0057] : Comprehensive confidence score for defect identification Defect types predicted by the network The corresponding probability value, The maximum severity rating set by the system.

[0058] Preferably, the conveying and positioning module in step S1 further includes a drying conveyor line. A detection station is provided on one side of the drying conveyor line. A clamping housing is fixedly connected to the lower surface of the hanging rail of the drying conveyor line. A clamping turntable is rotatably connected to one side surface of the clamping housing. A hook for hooking the hub is fixedly connected to the inner surface of the clamping turntable. A guide groove and a movable cavity are respectively opened inside the clamping turntable. One end of the two guide grooves is fixedly connected to the inside of the movable cavity. A bidirectional adjusting screw is rotatably connected inside the movable cavity. An L-shaped clamping block with a flexible rubber outer surface is threaded onto the outer surface of each end of the bidirectional adjusting screw. The outer surface of the clamping block is slidably engaged with the inner wall of the guide groove.

[0059] Preferably, the clamping housing has a drive cavity inside, and two drive motors are installed inside the drive cavity. A rotating shaft is rotatably connected inside the drive cavity via bearings. The free end surface of the rotating shaft is fixedly connected to the inside of the clamping turntable. A worm sleeve is rotatably fitted onto the outer surface of the rotating shaft, and one end of the worm sleeve extends into the interior of the movable cavity. An adjusting worm wheel is fixedly fitted onto the outer surface of the middle part of the bidirectional adjusting screw, and the adjusting worm wheel meshes with the worm sleeve. A driving gear is fixedly fitted onto the outer surface of the output shaft of one of the drive motors, and a driven gear is fixedly fitted onto the outer surface of the rotating shaft. The driving gear meshes with the driven gear. An adjusting worm is fixedly connected to the outer surface of the output shaft of the other drive motor via a coupling. A worm wheel rod with two worm wheels is rotatably connected inside the drive cavity. The two worm wheels on the worm wheel rod mesh with the worm sleeve and the adjusting worm, respectively.

[0060] Preferably, the dynamic structured lighting unit further includes a linear guide rail fixedly connected to the upper surface of the detection station base plate. A support slider is slidably engaged on the outer surface of the linear guide rail. A regular polygonal frame is fixedly connected to the upper surface of the support slider. Parallel rods are fixedly connected in a circular array on the inner surface of the regular polygonal frame. A grooved plate is fixedly connected to the inner surface of a plurality of parallel rods, and a linkage shaft is rotatably connected to them. Adjacent surfaces of adjacent linkage shafts are connected by universal joints. A rack rod is slidably engaged on the inner surface of the grooved plate. An LED light panel is fixedly connected to the side of the rack rod near the center point of the regular polygonal frame. The LED light panel has an arc-shaped design.

[0061] Preferably, a linkage gear is fixedly sleeved on the outer surface of the linkage shaft, and the linkage gear meshes with the corresponding rack rod. A linkage motor is fixedly installed on the surface of one of the parallel rods, and a drive gear is fixedly connected to the outer surface of the output shaft of the linkage motor. A synchronization gear is fixedly sleeved on the outer surface of one of the linkage shafts, and the synchronization gear meshes with the drive gear.

[0062] Preferably, grating sensors are symmetrically distributed and fixedly installed on the inner surface of the regular polygonal frame. Rotating screws for rotation are installed on both sides of the linear guide rail via support plates. Threaded sleeves are threaded onto the outer surface of the rotating screws. The outer surface of the threaded sleeves is fixedly connected to the outer surface of the regular polygonal frame. Linkage housings are installed on the outer surfaces of the two rotating screws. A bevel gear set is installed on the outer surface of the linkage housings. A reduction motor is fixedly connected to the upper surface of the base plate of the detection station. The outer surface of the output shaft of the reduction motor is fixedly connected to the inner surface of the driving bevel gear of the bevel gear set.

[0063] Preferably, the visual inspection system further includes a detection method, comprising the following steps:

[0064] Step 1: The hooks on the clamping turntable hook onto the wheel hub. One of the drive motors starts and drives the bidirectional adjusting screw to rotate through the worm gear, worm wheel, and worm sleeve transmission. This allows the wheel hub to be inspected to be conveyed to the inspection station via the drying conveyor line. The two L-shaped clamping blocks on the bidirectional adjusting screw move synchronously towards each other in the guide groove, firmly supporting the wheel hub from inside the center hole. This worm gear transmission mechanism has a self-locking function to ensure that the clamping force will not loosen after power failure, making it safe and reliable. When inspecting wheel hub defects, the other drive motor starts and drives the rotating shaft and the entire clamping turntable to rotate through the meshing of the drive gear and driven gear, adjusting the first surface of the wheel hub to be inspected to the initial position facing the sensor.

[0065] Step 2: The six-axis robot in the intelligent collaborative motion control module adaptively plans the optimal scanning path and posture based on the 3D model of the wheel hub. The six-axis robot drives the industrial area array camera and high-speed linear laser 3D scanner rigidly integrated on its end flange to move to the first scanning point. At this time, the reduction motor synchronously drives two rotating screws through the bevel gear set, pushing the threaded sleeve and driving the entire dynamic structured lighting unit to move along the linear guide to the preset optimal working distance. The grating sensor detects the position in real time to realize closed-loop control. The linkage motor starts and transmits the motion to all linkage gears through the transmission system composed of drive gears, synchronous gears, linkage shafts and universal joints. This drives all rack bars to slide synchronously in the groove plate, thereby adjusting the arc-shaped LED light panel to the preset radial position to form a specific lighting mode.

[0066] Step 3: After the six-axis robot reaches the predetermined scanning pose, it sends a hardware trigger signal to strictly and synchronously control the dynamic structured lighting unit to switch to the next lighting mode and trigger the industrial area array camera to take a high-definition 2D image. Then, it records the contour point cloud data obtained by the high-speed line laser 3D scanner at that position.

[0067] Step 4: Use the pre-calibrated rigid body transformation matrix Transform the 3D point cloud coordinate system into a 2D camera coordinate system to achieve... Spatial alignment is performed, and combined with hardware-triggered timestamps, spatiotemporal registration is completed. The normal vector of each pixel on the wheel hub surface is calculated using the registered 3D point cloud.

[0068] Step 5: Based on the Phong lighting model, and combining the position of the curved LED light panel and the camera position, predict the specular highlight intensity at each point in the image. The pixel value compensation algorithm is executed, which compares images under different lighting modes. If the predicted value of a certain pixel in the highlight mode is lower than the predicted value in the highlight mode, the pixel value is lowered. Exceeding the threshold Then, the pixel value with the richest diffuse reflection component is selected from other non-highlight mode images and replaced, finally generating a high-quality fused image without reflective distortion;

[0069] Step 6: On the fused image, the traditional algorithms of edge detection, texture analysis and color difference calculation in CIELAB color space are run in parallel to quickly delineate suspected defect areas and filter out false defects based on geometric features such as area and aspect ratio. Then, each ROI is fed into a deep learning convolutional neural network. This network adopts a multi-task learning framework and outputs the probability of defect type and severity level score at the same time, and calculates the comprehensive confidence score to achieve accurate classification and rating.

[0070] Step 7: The system integrates all defect information, determines the quality level of the wheel hub, and generates a visual report with a defect map on the human-machine interface. For cases where the system misjudges or misses detection, the operator can manually annotate them through the interface. These annotation data are stored in the self-learning database and used to periodically incrementally train the hybrid intelligent detection model, enabling the system to have the ability to continuously optimize.

[0071] The beneficial effects of this invention are as follows:

[0072] 1. By setting up a multimodal data acquisition and fusion system, the reflective distortion of the wheel hub surface is effectively eliminated, improving the accuracy of defect detection. During the adjustment process, the system combines a high-resolution 2D vision unit, a high-precision 3D contour scanning unit, and a dynamic structured lighting unit to simultaneously acquire multi-dimensional information of the wheel hub surface. It also uses the Phong lighting model and pixel value compensation algorithm to perform lighting compensation and reflection elimination on 2D images under different lighting modes, generating standard images without reflective distortion. This provides clean and reliable input data for subsequent defect identification, significantly reducing the false positive and false negative rates.

[0073] 2. By setting up a two-level defect identification architecture and a closed-loop learning mechanism, efficient and accurate defect classification and rating are achieved, and continuous optimization capabilities are available. During the adjustment process, a first-level screening module and a second-level fine inspection module are used to balance detection efficiency and accuracy. Then, the human-computer interaction and closed-loop learning module allow manual annotation of detection results, and the hybrid intelligent detection model is incrementally trained through a self-learning database, enabling the system to continuously adapt to new defect patterns and improve long-term performance.

[0074] 3. By setting up a transmission and positioning module and an adjustable dynamic lighting system, the system ensures stable positioning of the wheel hub and rapid switching between various lighting conditions, enhancing the system's stability and adaptability. During adjustment, the transmission and positioning module uses a worm gear transmission mechanism to achieve wheel hub clamping and rotation, with a self-locking function to ensure that the clamping force does not loosen after power failure, ensuring safety and reliability. The dynamic structured lighting unit uses linear guide rails, a regular polygonal frame, and synchronously driven LED light panels to achieve rapid switching of lighting modes and precise adjustment of the illumination angle, optimizing the lighting conditions for defect detection and improving the system's adaptability to different wheel hub models. Attached Figure Description

[0075] Figure 1 This is a schematic diagram of a visual inspection system for the paint quality of wheel hubs proposed in this invention;

[0076] Figure 2 This is a detailed flowchart of the multimodal data fusion of a visual inspection system for wheel hub paint quality proposed in this invention;

[0077] Figure 3 This is a diagram of a two-level defect detection architecture for a visual inspection system for wheel hub paint quality proposed in this invention.

[0078] Figure 4 This is a block diagram of the closed-loop learning mechanism of a visual inspection system for wheel hub paint quality proposed in this invention.

[0079] Figure 5 This is a three-dimensional view of the inspection station structure of a visual inspection system for wheel hub paint quality proposed in this invention.

[0080] Figure 6 This is a three-dimensional view of the regular polygonal frame structure of a visual inspection system for wheel hub paint quality proposed in this invention;

[0081] Figure 7 This is a three-dimensional view of the hook structure of a visual inspection system for wheel hub paint quality proposed in this invention;

[0082] Figure 8 This is a three-dimensional view of the clamping housing structure of a visual inspection system for wheel hub paint quality proposed in this invention;

[0083] Figure 9 This is a three-dimensional view of the worm gear sleeve structure of a visual inspection system for wheel hub paint quality proposed in this invention;

[0084] Figure 10 This is a three-dimensional view of the LED light panel structure of a visual inspection system for wheel hub paint quality proposed in this invention.

[0085] Figure 11This is a three-dimensional view of the grating sensor structure of a visual inspection system for wheel hub paint quality proposed in this invention.

[0086] Figure 12 This is a three-dimensional view of a six-axis robot structure for a visual inspection system for wheel hub paint quality proposed in this invention.

[0087] In the diagram: 1. Drying conveyor line; 2. Inspection station; 3. Clamping housing; 4. Clamping turntable; 5. Hook; 6. Guide groove; 7. Movable cavity; 8. Bidirectional adjusting screw; 9. Clamping block; 10. Drive cavity; 11. Drive motor; 12. Rotating shaft; 13. Worm sleeve; 14. Adjusting worm wheel; 15. Driving gear; 16. Driven gear; 17. Adjusting worm; 18. Worm wheel rod; 19. Linear guide rail; 20. Support slider; 21. Positive 21. Polygonal frame; 22. Parallel rod; 23. Groove plate; 24. Linkage shaft; 25. Rack and pinion; 26. LED light panel; 27. Linkage gear; 28. Linkage motor; 29. ​​Drive gear; 30. Synchronization gear; 31. Grating sensor; 32. Rotating screw; 33. Threaded sleeve; 34. Linkage housing; 35. Bevel gear set; 36. Gear motor; 37. Industrial area scan camera; 38. Laser 3D scanner; 39. Six-axis robot. Detailed Implementation

[0088] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0089] Reference Figures 1-12 A visual inspection system for the quality of wheel rim paint, the visual inspection system comprising:

[0090] S1, the conveying and positioning module, is used to stably convey the wheel hub to be tested to the testing station 2, and to clamp and rotate it.

[0091] S2, Multimodal Data Acquisition Module, used to synchronously acquire multi-dimensional information of the wheel hub surface. This module includes:

[0092] A high-resolution 2D vision unit is used to acquire color image information of the wheel hub surface.

[0093] A high-precision 3D contour scanning unit is used to acquire three-dimensional point cloud data of the wheel hub surface.

[0094] The dynamic structured lighting unit contains multiple independently controllable light sources, which can switch between multiple different lighting modes according to control commands for sequential shooting.

[0095] S3, Intelligent Cooperative Motion Control Module, is used to drive the multimodal data acquisition module to move along a predetermined path and adjust its observation attitude.

[0096] S4, Data Processing and Defect Identification Module, which is electrically connected to the multimodal data acquisition module, includes:

[0097] The data preprocessing and registration unit is used to accurately align 2D images, 3D point clouds, and corresponding lighting pattern information acquired at the same location in space and time.

[0098] The multimodal data fusion unit uses 3D point cloud data to calculate the surface normal and curvature geometry information corresponding to each pixel, and based on the geometry information, performs illumination compensation and reflection elimination on 2D images under different lighting modes to generate a standard image without reflection distortion.

[0099] The hybrid intelligent detection model unit is used to detect defects in standard images, including: quickly locating suspected defect areas through a primary algorithm, and then performing fine classification and rating of suspected defect areas through a deep learning convolutional neural network.

[0100] S5, the human-computer interaction and closed-loop learning module, is used to visualize the detection results and allow manual annotation of the detection results. The annotated data is then sent to the self-learning database for incremental training and optimization of the hybrid intelligent detection model.

[0101] The 2D vision unit includes an industrial area scan camera 37, the 3D contour scanning unit includes a high-speed linear laser 3D scanner 38, and the intelligent collaborative motion control module includes a six-axis robot 39. The industrial area scan camera 37 and the laser 3D scanner 38 are integrated on its end flange. The six-axis robot 39 adaptively plans the optimal scanning path and posture based on the wheel hub model. Each time the six-axis robot 39 moves to a new pose, the lighting system switches to a different mode, and the sensors simultaneously acquire data. This collaborative working mechanism is a prerequisite for generating 2D image sequences and corresponding 3D point clouds under multiple lighting modes.

[0102] In step S4, the spatial alignment of data preprocessing and registration units is achieved in the following way:

[0103] The rigid transformation matrix between the 2D vision unit and the 3D contour scanning unit is obtained through joint calibration. This makes it possible for points in the 3D contour scanning unit coordinate system Its corresponding point in the 2D visual unit coordinate system Calculated using the following formula:

[0104] Then, the transformed 3D point cloud is projected onto the 2D image plane using a camera model to establish pixel-level correspondences. and These are the homogeneous coordinates of the point in the 3D contour scanning unit coordinate system and the 2D vision unit coordinate system, respectively. , , These are the three-dimensional spatial coordinates of a point in the 3D contour scanning unit coordinate system. , , It refers to the three-dimensional spatial coordinates of a point in the 2D visual unit coordinate system, while the transformation matrix... It is a 4x4 rigid body transformation matrix that includes rotation and translation transformations.

[0105] Time alignment is achieved in the following ways:

[0106] As the six-axis robot moves to each scanning position, the control module sends a hardware trigger signal to synchronously control the switching mode of the dynamic structured lighting unit, the image acquisition by the 2D vision unit, and the recording of contour data acquired by the 3D contour scanning unit. All data acquired at the same position are assigned the same position identifier, and the rigid body transformation matrix is ​​obtained through joint calibration. and using the formula By establishing a pixel-level correspondence between the 3D point cloud coordinate system and the 2D camera coordinate system, and using hardware trigger signals, the illumination switching, 2D shooting, and 3D scanning are synchronously controlled at each scanning pose of the robot. A unified position identifier is assigned to the data at the same moment. Precise spatiotemporal alignment is the prerequisite and foundation for the success of subsequent multimodal data fusion algorithms. This scheme ensures that 2D texture information and 3D geometric information can be accurately matched.

[0107] The multimodal data fusion unit in step S4 includes the following steps:

[0108] Calculate the normal vector of each point on the surface using the registered 3D point cloud data. .

[0109] Unit normal vector based on surface points The unit direction vector from the surface point to the light source. And the viewpoint direction vector from the surface point to the optical center of the 2D visual unit. The light intensity of the surface is predicted based on the reflection model.

[0110] The reflection model is a Phong lighting model that combines diffuse reflection and specular reflection components, and its expression is as follows:

[0111] ,in:

[0112] This represents the total light intensity predicted by the model.

[0113] Indicates ambient light intensity. This represents the summation of contributions from multiple light sources.

[0114] Indicates the intensity of a single light source. denoted as the specular reflection coefficient of the surface material.

[0115] To be based on the direction of incident light and normal vector The calculated direction vector of the reflected light.

[0116] The gloss index is used to characterize surface smoothness. denoted as the diffuse reflectance coefficient of the surface material.

[0117] By analyzing multiple 2D images acquired under different lighting modes, and using a reflection model to identify and separate specular highlights and diffuse reflection components, a pixel value compensation algorithm is then used to eliminate highlight interference, ultimately generating a fused image. This model decomposes total illumination into ambient light, diffuse reflection, and specular reflection components. Through this model, the specular highlight components in the image can be quantitatively analyzed and separated. Compared with simple image processing solutions, this solution is based on physical principles and can more fundamentally and accurately predict and identify highlight areas, providing a theoretical basis and implementation path for completely eliminating highlight interference and significantly improving the accuracy of subsequent defect identification.

[0118] Highlight interference is eliminated through a pixel value compensation algorithm, and a fused image is generated, including the results for each pixel location. Perform the following steps:

[0119] For each lighting mode Calculate the specular highlight intensity based on the reflection model. .

[0120] Will With preset threshold In comparison, when > Then mark the image medium pixel These are highlight pixels.

[0121] Define a set That is, all pixels in the unexposed image A set of values.

[0122] when If not empty, then ,Right now From Select a value, including the selection The value with the highest brightness in the middle; another strategy is to select a set. The pixel value of the image with the largest diffuse reflectance component, the estimated diffuse reflectance component value. ,in:

[0123] , It is a function that finds the set of functions.

[0124] when If empty, then If a pixel is a highlight in all images, then we can estimate its value using spatial interpolation. It is a pixel-based Surrounding Neighborhood The interpolation function for non-highlight pixel values ​​can actively and adaptively generate high-quality distortion-free fused images by transforming physical model predictions into specific image inpainting algorithms. Its beneficial effect is that it provides clean input data for AI defect detection models, greatly reducing false positives and false negatives.

[0125] By setting up a multimodal data acquisition and fusion system, reflective distortion on the wheel hub surface is effectively eliminated, improving the accuracy of defect detection. During the adjustment process, the system combines a high-resolution 2D vision unit, a high-precision 3D contour scanning unit, and a dynamic structured lighting unit to simultaneously acquire multi-dimensional information of the wheel hub surface. It also uses the Phong lighting model and pixel value compensation algorithm to perform illumination compensation and reflection elimination on 2D images under different lighting modes, generating standard images without reflective distortion. This provides clean and reliable input data for subsequent defect identification, significantly reducing the false positive and false negative rates.

[0126] In step S4, the hybrid intelligent detection model unit employs a two-level progressive detection architecture to identify defects in the fused, non-reflective distortion image, including:

[0127] The primary filtering module is configured to extract edge features, texture features, and color consistency features from the fused image in parallel, and quickly locate all suspected defective regions of interest (ROIs) based on adaptive threshold segmentation and geometric feature filtering. Color consistency detection is achieved in the CIELAB color space by calculating local color differences.

[0128] And the final ROI set is composed of Confirmed, among which:

[0129] :Location The color difference value at that location.

[0130] , , :Location The brightness, red-green hue, and yellow-blue hue components in the CIELAB color space.

[0131] , , : The corresponding component of the reference standard area in the CIELAB color space.

[0132] The final set of suspected defective areas.

[0133] Candidate region :area The area.

[0134] Minimum area threshold :area The aspect ratio.

[0135] , The minimum and maximum thresholds for aspect ratio. : Logical AND operator.

[0136] The secondary fine inspection module is configured to use a deep convolutional neural network to refine the processing of each ROI. This network employs a multi-task learning framework and simultaneously outputs the defect type classification probability. and severity rating The final detection result is generated by combining the type confidence score and the severity score.

[0137] ,in:

[0138] : Predicted probability distribution of defect types : Normalized exponential function, used to convert network output into a probability distribution.

[0139] : Weight matrix of the defect type classification layer The feature vector output by the last layer of a deep convolutional neural network. : Bias vector of the defect type classification layer.

[0140] : Predicted defect severity rating Weight matrix of the severity level regression layer Bias vector of the severity level regression layer.

[0141] : Comprehensive confidence score for defect identification Defect types predicted by the network The corresponding probability value, The system has a preset upper limit for severity rating. This architecture achieves a perfect balance between efficiency and accuracy. The first-level screening quickly locates the problem area, reducing the burden on the deep learning model. The second-level fine inspection ensures the accuracy of classification and rating. Multi-task learning and comprehensive confidence calculation make the detection results more reliable and more in line with actual quality inspection needs.

[0142] By setting up a two-level defect identification architecture and a closed-loop learning mechanism, efficient and accurate defect classification and rating are achieved, and continuous optimization capabilities are available. During the adjustment process, a first-level screening module and a second-level fine inspection module are used to balance detection efficiency and accuracy. Then, the human-computer interaction and closed-loop learning module allow manual annotation of detection results, and the hybrid intelligent detection model is incrementally trained through a self-learning database, enabling the system to continuously adapt to new defect patterns and improve long-term performance.

[0143] The conveying and positioning module in step S1 also includes a drying conveyor line 1. A detection station 2 is set on one side of the drying conveyor line 1. A clamping housing 3 is fixedly connected to the lower surface of the hanging rail of the drying conveyor line 1. A clamping turntable 4 is rotatably connected to one side surface of the clamping housing 3. A hook 5 for hooking the hub is fixedly connected to the inner surface of the clamping turntable 4. Guide grooves 6 and movable cavities 7 are respectively opened inside the clamping turntable 4. One end of each guide groove 6 is fixedly connected to the interior of the movable cavity 7. A bidirectional adjusting screw 8 is rotatably connected inside the movable cavity 7. Both ends of the rod 8 are threaded with L-shaped clamping blocks 9 with flexible rubber outer surfaces. The outer surfaces of the clamping blocks 9 slide and engage with the inner wall of the guide groove 6. The drying conveyor line 1 transports the hub. The infrared sensor detects whether the hub has reached the detection station 2. At the same time, when the hub is transported on the drying conveyor line 1, the hub is placed on the hook 5. At this time, the bidirectional adjusting screw 8 rotates, so that the two L-shaped clamping blocks 9 move synchronously in opposite directions or back to back in the guide groove 6, and are tightened from the center hole of the hub. The flexible rubber outer surface prevents scratches on the paint surface.

[0144] The clamping housing 3 has a drive cavity 10 inside, which houses two drive motors 11. A rotating shaft 12 is rotatably connected to the drive cavity 10 via bearings. The free end of the rotating shaft 12 is fixedly connected to the inside of the clamping turntable 4. A worm sleeve 13 is rotatably fitted onto the outer surface of the rotating shaft 12, with one end of the worm sleeve 13 extending into the interior of the movable cavity 7. An adjusting worm wheel 14 is fixedly fitted onto the outer surface of the middle part of the bidirectional adjusting screw 8, meshing with the worm sleeve 13. A drive gear 15 is fixedly fitted onto the outer surface of the output shaft of one of the drive motors 11, and a driven gear 16 is fixedly fitted onto the outer surface of the rotating shaft 12, meshing with the drive gear 15. An adjusting worm 17 is fixedly connected to the outer surface of the output shaft of the other drive motor 11 via a coupling. A worm wheel rod 18 with two worm wheels is rotatably connected inside the drive cavity 10, with the two worm wheels on the worm wheel rod 18 respectively meshing with… The worm sleeve 13 and the adjusting worm 17 mesh to achieve the clamping and rotation of the hub. One of the drive motors 11 meshes with the drive gear 15 and the driven gear 16 to rotate the shaft 12. The shaft 12 then drives the clamping turntable 4 to rotate. This chain drives the clamped hub to rotate so that the sensors carried by the robot can scan its entire circumference. Then, the other drive motor 11 controls the adjustment worm 17 to rotate, so that the adjustment worm 17 drives the worm wheel rod 18 to rotate through the meshing of one of the worm wheels. In turn, the other worm wheel on the worm wheel rod 18 drives the worm sleeve 13 to rotate, which in turn drives the adjusting worm wheel 14 to rotate, thereby realizing the rotation of the adjusting screw. This chain drives the bidirectional adjusting screw 8 to make the two L-shaped clamping blocks 9 move synchronously towards or away from each other, thereby tightening or loosening the hub from the center hole. The two drive motors 11 independently control the rotation and clamping actions without interfering with each other.

[0145] The dynamic structured lighting unit also includes a linear guide rail 19 fixedly connected to the upper surface of the base plate of the inspection station 2. A support slider 20 is slidably engaged on the outer surface of the linear guide rail 19. A regular polygonal frame 21 is fixedly connected to the upper surface of the support slider 20. Parallel rods 22 are fixedly connected in a circular array on the inner surface of the regular polygonal frame 21. A grooved plate 23 is fixedly connected to the inner surface of multiple parallel rods 22, and a linkage shaft 24 is rotatably connected to it. Adjacent surfaces of adjacent linkage shafts 24 are connected by universal joints. A rack rod 25 is slidably engaged on the inner surface of the grooved plate 23. An LED light panel 26 is fixedly connected to the side surface of the rack rod 25 near the center point of the regular polygonal frame 21. The LED light panel 26 has an arc-shaped design. The entire lighting unit is mounted on the linear guide rail 19 via the support slider 20 and can move as a whole. Each LED light panel 26 is independently mounted via the rack rod 25. The grooved plate 23 can move radially. By moving the rack 25 radially, the distance and illumination angle of each LED light panel 26 to the hub surface can be changed. For example, shrinking the LED light panel 26 inward makes it easier to achieve low-angle grazing light, which is used to highlight surface micro-geometric defects such as scratches and dents. Expanding the LED light panel 26 outward makes it easier to form uniform diffuse light, which is used to detect color and texture. At the same physical position, multiple images can be acquired by switching different lighting modes. For example, when the six-axis robot 39 pauses after scanning a pose, the lighting system can quickly switch between different preset forms such as the overall position A + the radial position X of the light panel and the overall position A + the radial position Y of the light panel. This provides the 2D camera with multiple lighting conditions such as bright field, dark field, and low-angle light in a very short time. This is a prerequisite for subsequent multimodal data fusion to eliminate reflections.

[0146] A linkage gear 27 is fixedly sleeved on the outer surface of the linkage shaft 24. The linkage gear 27 meshes with the corresponding rack 25. A linkage motor 28 is fixedly mounted on the surface of one of the parallel rods 22. A drive gear 29 is fixedly connected to the outer surface of the output shaft of the linkage motor 28. A synchronous gear 30 is fixedly sleeved on the outer surface of one of the linkage shafts 24. The synchronous gear 30 meshes with the drive gear 29. The drive gear 29 is controlled by a linkage motor 28, which drives the synchronous gear 30 and a linkage shaft 24 to rotate. The motion is then transmitted to all the linkage shafts 24 arranged in a ring through a universal joint. The linkage gear 27 on each shaft rotates synchronously, thereby meshing and driving all the racks 25 and LED light panels 26 to move radially synchronously. This ensures that the entire ring lighting field always maintains perfect symmetry, and the light is evenly projected onto the hub from all directions.

[0147] A grating sensor 31 is symmetrically mounted on the inner surface of the regular polygonal frame 21. Rotating screws 32 are mounted on both sides of the linear guide rail 19 via support plates. Threaded sleeves 33 are threaded onto the outer surface of the rotating screws 32. The outer surface of the threaded sleeves 33 is fixedly connected to the outer surface of the regular polygonal frame 21. Linkage housings 34 are mounted on the outer surfaces of the two rotating screws 32. A bevel gear set 35 is mounted on the outer surface of the linkage housings 34. A geared motor 36 is fixedly connected to the upper surface of the base plate of the detection station 2. The outer surface of the output shaft of the geared motor 36 is fixedly connected to the inner surface of the active bevel gear of the bevel gear set 35. The system integrates grating sensors 31 for real-time detection of the absolute position of the regular polygonal frame 21. At the same time, the geared motor 36 drives the rotating screws 32 on both sides through the bevel gear set 35, which in turn pushes the entire frame to move through the threaded sleeves 33. By adjusting the distance, the size and intensity of the light spot irradiated on the hub surface can be optimized.

[0148] By incorporating a transmission and positioning module and an adjustable dynamic lighting system, the system ensures stable wheel hub positioning and rapid switching between various lighting conditions, enhancing its stability and adaptability. During adjustment, the transmission and positioning module utilizes a worm gear transmission mechanism to clamp and rotate the wheel hub, featuring a self-locking function to ensure that the clamping force remains stable even after power failure, ensuring safety and reliability. The dynamic structured lighting unit, through linear guide rails 19, a regular polygonal frame 21, and synchronously driven LED light panels 26, enables rapid switching of lighting modes and precise adjustment of the illumination angle, optimizing the lighting conditions for defect detection and improving the system's adaptability to different wheel hub models.

[0149] Reference Figures 1-12 A visual inspection method for the quality of wheel rim paint, the inspection method steps include:

[0150] Step 1: The hooks 5 on the clamping turntable 4 hook onto the wheel hub. One of the drive motors 11 starts and drives the bidirectional adjusting screw 8 to rotate through the adjusting worm 17, worm wheel 18, and worm sleeve 13. This allows the wheel hub to be inspected to be conveyed to the inspection station 2 via the drying conveyor line 1. The two L-shaped clamping blocks 9 on the bidirectional adjusting screw 8 move synchronously towards each other in the guide groove 6, firmly supporting the wheel hub from the center hole. This worm gear transmission mechanism has a self-locking function to ensure that the clamping force will not loosen after power failure, making it safe and reliable. When inspecting wheel hub defects, the other drive motor 11 starts and drives the rotating shaft 12 and the entire clamping turntable 4 to rotate through the meshing of the drive gear 15 and the driven gear 16, adjusting the first surface of the wheel hub to be inspected to the initial position facing the sensor.

[0151] Step 2: The six-axis robot 39 in the intelligent collaborative motion control module adaptively plans the optimal scanning path and posture based on the three-dimensional model of the wheel hub. The six-axis robot 39 drives the industrial area array camera 37 and the high-speed linear laser 3D scanner 38 rigidly integrated on its end flange to move to the first scanning point. At this time, the reduction motor 36 drives the two rotating screws 32 synchronously through the bevel gear set 35, pushes the threaded sleeve 33, and drives the entire dynamic structured lighting unit to move along the linear guide rail 19 to the preset optimal working distance. The grating sensor 31 detects the position in real time to realize closed-loop control. The linkage motor 28 starts and transmits the motion to all linkage gears 27 through the transmission system composed of drive gear 29, synchronous gear 30, linkage shaft 24 and universal joint. This drives all rack rods 25 to slide synchronously in the groove plate 23, thereby adjusting the arc-shaped LED light panel 26 to the preset radial position to form a specific lighting mode.

[0152] Step 3: After the six-axis robot 39 reaches the predetermined scanning pose, it sends a hardware trigger signal to strictly and synchronously control the dynamic structured lighting unit to switch to the next lighting mode and trigger the industrial area array camera 37 to capture a high-definition 2D image. Then, it records the contour point cloud data acquired by the high-speed line laser 3D scanner 38 at that position.

[0153] Step 4: Use the pre-calibrated rigid body transformation matrix Transform the 3D point cloud coordinate system into a 2D camera coordinate system to achieve... Spatial alignment is performed, and combined with hardware-triggered timestamps, spatiotemporal registration is completed. The normal vector of each pixel on the wheel hub surface is calculated using the registered 3D point cloud.

[0154] Step 5: Based on the Phong illumination model, combined with the position of the curved LED light panel 26 and the camera position, predict the specular highlight intensity of each point in the image, and execute the pixel value compensation algorithm. That is, compare the images under different lighting modes. If the predicted value of a certain pixel in the specular mode exceeds the threshold... Then, the pixel value with the richest diffuse reflection component is selected from other non-highlight mode images and replaced, finally generating a high-quality fused image without reflective distortion;

[0155] Step 6: On the fused image, the traditional algorithms of edge detection, texture analysis and color difference calculation in CIELAB color space are run in parallel to quickly delineate suspected defect areas and filter out false defects based on geometric features such as area and aspect ratio. Then, each ROI is fed into a deep learning convolutional neural network. This network adopts a multi-task learning framework and outputs the probability of defect type and severity level score at the same time, and calculates the comprehensive confidence score to achieve accurate classification and rating.

[0156] Step 7: The system integrates all defect information, determines the quality level of the wheel hub, and generates a visual report with a defect map on the human-machine interface. For cases where the system misjudges or misses detection, the operator can manually annotate them through the interface. These annotation data are stored in the self-learning database and used to periodically incrementally train the hybrid intelligent detection model, enabling the system to have the ability to continuously optimize.

[0157] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A visual inspection system for wheel hub paint bake quality, characterized by: The visual detection system comprises: S1, a conveying and positioning module, for stably conveying a hub to be detected to a detection station (2) and clamping and rotating the hub; S2, a multi-modal data acquisition module, for synchronously acquiring multi-dimensional information of a surface of the hub, the module comprising: a high-resolution 2D vision unit, for acquiring color image information of the surface of the hub; a high-precision 3D profile scanning unit, for acquiring three-dimensional point cloud data of the surface of the hub; a dynamic structured illumination unit, comprising a plurality of independently controllable light sources, so that the unit switches a plurality of different illumination modes according to a control instruction to perform sequential shooting; S3, an intelligent cooperative motion control module, for driving the multi-modal data acquisition module to move along a predetermined path and adjusting an observation posture of the module; S4, a data processing and defect identification module, electrically connected with the multi-modal data acquisition module, the module comprising: a data preprocessing and registration unit, for accurately aligning a 2D image, 3D point cloud and corresponding illumination mode information acquired at the same position in space and time; a multi-modal data fusion unit, for calculating a surface normal and curvature geometric information corresponding to each pixel point by using the 3D point cloud data, and performing illumination compensation and reflection elimination on the 2D image under different illumination modes based on the geometric information, to generate a standard image without reflection distortion; a hybrid intelligent detection model unit, for performing defect detection on the standard image, comprising: quickly locating a suspected defect area by a primary algorithm, and then performing fine classification and rating on the suspected defect area by a deep learning convolutional neural network; S5, a man-machine interaction and closed-loop learning module, for visually displaying detection results and allowing manual labeling of the detection results, and sending the labeled data to a self-learning database to perform incremental training and optimization on the hybrid intelligent detection model.

2. A visual inspection system for wheel hub paint quality according to claim 1, characterized in that: The 2D vision unit comprises an industrial area array camera (37), the 3D profile scanning unit comprises a high-speed line laser 3D scanner (38), the intelligent cooperative motion control module comprises a six-axis robot (39), an end flange plate of the six-axis robot (39) is integrated with the industrial area array camera (37) and the laser 3D scanner (38), and the six-axis robot (39) adaptively plans an optimal scanning path and posture according to a hub model.

3. A visual inspection system for wheel hub paint quality according to claim 2, characterized in that: The spatial alignment of the data preprocessing and registration unit in step S4 is realized by the following manner: The rigid body transformation matrix between the 2D vision unit and the 3D profile scanning unit is obtained by joint calibration such that for a point in the 3D profile scanning unit coordinate system its corresponding point in the 2D vision unit coordinate system is calculated by the following equation: and then project the transformed 3D point cloud to the 2D image plane by camera model to establish pixel-level correspondence, where and are the homogeneous coordinates of the point in the 3D profile scanning unit coordinate system and the 2D vision unit coordinate system, respectively, , , is the three-dimensional space coordinate of the point in the 3D profile scanning unit coordinate system, , , is the three-dimensional space coordinate of the point in the 2D vision unit coordinate system, and the transformation matrix is a 4x4 rigid transformation matrix containing rotation and translation transformation; The time alignment is realized by the following manner: When the six-axis robot (39) moves to each scanning position, a hardware trigger signal is sent by a control module, the dynamic structured illumination unit is controlled to switch modes, the 2D vision unit is controlled to acquire images, and profile data acquired by the 3D profile scanning unit is recorded, and all data acquired at the same position are given the same position identifier.

4. A visual inspection system for wheel hub paint quality according to claim 3, characterized in that: The multi-modal data fusion unit in step S4 comprises the following steps: calculating normal vectors for points of the surface using the registered 3D point cloud data ; a unit normal vector based on the surface point a unit direction vector from the surface point to the light source and a view direction vector from the surface point to the 2D visual cell's optical center predicting the surface's illumination intensity according to a reflection model The reflection model is a Phong illumination model combining a diffuse reflection component and a specular reflection component, and is expressed as follows: wherein: total light intensity as predicted by the model; represents an ambient light intensity, represents a sum over contributions of a plurality of light sources; representing the intensity of the individual light sources, is the specular reflection coefficient of the surface material; to calculate the reflected light direction vector and normal vector to calculate the reflected light direction vector high light index for characterizing the surface smoothness, diffuse reflection coefficient of the surface material; By analyzing a plurality of 2D images collected under different illumination modes, and using the reflection model to identify and separate the specular highlight and diffuse reflection components, and then eliminating the highlight interference through a pixel value compensation algorithm, the fusion image is finally generated.

5. A visual inspection system for wheel hub paint quality according to claim 4, characterized in that: The high light interference is eliminated by the pixel value compensation algorithm, and a fusion image is generated, including the following steps for each pixel position : For each lighting mode , calculate the specular highlight intensity according to the reflection model ; Will With preset threshold In comparison, when > Then mark the image medium pixel Highlight pixels; Definition Set , the set of values of all pixels in the image that are not overexposed ; When is not empty, then i.e. one value can be chosen from including choosing the value with the highest luminance in another strategy is to choose the pixel value in the image with the largest diffuse reflectance component in the set where: , is a function that aggregates functions; When is empty, then i.e. all images have the pixel as highlight, we can estimate the value of the pixel by spatial interpolation, where is an interpolation function based on the non-highlight pixel values in the neighborhood around the pixel. ​ 6. A visual inspection system for wheel hub paint quality according to claim 5, characterized in that: In step S4, the mixed intelligent detection model unit adopts a two-stage progressive detection architecture to identify defects in the fused non-reflection distortion image, including: The first screening module is configured to quickly locate all suspected defect regions ROI by extracting edge features, texture features and color consistency features of the fusion image in parallel, and based on adaptive threshold segmentation and geometric feature filtering, wherein the color consistency detection is realized by calculating the local color difference in the CIELAB color space: and the final ROI set is determined by wherein: : position color difference value at the position , , : position L*, a*, b* components of lightness, red-green color, yellow-blue color in CIELAB color space; , , : corresponding components of the reference standard area in the CIELAB color space; : final set of suspected defect areas; : candidate region, : region area; : minimum area threshold, : area : aspect ratio; , : minimum and maximum threshold of aspect ratio, : logical AND operator; The secondary fine detection module is configured to use a deep convolutional neural network to perform fine processing on each ROI, and the network uses a multi-task learning framework to simultaneously output defect type classification probability: and severity level score and generate a final detection result by combining the type confidence and severity score: wherein: : predictive probability distribution of defect type, : normalizing exponential function for converting network output to probability distribution; : weight matrix of the defect type classification layer, : feature vector outputted by the last layer of the deep convolutional neural network, : bias vector of the defect type classification layer; : predicted defect severity score, : weight matrix of the severity regression layer, bias vector of the severity regression layer; : a comprehensive confidence score of defect identification, : a defect type predicted by the network a corresponding probability value, : a system preset upper limit value of severity level score.

7. A visual inspection system for wheel hub paint quality according to claim 6, characterized in that: The conveying and positioning module in step S1 further includes a drying conveying line (1), one side of the drying conveying line (1) is provided with a detection station (2), the lower surface of the hanging rail of the drying conveying line (1) is fixedly connected with a clamping shell (3), one side surface of the clamping shell (3) is rotatably connected with a clamping turntable (4), the inner surface of the clamping turntable (4) is fixedly connected with a bent hook (5) for hooking the hub, the inner part of the clamping turntable (4) is respectively provided with a guide groove (6) and a movable cavity (7), one end of the two guide grooves (6) is fixedly communicated with the inside of the movable cavity (7), the inside of the movable cavity (7) is rotatably connected with a bidirectional adjusting screw (8), the outer surfaces of the two ends of the bidirectional adjusting screw (8) are respectively threadedly sleeved with L-shaped clamping blocks (9) with flexible rubber outer surfaces, and the outer surfaces of the clamping blocks (9) are slidably connected with the inner walls of the guide grooves (6); The inside of the clamping shell (3) is provided with a driving cavity (10), two driving motors (11) are installed in the inside of the driving cavity (10), a rotating shaft (12) is rotatably connected in the inside of the driving cavity (10), the free end surface of the rotating shaft (12) is fixedly connected with the inside of the clamping turntable (4), a worm sleeve (13) is rotatably sleeved with the outer surface of the rotating shaft (12), one end of the worm sleeve (13) extends into the inside of the movable cavity (7), an adjusting worm wheel (14) is fixedly sleeved with the outer surface of the middle part of the bidirectional adjusting screw (8), the adjusting worm wheel (14) is engaged with the worm sleeve (13), the outer surface of the output shaft of one of the driving motors (11) is fixedly sleeved with a driving gear (15), the outer surface of the rotating shaft (12) is fixedly sleeved with a driven gear (16), the driving gear (15) is engaged with the driven gear (16), and the output shaft of the other driving motor (11) is fixedly connected with an adjusting worm (17) through a shaft coupling.

8. A visual inspection system for wheel hub paint quality according to claim 7, characterized in that: The dynamic structured lighting unit further comprises a linear guide rail (19) fixedly connected to the upper surface of the bottom plate of the detection station (2), the outer surface of the linear guide rail (19) is slidably connected with a supporting sliding block (20), the upper surface of the supporting sliding block (20) is fixedly connected with a regular polygon frame (21), the inner surface of the regular polygon frame (21) is fixedly connected with a plurality of parallel rods (22) arranged in an annular array, the inner surfaces of the plurality of parallel rods (22) are respectively fixedly connected with a groove plate (23) and rotatably connected with a linkage shaft (24), adjacent surfaces of adjacent linkage shafts (24) are connected through a universal joint, the inner surface of the groove plate (23) is slidably connected with a rack rod (25), the side surface of the rack rod (25) close to the center point of the regular polygon frame (21) is fixedly connected with an LED light plate (26), and the LED light plate (26) is designed in an arc shape.

9. A visual inspection system for wheel hub paint quality according to claim 8, characterized in that: The outer surface of the linkage shaft (24) is fixedly sleeved with a linkage gear (27), the linkage gear (27) is engaged with the corresponding rack rod (25), the surface of one of the parallel rods (22) is fixedly installed with a linkage motor (28), the output shaft of the linkage motor (28) is fixedly connected with a driving gear (29) on the outer surface, the outer surface of one of the linkage shafts (24) is fixedly sleeved with a synchronous gear (30), and the synchronous gear (30) is engaged with the driving gear (29); The inner surface of the regular polygon frame (21) is fixedly installed with a grating sensor (31) arranged in a symmetrical manner, both sides of the linear guide rail (19) are installed with a rotating screw rod (32) rotating through a supporting plate, the outer surface of the rotating screw rod (32) is threadedly sleeved with a threaded sleeve (33), the outer surface of the threaded sleeve (33) is fixedly connected with the outer surface of the regular polygon frame (21), the outer surfaces of the two rotating screw rods (32) are installed with a linkage housing (34), the outer surface of the linkage housing (34) is installed with a bevel gear set (35), the bottom plate of the detection station (2) is fixedly connected with a reduction motor (36), and the output shaft of the reduction motor (36) is fixedly connected with the inner surface of the driving bevel gear of the bevel gear set (35).

10. The visual inspection system for wheel hub paint quality according to claim 7, wherein: The visual detection system further comprises a detection method, and the steps include the following: Step one, the hook (5) on the clamping turntable (4) hooks the hub, one of the drive motors (11) starts, through the transmission of the worm (17), the worm gear rod (18), the worm sleeve (13), finally drives the bidirectional adjusting screw (8) to rotate, so that the hub to be detected is conveyed to the detection station (2) through the drying conveying line (1), the two L-shaped clamping blocks (9) on the bidirectional adjusting screw (8) move synchronously in the guide groove (6), and the hub is tightly supported from the inside of the center hole of the hub. When the hub defect is detected, the other drive motor (11) starts, the driving gear (15) is engaged with the driven gear (16), the rotating shaft (12) and the whole clamping turntable (4) are driven to rotate, and the first detection surface of the hub is adjusted to the initial position opposite to the sensor; Step two, the six-axis robot (39) in the intelligent cooperative motion control module adaptively plans the optimal scanning path and posture according to the hub three-dimensional model, the six-axis robot (39) drives the rigidly integrated industrial area array camera (37) and high-speed line laser 3D scanner (38) on the end flange plate to move to the first scanning point, at this time, the reduction motor (36) synchronously drives two rotating screws (32) through the bevel gear set (35), pushes the threaded tube sleeve (33), and drives the whole dynamic structured illumination unit to move to the preset optimal working distance along the linear guide rail (19). The grating sensor (31) detects the position in real time to realize closed-loop control. The linkage motor (28) starts, and through the transmission system composed of the driving gear (29), the synchronous gear (30), the linkage shaft (24) and the universal joint group, the motion is transmitted to all linkage gears (27), driving all rack rods (25) to slide synchronously in the groove plate (23), so as to adjust the arc-shaped LED lamp plate (26) to the preset radial position, forming a specific illumination mode; Step three, after the six-axis robot (39) reaches the predetermined scanning posture, a hardware trigger signal is sent out, the dynamic structured illumination unit is strictly and synchronously controlled to switch to the next illumination mode, and the industrial area array camera (37) is triggered to shoot a high-definition 2D image, and then the profile point cloud data acquired by the high-speed line laser 3D scanner (38) at the position is recorded; Step four, using the rigid transformation matrix calibrated in advance , the 3D point cloud coordinate system is unified to the 2D camera coordinate system, the spatial alignment is realized , the space-time registration is completed in combination with the hardware triggered timestamp, and the normal vector of each pixel point on the hub surface is calculated by using the registered 3D point cloud. Step 5: Based on the Phong lighting model, and combining the position of the curved LED light panel (26) and the camera position, predict the specular highlight intensity at each point in the image. The pixel value compensation algorithm is executed, which compares images under different lighting modes. If the predicted value of a certain pixel in the highlight mode is lower than the predicted value in the highlight mode, the pixel value is lowered. Exceeding the threshold Then, the pixel value with the richest diffuse reflection component is selected from other non-highlight mode images and replaced, finally generating a high-quality fused image without reflective distortion; Step six, run the traditional algorithms of edge detection, texture analysis and color difference calculation in CIELAB color space on the fused image in parallel, quickly circumscribe the suspected defect area, and filter false defects according to area, aspect ratio and other geometric characteristics, then send each ROI into the deep learning convolutional neural network, which adopts a multi-task learning framework to simultaneously output defect type probability and severity rating score, and calculates comprehensive confidence to realize accurate classification and rating; Step seven, the system integrates all defect information to determine the hub quality grade, and generates a visual report with a defect map on the human-computer interaction interface. For cases of system misjudgment or missed detection, the operator manually marks through the interaction interface, and these marking data are stored in the self-learning database for regular incremental training of the hybrid intelligent detection model, so that the system has the ability of continuous optimization.

Images