Automobile hub back cavity coating quality detection device and method based on vision

Abstract

In order to solve the problems of low hub image acquisition precision and low recognition accuracy in the prior art, the invention provides an automobile hub back cavity coating quality detection device and method based on vision, and the detection device comprises a transmission structure, a clamping rotation structure, an image acquisition structure and an image processing structure; the clamping and rotating structure is arranged over the conveying structure and used for clamping the hub and rotating the hub along the axis of the hub. The image acquisition structure is arranged right below the clamping and rotating structure and is arranged corresponding to the clamping and rotating structure; the image acquisition structure comprises at least one line scanning camera and is used for acquiring an image of the back surface of the hub; the image processing structure is in communication connection with the image obtaining structure, and the image processing structure adopts a hub back cavity real-time defect detection model obtained through training and obtained based on a Yolov5 feature fusion network to process an image of the back face of the hub; an image with a higher dynamic range and higher precision and resolution is obtained, and the accuracy of subsequent hub defect detection is improved.

Description

Technical Field

[0001] The present invention relates to the field of image detection, and particularly to a vision-based device and method for detecting the painting quality of the back cavity of an automotive wheel hub. Background Art

[0002] In the manufacturing process of automotive wheel hubs, it is very important to inspect the surface defects on the back side, as this directly affects the safety and service life of the wheel hubs. Although the traditional manual inspection method is simple and intuitive, there are many problems. For example, the manual inspection speed is slow, which is not suitable for large-scale production. Moreover, the human eyes are prone to fatigue and the attention will be distracted, resulting in missed inspections or misinspections, affecting the accuracy and consistency of the detection, etc.

[0003] In recent years, with the development of computer vision and artificial intelligence technologies, automatic detection systems based on machine vision have begun to receive attention and have gradually become a new choice for improving the detection efficiency and accuracy of wheel hub surface defects, such as the content described in the patent with the application publication number CN117095002A. These systems capture images of the wheel hub surface through high-resolution cameras, and then use advanced image processing algorithms to analyze these images, automatically identifying and marking defects such as scratches, cracks, and depressions, greatly improving the objectivity and speed of the detection. However, most of the traditional camera imaging relies on 2D cameras for planar imaging. On the one hand, the obtained images have low pixels. On the other hand, it is difficult for a single camera to comprehensively obtain images of all directions of the back cavity of the wheel hub. Usually, multiple images need to be taken and combined, increasing the workload of image processing operations. Therefore, a device and method for detecting the painting quality of the back cavity of an automotive wheel hub with high dynamic range, high precision, and high resolution are needed. Summary of the Invention

[0004] The purpose of the present invention is to solve the deficiencies of the prior art and provide a vision-based device and detection method for detecting the painting quality of the back cavity of an automotive wheel hub.

[0005] To solve the above problems, the present invention adopts the following technical solutions: A vision-based device for detecting the painting quality of the back cavity of an automotive wheel hub includes a transmission structure for transporting the wheel hub; it also includes a clamping and rotating structure, an image acquisition structure, and an image processing structure; wherein the clamping and rotating structure is arranged directly above the transmission structure, used for clamping the wheel hub and rotating the wheel hub along the axis of the wheel hub; the image acquisition structure is arranged directly below the clamping and rotating structure, corresponding to the clamping and rotating structure; the image acquisition structure includes at least one line-scan camera, used for acquiring images of the back side of the wheel hub; the image processing structure is communicatively connected to the image acquisition structure, and the image processing structure uses a real-time defect detection model of the back cavity of the wheel hub obtained by training based on the Yolov5 feature fusion network to process the images of the back side of the wheel hub.

[0006] Further, the image acquisition structure includes a camera module, a lifting control device, and a lighting device; wherein the camera module is disposed on the lifting control device that moves in the vertical direction; a lighting device is further disposed around the camera module; the camera module includes three line-scan cameras, and the three line-scan cameras are arranged side by side, and the respective orientation angles of the three line-scan cameras are set angles.

[0007] Further, the lifting control device includes a lead screw transmission mechanism for controlling lifting and a servo motor.

[0008] Further, the lighting device includes lighting lamps, and the lighting lamps are distributed on both sides of the camera module.

[0009] Further, the clamping and rotating structure includes a support frame, a rotating device, a lifting device, and a clamping device; wherein the lifting device is fixedly disposed on the support frame, and the end of the lifting device is further fixedly connected to the rotating device; the rotating output end of the rotating device is connected to the clamping device; the clamping device includes pneumatically parallel clamps.

[0010] A method for detecting the painting quality of the back cavity of an automotive wheel hub based on vision, the detection method is based on the above detection device, and the detection method includes the following steps: Step 1: After the transmission structure transports the wheel hub to be detected to the detection station directly below the image acquisition device, it stops operating; Step 2: The clamping and rotating structure moves downward, and controls the pneumatically parallel clamps to clamp the wheel hub, and then lifts it to a set height; Step 3: The lifting control device in the image acquisition structure controls the camera to move to a set height; Step 4: The clamping and rotating structure controls the wheel hub to rotate, and at the same time, the camera module and the lighting device in the image acquisition structure start to work, and scan and acquire the image of the back cavity of the bottom side of the wheel hub; Step 5: Transmit the scanned image to the image processing structure, and use the defect detection model based on the Yolov5 feature fusion network that has completed training and deployment to detect the wheel hub image, and judge whether there are defects in the back cavity of the wheel hub, and end the step.

[0011] Further, controlling the camera to move to a set height in the step 3 specifically includes the following steps: Step 31: The lifting control device controls the camera to move to a preset height and then acquires the initial image of the wheel hub; Step 32: Preprocess the initial image; Step 33: Use the Canny edge detection algorithm to extract the edges in the image; Step 34: Perform a closing operation on the edges of the wheel hub; Step 35: Find all closed contours in the image and select the largest contour as the hub edge contour; Step 36: Use the Hough circle detection method to obtain the radius of the hub edge contour and calculate its area; Step 37: Adjust the height of the camera according to the preset table of the comparison between the circle area and the height adjustment.

[0012] Further, in the said step 32, the preprocessing of the initial image includes converting it into a grayscale image and Gaussian blur processing.

[0013] Further, the training process of the defect detection model in the said step 5 includes the following steps: Step 51: Obtain the image dataset of various hub back cavities for training and establish a hub defect database; Step 52: Classify and label the defect data and segment and label the data of the image data according to the hub defect type and position in the image; Step 53: Divide the labeled image data into a training dataset and a test dataset according to a set ratio; Step 54: Preprocess the images in the training dataset to expand the number of data samples; Step 55: Build a model based on the YOLOv5 feature fusion network; Step 56: Use the learning rate adaptation method to train and test the model with the training dataset and the test dataset; Step 57: Deploy the tested model to the image processing structure to end the step.

[0014] Further, in the said step 55, the process of building the model is as follows: Obtain the feature fusion network of YOLOv5; Improve the upsampling method of the feature fusion network, including embedding content-aware adaptive feature fusion into the feature fusion network of YOLOv5 to construct the CARAFE-PANET feature fusion network; Improve the downsampling method of the convolutional layer PyConv, including setting multiple convolutional kernels of different sizes in the convolutional layer for feature extraction operations on the image in the same layer, and the convolutional kernels are connected in the form of grouped convolution in the convolutional layer; Add the CBAM attention mechanism; Improve the confidence cross-entropy loss function, classification cross-entropy loss function, and mean square error loss function of regression prediction in the YOLOv5 algorithm. The improved loss function is: , Among them, p represents the output value of the activation function Sigmoid; 0 ≤ ɑ ≤ 1 is the class weight factor; y is the label value; γ ≥ 0 is the focusing parameter; ɑ and γ are fixed values; represents the adjustment factor.

[0015] The beneficial effects of the present invention are as follows: By using a line-scan camera to scan and obtain pictures of the hub back cavity, compared with traditional optical cameras, images with a higher dynamic range, higher precision, and higher resolution can be obtained, improving image clarity and thus the accuracy of subsequent hub defect detection; By adopting a deep learning model to detect defects in hub images, on the one hand, the detection efficiency can be improved, and on the other hand, problems of false detection or missed detection caused by fatigue and other reasons in traditional manual detection can be avoided. Moreover, as the training progress of the learning model, it can also be used to simultaneously detect different hub defects, which is more efficient; By improving the upsampling, downsampling, and loss function of the YOLOv5 model and introducing an attention mechanism, small target defects on the hub can be better detected. Since the defect area on the hub is usually small, the model parameters need to be adjusted to adapt to specific detection requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 is a schematic diagram of the overall structure of Embodiment 1 of the present invention; Figure 2 is a schematic diagram of the image acquisition structure of Embodiment 1 of the present invention; Figure 3 is a schematic diagram of the CARAFE-PANET feature fusion network of Embodiment 1 of the present invention; Figure 4 is a schematic diagram of the PyConv network structure of Embodiment 1 of the present invention.

[0017] Description of the drawing reference numerals: rotating device 1, lifting device 2, clamping device 3, image acquisition structure 4, line-scan camera 5, lighting device 6, lifting control device 7. DETAILED DESCRIPTION OF THE INVENTION

[0018] The following specific examples illustrate the implementation manners of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific implementation manners. Various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, without conflict, the following embodiments and the features in the embodiments can be combined with each other.

[0019] It should be noted that the illustrations provided in the following embodiments only schematically illustrate the basic concept of the present invention. Therefore, only the components related to the present invention are shown in the figures, rather than being drawn according to the number, shape, and size of the components in actual implementation. The types, quantities, and proportions of the components in actual implementation can be arbitrarily changed, and the layout type of the components may also be more complex. Embodiment

[0020] As Figure 1 , 2 shown, a vision-based automotive wheel hub back cavity painting quality detection device includes a transmission structure for transmitting the wheel hub. In this example, a roller conveyor belt is used to transmit the wheel hub to be detected. The detection device further includes a clamping and rotating structure, an image acquisition structure 4, and an image processing structure; wherein the clamping and rotating structure is arranged directly above the transmission structure, used to clamp the wheel hub and rotate the wheel hub along the axis of the wheel hub; the image acquisition structure 4 is arranged directly below the clamping and rotating structure, corresponding to the clamping and rotating structure; the image acquisition structure 4 includes at least one line scan camera 5, used to acquire images of the wheel hub back cavity; the image processing structure is communicatively connected to the image acquisition structure 4, and the image processing structure uses a real-time defect detection model of the wheel hub back cavity obtained by training based on the Yolov5 feature fusion network to process the images of the back of the wheel hub, realizing the detection of defects in the wheel hub back cavity.

[0021] A blocking structure is also arranged on the roller conveyor belt, and the blocking structure includes a lifting mechanism that can be lifted and a blocking plate. The blocking plate is arranged at the movable end of the lifting mechanism, and the lifting structure is fixedly arranged below the roller conveyor belt.

[0022] The image acquisition structure 4 includes a camera module, a lifting control device 7, and a lighting device 6; wherein the camera module is arranged on the lifting control device 7 that moves in the vertical direction; a lighting device 6 is also arranged around the camera module; the camera module includes three line scan cameras 5, and the three line scan cameras 5 are arranged side by side. The respective orientation angles of the three line scan cameras 5 are set angles. Among them, the direction of the line scan camera 5 in the middle faces directly upward, and the orientations of the line scan cameras 5 on both sides face away from the middle line scan camera 5 and form set angles with the vertical direction; in this way, with the cooperation of the three line scan cameras 5, rotating the wheel hub can obtain images of the front wheel hub back cavity faster.

[0023] The lifting control device 7 includes a lead screw drive mechanism for controlling lifting and a servo motor, and the height of the camera module on the lead screw drive mechanism is precisely controlled by the servo motor. An L-shaped support plate is fixedly arranged on the slider of the lead screw drive mechanism. One side of the support plate is fixedly connected to the slider, and the other side of the support plate is in the horizontal plane. The upper surface of the support plate is used to arrange the camera module.

[0024] The lighting device 6 includes a lighting lamp. The lighting lamp is a strip light source as a whole and is distributed on both sides of the camera module. The lighting lamp is also fixed on the support plate of the lifting control device 7. The lighting lamp also cooperates with the roller conveyor belt. Specifically, the lighting lamp is located between adjacent rollers of the roller conveyor belt, facilitating the lighting lamp to transmit its light to the back cavity of the hub located above the roller conveyor belt, achieving a better lighting effect.

[0025] The clamping and rotating structure includes a support frame, a rotating device 1, a lifting device 2, and a clamping device 3. The lifting device 2 is fixedly arranged on the support frame, and the end of the lifting device 2 is also fixedly connected to the rotating device 1. The rotating output end of the rotating device 1 is connected to the clamping device 3. The clamping device 3 includes pneumatically parallel clamps. When operating, the clamps on both sides move towards the middle part simultaneously, thereby clamping the edge of the hub and achieving clamping of the hub.

[0026] A method for detecting the painting quality of the back cavity of an automotive hub based on vision. The detection method is based on the above detection device, and the detection method includes the following steps: Step 1: After the transmission structure transports the hub to be detected to the detection station directly below the image acquisition device, it stops operating. Step 2: The clamping and rotating structure moves downward, and the pneumatically parallel clamps are controlled to clamp the hub, and then it is lifted upward to a set height. Step 3: The lifting control device 7 in the image acquisition structure 4 controls the camera to move to a set height. Step 4: The clamping and rotating structure controls the hub to rotate. At the same time, the camera module and the lighting device 6 in the image acquisition structure 4 start to work, and an image of the back cavity on the bottom side of the hub is scanned and acquired. Step 5: The scanned image is transmitted to the image processing structure, and the defect detection model based on the Yolov5 feature fusion network that has completed training and deployment detects the hub image to determine whether there are defects in the back cavity of the hub, and the step ends.

[0027] Controlling the camera to move to a set height in Step 3 specifically includes the following steps: Step 31: The lifting control device 7 controls the camera to move to a preset height and then acquires the initial image of the hub. Step 32: Preprocess the initial image. Step 33: Use the Canny edge detection algorithm to extract the edges in the image. Step 34: Perform a closing operation on the edges of the hub image. Step 35: Find all closed contours in the image and select the largest contour as the hub edge contour. Step 36: Use the Hough circle detection method to obtain the radius of the hub edge contour and calculate its area; Step 37: Adjust the height of the camera according to the preset table of the comparison between the circle area and the height adjustment.

[0028] In the said Step 32, the preprocessing of the initial image includes converting it into a grayscale image and performing Gaussian blur processing. The former is to reduce the calculation amount and retain the edge information of the hub, and the latter is to smooth the image and reduce the interference of noise.

[0029] In the said Step 34, the closing operation of the edge of the hub image is realized by the method of dilation followed by erosion.

[0030] In the said Step 37, the height of the camera is negatively correlated with the area. Because the larger the area of the detected hub image indicates that it is too close to the hub. In order to ensure that a relatively complete hub image can be obtained as much as possible, the camera is controlled to descend to reduce the area of the hub image.

[0031] The training process of the defect detection model in the said Step 5 includes the following steps: Step 51: Obtain the image data set of various hub back cavities for training and establish a hub defect database; Step 52: Classify and label the defect data and segment and label the data of the image data according to the hub defect type and position in the image; Step 53: Divide the labeled image data into a training data set and a test data set according to a set ratio; Step 54: Preprocess the images in the training data set to expand the number of data samples; Step 55: Build a model based on the Yolov5 feature fusion network; Step 56: Use the learning rate adaptation method to train and test the model using the training data set and the test data set; Step 57: Deploy the tested model to the image processing structure to end the step.

[0032] In the said Step 54, the preprocessing of the image includes image enhancement by randomly adjusting the image brightness, saturation, and contrast of the hub back cavity image, and expanding the number of samples by adding cropping, flipping the hub back cavity pictures left and right and up and down.

[0033] In the said Step 55, the process of building the model is as follows: Obtain the feature fusion network of YOLOv5; the nearest neighbor interpolation algorithm used by the traditional YOLOv5 feature fusion network for upsampling the high-order feature map simply copies the nearest neighbor pixels, with a small perception range, ignoring the influence of other adjacent pixel points, which will cause image blurring and jaggedness, and cannot obtain more semantic information about small targets.

[0034] To improve the object detection performance of YOLOv5, the upsampling method of the feature fusion network is improved, namely Content-Aware ReAssembly of FEatures (CARAFE); CARAFE can predict the optimal upsampling kernel for each pixel position by analyzing the input feature map, and then recombine the features according to the predicted upsampling kernel, thus realizing upsampling based on the semantic information of the input feature map; in this way, not only can the upsampling strategy be dynamically adjusted to adapt to different feature contents, but also the information loss and blurring phenomenon caused by the fixed upsampling kernel in the traditional feature fusion network are avoided. By Figure 3 embedding the content-aware adaptive feature fusion into the feature fusion network of YOLOv5 through the structure shown in the figure, a CARAFE-PANET feature fusion network is constructed; the CARAFE-PANET feature fusion network can capture the spatial relationship between features during the upsampling process, expand the receptive field, and retain semantic information at the same time, so as to generate a clearer and more detailed feature map.

[0035] For the situation where the sizes of image defects are unevenly distributed and multiple defects are distributed in the same image, the downsampling method of the convolutional layer PyConv is selected to improve the overall accuracy of the model in the presence of multiple defects in the same image, and at the same time, the overall model is continuously optimized to improve the model in the direction of lightweight. The PyConv network simultaneously uses multiple convolutional kernels of different sizes, such as 3×3, 5×5, 7×7, etc., to perform feature extraction operations in the same layer, and each convolutional kernel is connected in the form of grouped convolution in PyConv. The application of multi-scale convolutional kernels enables the model to capture the details at all levels in the image in detail, thus significantly improving the feature extraction ability of the neural network model; thanks to the structure of grouped convolution, PyConv enhances the ability to process complex image scenes while keeping the number of model parameters and the computational amount similar to those of the traditional convolutional layer. Therefore, when using PyConv to build a computer vision model, it is possible to focus on optimizing the network structure without significantly increasing the computational cost. The PyConv network structure is as Figure 4 shown.

[0036] Adding the CBAM attention mechanism to the feature fusion network can learn weight information more reasonably, ignore irrelevant information, save computing resources, improve the detection performance of the network, enhance the detection ability of small target features in defective images, and reduce the impact of the background on detection, so as to more accurately identify and locate defective targets. CBAM enables the network to simultaneously focus on the channel and spatial dimensions, which is more conducive to the extraction of defective target features. First, the input feature map will be subjected to max pooling and average pooling simultaneously, compressing the feature map in the spatial dimension to generate two one-dimensional feature vectors, concentrating the attention of the network on the channel dimension, and enhancing the learning ability of the channel weight sum. After passing through a multilayer perceptron (MLP) and a Sigmoid activation operation, the features input to the spatial attention module are generated. In the spatial attention module, the feature map will be subjected to max pooling and average pooling in sequence, compressing the channels of the feature map to obtain two feature maps of H×W×1, prompting the attention of the network to focus on the spatial dimension. After a 7×7 convolution to aggregate channel information to generate a 2D feature map, the spatial feature map is finally obtained through a convolution operation.

[0037] Improve the confidence cross-entropy loss function, classification cross-entropy loss function, and mean square error loss function for regression prediction in the YOLOv5 algorithm. The improved loss functions are as follows:

[0038] Among them, p represents the output value of the activation function Sigmoid; 0≤ɑ≤1 is the class weight factor; y is the label value; γ≥0 is the focusing parameter; ɑ and γ are fixed values; represents the adjustment factor. As p increases, is smaller, lg p is larger, that is, the adjustment factor reduces the loss contribution of high-probability targets and at the same time expands the perception range of small targets. Such an adjustment helps to improve the accuracy of small target detection. Since the number of samples of some categories in the training data is relatively small, the performance of the model in predicting these categories is not ideal. Focal-Loss adjusts the weights of the loss function, enabling the model to pay more attention to difficult-to-classify samples, thereby improving the prediction accuracy of the model for minority classes.

[0039] In the implementation process, by using the line-scan camera 5 to scan and obtain pictures of the back cavity of the wheel hub, compared with traditional optical cameras, images with higher dynamic range, higher precision and resolution can be obtained, improving the image clarity, and further improving the accuracy of subsequent wheel hub defect detection; by adopting a deep learning model to detect defects in the wheel hub images, on the one hand, the detection efficiency can be improved, and on the other hand, the problems of false detection or missed detection caused by fatigue and other reasons in traditional manual detection can be avoided. Moreover, with the training progress of the learning model, it can also be used to detect different wheel hub defects simultaneously, which is more efficient; by improving the upsampling, downsampling and loss function of the YOLOv5 model and introducing an attention mechanism, small target defects on the wheel hub can be better detected. Since the defect area on the wheel hub is usually small, the model parameters need to be adjusted to adapt to specific detection requirements.

[0040] The above description is only a specific example of the present invention and does not constitute any limitation to the present invention. Obviously, for professionals in the field, after understanding the content and principle of the present invention, various modifications and changes in form and details may be made without departing from the principle and structure of the present invention. However, these corrections and changes based on the idea of the present invention are still within the protection scope of the claims of the present invention.

Claims

1. A vision-based quality inspection device for the coating of the back cavity of an automotive wheel hub, comprising a transmission structure for transmitting the wheel hub; characterized in that, It further includes a clamping and rotating structure, an image acquisition structure (4), and an image processing structure; Among them, the clamping and rotating structure is arranged directly above the transmission structure and is used to clamp the wheel hub and rotate the wheel hub along the axis of the wheel hub; the image acquisition structure (4) is arranged directly below the clamping and rotating structure and is correspondingly arranged with the clamping and rotating structure; the image acquisition structure (4) includes at least one line-scan camera (5) for acquiring images of the back of the wheel hub; the image processing structure is communicatively connected to the image acquisition structure (4), and the image processing structure uses a real-time defect detection model for the back cavity of the wheel hub obtained by training based on the Yolov5 feature fusion network to process the images of the back of the wheel hub.

2. The visual-based automobile wheel hub back cavity painting quality detection device according to claim 1, wherein The image acquisition structure (4) includes a camera module, a lifting control device (7), and a lighting device (6); among them, the camera module is arranged on the lifting control device (7) that moves in the vertical direction; a lighting device (6) is also arranged around the camera module; the camera module includes three line-scan cameras (5), and the three line-scan cameras (5) are arranged side by side, and the respective orientation angles of the three line-scan cameras (5) are set angles.

3. The visual-based automobile wheel hub back cavity painting quality detection device according to claim 2, characterized in that, The lifting control device (7) includes a lead screw transmission mechanism for controlling lifting and a servo motor.

4. The visual-based automotive wheel hub back cavity painting quality detection device according to claim 2, wherein The lighting device (6) includes lighting lamps, and the lighting lamps are distributed on both sides of the camera module.

5. The visual-based automobile wheel hub back cavity painting quality detection device according to claim 1, wherein The clamping and rotating structure includes a support frame, a rotating device (1), a lifting device (2), and a clamping device (3); among them, the lifting device (2) is fixedly arranged on the support frame, and the end of the lifting device (2) is also fixedly connected to the rotating device (1); the rotating output end of the rotating device (1) is connected to the clamping device (3); the clamping device (3) includes pneumatically parallel clamps.

6. A vision-based method for detecting the painting quality of the back cavity of an automotive wheel hub, characterized in that, The detection method is based on the detection device according to any one of claims 1 to 5, and the detection method includes the following steps: Step 1: After the transmission structure transports the wheel hub to be detected to the detection station directly below the image acquisition device, it stops operating; Step 2: The clamping and rotating structure moves downward, and controls the pneumatically parallel clamps to clamp the wheel hub, and then lifts it to a set height; Step 3: The lifting control device (7) in the image acquisition structure (4) controls the camera to move to a set height; Step 4: The clamping and rotating structure controls the wheel hub to rotate, and at the same time, the camera module and the lighting device (6) in the image acquisition structure (4) start to work, and scan and acquire images of the back cavity of the bottom side of the wheel hub; Step 5: Transmit the scanned images to the image processing structure, and use the defect detection model based on the Yolov5 feature fusion network that has completed training and deployment to detect the wheel hub images, and judge whether there are defects in the back cavity of the wheel hub, and end the step.

7. A visual-based method for detecting the painting quality of the back cavity of an automotive wheel hub, characterized in that, In step 3, controlling the camera to move to a set height specifically includes the following steps: Step 31: The lifting control device (7) controls the camera to move to a preset height and then acquires the initial image of the wheel hub; Step 32: Preprocess the initial image; Step 33: Use the Canny edge detection algorithm to extract the edges in the image; Step 34: Perform a closing operation on the edges of the wheel hub image; Step 35: Find all closed contours in the image and select the largest contour as the wheel hub edge contour; Step 36: Use the Hough circle detection method to obtain the radius of the wheel hub edge contour and calculate its area; Step 37: Adjust the height of the camera according to the preset table of the comparison between the circle area and the height adjustment.

8. A vision-based method for detecting the painting quality of the back cavity of an automotive wheel hub according to claim 7, characterized in that, In the said Step 32, the preprocessing of the initial image includes converting it into a grayscale image and Gaussian blur processing.

9. A vision-based method for detecting the painting quality of the back cavity of an automotive wheel hub according to claim 6, characterized in that The training process of the defect detection model in the said Step 5 includes the following steps: Step 51: Obtain an image data set of various wheel hub back cavities for training and establish a wheel hub defect database; Step 52: Classify and label the defect data and segment and label the data of the image data according to the types and positions of the wheel hub defects in the image; Step 53: Divide the labeled image data into a training data set and a test data set according to a set ratio; Step 54: Preprocess the images in the training data set to expand the number of data samples; Step 55: Build a model based on the Yolov5 feature fusion network; Step 56: Use the learning rate adaptive method to train and test the model using the training data set and the test data set; Step 57: Deploy the tested model into the image processing structure to end the step.

10. A method for detecting the painting quality of the back cavity of an automotive wheel hub based on vision according to claim 9, characterized in that, In the said Step 55, the process of building the model is as follows: Obtain the feature fusion network of YOLOv5; Improve the upsampling method of the feature fusion network, including embedding content-aware adaptive feature fusion into the feature fusion network of YOLOv5 to construct the CARAFE-PANET feature fusion network; Improve the downsampling method of the convolutional layer PyConv, including setting multiple convolutional kernels of different sizes in the convolutional layer for feature extraction operations on the image in the same layer, and the convolutional kernels are connected in the form of grouped convolution in the convolutional layer; Add the CBAM attention mechanism; Improve the confidence cross-entropy loss function, classification cross-entropy loss function, and mean square error loss function for regression prediction in the YOLOv5 algorithm. The improved loss function is: ， Among them, p represents the output value of the activation function Sigmoid; 0 ≤ ɑ ≤ 1 is the class weight factor; y is the label value; γ ≥ 0 is the focusing parameter; ɑ and γ are fixed values; represents the adjustment factor.

Images