Automobile part quality intelligent detection method based on machine vision

Abstract

The invention discloses an intelligent automobile part quality detection method based on machine vision, and relates to the related field of machine vision and quality detection, and the method comprises the steps: S1, collecting an automobile part surface image through an image collection system; s2, a multi-scale Retinex algorithm based on guided filtering is adopted to pre-process the surface image of the automobile part, and the image quality is improved; s3, using a YOLOv11 instance segmentation model to carry out intelligent detection on the preprocessed automobile part surface image, and identifying and positioning a defect area of the image; and S4, obtaining a defect detection result of the surface of the automobile part, and generating a quality detection report. The problem that in the prior art, the quality detection precision of automobile parts is low is solved. According to the invention, the improved image enhancement algorithm is adopted to improve the image quality, and the learning ability of the neural network is utilized to improve the stability and robustness of quality detection. According to the intelligent quality detection method, the maintenance work of automobile part production is more targeted, and the maintenance and repair cost is effectively reduced.

Description

Technical Field

[0001] The invention relates to the technical field of machine vision and quality detection, and mainly to a method for intelligently detecting the quality of automobile parts by using machine vision and a neural network model at an automobile parts production site. Background Art

[0002] In the production workshop of the automated assembly line of auto parts, the quality inspection of auto parts is extremely important. Traditional manual visual inspection methods have many shortcomings, such as low efficiency, large subjective errors, and inconsistent inspection standards, which make it difficult to meet the high precision and high efficiency requirements of modern automobile manufacturing. Machine vision technology uses the bionic human visual perception mechanism, with the help of hardware devices such as industrial cameras and optical sensors, to obtain digital images of target objects, and combines advanced image processing technology to achieve accurate decision-making and judgment of object shape, texture and spatial relationship. Through multi-dimensional feature extraction and real-time comparison and analysis, the machine vision system can efficiently complete the accurate identification and automatic detection of accessory defects, significantly improving production efficiency and product quality consistency, and providing strong support for the intelligent upgrade of the automobile manufacturing industry.

[0003] Traditional visual inspection-based quality inspection methods for automotive parts rely on artificially designed features and use image processing technology, which has poor adaptability to lighting changes and complex backgrounds. Its general steps include image acquisition, preprocessing, segmentation, feature extraction, and classification recognition. Among them, image segmentation often uses algorithms such as threshold segmentation, edge detection, and region growing. However, the defect morphology of automotive parts is complex and diverse, and the morphological characteristics of defects of different categories are significantly different, which limits the universality of traditional detection algorithms and makes development extremely difficult. In addition, factors such as the consistency of the position of accessories, the stability of lighting, and the effectiveness of the detection algorithm will directly affect the quality of image acquisition and detection performance. Therefore, how to develop a stable, reliable, and robust visual inspection-based automotive parts quality inspection method to effectively deal with the interference of external adverse environments is a key issue that needs to be solved urgently. Summary of the invention

[0004] The present invention aims to solve the deficiencies of the prior art and proposes an intelligent detection method for automobile parts quality based on machine vision. The method focuses on overcoming the limitations of the prior art in automobile parts quality detection, especially the problem of low detection accuracy caused by light interference and poor image quality. By introducing a neural network model, its powerful learning ability is fully utilized, and the stability and robustness of the detection are significantly improved, thereby providing an efficient and reliable solution for automobile parts quality detection.

[0005] To achieve the above object, the present invention provides an intelligent detection method for automobile parts quality based on machine vision, which specifically includes the following steps:

[0006] S1: Use an image acquisition system to acquire the surface image of auto parts;

[0007] S2: Preprocess the surface image of auto parts using a multi-scale Retinex algorithm based on guided filtering to improve the image quality;

[0008] S3: Use the YOLOv11 instance segmentation model to perform intelligent detection on the preprocessed surface image of auto parts, and identify and locate the defective areas of the image;

[0009] S4: Obtain the defective detection result of the auto parts surface and generate a quality inspection report.

[0010] Furthermore, the acquisition system for the surface image of auto parts includes:

[0011] Industrial camera: Used to capture the surface image of auto parts and placed at the center of the light source device.

[0012] Light source device: Used to provide stable lighting conditions to ensure the clarity and consistency of image acquisition. The type and intensity of the light source are adjusted according to the detection requirements.

[0013] Image acquisition card: Used to receive the image data captured by the camera and transfer it to the computer system.

[0014] Bracket and fixing device: Used to fix the camera and the light source to ensure the stability and accuracy of their positions.

[0015] Furthermore, use a multi-scale Retinex image enhancement algorithm based on guided filtering to preprocess the image. The functions of this method are as follows:

[0016] In the improved Retinex image enhancement algorithm, the guided filter kernel function is used to replace the traditional Gaussian kernel function to obtain a clearer reflection image, and the contrast is further stretched through contrast-limited adaptive histogram equalization. The guided filter combines the concepts of Gaussian filtering and guided diffusion, and can effectively remove noise and retain edge information. It uses the local structure information of the image to filter the image, while avoiding the halo effect and detail loss that may be caused by Gaussian filtering.

[0017] Furthermore, in order to improve the diversity of the dataset and the generalization ability of the model, the present invention uses data augmentation methods in different ways to expand the dataset, including: rotation, horizontal and vertical flipping, and brightness adjustment.

[0018] Furthermore, the YOLOv11 instance segmentation model has made significant improvements in architecture and training methods based on previous YOLO versions. It integrates improved model structure design, enhanced feature extraction techniques, and optimized training methods. YOLOv11 includes an improved Backbone, Neck, and Head, and the detailed structures of each network are as follows:

[0019] The Backbone enhances the model's feature extraction ability through CBS, C3K2, SPPF, and C2PSA modules. Among them, both the CBS and C3K2 modules adopt a bottleneck structure;

[0020] Furthermore, the bottleneck structure consists of three convolutional layers. Through the first convolutional layer, the number of channels of the input feature map is halved first. The second convolutional layer is used to extract features, and the last convolutional layer restores the number of channels and adds it to the original input feature map for output;

[0021] Furthermore, the convolutional layer includes convolutional operations, batch normalization, and activation functions;

[0022] Furthermore, the C3K2 module is the core of the Backbone. It optimizes the information flow in the network by splitting the feature map and applying a series of small-kernel convolutions, which is faster and has lower computational cost than large-kernel convolutions. By processing smaller independent feature maps and merging them after several convolutions, the C3K2 module uses fewer parameters to improve feature representation compared to the C2f module of YOLOv8.

[0023] Furthermore, the SPFF module performs max-pooling operations on the input feature map through multiple pooling kernels of different sizes, thereby extracting feature information of different scales. By fusing multi-scale features, the SPFF module can better capture the detailed features of the surface images of small automotive parts. Compared with the traditional Spatial Pyramid Pooling (SPP), SPFF reduces the computational amount through optimized pooling operations while maintaining high detection accuracy.

[0024] Furthermore, the C2PSA module uses two PSA (Partial Spatial Attention) modules, which operate on different branches of the feature map and are then connected. By applying spatial attention to the extracted feature map, it refines the model's ability to selectively focus on regions of interest. This makes YOLOv11 superior to previous YOLO versions in scenarios where surface defects of automotive parts are accurately detected.

[0025] The Neck is located between the Backbone and the Head, and uses the Upsample module and Contact module to achieve feature fusion and enhancement;

[0026] Furthermore, the Upsample module is an upsampling module, and the upsampling method used is nearest neighbor interpolation;

[0027] Furthermore, the Contact module represents a concatenation operation that concatenates features of different scales along the channel dimension.

[0028] The Head is the decision-making part of the model, responsible for generating the final detection results; the head network receives multi-scale feature maps from the neck network and processes bounding box regression and class prediction separately through decoupled detection heads; the bounding box regression prediction branch uses the CBS module to process features, and the class prediction branch uses the DSC module based on depthwise separable convolution to process features, reducing the computational amount and improving efficiency.

[0029] Furthermore, for each position on each feature map, the head network of YOLOv11 generates bounding box coordinates, class probabilities, and confidence scores. Redundant bounding boxes are removed through non-maximum suppression, and finally the detection results are output to generate a detection report.

[0030] The present invention discloses the following beneficial effects:

[0031] In the process of quality inspection of automotive parts based on visual detection, the image quality is crucial for the detection effect. Therefore, the present invention introduces a multi-scale Retinex image enhancement algorithm based on guided filtering in the image preprocessing stage. This algorithm has good edge-preserving ability and can effectively enhance the details and contrast of the surface image of automotive parts, making the defects on the surface image of the parts clearer and more complete, thus providing strong support for subsequent defect recognition. When selecting a detection model, the present invention adopts the advanced YOLOv11 instance segmentation model. This model can not only achieve high-precision detection with lower computational complexity but also provide an efficient technical guarantee for the intelligent detection of the quality of automotive parts. Through this intelligent quality detection method, enterprises can accurately analyze the causes of defects, which can not only make the maintenance work of the production line more targeted but also effectively reduce the maintenance cost and overhaul cost. Brief Description of the Drawings

[0032] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings of the embodiments of the present invention will be briefly introduced below. Flowcharts are used in this application to illustrate the operations performed by the system according to the embodiments of the present application. It should be understood that the operations described above or below do not necessarily have to be performed precisely in sequence. On the contrary, various steps can be performed in reverse order or simultaneously as needed. At the same time, other operations can also be added to these processes, or one or several steps can be removed from these processes.

[0033] Figure 1 It is a schematic flowchart of the intelligent quality detection method for automotive parts based on machine vision provided by the embodiments of this application.

[0034] Figure 2 It is a schematic structural diagram of the YOLOv11 instance segmentation model.

[0035] Explanation of reference numerals: Figure 2 The output feature size of each layer module is marked, describing the number of channels, width, and height of the features. Detailed implementation manners

[0036] The above description is only an overview of the technical solution of this application. In order to be able to understand the technical means of this application more clearly, it can be implemented according to the content of the specification. And in order to make the above and other purposes, features, and advantages of this application more obvious and understandable, the following specifically gives the detailed implementation manners of this application.

[0037] In order to make the purpose, technical solution, and advantages of this application clearer, the following will further describe this application in detail with reference to the accompanying drawings. The described embodiments should not be regarded as limitations of this application. All other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of this application.

[0038] In the following description, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or server that includes a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but may include other steps or modules that are not clearly listed or are inherent to these processes, methods, products, or devices. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the technical field to which this application belongs. The terms used herein are only for the purpose of describing the embodiments of this application.

[0039] Embodiment 1. This application provides an intelligent quality inspection method for automotive parts based on machine vision, as Figure 1 shown. The method includes the following steps:

[0040] S1: Use an image acquisition system to acquire the surface image of the automotive part;

[0041] S2: Adopt a multi-scale Retinex algorithm based on guided filtering to preprocess the surface image of the automotive part to improve the image quality;

[0042] S3: Use the YOLOv11 instance segmentation model to perform intelligent detection on the preprocessed surface image of the automotive part to identify and locate the defective areas in the image;

[0043] S4: Obtain the defective detection result on the surface of the automotive part and generate a quality inspection report.

[0044] The surface image acquisition system of the automotive part includes:

[0045] Industrial camera: Used to capture the surface images of automotive parts and placed at the center of the light source device.

[0046] Light source device: Used to provide stable lighting conditions to ensure the clarity and consistency of image acquisition. The type and intensity of the light source are adjusted according to the detection requirements.

[0047] Image acquisition card: Used to receive the image data captured by the camera and transmit it to the computer system.

[0048] Bracket and fixing device: Used to fix the camera and the light source to ensure the stability and accuracy of their positions.

[0049] The detailed process of the multi-scale Retinex image enhancement algorithm based on guided filtering is as follows:

[0050] The traditional Retinex algorithm uses Gaussian filtering to estimate the illumination component, but Gaussian filtering will cause the halo effect and loss of edge information. In the present invention, guided filtering is used to replace Gaussian filtering. Guided filtering combines spatial geometry and brightness similarity, can better preserve edge information, and has higher computational efficiency. The principles involved in the multi-scale Retinex algorithm include:

[0051] The Retinex theory holds that an image is composed of an illumination image and a reflection image. The illumination image refers to the information of the incident component of an object, denoted by L(x, y); the reflection image refers to the reflected part of the object, denoted by R(x, y). The original image is denoted by I(x, y), and the formula is expressed as: I(x, y) = L(x, y) * R(x, y)

[0052] The illumination vector of the image: L(x, y) = I(x, y) * F(x, y), where F(x, y) is the guided filtering kernel function. Guided filtering uses the local linear relationship of the guidance image for filtering, avoiding complex weight calculations, being similar to mean filtering in flat areas, but retaining details in edge areas.

[0053] The reflection component of the image: R(x, y) = logI(x, y) - logL(x, y).

[0054] The multi-scale Retinex algorithm sums the Retinex results of multiple scales weighted, and the formula is:

[0055]

[0056] Among them, ω n represents the weight coefficient, N represents the number of scales, F n (x, y) represents the guided filtering kernel function of the nth scale;

[0057] To improve the visual effect of the image, color restoration and normalization processing of the image are required. A color restoration factor C is introduced to reduce color deviation. The calculation formula of the restoration factor C is as follows:

[0058]

[0059] where C i (x, y) is the color restoration function of the i-th color channel (R, G, B), which is used to adjust the proportional relationship of the color channels; I i (x, y) represents the original image of the i-th color channel; a is the adjustment factor and b is the gain constant. By introducing the color restoration factor, the multi-scale Retinex algorithm can effectively reduce the color distortion problem that may occur after image enhancement, making the image color more natural.

[0060] The final enhanced image can be calculated by the following formula:

[0061]

[0062] where represents the image of the i-th color channel after enhancement, and R i (x, y) represents the reflection image of the i-th color channel.

[0063] Finally, the enhanced image is normalized so that its pixel value range is within [0, 255].

[0064] To obtain a reflection image with higher quality, the adaptive histogram equalization method with contrast limitation is used to further stretch the contrast. The detailed process of this method is as follows:

[0065] The input image is segmented into multiple non-overlapping small blocks, and the size of the small blocks is preset in advance;

[0066] For each small block, the histogram of the gray value is calculated separately, and the histogram of each small block is cropped. If the number of pixels at a certain gray level exceeds the contrast limitation threshold, the excess part is cropped, and the number of pixels in the excess part is evenly distributed to other gray levels, thereby limiting the excessive enhancement of the contrast;

[0067] The cropped histogram is equalized so that the pixel distribution within each small block is more uniform;

[0068] The processed small blocks are merged into a complete image by interpolation method to avoid the splicing traces between small blocks;

[0069] The finally obtained image has a significant enhancement in local contrast, especially able to enhance the details of the dark and bright parts.

[0070] To improve the diversity of the dataset and the generalization ability of the model, the present invention performs data augmentation on the preprocessed images in different ways:

[0071] Rotation: The image is rotated by ±5° to simulate the surface defects of auto parts at different angles.

[0072] Horizontal and vertical flipping: The image is randomly flipped horizontally and vertically to simulate the uncertain placement direction of auto parts in the actual detection process.

[0073] Brightness adjustment: The brightness of the image is randomly adjusted within the range of ±5% to simulate the detection environment under different lighting conditions.

[0074] Example 2. This application proposes to use the superior performance YOLOv11 instance segmentation model to achieve intelligent detection of surface defects of auto parts. The model is trained with a large dataset of labeled images of surface defects of auto parts, enabling it to accurately identify and locate various defects on the surface of auto parts. The detailed structure of the model is as Figure 2 shown. The YOLOv11 instance segmentation model has made significant improvements in architecture and training methods based on the previous YOLO versions. It integrates improved model structure design, enhanced feature extraction technology, and optimized training methods. YOLOv11 includes an improved backbone network, neck network, and head network.

[0075] The Backbone enhances the model's feature extraction ability through CBS, C3K2, SPPF, and C2PSA modules. Among them, both the CBS and C3K2 modules adopt a bottleneck structure;

[0076] The bottleneck structure consists of three convolutional layers. Through the first convolutional layer, the number of channels of the input feature map is first halved, the second convolutional layer is used to extract features, and the last convolutional layer restores the number of channels and adds it to the original input feature map for output;

[0077] The convolutional layer includes convolutional operations, batch normalization, and activation functions;

[0078] The C3K2 module is the core of the backbone network. It optimizes the information flow in the network by splitting the feature map and applying a series of small kernel convolutions, which is faster and has lower computational cost than large kernel convolutions. By processing small independent feature maps and merging them after several convolutions, the C3K2 module uses fewer parameters to improve the feature representation compared with the C2f module of YOLOv8;

[0079] The SPFF module performs max - pooling operations on the input feature map through multiple pooling kernels of different sizes, thereby extracting feature information of different scales. By fusing multi - scale features, the SPFF module can better capture the detailed features of the surface image of small automotive parts. Compared with the traditional Spatial Pyramid Pooling (SPP), SPFF reduces the computational cost through optimizing the pooling operation while maintaining a high detection accuracy;

[0080] The C2PSA module uses two PSA (Partial Spatial Attention) modules, which operate on different branches of the feature map and then are connected. By applying spatial attention on the extracted feature map, it refines the model's ability to selectively focus on regions of interest. This makes YOLOv11 superior to previous versions of YOLO in scenarios of accurately detecting surface defects of automotive parts;

[0081] Neck is located between the backbone network and the head network, and applies the Upsample module and the Contact module to achieve feature fusion and enhancement;

[0082] The Upsample module is an up - sampling module, and the up - sampling method used is nearest - neighbor interpolation;

[0083] The Contact module represents a concatenation operation, which concatenates features of different scales along the channel dimension;

[0084] Head is the decision - making part of the model, responsible for generating the final detection results; the head network receives multi - scale feature maps from the neck network and separately processes bounding - box regression and class prediction through decoupled detection heads; the bounding - box regression prediction branch processes features using the CBS module, and the class prediction branch processes features using the DSC module based on depth - wise separable convolution to reduce the computational cost and improve efficiency.

[0085] In this embodiment, using the YOLOv11 instance segmentation model to process the pre - processed image, the steps of identifying and locating the defect area of the image are as follows:

[0086] Adjust the pre - processed surface image of the automotive part into an input image with a size of 3×640×640 (image size is 640×640, number of channels is 3) to meet the size requirements of the input features of the YOLOv11 instance segmentation model, and perform normalization processing;

[0087] The surface image of the automotive parts is first fed into the Backbone network to extract high-level feature representations. In the Backbone network, the image first passes through 2 CBS modules to remove the noise interference therein; then, the image undergoes a combined processing of multiple C3K2 modules and CBS modules to extract low-level features such as edges, textures, and color gradients, and further extract features such as the shape, corners, and line segments of the image to obtain high-level feature representations; the C3K2 module uses a smaller convolutional kernel to extract independent feature maps and merge them, which can expand the receptive field area of the convolution to obtain the global relationship of the features; then, the high-level features are fed into the SPFF module for multi-scale max pooling operation, which can obtain richer multi-scale feature information while reducing the computational amount; finally, the C2PSA module applies spatial attention to enhance the attention degree to the key information of the features and strengthen the feature extraction ability of the model for detailed information;

[0088] Three feature maps with different sizes in the middle layer of the Backbone network are fed into the Neck network. Through a series of upsampling and splicing operations, combined with the CBS module and the C3K2 module to fuse feature maps of different scales, the high-level features such as the category and defect location of the defective image can be fully obtained;

[0089] The three Head networks connected by the Neck network are responsible for outputting prediction results according to the obtained feature information. The specific prediction process includes:

[0090] The high-level features are divided into several network units, and each grid unit predicts multiple candidate boxes; perform bounding box prediction and category prediction for each candidate box, and perform mask prediction for each detected defect area; the bounding box is a rectangular box used to frame the defects in the image, and the bounding box prediction means that the model will learn how to adjust the position, size, and confidence of the candidate box to generate the final bounding box to more accurately locate the defects in the image.

[0091] In this embodiment, YOLOv11 uses the MPDIoU loss function instead of the traditional IoU variant as the regression prediction loss of the bounding box to optimize the bounding box localization accuracy. The formula is as follows:

[0092]

[0093] Among them, A∩B is the intersection area of the predicted box and the ground truth box; A∪B is the union area of the predicted box and the ground truth box; d1 is the Euclidean distance between the upper left corner points of the predicted box and the ground truth box; d2 is the Euclidean distance between the lower right corner points of the predicted box and the ground truth box; w, h are the width and height of the image;

[0094] The DSC module of the Head network uses depthwise separable convolution to perform channel-wise convolution and pointwise convolution on images. Only performing convolution operations on some channels can reduce the number of parameters and achieve cross-channel communication of feature information;

[0095] Finally, non-maximum suppression is performed on the candidate boxes predicted by the model to remove redundancy, and the bounding boxes, categories, and instance masks of each defect are obtained. Analyze the prediction results of the model to determine the defect areas on the surface of auto parts and generate a quality inspection report.

[0096] The above specific implementation manners do not constitute a limitation on the protection scope of this application. Those skilled in the art should understand that various modifications, combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principle of this application shall be included within the protection scope of this application. In some cases, the actions or steps recorded in this application can be executed in a different order from that in the embodiments and still achieve the desired results. Additionally, the processes depicted in the drawings do not necessarily require the specific order or sequential order shown to achieve the desired results. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

Claims

1. An intelligent quality inspection method for automotive parts based on machine vision, characterized in that, The method includes: S1: Collect the surface image of the auto parts using an image acquisition system; S2: Preprocess the surface image of the auto parts using the multi-scale Retinex algorithm based on guided filtering to improve the image quality; S3: Use the YOLOv11 instance segmentation model to intelligently detect the preprocessed surface image of the auto parts, and identify and locate the defect areas in the image; S4: Obtain the defect detection result of the surface of the auto parts and generate a quality inspection report.

2. The intelligent quality inspection method for automotive parts based on machine vision according to claim 1, characterized in that The image acquisition system includes: Industrial camera: Used to capture the surface image of the auto parts, placed at the center of the light source device; Light source device: Used to provide stable lighting conditions to ensure the clarity and consistency of image acquisition; The type and intensity of the light source are adjusted according to the detection requirements; Image acquisition card: Used to receive the image data collected by the camera and transmit it to the computer system; Bracket and fixing device: Used to fix the camera and the light source to ensure the stability and accuracy of their positions.

3. The intelligent quality inspection method for automotive parts based on machine vision according to claim 1, characterized in that Based on the traditional multi-scale Retinex image enhancement method, the multi-scale Retinex algorithm based on guided filtering introduces a guided filtering kernel function to replace the traditional Gaussian kernel function; And further stretch the contrast through contrast-limited adaptive histogram equalization to provide more effective surface defect features of auto parts for the detection model.

4. The intelligent quality inspection method for automotive parts based on machine vision according to claim 3, characterized in that, The steps of the multi-scale Retinex algorithm based on guided filtering are as follows: The Retinex theory believes that an image is composed of an illumination image and a reflection image; The illumination image refers to the information of the incident component of an object, represented by L(x,y); The reflection image refers to the reflected part of an object, represented by R(x,y); The original image is represented by I(x,y), and the formula is expressed as: I(x,y) = L(x,y) * R(x,y); First, calculate and extract the illumination vector of the image: L(x,y) = I(x,y) * F(x,y), where F(x,y) is the guided filtering kernel function; Based on the illumination component, calculate the reflection component of the image: R(x,y) = logI(x,y) - logL(x,y); The multi-scale Retinex algorithm sums the Retinex results of multiple scales weighted, and the formula is: Among them, ω n represents the weight coefficient, N represents the number of scales, and F n (x, y) represents the guiding filter kernel function of the nth scale; the multi-scale reflection image of the image is obtained based on the multi-scale Retinex algorithm formula; Introduce a color restoration factor C to perform color restoration on the image. The calculation formula of the restoration factor C is as follows: Among them, C i (x, y) is the color restoration function of the i-th color channel (R, G, B), which is used to adjust the proportional relationship of the color channels; I i (x, y) represents the original image of the i-th color channel; a is the adjustment factor and b is the gain constant; The final enhanced image is calculated by the following formula: Among them, represents the image of the i-th color channel after enhancement, and R i (x, y) represents the reflected image of the i-th color channel; Finally, normalize the enhanced image so that its pixel value range is within [0, 255].

5. The intelligent quality inspection method for automotive parts based on machine vision according to claim 3, wherein, The steps of the contrast-limited adaptive histogram equalization are as follows: Divide the input image into multiple non-overlapping small blocks, and the size of the small blocks is preset in advance; Calculate the histogram of the gray values for each small block separately, and crop the histogram of each small block; If the number of pixels at a certain gray level exceeds the contrast limit threshold, then crop the excess part, and evenly distribute the excess number of pixels to other gray levels; Equalize the cropped histogram; Merge the processed small blocks into a complete image through an interpolation method.

6. The intelligent quality inspection method for automotive parts based on machine vision according to claim 1, characterized in that, The YOLOv11 instance segmentation model includes a backbone network, a neck network, and a head network.

7. The intelligent quality inspection method for automotive parts based on machine vision according to claim 6, characterized in that, The detailed structure of the backbone network of the YOLOv11 instance segmentation model is as follows: The backbone network enhances the model's feature extraction ability through CBS, C3K2, SPPF, and C2PSA modules; among them, both the CBS and C3K2 modules adopt a bottleneck structure; The bottleneck structure consists of three convolutional layers. First, the number of channels of the input feature map is halved through the first convolutional layer, features are extracted using the second convolutional layer, and the number of channels is restored in the last convolutional layer and added to the original input feature map for output; The convolutional layer includes convolutional operations, batch normalization, and activation functions; The C3K2 module is the core of the backbone network, which optimizes the information flow in the network by splitting the feature map and applying a series of small kernel convolutions; The SPFF module performs max-pooling operations on the input feature map through multiple pooling kernels of different sizes to extract feature information of different scales; The C2PSA module uses two partial spatial attention modules, which operate on different branches of the feature map and are then connected; the ability of the model to selectively focus on regions of interest is refined by applying spatial attention to the extracted feature map.

8. The intelligent quality inspection method for automotive parts based on machine vision according to claim 6, wherein The detailed structure of the neck network of the YOLOv11 instance segmentation model is as follows: The neck network is located between the backbone network and the head network, and uses the Upsample module and the Contact module to achieve feature fusion and enhancement; The Upsample module is an upsampling module, and the upsampling method used is nearest neighbor interpolation; The Contact module represents a concatenation operation, which concatenates features of different scales along the channel dimension.

9. The intelligent quality inspection method for automotive parts based on machine vision according to claim 6, characterized in that, The detailed structure of the head network of the YOLOv11 instance segmentation model is as follows: The head network is the decision-making part of the model, responsible for generating the final detection results; the head network receives multi-scale feature maps from the neck network and processes bounding box regression and class prediction separately through decoupled detection heads; the bounding box regression prediction branch uses the CBS module to process features, and the class prediction branch uses the DSC module based on depthwise separable convolution to process features.

10. The intelligent quality inspection method for automotive parts based on machine vision according to claim 6, characterized in that, For the YOLOv11 instance segmentation model, data augmentation is adopted during the training of the network to improve the diversity of data samples and enhance the generalization ability of the model. The data augmentation methods include: rotation, horizontal and vertical flipping 、 Brightness adjustment.

Images