A method for detecting the stamping quality of a bumper fixing bracket

Abstract

The present invention relates to the field of image data processing technology, and more specifically, to a method for detecting the stamping quality of a bumper fixing bracket, comprising: obtaining a surface image of the bumper fixing bracket; determining the suspected noise level of each pixel point; determining the position sensitivity of each pixel point; determining the window requirement level of each pixel point; determining the adaptive window side length of each pixel point; and filtering the surface image using a guided filtering algorithm based on the adaptive window side length to achieve stamping quality detection of the bumper fixing bracket. The present invention improves the accuracy of the guided filtering algorithm by dynamically adjusting the filter window size of each target pixel point, avoids defects such as over-smoothing or insufficient noise suppression caused by a fixed window, ensures that the details of the burr defect are retained, and reduces the interference between noise and real defects while denoising, thereby improving the accuracy of subsequent quality inspections.

Description

Technical Field

[0001] The present invention relates to the technical field of image data processing, and more particularly to a method for detecting the stamping quality of a bumper fixing bracket. Background Art

[0002] In the automotive manufacturing industry, especially in the production of body parts, the bumper bracket is one of the most important body components, and its production quality plays a vital role in the safety and appearance of the entire vehicle. Burr defects are an inevitable and common problem during the stamping process of the bumper bracket. These burrs not only affect the functionality of the bracket, but may also have an impact on subsequent assembly and use. However, noise in the image of the bumper bracket can interfere with the identification of potential burr defects, resulting in reduced accuracy in subsequent defect detection. The guided filtering algorithm, as a boundary-preserving filtering method, can effectively smooth the image while retaining details and boundary information, improving the accuracy and reliability of burr defect detection in the bumper bracket and reducing false positives and missed negatives.

[0003] However, when using the guided filtering algorithm to denoise the surface image of the bumper bracket, a fixed-size window is selected for each pixel in the surface image. The size of the window controls the smoothness of the guided filtering algorithm when filtering this pixel. When the window is larger, the filtering is smoother but the boundary details may be more blurred. When the window is smaller, the boundary details can be better preserved, but the denoising effect is relatively poor. Therefore, if a larger window is selected for the pixel points on the bumper fixing bracket that are potentially located at the burr defect, the filtered image will lose the detailed characteristics of the burr defect, and all potential burr defects in the image cannot be accurately detected, resulting in missed detections in subsequent defect detections; and if a smaller window is selected, the noise in the image will be insufficiently suppressed, so that the noise is not effectively filtered out, and the residual noise will be confused with the characteristics of the burr defect, making it difficult to accurately distinguish between real burr defects and noise interference in the subsequent defect detection process, resulting in missed reports (such as missing the real burr defect due to noise masking) or false alarms (such as misjudging noise as burrs), thereby affecting the accuracy and reliability of quality inspection of the bumper fixing bracket stamping. Summary of the Invention

[0004] In order to solve the problem that when a surface image of a bumper fixing bracket is subjected to denoising using a guided filtering algorithm, a fixed-size window is used. When the window is large, burr details are easily lost, resulting in missed detections. When the window is small, denoising is incomplete, noise is confused with defect features, and missed or false positives are caused, thereby affecting the accuracy and reliability of the final quality inspection results, the present invention proposes a method for inspecting the stamping quality of a bumper fixing bracket. The method comprises the following steps:

[0005] A surface image of a bumper mounting bracket is obtained; any pixel in the surface image is marked as a target pixel, defective pixels in the target pixel's neighborhood are obtained, and the suspected noise level of the target pixel is determined based on the sum of the grayscale value differences between the target pixel and its neighborhood pixels and the number of defective pixels; the position sensitivity of the target pixel's region is determined based on the standard deviation of the gradient amplitudes of all pixels with gradient changes within the target pixel's region and the distance between the target pixel and pixels located in the target pixel's basic direction and at the boundary of the surface image; the window requirement of the target pixel is determined based on the suspected noise level and the position sensitivity, as well as a preset value adjustment coefficient; the initial window side length in the guided filtering algorithm is weighted based on the window requirement to obtain an adaptive window side length for the target pixel; based on the adaptive window side length, the surface image is filtered using the guided filtering algorithm, and then defect detection is performed on the filtered surface image of the bumper mounting bracket using a deep learning algorithm to complete the stamping quality inspection of the bumper mounting bracket.

[0006] The present invention realizes the adaptive configuration of the filter window size by dynamically adjusting the window side length of the guided filtering algorithm according to the suspected noise level and position sensitivity of the target pixel point, so that the guided filtering algorithm can not only effectively suppress noise and improve the denoising effect, but also maintain detail information at the boundary of the burr defect to avoid blurring of details; dynamic adjustment of the window size can accurately distinguish between real burr defects and image noise, reduce false alarms caused by noise interference, and avoid the situation where defect details are lost and missed detection due to large windows, thereby improving the accuracy and reliability of defect detection; the use of adaptive guided filtering combined with deep learning for defect detection fully retains the defect characteristics and removes noise, greatly improving the recognition ability of the stamped surface defects of the bumper fixing bracket and the overall quality of detection.

[0007] Furthermore, the surface image is obtained by graying the collected original image of the bumper fixing bracket, and then using a semantic segmentation model to extract the grayed original image of the bumper fixing bracket to obtain a surface image of the bumper fixing bracket.

[0008] Furthermore, the suspected noise level:

[0009] Where, For the The suspected noise level of each pixel, For the The sum of the differences in the grayscale values ​​of a pixel and all its neighboring pixels, For the The number of defective pixels in the neighborhood of a pixel, For the The difference between the gray value of a pixel and the maximum gray value of all pixels in the surface image, For the The difference between the gray value of a pixel and the minimum gray value of all pixels in the surface image, is a hyperparameter, is the standard normalization function, is the minimum function.

[0010] By introducing the sum of the grayscale differences between a pixel and its neighbors, the number of defective pixels in the neighborhood, and the combined calculation of the grayscale differences between the pixel and the maximum and minimum grayscale values ​​of the entire image, the present invention can more comprehensively and accurately reflect the noise characteristics of the pixel and scientifically measure the degree of suspected noise. By using normalization processing and the introduction of a minimum value function, the instability of the calculation caused by zero or extremely small grayscale differences is avoided, making the assessment of the suspected noise level more accurate, effectively distinguishing between real defects and noise interference, and improving the reliability of noise detection.

[0011] Furthermore, the defective pixel points are obtained by recording the neighboring pixel points of the target pixel point whose grayscale difference with the target pixel point is greater than a preset defect threshold as defective pixel points.

[0012] Furthermore, the basic directions include horizontally left, horizontally right, vertically upward, and vertically downward.

[0013] Furthermore, the position sensitivity satisfies:

[0014] Where, For the The position sensitivity of the area to which the pixel belongs, No. The standard deviation of the gradient amplitude of all pixels with gradient changes in the area to which the pixel belongs, Respectively The distance between the pixel point and the pixel point located in the horizontal left, horizontal right, vertical upward, and vertical downward directions of the target pixel point and at the boundary of the surface image, is the natural exponential function, is the minimum function.

[0015] The present invention quantifies the position sensitivity of the target pixel in the spatial structure by combining the standard deviation of the gradient amplitude of the target pixel and the minimum value of the distance from the pixel to the image boundary in multiple directions, which can accurately reflect the importance of the pixel at the image boundary and gradient change. It uses the exponential decay function to combine the gradient change and the boundary distance to highlight the areas near the image boundary with large gradient changes, effectively improving the focus on areas with rich details or complex structures, and preventing key areas from being blurred during filtering.

[0016] Furthermore, the distance is Euclidean distance.

[0017] Furthermore, the window requirement degree satisfies:

[0018] Where, For the The degree of window requirement of pixels, For the The suspected noise level of each pixel, For the The position sensitivity of each pixel, is the preset value adjustment coefficient, is the standard normalization function.

[0019] The window requirement degree of the present invention reflects the comprehensive demand of the pixel points for the filter window size by combining the suspected noise level and the position sensitivity, and realizes the reasonable trade-off of multi-dimensional information; the normalization function and the value adjustment coefficient are used to normalize and fine-tune the results, so that the filter window size can be adaptively changed. It can not only increase the window in the area with high noise and insensitive to details to achieve effective smoothing, but also reasonably reduce the window in the boundary sensitive area to protect the image details.

[0020] Furthermore, the adaptive window side length satisfies:

[0021] Where, For the The adaptive window side length of pixels, is the initial window length in the guided filtering algorithm, For the The degree of window requirement of pixels, is a combinatorial function used to round non-odd numbers up to the nearest odd number.

[0022] The present invention rounds up the calculated adaptive window side length to the nearest odd number to ensure that the window size meets the requirements of the guided filtering algorithm for odd window side lengths, which helps to achieve symmetry and algorithm stability in the filtering process; combines the initial window side length with the window demand degree, and dynamically adjusts the filter window size corresponding to each pixel point, making the filtering operation more flexible and personalized, meeting the different noise suppression and detail protection requirements of different pixel points; adaptively adjusting the window side length can effectively avoid the problems of excessive smoothing or missing details caused by a fixed window, enhance the protection of burr defect details in the filtering process, and improve the noise suppression effect.

[0023] Furthermore, the deep learning algorithm adopts a convolutional neural network model.

[0024] The present invention has the following beneficial effects:

[0025] (1) The present invention can balance the requirements of denoising and boundary preservation according to the suspected noise level and position sensitivity of the image by dynamically adjusting the filter window size of each target pixel, thereby improving the accuracy of the guided filtering algorithm and avoiding the defects of excessive smoothing or insufficient noise suppression caused by a fixed window; a larger window can smooth the image but easily blurs the details, especially small defects such as burrs, while a smaller window helps to retain details but has a relatively poor denoising effect. The present invention ensures that the details of the burr defect are retained through adaptive window optimization, and reduces the interference between noise and real defects while denoising, thereby improving the accuracy of burr defect detection.

[0026] (2) The present invention comprehensively considers the suspected noise level, gradient change and position sensitivity of the target pixel points, adaptively adjusts the filter window, effectively enhances the denoising effect of the image, and ensures that the boundary features and burr defects are not weakened, avoids false alarms and missed alarms, and improves the accuracy of subsequent defect detection; through the deep learning algorithm combined with the high-quality image after guided filtering for defect detection, the accuracy of the detection results can be effectively improved, the missed detection and false alarm caused by noise masking or erroneous judgment can be reduced, and the reliability and stability of stamping quality detection can be improved.

[0027] (3) The present invention, through optimized image processing technology, can provide more accurate input images for subsequent defect detection, improve the efficiency of quality inspection, help reduce manual intervention, ensure the quality stability of automotive parts, and improve production efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 The present invention is a flowchart of a method for inspecting the quality of a bumper fixing bracket by stamping. DETAILED DESCRIPTION

[0029] The following is a clear and complete description of the technical solutions in the embodiments of the present invention. The described embodiments are part of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without making creative efforts are within the scope of protection of the present invention.

[0030] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0031] See also Figure 1 , which shows a flowchart of a method for inspecting the quality of a bumper fixing bracket stamping provided by one embodiment of the present invention, the method comprising the following steps:

[0032] S01: Acquire a surface image of the bumper fixing bracket.

[0033] It should be noted that the present invention utilizes a high-definition camera to capture the original image of the bumper fixing bracket.

[0034] Specifically, the surface image is obtained in the following manner:

[0035] The collected original image of the bumper fixing bracket is grayscaled, and then the semantic segmentation model is used to extract the grayscaled original image of the bumper fixing bracket to obtain a surface image of the bumper fixing bracket.

[0036] Implementers can select a semantic segmentation model based on specific implementation circumstances, for example, the FCN semantic segmentation model.

[0037] S02: Determine the suspected noise level of each pixel.

[0038] It should be noted that the more extreme the relative magnitude of each pixel's grayscale value in the image, the more likely it is to be affected by noise, and the higher the corresponding noise suspicion level. However, since bumper brackets are mostly black, the grayscale values ​​of their pixels are also very low, and noise pixels are usually random and sporadic. Therefore, the greater the difference in grayscale value between a pixel and its neighboring pixels, the more likely it is to be affected by noise, and the higher the corresponding noise suspicion level. However, there may be potential burr defects on the surface of the bumper fixing bracket, and burr defects are usually in the shape of slender spikes. Therefore, some pixels in the potential burr defect area may also have relatively extreme grayscale values ​​and a large grayscale value difference with their neighboring pixels. However, the pixels in the potential burr defect area usually have pixels with similar grayscale values ​​in a specific neighborhood direction. Therefore, in order to eliminate the influence of the pixels in the potential burr defect area on the quantitative noise suspicion level of a pixel point, that is, to distinguish between noise pixels and pixels in the potential burr defect area, the more pixels have grayscale differences between a pixel point and its neighboring pixels that exceed the defect threshold, the more likely this pixel point is to be affected by noise, the less likely it is to belong to the potential burr defect area, and the corresponding suspected noise level is greater.

[0039] Any pixel point in the surface image is recorded as the target pixel point, and the defective pixels in the neighborhood pixel points of the target pixel point are obtained. The suspected noise degree of the target pixel point is determined according to the sum of the differences in the grayscale values ​​of the target pixel point and the neighborhood pixel points of the target pixel point, as well as the number of the defective pixels.

[0040] Implementers can set the number of neighborhood pixels according to specific implementation circumstances, for example, 8 neighborhood pixels; if a pixel is located at the boundary of the surface image, the neighboring pixels of the pixel that exceed the image boundary are obtained by mirror reflection filling.

[0041] Specifically, the defective pixel points are obtained as follows:

[0042] Among the neighborhood pixels of the target pixel, the neighborhood pixels whose grayscale difference with the target pixel is greater than a preset defect threshold are recorded as defective pixels.

[0043] Implementers can set the defect threshold to, for example, 10, based on specific implementation circumstances.

[0044] Specifically, the suspected noise level:

[0045] ;

[0046] Where, For the The suspected noise level of each pixel, For the The sum of the differences in the grayscale values ​​of a pixel and all its neighboring pixels, For the The number of defective pixels in the neighborhood of a pixel, For the The difference between the gray value of a pixel and the maximum gray value of all pixels in the surface image, For the The difference between the gray value of a pixel and the minimum gray value of all pixels in the surface image, is a hyperparameter, is the standard normalization function, is the minimum function.

[0047] Implementers can set hyperparameters according to specific implementation conditions, for example, 0.001. The existence of hyperparameters is to prevent , which makes the formula meaningless.

[0048] in, The smaller the The more extreme the relative size of the grayscale value of a pixel in the surface image is, the more likely it is to be affected by noise, and the greater the corresponding suspected noise level; and vice versa. The larger the The more extreme the relative size of the grayscale value of a pixel in the surface image is, the greater its credibility is, indicating that it is more likely to be affected by noise and the corresponding suspected noise level is greater; vice versa. The larger the The greater the possibility that a pixel is affected by noise, the greater its credibility is, the smaller the possibility that it belongs to the potential burr defect area is, and the greater the corresponding suspected noise level is; vice versa.

[0049] S03: Determine the position sensitivity of each pixel.

[0050] It should be noted that since the bumper fixing bracket may have burr defects, in order to prevent the guided filtering algorithm from over-smoothing the potential burr defects in the surface image, thereby affecting the final defect detection results. Therefore, this step will continue to analyze the texture features of the area to which each pixel belongs, and calculate the position sensitivity of the area to which each pixel belongs. This indicator is used to reflect the possibility of the existence of burr defects at the position of the pixel. Afterwards, based on the position sensitivity and suspected noise level corresponding to each pixel, the window requirement of each pixel is calculated, which is used to calculate the window side length of each pixel in the subsequent steps; then, when analyzing and calculating the position sensitivity of each pixel, since burr defects usually exist at the outer contour edge of the bracket, the pixels in the outer contour edge area should retain as many detailed features as possible. For these pixels, a smaller window is required for filtering, so the window requirement will also be smaller.

[0051] The position sensitivity of the area to which the target pixel belongs is determined based on the standard deviation of the gradient amplitudes of all pixels with gradient changes in the area to which the target pixel belongs, and the distance between the target pixel and the pixels located in the basic direction of the target pixel and at the boundary of the surface image.

[0052] Implementers can set the size of the area according to the specific implementation situation, for example, the area of ​​7×7 pixels centered on the target pixel; if a pixel is located at or near the boundary of the surface image, the part of the area to which the pixel belongs that exceeds the image boundary is obtained by mirror reflection filling.

[0053] Specifically, the basic directions include horizontally left, horizontally right, vertically upward, and vertically downward.

[0054] Specifically, the distance is the Euclidean distance.

[0055] Specifically, the position sensitivity satisfies:

[0056] ;

[0057] Where, For the The position sensitivity of the area to which the pixel belongs, No. The standard deviation of the gradient amplitude of all pixels with gradient changes in the area to which the pixel belongs (if there is no pixel with gradient change in the area to which a pixel belongs, then let , that is, not considering the existence of this indicator), Respectively The distance between the pixel point and the pixel point located in the horizontal left, horizontal right, vertical upward, and vertical downward directions of the target pixel point and at the boundary of the surface image, is the natural exponential function, is the minimum function.

[0058] in, The smaller the The greater the uniformity of the gradient amplitude of the pixel points with gradient changes in the area to which the pixel points belong, the greater the possibility that the area to which the pixel points belong is in the outer contour boundary area, the greater the possibility that there are potential burr defects in the area to which the pixel points belong, and the greater the sensitivity of the corresponding position; vice versa. The smaller the The greater the possibility that the area to which the pixel belongs is in the outer contour boundary area, the greater the credibility, the greater the possibility that there is a potential burr defect in the area to which the pixel belongs, and the greater the sensitivity of the corresponding position; vice versa.

[0059] S04: Determine the window requirement of each pixel.

[0060] It should be noted that the window requirement degree is a key indicator for balancing the image denoising effect and the retention of burr defect details, in order to overcome the problems of missed reports and false alarms caused by the fixed window size of traditional guided filtering. Therefore, this step constructs a dynamic window requirement evaluation mechanism by fusing the suspected noise degree and position sensitivity of the pixel points. The suspected noise degree is used to quantify the possibility of the pixel point being interfered by noise, while the position sensitivity reflects the probability of the existence of burr defects in the area where the pixel point is located. By combining the two indicators, it is possible to use a large window to enhance the denoising effect in high-noise areas, and to adaptively reduce the window in potential burr defect areas to retain details, thereby achieving the optimal balance between denoising and feature protection, and providing an accurate basis for the dynamic adjustment of the subsequent guided filter window size.

[0061] The window requirement of the target pixel is determined according to the suspected noise level, the position sensitivity level, and a preset value adjustment coefficient.

[0062] Specifically, the window requirement degree satisfies:

[0063] ;

[0064] Where, For the The degree of window requirement of pixels, For the The suspected noise level of each pixel, For the The position sensitivity of each pixel, is the preset value adjustment coefficient, is the standard normalization function.

[0065] Implementers can set the value adjustment coefficient according to the specific implementation situation, for example,

[0066] in, The larger the The greater the possibility that a pixel is affected by noise, the larger the window should be set to filter the pixel to achieve a greater degree of smoothing, so the larger the window requirement for the pixel. The larger the The greater the possibility that the area to which a pixel belongs is located at the edge of the outer contour, the greater the possibility that there is a potential burr defect in the area to which the pixel belongs. Therefore, a smaller window is required for filtering the area to retain more detail information, so the window requirement for the pixel will also be smaller.

[0067] S05: Determine the side length of the adaptive window for each pixel.

[0068] It should be noted that the greater the window requirement for each pixel, the more likely this pixel is to be affected by noise, and the more likely it is to be located in an area with simple texture and less prone to burrs. In this case, a larger window should be set for this pixel to perform a larger filtering to improve the denoising effect. The smaller the window requirement for each pixel, the less likely this pixel is to be affected by noise, and the more likely it is to be located in an area with complex texture and prone to burrs. In this case, a smaller window should be set for this pixel to retain details as much as possible.

[0069] According to the degree of window demand, the initial window side length in the guided filtering algorithm is weighted to obtain the adaptive window side length of the target pixel point.

[0070] Specifically, the adaptive window side length satisfies:

[0071] ;

[0072] Where, For the The adaptive window side length of pixels, is the initial window length in the guided filtering algorithm, For the The degree of window requirement of pixels, is a combinatorial function used to round non-odd numbers up to the nearest odd number.

[0073] Implementers can set the initial window side length in the guided filtering algorithm according to specific implementation conditions, for example, 7.

[0074] S06: Based on the adaptive window side length, a guided filtering algorithm is used to filter the surface image to achieve stamping quality inspection of the bumper fixing bracket.

[0075] It should be noted that after completing the adaptive calculation of the window size for each pixel point in the surface image, the optimized guided filtering algorithm will be used in this step to filter the surface image of the bumper fixing bracket, and then the potential burr defects on the bracket surface will be detected based on the filtered surface image of the bumper fixing bracket.

[0076] Based on the adaptive window side length, the surface image is filtered using a guided filtering algorithm, and then defect detection is performed on the filtered surface image of the bumper fixing bracket through a deep learning algorithm to complete the stamping quality inspection of the bumper fixing bracket.

[0077] Implementers can set the regularization parameter in the guided filtering algorithm according to specific implementation conditions, for example, 0.05.

[0078] Specifically, the deep learning algorithm adopts a convolutional neural network model (CNN).

[0079] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for detecting the quality of a bumper fixing bracket by stamping, characterized in that: include: Acquire a surface image of the bumper fixing bracket; Any pixel in the surface image is recorded as a target pixel, and the neighboring pixels of the target pixel whose grayscale difference with the target pixel is greater than a preset defect threshold are recorded as defective pixels. The defective pixels in the neighboring pixels of the target pixel are obtained, and the suspected noise level of the target pixel is determined based on the sum of the grayscale value differences between the target pixel and the neighboring pixels of the target pixel and the number of defective pixels; suspected noise level: ； Where, For the The suspected noise level of each pixel, For the The sum of the differences in the grayscale values ​​of a pixel and all its neighboring pixels, For the The number of defective pixels in the neighborhood of a pixel, For the The difference between the gray value of a pixel and the maximum gray value of all pixels in the surface image, For the The difference between the gray value of a pixel and the minimum gray value of all pixels in the surface image, is a hyperparameter, is the standard normalization function, is the minimum function; Determine the position sensitivity of the region to which the target pixel belongs based on the standard deviation of the gradient amplitudes of all pixels with gradient changes in the region to which the target pixel belongs, and the distance between the target pixel and pixels located in the basic direction of the target pixel and at the boundary of the surface image; The basic directions include horizontal left, horizontal right, vertical upward, and vertical downward; the area is constructed with the target pixel as the center; If a pixel is located at or near the boundary of the surface image, the portion of the area to which the pixel belongs that exceeds the image boundary is obtained by mirror reflection filling; Determine the window requirement for the target pixel based on the suspected noise level, position sensitivity, and preset value adjustment coefficients; According to the degree of window demand, the initial window side length in the guided filtering algorithm is weighted to obtain the adaptive window side length of the target pixel; Based on the adaptive window side length, the guided filtering algorithm is used to filter the surface image. Then, the deep learning algorithm is used to detect defects on the filtered surface image of the bumper bracket to complete the stamping quality inspection of the bumper bracket.

2. A method for inspecting the quality of a bumper fixing bracket by stamping according to claim 1, characterized in that: The surface image is obtained as follows: The collected original image of the bumper fixing bracket is grayscaled, and then the semantic segmentation model is used to extract the grayscaled original image of the bumper fixing bracket to obtain a surface image of the bumper fixing bracket.

3. A method for inspecting the quality of a bumper fixing bracket by stamping according to claim 1, characterized in that: The basic directions include horizontally left, horizontally right, vertically upward, and vertically downward.

4. A method for inspecting the quality of a bumper fixing bracket by stamping according to claim 3, characterized in that: The position sensitivity satisfies: ； Where, For the The position sensitivity of the area to which the pixel belongs, No. The standard deviation of the gradient amplitude of all pixels with gradient changes in the area to which the pixel belongs, Respectively The distance between the pixel point and the pixel point located in the horizontal left, horizontal right, vertical upward, and vertical downward directions of the target pixel point and at the boundary of the surface image, is the natural exponential function, is the minimum function.

5. A method for inspecting the quality of stamping of a bumper fixing bracket according to claim 1 or 4, characterized in that: The distance is the Euclidean distance.

6. The method for inspecting the quality of a bumper fixing bracket by stamping according to claim 1, characterized in that: The window requirement degree meets: ； Where, For the The degree of window requirement of pixels, For the The suspected noise level of each pixel, For the The position sensitivity of each pixel, is the preset value adjustment coefficient, is the standard normalization function.

7. The method for inspecting the quality of stamping of a bumper fixing bracket according to claim 1, characterized in that: The adaptive window side length satisfies: ； Where, For the The adaptive window side length of pixels, is the initial window length in the guided filtering algorithm, For the The degree of window requirement of pixels, is a combinatorial function used to round non-odd numbers up to the nearest odd number.

8. The method for inspecting the quality of a bumper fixing bracket by stamping according to claim 1, characterized in that: The deep learning algorithm adopts a convolutional neural network model.

Images