An intelligent identification method for surface cracks in the vibration test of a radiator

Abstract

The present invention relates to the field of image processing, and particularly to an intelligent identification method for surface cracks in a radiator vibration test. The method includes: collecting real-time images of the radiator surface and preprocessing the images to obtain a grayscale image of the radiator; calculating the smoothness degree, the density of key points in the neighborhood, and the edge degree of each pixel point of the grayscale image, and taking the product of the smoothness degree, the key point density, and the edge degree as the preference degree of the pixel point; obtaining an adaptive corner threshold based on the preference degree, and performing corner detection on the grayscale image according to the adaptive corner threshold to detect all key points; based on the detected key points, using the LKT optical flow method to achieve crack detection of the radiator surface. According to the adaptive corner threshold, the present invention can obtain a sufficient number of evenly distributed corners in different regions of the radiator, thereby improving the accuracy and robustness of crack detection.

Description

Technical Field

[0001] The present invention relates to the field of image processing, and more specifically, to an intelligent surface crack recognition method in a radiator vibration test. Background Art

[0002] With the continuous progress of the automotive industry, the performance and complexity of vehicles are increasing. The engine is the "heart" of the vehicle, and its thermal management system is crucial to ensure the normal operation of the engine. The radiator, as the core component of this thermal management system, is responsible for transferring the heat generated by the engine to the coolant, and then dissipating it to the outside through the cooling system, thereby keeping the engine working within a suitable temperature range.

[0003] Since cars are subject to various road conditions and driving conditions during driving, the engine and its accessories (including the radiator) often face a complex vibration environment. This long-term, high-frequency, multi-directional mechanical stress will have a certain impact on the radiator, causing tiny cracks on its surface. Over time, the cracks will gradually expand and eventually lead to coolant leakage. Once the coolant leaks, it will not only affect the heat dissipation efficiency of the radiator, but may also cause the engine to overheat, leading to a series of more serious failures.

[0004] Existing technologies, such as the use of LKT optical flow method, perform crack detection on the surface image of the radiator. When cracks appear on the surface of the radiator, the pixels around the crack area will move relative to each other due to the generation and development of the cracks. By calculating the optical flow of these pixels, the existence and expansion of the cracks can be detected. However, since the structure of the radiator surface is relatively regular and its surface image has a strong regularity, it is difficult for the optical flow method to track and obtain a sufficient number of evenly distributed corner points as key points in the radiator surface image, resulting in poor crack detection results. Summary of the invention

[0005] In order to solve the above-mentioned technical problem that it is difficult to track and obtain a sufficient number of evenly distributed corner points as key points in the radiator surface image, resulting in poor crack detection effect, the present invention provides the following technical solution.

[0006] A method for intelligently identifying surface cracks in a radiator vibration test, comprising:

[0007] Collect the radiator surface image in real time, and pre-process the image to obtain the radiator grayscale image;

[0008] Calculate the smoothness degree of each pixel point in the grayscale image, the density of key points in the neighborhood, and the edge degree, and take the product of the smoothness degree, the density of key points in the neighborhood, and the edge degree as the preference degree of the pixel point; obtain an adaptive corner threshold based on the preference degree, and perform corner detection on the grayscale image according to the adaptive corner threshold to detect all key points;

[0009] Based on the detected key points, use the LKT optical flow method to realize crack identification on the surface of the radiator.

[0010] By calculating the smoothness degree, the density of key points in the neighborhood, and the edge degree, and taking the product of the three as the preference degree to determine the importance of pixel points, the present invention can more comprehensively evaluate the characteristics of pixel points. It can not only screen out pixel points with obvious corner features, but also ensure to a certain extent that the distribution of corners in different regions is relatively uniform. Because the three characteristics of smoothness degree, key point density, and edge degree describe the characteristics of pixel points and their surrounding regions from different angles, considering them comprehensively can better capture potential corners, rather than relying solely on a single feature, thus increasing the possibility of obtaining a sufficient number of corners;

[0011] Obtain an adaptive corner threshold based on the preference degree. Compared with a fixed threshold, the adaptive threshold can be dynamically adjusted according to the local features of the image, and thus better adapt to the characteristics of images in different regions, avoiding the omission or misjudgment of corners in some regions due to improper threshold setting, and thus more accurately detecting corners, which helps to obtain a sufficient number of corners with relatively uniform distribution;

[0012] Based on all the obtained high-quality corners, it can better reflect the actual deformation and movement of the radiator surface, thereby improving the accuracy and robustness of crack detection.

[0013] Preferably, after obtaining the grayscale image of the radiator, it further includes:

[0014] According to a preset initial corner threshold, use a corner detection algorithm to process the grayscale image to obtain all initial key points in the grayscale image.

[0015] The preset initial corner threshold may not be completely applicable to all grayscale images of radiators. By obtaining the initial key points and observing their distribution and quantity, the threshold can be verified and adjusted. If the number of initial key points is too large or too small, or the distribution is unreasonable, the threshold can be increased or decreased accordingly to obtain a more suitable corner detection result.

[0016] Preferably, the corner detection algorithm is the Shi-Tomasi corner detection algorithm.

[0017] Preferably, the process of obtaining the smoothness degree includes:

[0018] Get the gradient value and gradient direction of the pixel point in the grayscale image, calculate the average gradient value in the set neighborhood of the pixel point, normalize the average gradient value using a logarithmic function, and get the smoothness of the pixel point.

[0019] The smoothness information can help determine the distribution of corner points in different areas. In relatively smooth areas, there may be fewer corner points, while in areas with rich textures and complex edges, the corner points may be denser.

[0020] Preferably, the process of obtaining the density of key points in the neighborhood includes:

[0021] The number of pixels in the neighborhood of the pixel point that have been detected as initial key points is counted, and the ratio of the number of pixels in the neighborhood of the pixel point that have been detected as initial key points to the total number of pixels in the neighborhood of the pixel point is calculated and normalized to obtain the density of key points in the neighborhood of the pixel point.

[0022] In the corner detection algorithm, there may be some false detections or noise points being treated as corner points. By analyzing the density of key points, corner points located in high-density areas can be considered more reliable corner points because they are supported by more similar key points around them, while isolated corner points or corner points in low-density areas may be noise or false detection points.

[0023] Preferably, the process of obtaining the edge degree includes:

[0024] Based on the gradient direction, the angularly similar pixels of the pixels are screened out, the absolute difference between the gradient value of the pixel and the gradient values of the angularly similar pixels is calculated, and the absolute difference is normalized. Then, the normalized values of all angularly similar neighboring pixels are accumulated, and the accumulated results are averaged to obtain the average similarity in gradient amplitude between the pixel and its angularly similar neighboring pixels, which is used as the edge degree of the pixel.

[0025] In the corner detection algorithm, we usually hope to find those points that are both in the edge-significant area and have unique features as corner points. By calculating the edge degree of the pixel points, we can pre-screen those areas with a high probability of being in the edge, and then further search and determine the corner points in these areas. This can greatly reduce the scope of corner point search and improve the efficiency and accuracy of corner point extraction.

[0026] Preferably, the adaptive corner point threshold satisfies the relationship:

[0027] ; In the formula, is the adaptive corner point threshold corresponding to the pixel point, is the preference degree of the pixel point, is the initial corner threshold.

[0028] By introducing the preference degree of pixel points, the corresponding corner threshold can be adjusted according to the characteristics of each pixel point. For those pixel points that are more likely to become corners (i.e., pixel points with a higher preference degree), their adaptive corner threshold will be relatively small, which makes these pixel points more likely to meet the threshold condition during the corner judgment process, so as to be accurately detected as corners, reducing the misjudgment or missed judgment situations that may be caused by a unified threshold.

[0029] Preferably, the corner detection of the grayscale image according to the adaptive corner threshold includes:

[0030] Calculate the covariance matrix of the pixel values within the neighborhood range of each pixel point of the grayscale image. For the covariance matrix of each pixel point, perform eigenvalue decomposition to obtain eigenvalues and corresponding eigenvectors; record the ratio of the minimum eigenvalue of the pixel point to the adaptive corner threshold as the corner ratio; screen the pixel points whose corner ratio is greater than a preset threshold, and take the pixel point with the largest corner ratio as the key point, and recalculate the adaptive corner threshold of other pixel points within the neighborhood range, and gradually find all key points.

[0031] Preferably, the process of obtaining the smoothness degree includes:

[0032] Take the reciprocal of the maximum gradient magnitude among the neighborhood pixel points of the pixel point as the smoothness degree of the pixel point.

[0033] Preferably, the process of obtaining the key point density within the neighborhood includes:

[0034] Divide the preset neighborhood of the pixel point into multiple sub-neighborhoods. For each sub-neighborhood, count the number of pixel points that have been detected as initial key points and the total number of pixel points within the sub-neighborhood, and take the ratio of the number of pixel points that have been detected as initial key points to the total number of pixel points within the sub-neighborhood as the key point density within the sub-neighborhood, and average the key point densities within all sub-neighborhoods to obtain the key point density within the neighborhood of the pixel point.

[0035] The beneficial effects of the present invention are:

[0036] By calculating the smoothness degree, the key point density within the neighborhood, and the edge degree of each pixel point of the grayscale image, and taking the product of the three as the preference degree of the pixel point, various characteristic information of the pixel points in the image is comprehensively considered. This multi-dimensional feature fusion method can more comprehensively and accurately reflect the characteristics of the pixel points that may belong to the crack area, avoiding misjudgment caused by a single feature. Then, based on the preference degree of the pixel points, the threshold of corner detection is dynamically adjusted, so that the corner detection can better adapt to the image features of different regions, avoiding misdetection or missed detection caused by a fixed threshold, and further improving the accuracy of crack detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is the flowchart of the method from step S1 to step S3 in the intelligent identification method of surface cracks in the radiator vibration test of the embodiment of the present invention. Specific embodiments

[0038] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments.

[0039] Refer to Figure 1 , an intelligent identification method of surface cracks in a radiator vibration test includes steps S1 to S3, which are specifically as follows:

[0040] S1: Collect the surface image of the radiator, and preprocess the image to obtain the grayscale image of the radiator.

[0041] In one embodiment, place the radiator in a suitable position, and use a high-resolution industrial camera to capture the surface image of the radiator in real time. Since the captured image may be affected by environmental noise during the shooting process, it is necessary to denoise the captured surface image (such as using methods such as mean filtering and Gaussian filtering to remove noise).

[0042] Further, convert the above-mentioned captured surface image of the radiator into a grayscale image.

[0043] After obtaining the grayscale image, according to experience or pre-experimental results, set a suitable initial corner threshold, which is initially set to 0.04 in the embodiment of the present invention; then use a corner detection algorithm (such as Shi-Tomasi corner detection) to process the above-mentioned captured grayscale image to obtain the initial corners in the grayscale image, and mark them as initial key points, and record the position information (such as row number and column number) of all pixel points determined to be initial key points. The corner detection algorithm is a prior art and will not be elaborated here.

[0044] S2: Calculate the smoothness, the density of key points in the neighborhood, and the edge degree of each pixel point in the grayscale image, and take the product of the smoothness, the density of key points in the neighborhood, and the edge degree as the preference degree of the pixel point; obtain an adaptive corner threshold based on the preference degree, and perform corner detection on the grayscale image according to the adaptive corner threshold to detect all key points.

[0045] Although the initial key points in the grayscale image are obtained in S1 above, the initial corner threshold is set based on experience or preset values, and it may not be able to adapt to the characteristics of all regions in the radiator surface image. Especially in the radiator surface image, there are regions with different features such as smooth regions, textured regions, and edges. Using a fixed initial corner threshold may result in insufficient corner points in some regions (such as smooth regions), missing important feature points, while in other regions (such as regions with complex textures), there are too many corner points, causing information redundancy and difficulties in subsequent processing.

[0046] In smooth regions, the gradient change of pixel points is small. If a fixed corner threshold is used, it may be difficult to detect corner points. At the same time, in order to avoid over-concentration of corner points, it is also necessary to consider the number of existing key points in the neighborhood of each pixel point.

[0047] In one embodiment, first, for each pixel point in the grayscale image, convolution operations are respectively performed using Sobel operators in four directions ( , , , ) to obtain the gradient values of each pixel point in the four directions. Further compare the magnitudes of the gradient values in these four directions, and take the largest gradient value as the gradient value of the pixel point, and the direction corresponding to the largest gradient value as the gradient direction of the pixel point.

[0048] Next, for each pixel point in the grayscale image, all pixel points in its 9×9 neighborhood are obtained, and these pixel points are marked as neighborhood pixel points. Calculate the average gradient value within the neighborhood of the pixel point, and then normalize the average gradient value using a logarithmic function to obtain the smoothness of the pixel point. The smoothness reflects the texture change situation of the region where the pixel point is located. In smooth regions, the gradient values of pixel points change relatively smoothly, while corner points are usually located in regions with large gradient changes in the image. By calculating the smoothness, smooth regions and regions that may contain corner points can be distinguished.

[0049] Exemplarily, the smoothness of the above pixel point is expressed by the relational expression:

[0050]

[0051] In the formula, is the smoothness of the pixel point, is the average gradient value within the neighborhood of the pixel point.

[0052] When is small, is also small, then is close to 1, indicating that the neighborhood of the pixel point is relatively smooth; conversely, the gradient change in the neighborhood of the pixel point is complex.

[0053] Then, count the number of pixels that have been detected as initial key points within the neighborhood of the pixel, calculate the ratio of the number of pixels that have been detected as initial key points within the neighborhood of the pixel to the total number of neighboring pixels of the pixel and then normalize it to obtain the key point density within the neighborhood of the pixel. The key point density measures the proportion of pixels that have been determined as key points within the neighborhood. If there are already many key points within the neighborhood of a pixel, then the likelihood of that point becoming a new key point is relatively low because corner points are usually scattered in the image. Considering the key point density can avoid detecting too many corner points in a local area and improve the accuracy of corner point detection.

[0054] Exemplarily, the key point density within the neighborhood of the above pixel is expressed by the relational expression:

[0055]

[0056] In the formula, is the key point density within the neighborhood of the pixel, is the number of pixels that have been detected as initial key points within the neighborhood of the pixel, is the total number of neighboring pixels of the pixel.

[0057] It should be noted that the edge pixels in the grayscale image have special properties, and the gradient values of adjacent pixels are similar. If a fixed threshold is directly used for corner point detection, these edge pixels may be misselected as corner points, resulting in errors in the optical flow method detection.

[0058] Therefore, by considering the degree of edge, the corner point detection threshold can be appropriately adjusted in the edge area to avoid selecting edge pixels. Furthermore, by comprehensively considering the smoothness degree, key point density, and edge degree of the pixel, calculate the preference degree of each pixel, and adaptively adjust the corner point threshold according to the preference degree. In this way, pixels with a high preference degree (i.e., pixels in the smooth area, with a low key point density and in a non-edge area) will have a lower corner point threshold and are more likely to be detected as corner points; while pixels with a low preference degree will have a higher corner point threshold, reducing the possibility of false detection.

[0059] In another embodiment, compared with using the average value of the gradients of all neighboring pixels, using the maximum value for simplified calculation, specifically, taking the reciprocal of the maximum value of the gradient magnitudes among the neighboring pixels of the pixel as the smoothness degree of the pixel, that is, the relational expression is satisfied as:

[0060]

[0061] In the formula, is the smoothness degree of the pixel, is the maximum value of the gradient magnitudes among the neighboring pixels of the pixel, is the total number of neighboring pixels of the pixel point. Among them, adding 1 to the denominator of the formula is to avoid the denominator being 0, and adding 1 to the denominator can smooth the maximum gradient value and avoid the small change of the gradient value having too much influence on the calculation of the smoothness degree.

[0062] In another embodiment, considering the distribution of key points in the neighborhood, the neighborhood of the pixel point is further divided into multiple sub-neighborhoods. For example, a 9×9 neighborhood can be divided into 9 3×3 sub-neighborhoods.

[0063] For each sub-neighborhood, count the number of pixels that have been detected as initial key points and the total number of pixels in the sub-neighborhood, and take the ratio of the number of pixels that have been detected as initial key points to the total number of pixels in the sub-neighborhood as the key point density in the sub-neighborhood. Average the key point densities in all sub-neighborhoods to obtain the key point density in the neighborhood of the pixel point.

[0064] In one embodiment, for each pixel point of the grayscale image, filter out the neighboring pixels whose difference from its gradient angle is not (in this embodiment, by selecting pixels with a gradient direction difference of , features orthogonal to the gradient direction of the current pixel point can be better captured, thereby improving the accuracy of corner detection), and record them as angle-similar pixels. For example, if the gradient direction of this pixel point is , then pixels with a gradient direction within the range of can be regarded as angle-similar pixels.

[0065] Then calculate the absolute difference between the gradient value of this pixel point and the gradient values of each angle-similar pixel point, normalize this absolute difference, further accumulate the normalized values of all angle-similar neighboring pixels, and average the accumulated result to obtain the average similarity degree of this pixel point and its angle-similar neighboring pixels in terms of gradient magnitude, which is used as the edge degree of this pixel point. The edge degree focuses on whether the pixel point is located at the edge of the image. Pixels at the edge often have larger gradient values, but these points are not necessarily corner points. By considering the edge degree, pixels located at the edge but not corner points can be further filtered out, reducing the false detection rate.

[0066] According to the above operations, the smoothness degree, the key point density in the neighborhood, and the edge degree of all pixel points in the grayscale image can be obtained, and the product of these three factors is used as the preference degree of the pixel point.

[0067] The image features in different regions vary greatly, and using a fixed corner threshold may not accurately detect all corners. By calculating the preference degree of each pixel and adjusting the corner threshold according to the preference degree, the corner detection can be made more adaptive to the local features of the image. For pixels with a high preference degree, the possibility of them becoming corners is relatively large. Therefore, the corner threshold can be appropriately reduced to improve the sensitivity of corner detection; while for pixels with a low preference degree, the corner threshold can be increased to reduce the false detection rate.

[0068] Exemplarily, the expression for obtaining the adaptive corner threshold based on the preference degree is as follows:

[0069]

[0070] In the formula, is the adaptive corner threshold corresponding to the pixel, is the preference degree of the pixel, is the initial corner threshold.

[0071] Furthermore, corner detection is performed on the grayscale image according to the adaptive corner threshold, and all key points are detected. The specific process is as follows:

[0072] Calculate the covariance matrix of all pixels in the image, perform eigenvalue decomposition on the covariance matrix of each pixel to obtain the eigenvalues and the corresponding eigenvectors, and select the minimum eigenvalue; record the ratio of the minimum eigenvalue to the adaptive corner threshold corresponding to the pixel calculated above as the corner ratio. If this ratio is greater than 1, it indicates that the pixel may be a corner; from all pixels with a corner ratio greater than 1, select the pixel with the largest ratio as the key point; for the newly selected key point, recalculate the adaptive corner thresholds of other pixels within its neighborhood range; continue to find new key points until the corner ratios of all pixels are less than 1; the finally obtained corner set is all the required key points.

[0073] S3: Based on the detected key points, use the LKT optical flow method to achieve crack recognition of the cracks on the radiator surface.

[0074] Under normal circumstances, the movement trajectories of the key points on the radiator surface should be relatively regular and consistent during periodic vibration. However, if there are cracks on the radiator surface, the movement characteristics of the area near the cracks will change due to the influence of the cracks, resulting in abnormal movement trajectories of the key points. By analyzing the movement trajectories of the key points, detect whether there is local movement abnormality caused by cracks, so as to judge whether there are cracks on the radiator surface.

[0075] For all the key points obtained according to the above S2, the LKT optical flow method is used to track the movement trajectories of these key points during the periodic vibration of the radiator. If there are cracks on the surface of the radiator, during the vibration, the movement state of the material near the cracks will be different from that of the normal area. For example, the cracks may cause elastic deformation of local materials, discontinuous displacements, etc. By analyzing the movement trajectories of the key points, local movement anomalies caused by cracks can be found. For example, the points near the cracks may exhibit different movement directions, speeds, or accelerations from the surrounding points.

[0076] Based on the above-detected local movement anomalies, it can be inferred whether there are cracks on the surface of the radiator. If the movement trajectories of the key points in a certain area deviate significantly from the normal range, or there are discontinuous movement changes, it is considered that there are cracks in this area, and the position and scope of the cracks are roughly determined, providing a basis for subsequent technicians to carry out repairs.

[0077] It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several deformations and improvements can still be made, and these all belong to the protection scope of the present invention. Therefore, the protection scope of this invention patent shall be subject to the appended claims.

Claims

1. An intelligent identification method for surface cracks in the vibration test of a radiator, characterized in that, Including: Collecting the surface image of the radiator in real time and preprocessing the image to obtain a grayscale image of the radiator; Calculating the smoothness, the density of key points in the neighborhood, and the edge degree of each pixel point in the grayscale image, taking the product of the smoothness, the density of key points in the neighborhood, and the edge degree as the preference degree of the pixel point; obtaining an adaptive corner threshold based on the preference degree, and performing corner detection on the grayscale image according to the adaptive corner threshold to detect all key points; Based on the detected key points, using the LKT optical flow method to realize crack identification of the cracks on the surface of the radiator; The process of obtaining the edge degree includes: Screening out the angle-similar pixel points of the pixel point based on the gradient direction, calculating the absolute difference between the gradient value of this pixel point and the gradient values of each angle-similar pixel point, and performing normalization processing on the absolute difference, then performing cumulative addition on all angle-similar neighborhood pixel points after normalization processing, and performing average processing on the cumulative result to obtain the average similarity degree of this pixel point and its angle-similar neighborhood pixel points in terms of gradient amplitude as the edge degree of this pixel point; Performing corner detection on the grayscale image according to the adaptive corner threshold includes: Calculating the covariance matrix of the pixel values within the neighborhood range of each pixel point in the grayscale image, performing eigenvalue decomposition on the covariance matrix of each pixel point to obtain eigenvalues and corresponding eigenvectors; recording the ratio of the minimum eigenvalue of the pixel point to the adaptive corner threshold as the corner ratio; screening out the pixel points with the corner ratio greater than a preset threshold, and taking the pixel point with the largest corner ratio as the key point, and recalculating the adaptive corner threshold of other pixel points within the neighborhood range, and gradually finding all key points.

2. The intelligent identification method for surface cracks in the vibration test of a radiator according to claim 1, characterized in that, After obtaining the grayscale image of the radiator, it further includes: Processing the grayscale image using a corner detection algorithm according to a preset initial corner threshold to obtain all initial key points in the grayscale image.

3. The intelligent identification method for surface cracks in the vibration test of a radiator according to claim 2, characterized in that, The corner detection algorithm is the Shi-Tomasi corner detection algorithm.

4. The intelligent identification method for surface cracks in the vibration test of a radiator according to claim 3, characterized in that, The process of obtaining the smoothness includes: Obtaining the gradient value and gradient direction of the pixel points in the grayscale image, calculating the average gradient value within the set neighborhood of the pixel point, and normalizing the average gradient value using a logarithmic function to obtain the smoothness of the pixel point.

5. The intelligent identification method for surface cracks in the vibration test of a radiator according to claim 4, characterized in that, The process of obtaining the density of key points in the neighborhood includes: Counting the number of pixel points that have been detected as initial key points within the neighborhood of the pixel point, calculating the ratio of the number of pixel points that have been detected as initial key points within the neighborhood of the pixel point to the total number of neighborhood pixel points of the pixel point and then normalizing it to obtain the density of key points in the neighborhood of the pixel point.

6. The intelligent identification method for surface cracks in the vibration test of a radiator according to claim 5, wherein The adaptive corner threshold satisfies the relational expression: ; where, is the adaptive corner threshold corresponding to the pixel point, is the preference degree of the pixel point, is the initial corner threshold.

7. The intelligent recognition method for surface cracks in the vibration test of a radiator according to claim 3, characterized in that The process of obtaining the smoothness includes: Taking the reciprocal of the maximum value of the gradient amplitude among the neighborhood pixel points of the pixel point as the smoothness of the pixel point.

8. The intelligent identification method for surface cracks in the vibration test of a radiator according to claim 1, characterized in that, The process of obtaining the density of key points in the neighborhood includes: Divide the preset neighborhood of a pixel into multiple sub-neighborhoods. For each sub-neighborhood, count the number of pixels that have been detected as initial key points and the total number of pixels in the sub-neighborhood. Take the ratio of the number of pixels that have been detected as initial key points to the total number of pixels in the sub-neighborhood as the key point density in the sub-neighborhood. Average the key point densities in all sub-neighborhoods to obtain the key point density within the pixel neighborhood.

Images