Automobile side window electroplated part integrity detection algorithm

Abstract

The invention relates to a part integrity detection algorithm for an automobile side window electroplated part, which comprises the following steps of: positioning each part and a welding spot of the automobile electroplated part by adopting multiple exposure, acquiring a template picture, and generating a template matching model through the high-exposure template picture; training a target detection model by using the low-exposure template picture; obtaining a homography matrix of feature point changes of the picture of the to-be-tested piece and the template picture, and calculating to obtain a welding spot ROI position of the picture of the to-be-tested piece; and when the detection score value is lower than the threshold value, it is indicated that the welding spot is well welded. Performing foam detection on a high-exposure image, detecting other parts by using a trained target detection model, performing threshold processing and non-maximum suppression on a model reasoning result, finally summarizing all AI detection results, welding spot detection results and foam detection results, and judging OK and NG results of electroplated parts. According to the method, the interference of unstable illumination in a dark box-free environment can be overcome, and meanwhile, relatively high algorithm accuracy is ensured.

Description

Technical Field

[0001] This invention relates to the field of automotive technology, and in particular to an algorithm for detecting the integrity of electroplated parts for automotive side windows. Background Technology

[0002] In the automotive industry, cars have various electroplated parts for side windows. Their base plates require multi-layered welding processes, and the large number and variety of parts make it easy for workers to miss welds, have weak welds, or incorrectly install parts, leading to quality issues and controversies. To address this problem, the industry commonly uses visual inspection solutions for quality control. Because this needs to be integrated with assembly line work, manual triggering inspection using area scan cameras and light sources is typically employed. Visual analysis of the assembled electroplated parts is required, including matching, correction, positioning, segmentation, and other visual tasks. Finally, the sample is judged as OK or NG, and the location and type of NG parts are identified, effectively assisting workers in their assembly process, improving efficiency, and preventing NG parts from being shipped out. Generally, there are three main types of visual inspection algorithms for integrity: one uses traditional image processing algorithms to find image features to distinguish weld points; another uses AI models to collect image data, which, after training, form a non-linear logical black box for inference; and the third directly uses 3D vision to process point cloud data for judgment. Each of these three methods has its own advantages and disadvantages: The former requires high imaging stability, but this assembly line production method often cannot achieve a darkroom environment. It is often accompanied by inconsistent workpiece placement, unevenness, varying reflection angles, and external ambient light interference. This poses a significant challenge to the robustness of traditional algorithms, especially template matching algorithms, which are prone to failing to identify or determine components under varying lighting conditions. Furthermore, due to the large variety and quantity of parts on electroplated parts, numerous product models need to be constructed when inputting samples, requiring substantial computing resources. The detection process also demands significant computational resources, limiting the detection cycle time. Overall, the computing power requirement increases proportionally with the variety of products. The other AI algorithm, while improving model robustness and anti-interference performance through image enhancement, also has the advantage of a fixed computational time complexity due to the darkroom nature of the AI ​​model, unaffected by the number of products to be detected. However, its disadvantage is that the computation time is positively correlated with image size. If the images to be detected are large and accuracy cannot be reduced, the computing power requirement is extremely high. In addition, AI algorithms require the collection of a large number of data images, and the model's performance depends on the fit between the training samples and the actual detection scene, as well as the model's convergence. Furthermore, algorithm updates also rely on data; if similar samples change significantly, the server needs to retrain the model, making it difficult to take effect quickly. The biggest problem with the third type of 3D point cloud vision is cost; the hardware is more expensive than ordinary area scan cameras, which is inherently unacceptable from a customer's perspective.

[0003] Another characteristic of this inspection scenario is that the parts on the electroplated automotive components are not on a single plane but are three-dimensional with height differences, requiring consideration of camera accuracy and depth of field. The inspection areas include: nail posts, sol-gel components, and corner posts.

[0004] The process includes the inspection of circular foam, strip foam, and clips, as well as the inspection of weld points. Compared to parts with a fixed appearance, the state of weld points after welding is not fixed. Due to different worker experience, there may be weld deviations, incomplete welds, or over-welding. The ideal state after welding is a relatively flat circular bulge, but improper welding may cause dents or more burrs on the edges.

[0005] Another challenge in imaging is the reflective properties of electroplated automotive parts. Because there's no darkroom environment and workers place the parts differently each time, the positions of the reflections vary, and the resulting uneven lighting and shadows can easily interfere with the inspection results. Summary of the Invention

[0006] The purpose of this invention is to provide an integrity detection algorithm for electroplated automotive side window parts, which can overcome the interference of unstable lighting in a darkroom environment, while ensuring high algorithm accuracy and relatively low cost.

[0007] To achieve the above objectives, the present invention provides the following technical solution: an integrity detection algorithm for electroplated automotive side window parts, comprising the following steps:

[0008] Step S1: Use multiple exposure to locate each part and weld point of the automotive electroplating part and collect template images. Mark the parts and weld point positions on the template images using the software UI, record the outer contour area of ​​the parts and weld points, save the generated template images and template image areas, and then generate a template matching model using the high-exposure template images to record the ROI of each part and weld point.

[0009] Step S2: Crop the rectangles into smaller images on the low-exposure template image, and then stitch the smaller images together to form a larger image for training the object detection model;

[0010] Step S3: During the detection, use the same template matching algorithm to obtain the workpiece feature points in the image of the part under test, obtain the homography matrix of the feature point changes between the image of the part under test and the template image, eliminate a small number of invalid points, and use the previously saved weld point ROI contour points to perform matrix multiplication to obtain the weld point ROI position in the image of the part under test.

[0011] Step S4: After locating the ROI outline of each component in the image of the part to be tested, first obtain the maximum inscribed circle diameter and center point coordinates of the ROI, and then construct a rotated rectangle;

[0012] Step S5: Then extract the pixels within each rotating rectangle to obtain the gradient points along the gradient change direction. Calculate the Euclidean distance from these points to the center of the circle, then divide by the radius of the circle, normalize to the range of 0 to 1, then take the complement of 1, raise it to the cube, and sum the values ​​of all points to obtain the final detection score of the solder joint. When the detection score is lower than the threshold, it means that the solder joint has been successfully soldered.

[0013] Step S6: Perform foam detection on the high-exposure image, perform Fourier transform, extract the high-frequency part to get the noise area, and then perform region dilation operation to finally obtain the circular and strip areas of the foam.

[0014] Step S7: Use the trained target detection model to detect other parts, perform thresholding and non-maximum suppression on the model inference results, and finally summarize all AI detection results, solder joint detection results and foam detection results to determine whether the electroplated parts are OK or NG.

[0015] Furthermore, in step S1, the method for saving the template image and the template image area can be a scatter plot of the outer contour.

[0016] Furthermore, the template matching model in step S1 is further defined as follows: based on the image imaging characteristics, if there is obvious backlighting, template matching based on contour query is used; if it is a bright field and the workpiece texture is clear, template matching based on descriptor is used.

[0017] Furthermore, in step S2, the method of collaging is to use the skyline algorithm to collage the image and obtain a large image with the least redundant area. When the last image is not filled, a greedy algorithm is used to continue filling it. The material is a small image that has been rotated 90 degrees and filled.

[0018] Furthermore, the method for obtaining the homography matrix in step S3 is to map the feature points of the workpiece in the test image to the feature points of the workpiece in the template image using the RANSAC algorithm.

[0019] Furthermore, the method for eliminating a small number of invalid points in step S3 is to use the outlier fitting algorithm ransac.

[0020] Furthermore, in step S4, the long side of the rotation matrix is ​​0.6 times the diameter of the inscribed circle, the short side is 0.33 times the long side, the rotation direction can be customized, mainly based on the direction of the shadow change of the product column, and the coordinates of the center point of the rotation rectangle are the coordinates of the circle center.

[0021] Furthermore, the formula for calculating the detection score in step S5 is further as follows:

[0022]

[0023] Where n is the number of gradient points at the solder joint; p x p represents the X-axis coordinate of the gradient point. y Let c be the y-coordinate of the gradient point. x Let c be the X-axis coordinate of the center of the inscribed circle. y c is the Y-axis coordinate of the center of the inscribed circle; r p is the radius of the inscribed circle; dist Let f be the Euclidean distance from the point to the center of the circle. score This is the final score.

[0024] The beneficial effects of this invention are as follows: Structurally, the algorithm employs a multi-exposure method, selectively choosing appropriate exposure values ​​based on the material of the part being detected. It also uses template matching for initial positioning and cropping, reducing a significant amount of unnecessary computation. Furthermore, subsequent detection utilizes a multi-threaded concurrent strategy, employing a divide-and-conquer approach to improve detection efficiency. In terms of cost, the algorithm maintains high measurement accuracy even with simple, open-access equipment and tooling, satisfying rotation and translation invariance. It also avoids the need for complex morphological algorithms that consume substantial computational resources. The AI ​​inference part, due to the use of the skyline algorithm, minimizes the redundant area of ​​small-image patching into large-image processing, allowing the CPU to be used solely for ONNX model inference, thus meeting the detection cycle time.

[0025] In terms of computational complexity for the solder joints, it primarily involves linear calculations such as matrix multiplication, Euclidean distance, and gradient differentiation. It doesn't perform contour segmentation or connected component calculations, nor does it involve template matching for solder joints, making its detection efficiency very high. Furthermore, the algorithm cleverly considers factors such as a limited number of gradient points, gradient points being off-center, and the non-linear weighting of the exponent. This ensures that the final score is negatively correlated with the integrity of the pillar and is easily distinguishable by a fixed threshold.

[0026] Previous algorithms for inspecting automotive electroplated parts lacked a fixed framework. Due to the diverse types of parts involved, different exposure methods and inspection techniques each had their own shortcomings. Therefore, this invention employs a divide-and-conquer multi-threaded parallel processing approach after initial template matching for localization. Fourier transform is used to process the surface of foam with high-frequency pores, target detection is used to handle regular part appearance defects, and a self-developed algorithm formula is used to handle weld points with varying lighting conditions. Attached Figure Description

[0027] Figure 1 This is an optimized diagram of the jigsaw puzzle filled using a greedy algorithm;

[0028] Figure 2 This is a schematic diagram illustrating the gradient range for solder joint queries.

[0029] Figure 3 This is a schematic diagram of the gradient points of the solder joints;

[0030] Figure 4 This is a schematic diagram showing the effect of a product with OK solder joints.

[0031] Figure 5 This is a schematic diagram illustrating the effect of jigsaw puzzle inspection in the inspection of other parts;

[0032] Figure 6 This is a schematic diagram of the overall process of the present invention. Detailed Implementation

[0033] The invention will now be further described with reference to the accompanying drawings.

[0034] Please see Figures 1 to 5 The present invention provides an embodiment: an algorithm for detecting the integrity of electroplated automotive side window parts, comprising the following steps:

[0035] Step S1: Use multiple exposure to locate each part and weld point of the automotive electroplating part and collect template images. Mark the parts and weld point positions on the template images using the software UI, record the outer contour area of ​​the parts and weld points, save the generated template images and template image areas, and then generate a template matching model using the high-exposure template images to record the ROI of each part and weld point.

[0036] Step S2: Crop the rectangles into smaller images from the low-exposure template image, and then stitch the smaller images together to form a larger image for training the object detection model; the products detected by the object detection model are workpieces other than foam and solder joints that are detected using AI object detection.

[0037] Step S3: During inspection, the same template matching algorithm is used to obtain the workpiece feature points in the image of the part under test. The homography matrix of the feature point changes between the image of the part under test and the template image is obtained. A small number of invalid points are eliminated. The previously saved ROI contour points of the weld points are substituted into the matrix multiplication to obtain the ROI position of the weld points in the image of the part under test. When inspecting incoming materials, the ROI of each part and weld point needs to be extracted through template matching and then assigned to the corresponding detection algorithm.

[0038] Step S4: After locating the ROI outline of each component in the image of the part to be tested, first obtain the maximum inscribed circle diameter and center point coordinates of the ROI, and then construct a rotated rectangle;

[0039] Step S5: Then extract the pixels within each rotating rectangle to obtain the gradient points along the gradient change direction. Calculate the Euclidean distance from these points to the center of the circle, then divide by the radius of the circle, normalize to the range of 0 to 1, then take the complement of 1, raise it to the cube, and sum the values ​​of all points to obtain the final detection score of the solder joint. When the detection score is lower than the threshold, it means that the solder joint has been successfully soldered.

[0040] Step S6: Perform foam detection on the high-exposure image, perform Fourier transform, extract the high-frequency part to get the noise area, and then perform region dilation operation to finally obtain the circular and strip areas of the foam.

[0041] Step S7: Use the trained target detection model to detect other parts, perform thresholding and non-maximum suppression on the model inference results, and finally summarize all AI detection results, solder joint detection results and foam detection results to determine whether the electroplated parts are OK or NG.

[0042] Please continue reading. Figure 5 As shown, in one embodiment of the present invention, the method for saving the template image and the template image area in step S1 can be a scatter plot of the outer contour.

[0043] Please continue reading. Figure 5 As shown, in one embodiment of the present invention, the template matching model in step S1 is further defined as follows: based on the image imaging characteristics, if there is obvious backlight, template matching based on contour query is used; if it is a bright field and the workpiece texture is clear, template matching based on descriptor is used.

[0044] Please continue reading. Figure 5 As shown, in one embodiment of the present invention, the method of jigsaw puzzle in step S2 is to use the skyline algorithm to jigsaw puzzle and obtain a large image with the least redundant area. When the last image is not filled, a greedy algorithm is used to continue filling it. The material is a small image that has been rotated 90 degrees and filled.

[0045] Please continue reading. Figure 5 As shown, in one embodiment of the present invention, the method for obtaining the homography matrix in step S3 is to map the feature points of the workpiece in the test image to the feature points of the workpiece in the template image using the RANSAC algorithm.

[0046] Please continue reading. Figure 5 As shown, in one embodiment of the present invention, the method for eliminating a small number of invalid points in step S3 is to use the outlier fitting algorithm ransac.

[0047] Please continue reading. Figure 5 As shown, in one embodiment of the present invention, the long side of the rotation matrix in step S4 is 0.6 times the diameter of the inscribed circle, the short side is 0.33 times the long side, the rotation direction can be customized, mainly based on the shadow change direction of the product column, and the coordinates of the center point of the rotation rectangle are the coordinates of the circle center.

[0048] Please continue reading. Figure 5 As shown, in one embodiment of the present invention, the calculation formula for the detection score in step S5 is further as follows:

[0049]

[0050] Where n is the number of gradient points at the solder joint; p x p represents the X-axis coordinate of the gradient point. y Let c be the y-coordinate of the gradient point. x Let c be the X-axis coordinate of the center of the inscribed circle. y c is the Y-axis coordinate of the center of the inscribed circle; r p is the radius of the inscribed circle; dist Let f be the Euclidean distance from the point to the center of the circle. score This is the final score. Specific Implementation Example 1:

[0052] This invention relates to an algorithm for integrity detection of electroplated parts on automotive side windows. It features a self-developed imaging method, algorithm framework, and formula, effectively distinguishing between OK and NG parts, overcoming interference from unstable lighting in a darkroom environment, and ensuring high algorithm accuracy. To achieve the above objectives, the implementation process of this invention is as follows:

[0053] 1. The imaging design employs multiple exposures: low exposure for AI detection and recognition, and high exposure for weld points and foam parts. This is because different parts have varying degrees of reflectivity, requiring different levels of detail in feature texture and necessitating different processing. High-exposure images are used initially for locating various parts and weld points in automotive electroplated components, as the black plastic base of the electroplated part has strong light absorption. When inputting product information, users initially need to mark the parts and weld point locations on the template image using the software UI. This can be a rectangular or prototype-like closed area. The template image and the template image area are then saved, either as scattered points of the outer contour. A template matching model is generated from the template image. The template matching method depends on the image imaging characteristics. If there is significant backlighting, contour-based template matching is used; if it is bright field and the workpiece texture is clear, descriptor-based template matching is used. Algorithms such as SIFT and SURF are commonly used for electroplated parts. After storing the template matching model, the object detection model also needs to be trained. Under low exposure, rectangles representing small images of parts to be used for AI detection are cropped, and then these are stitched together to form a large image for training the AI ​​model. The reason for not directly training with small images is that, based on actual testing, the AI ​​model's processing efficiency for a single large image with the same number of small images is higher than processing multiple small images sequentially. Because the rectangles of different parts have varying lengths and widths, a combined stitching method is used. However, previous greedy algorithms tend to cause significant image redundancy, such as... Figure 1As shown, after optimization, more small images can be inserted into a large image of the same area. The problem with the greedy algorithm is that it finds a local optimum, without pre-recording the dimensions of all images. In the algorithm field, this is a two-dimensional rectangular bin packing problem, an NP-hard problem that can only obtain an approximate solution. Common algorithms include the Maximum Rectangles Algorithm and the Skyline Algorithm. Here, we use the Skyline Algorithm to obtain the large image with the least redundant area. The last image is not filled, so we use the greedy algorithm to continue filling it. The source material is a 90-degree rotated image that has already been filled with small images. This is to improve the training efficiency of the AI ​​model and avoid wasting resources in redundant areas. Moreover, the training of the object detection model itself also involves image enhancement similar to cutout and rotation / scaling.

[0054] 2. During inspection, the same template matching algorithm is used to obtain the feature points of the electroplated part in the inspection image. Mapping these feature points to the feature points of the template image using the RANSAC algorithm yields a homography matrix representing the changes between the two feature points. As is well known, the homography matrix has 8 degrees of freedom, theoretically providing a unique solution for 4 pairs of points. However, in practice, the number of detected feature points can vary from 200 to 300 even with good imaging. Therefore, the RANSAC outlier fitting algorithm is needed to ensure high-precision matching for the majority of points and eliminate a small number of invalid points, avoiding interference from their fitting accuracy. The resulting change matrix represents the positional relationship between the template and the part under test. Using the previously saved ROI contour points of each part and weld point, matrix multiplication is performed to obtain the corresponding positions in the image of the part under test.

[0055] 3. After locating the ROI contours of each component, multi-threaded parallel processing is used to detect each part. First, regarding the solder joints, the detection image used is high-exposure because the base plastic is a light-absorbing material. First, the maximum inscribed circle diameter and center point coordinates of the solder joint ROI are calculated. This algorithm is already implemented in the OpenCV open-source library. A two-sided scanning algorithm is used to calculate an approximate distance transformation matrix. This matrix records the minimum distance from all foreground pixels to background pixels. The radius of the inscribed circle is the maximum value in this matrix, and the center point coordinates are the coordinates of the maximum value. Then, a rotating rectangle is constructed. The long side of the matrix is ​​0.6 times the inscribed circle diameter, and the short side is 0.33 times the long side. The rotation direction can be customized, mainly based on the shadow change direction of the product pillar. This allows for the capture of as many gradient change points as possible when extracting the solder joint gradient later. For example... Figure 2 The rotation direction is 45 degrees counterclockwise. The coordinates of the center point of the rotated rectangle are the coordinates of the circle's center, and the rectangles are shifted to the left and right by 3 units of the shorter side length, so a total of 7 rectangles are constructed, as shown below. Figure 2 As shown.

[0056] 4. Then, extract the pixels within each rotated rectangle, perform a first-order Gaussian filter along its shorter side, and calculate the gradient change of the first derivative along its longer side. Finally, obtain the gradient points as shown below. Figure 3 As shown. These points are all along the gradient change direction and have a clear black-and-white boundary. After excluding gradient points not inside the circle, the Euclidean distance from these points to the circle's center is calculated, then divided by the circle's radius, normalized to the range of 0-1, then the complement of 1 is taken, the result is cubed, and the values ​​of all points are summed to obtain the final inspection score for the weld point. The formula is as follows:

[0057]

[0058] This formula considers multiple scenarios. The more gradient points appear, the greater the probability of a pillar being present. Therefore, the formula is designed to be cumulative. The more centrally located the gradient point, the higher the probability of a pillar being present. Therefore, normalization and inverted centrifugal distance are used, and gradient points outside the circle are excluded by using the maximum value of 0. Ultimately, a higher score for this weld joint indicates a missed weld or a weak weld, meaning the pillar is still basically present; a lower score indicates the weld joint has been successfully welded. Figure 4 The solder joints are okay, but their solder joint score is relatively low. This can be seen from... Figure 4 As can be seen, although the lighting still produces shadow changes and gradient points, the number of these points is small and they are close to the edge. Secondly, the algorithm performs Gaussian filtering on the region before calculating the gradient, so the overall calculated solder joint score is not high. Therefore, by setting a threshold, which is usually 2.4, it is possible to effectively distinguish between cold solder joints, missing solder joints and OK products.

[0059] 5. In the foam inspection section, because foam is a highly light-absorbing material, inspection can only be performed on high-exposure images. Whether it's circular or strip-shaped foam, the material characteristics of foam include regularly spaced pores. These tiny pores can be considered noise in the image. By performing a Fourier transform and extracting the high-frequency components, we identify the noise areas. Then, by performing region dilation, we obtain the circular and strip-shaped areas of the foam, which allows us to determine the integrity of the foam.

[0060] 6. For the inspection of other parts, such as nails and clips, AI object detection is used. The previously trained object detection model is used for inference. Low-exposure images are used here, and small images need to be stitched together to form a larger image. Figure 5 As shown, a skyline algorithm is used for image stitching. However, since this is for detection, redundant padding is no longer needed for the final large image. Finally, thresholding and non-maximum suppression are applied to the AI ​​model's inference results, i.e., post-processing. Finally, all AI detection results, solder joint detection results, and foam detection results are summarized to determine whether the electroplated part is OK or NG.

[0061] This invention's welding point algorithm does not follow the traditional image vision processing approach, which aims to determine the state of the column before or after welding. The pre-welding column is viewed in 3D, and the image is affected by the workpiece's placement, resulting in varying angles. The closer to the edge of the field of view, the larger the tilt angle. Furthermore, due to unstable lighting and the reflectivity of plastic materials, highly reflective and dark areas appear randomly. Therefore, identifying the pre-welding column using 2D template matching is unrealistic; AI algorithms might be able to achieve this, but require a large number of robust samples. Determining the post-welding state is even more difficult. Although the welding adhesive can reflect light, it cannot be directly determined by calculating highly reflective areas because it is still affected by lighting and the shooting angle, including surface texture shadows and potentially entirely dark areas. Therefore, this invention's algorithm only distinguishes, not identifies, employing a linear algorithm that uses minimal computation to distinguish complex scenes.

[0062] The above description is only a preferred embodiment of the present invention and should not be construed as a limitation of this application. All equivalent changes and modifications made in accordance with the scope of the patent application of the present invention should be covered by the present invention.

Claims

1. An algorithm for detecting the integrity of electroplated automotive side window parts, characterized in that: Includes the following steps: Step S1: Use multiple exposure to locate each part and weld point of the automotive electroplating part and collect template images. Mark the parts and weld point positions on the template images using the software UI, record the outer contour area of ​​the parts and weld points, save the generated template images and template image areas, and then generate a template matching model using the high-exposure template images to record the ROI of each part and weld point. Step S2: Crop the rectangles into smaller images on the low-exposure template image, and then stitch the smaller images together to form a larger image for training the object detection model; Step S3: During the inspection, the workpiece feature points in the image of the workpiece under test are obtained using the template matching algorithm. The homography matrix of the feature point changes between the image of the workpiece under test and the template image is obtained. A small number of invalid points are eliminated. The previously saved ROI contour points of the weld points are substituted into the matrix multiplication to obtain the ROI position of the weld points in the image of the workpiece under test. Step S4: After locating the ROI outline of each component in the image of the part to be tested, first obtain the maximum inscribed circle diameter and center point coordinates of the ROI, and then construct a rotated rectangle; Step S5: Then extract the pixels within each rotating rectangle to obtain the gradient points along the gradient change direction. Calculate the Euclidean distance from these points to the center of the circle, then divide by the radius of the circle, normalize to the range of 0 to 1, then take the complement of 1, raise it to the cube, and sum the values ​​of all points to obtain the final detection score of the solder joint. When the detection score is lower than the threshold, it means that the solder joint has been successfully soldered. Step S6: Perform foam detection on the high-exposure image, perform Fourier transform, extract the high-frequency part to get the noise area, and then perform region dilation operation to finally obtain the circular and strip areas of the foam. Step S7: Use the trained target detection model to detect other parts, perform thresholding and non-maximum suppression on the model inference results, and finally summarize all AI detection results, solder joint detection results and foam detection results to determine whether the electroplated parts are OK or NG.

2. The algorithm for detecting the integrity of electroplated automotive side window parts according to claim 1, characterized in that: In step S1, the template image and the template image area can be saved using a scatter plot of the outer contour.

3. The algorithm for detecting the integrity of electroplated automotive side window parts according to claim 1, characterized in that: The template matching model in step S1 is further defined as follows: based on the image imaging characteristics, if there is obvious backlight, template matching based on contour query is used; if it is a bright field and the workpiece texture is clear, template matching based on descriptor is used.

4. The algorithm for detecting the integrity of electroplated automotive side window parts according to claim 1, characterized in that: The method for creating the jigsaw puzzle in step S2 is to use the skyline algorithm to obtain a large image with the least redundant area. When the last image is not filled, a greedy algorithm is used to continue filling it. The material is a small image that has been rotated 90 degrees and filled.

5. The algorithm for detecting the integrity of electroplated automotive side window parts according to claim 1, characterized in that: The method for obtaining the homography matrix in step S3 is to map the feature points of the workpiece in the test image to the feature points of the workpiece in the template image using the RANSAC algorithm.

6. The integrity detection algorithm for electroplated automotive side window parts according to claim 1, characterized in that: The method for eliminating a small number of invalid points in step S3 is to use the outlier fitting algorithm ransac.

7. The algorithm for detecting the integrity of electroplated automotive side window parts according to claim 1, characterized in that: In step S4, the long side of the rotation matrix is ​​0.6 times the diameter of the inscribed circle, the short side is 0.33 times the diameter of the long side, the rotation direction can be customized, mainly based on the direction of the shadow change of the product column, and the coordinates of the center point of the rotation rectangle are the coordinates of the circle center.

8. The algorithm for detecting the integrity of electroplated automotive side window parts according to claim 1, characterized in that: The formula for calculating the detection score in step S5 is further as follows: Where n is the number of gradient points at the solder joint; p x p represents the X-axis coordinate of the gradient point. y Let c be the y-coordinate of the gradient point. x Let c be the X-axis coordinate of the center of the inscribed circle. y c is the Y-axis coordinate of the center of the inscribed circle; r Let p be the radius of the inscribed circle. dist Let f be the Euclidean distance from the point to the center of the circle. score This is the final score.

Images